Upload
others
View
20
Download
0
Embed Size (px)
Citation preview
HAL Id: hal-00768190https://hal.archives-ouvertes.fr/hal-00768190
Submitted on 21 Dec 2012
HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.
Panels with nonstationary multifactor error structuresG. Kapetanios, M. Hashem Pesaran, T. Yamagata
To cite this version:G. Kapetanios, M. Hashem Pesaran, T. Yamagata. Panels with nonstationary multifactor error struc-tures. Econometrics, MDPI, 2010, 160 (2), pp.326. �10.1016/j.jeconom.2010.10.001�. �hal-00768190�
Accepted Manuscript
Panels with nonstationary multifactor error structures
G. Kapetanios, M. Hashem Pesaran, T. Yamagata
PII: S0304-4076(10)00202-2DOI: 10.1016/j.jeconom.2010.10.001Reference: ECONOM 3409
To appear in: Journal of Econometrics
Received date: 15 June 2009Revised date: 14 April 2010Accepted date: 5 October 2010
Please cite this article as: Kapetanios, G., Hashem Pesaran, M., Yamagata, T., Panels withnonstationary multifactor error structures. Journal of Econometrics (2010),doi:10.1016/j.jeconom.2010.10.001
This is a PDF file of an unedited manuscript that has been accepted for publication. As aservice to our customers we are providing this early version of the manuscript. The manuscriptwill undergo copyediting, typesetting, and review of the resulting proof before it is published inits final form. Please note that during the production process errors may be discovered whichcould affect the content, and all legal disclaimers that apply to the journal pertain.
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Panels with Nonstationary Multifactor Error Structures�
G. KapetaniosQueen Mary, University of London
M. Hashem PesaranCambridge University and USC
T. YamagataUniversity of York
September 27, 2010
Abstract
The presence of cross-sectionally correlated error terms invalidates much inferentialtheory of panel data models. Recently, work by Pesaran (2006) has suggested a methodwhich makes use of cross-sectional averages to provide valid inference in the case of sta-tionary panel regressions with a multifactor error structure. This paper extends this workand examines the important case where the unobservable common factors follow unit rootprocesses. The extension to I(1) processes is remarkable on two counts. Firstly, it is ofgreat interest to note that while intermediate results needed for deriving the asymptoticdistribution of the panel estimators di¤er between the I(1) and I(0) cases, the �nal resultsare surprisingly similar. This is in direct contrast to the standard distributional resultsfor I(1) processes that radically di¤er from those for I(0) processes. Secondly, it is worthnoting the signi�cant extra technical demands required to prove the new results. Thetheoretical �ndings are further supported for small samples via an extensive Monte Carlostudy. In particular, the results of the Monte Carlo study suggest that the cross-sectionalaverage based method is robust to a wide variety of data generation processes and haslower biases than the alternative estimation methods considered in the paper.Keywords: Cross Section Dependence, Large Panels, Unit Roots, Principal Components, Com-
mon Correlated E¤ects.JEL-Classi�cation: C12, C13, C33.
�The authors thank Vanessa Smith and Elisa Tosetti, and seminar participants at Kyoto University, Univer-sity of Amsterdam, University of Nottingham and the annual conference at the Granger Centre for most helpfulcomments on a previous version of this paper. This version has also bene�ted greatly from constructive com-ments and suggestions by the Editor (Cheng Hsiao), an associate editor and three anonymous referees. HashemPesaran and Takashi Yamagata acknowledge �nancial support from the ESRC (Grant No. RES-000-23-0135).
1
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
1 Introduction
Panel data sets have been increasingly used in economics to analyze complex economic phenom-
ena. One of their attractions is the ability to use an extended data set to obtain information
about parameters of interest which are assumed to have common values across panel units.
Most of the work carried out on panel data has usually assumed some form of cross sectional
independence to derive the theoretical properties of various inferential procedures. However,
such assumptions are often suspect and as a result recent advances in the literature have focused
on estimation of panel data models subject to error cross sectional dependence.
A number of di¤erent approaches have been advanced for this purpose. In the case of
spatial data sets where a natural immutable distance measure is available the dependence is
often captured through �spatial lags�using techniques familiar from the time series literature.
In economic applications, spatial techniques are often adapted using alternative measures of
�economic distance�. This approach is exempli�ed in work by Lee and Pesaran (1993), Conley
and Dupor (2003), Conley and Topa (2002) and Pesaran, Schuermann, and Weiner (2004), as
well as the literature on spatial econometrics recently surveyed by Anselin (2001). In the case
of panel data models where the cross section dimension (N) is small (typically N < 10) and
the time series dimension (T ) is large the standard approach is to treat the equations from the
di¤erent cross section units as a system of seemingly unrelated regression equations (SURE)
and then estimate the system by the Generalized Least Squares (GLS) techniques.
The SURE approach is not applicable if the errors are correlated with the regressors and/or
if the panels under consideration have a large cross sectional dimension. This has led a number
of investigators to consider unobserved factor models, where the cross section error correlations
are de�ned in terms of the factor loadings. The use of unobserved factors also allows for certain
degree of correlation between the idiosyncratic errors and the unobserved factors. Use of factor
models is not new in economics and dates back to the pioneering work of Stone (1947) who
applied the principal components (PC) analysis of Hotelling to US macroeconomic time series
over the period 1922-1938 and was able to demonstrate that three factors (namely total income,
its rate of change and a time trend) explained over 97 per cent of the total variations of all the
17 macro variables that he had considered. Until recently, subsequent applications of the PC
approach to economic times series has been primarily in �nance. See, for example, Chamberlain
and Rothschild (1983), Connor and Korajzcyk (1986) and Connor and Korajzcyk (1988). But
more recently the unobserved factor models have gained popularity for forecasting with a large
number of variables as advocated by Stock and Watson (2002). The factor model is used very
2
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
much in the spirit of the original work by Stone, in order to summarize the empirical content
of a large number of macroeconomics variables by a small set of factors which, when estimated
using principal components, is then used for further modelling and/or forecasting. A related
literature on dynamic factor models has also been put forward by Forni and Reichlin (1998)
and Forni, Hallin, Lippi, and Reichlin (2000).
Recent uses of factor models in forecasting focus on consistent estimation of unobserved
factors and their loadings. Related theoretical advances by Bai and Ng (2002) and Bai (2003)
are also concerned with estimation and selection of unobserved factors and do not consider the
estimation and inference problems in standard panel data models where the objects of interest
are slope coe¢ cients of the conditioning variables (regressors). In such panels the unobserved
factors are viewed as nuisance variables, introduced primarily to model the cross section de-
pendencies of the error terms in a parsimonious manner relative to the SURE formulation.
Despite these di¤erences knowledge of factor models could still be useful for the analysis of
panel data models if it is believed that the errors might be cross sectionally correlated. Disre-
garding the possible factor structure of the errors in panel data models can lead to inconsistent
parameter estimates and incorrect inference. Coakley, Fuertes, and Smith (2002) suggest a pos-
sible solution to the problem using the method of Stock and Watson (2002). But, as Pesaran
(2006) shows, the PC approach proposed by Coakley, Fuertes, and Smith (2002) can still yield
inconsistent estimates. Pesaran (2006) suggests a new approach by noting that linear combi-
nations of the unobserved factors can be well approximated by cross section averages of the
dependent variable and the observed regressors. This leads to a new set of estimators, referred
to as the Common Correlated E¤ects (CCE) estimators, that can be computed by running
standard panel regressions augmented with the cross section averages of the dependent and
independent variables. The CCE procedure is applicable to panels with a single or multiple
unobserved factors and does not necessarily require the number of unobserved factors to be
smaller than the number of observed cross section averages.
In this paper we extend the analysis of Pesaran (2006) to the case where the unobserved
common factors are integrated of order 1, or I(1). Our analysis does not require an a priori
knowledge of the number of unobserved factors. It is only required that the number of un-
observed factors remains �xed as the sample size is increased. The extension of the results of
Pesaran (2006) to the I(1) case is far from straightforward and involves the development of new
intermediate results that could be of relevance to the analysis of panels with unit roots. It is
also remarkable in the sense that whilst the intermediate results needed for deriving the asymp-
totic distribution of the panel estimators di¤er between the I(1) and I(0) cases, the �nal results
are surprisingly similar. This is in direct contrast to the usual phenomenon whereby distribu-
3
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
tional results for I(1) processes are radically di¤erent to those for I(0) processes and involve
functionals of Brownian motion whose use requires separate tabulations of critical values.
It is very important to appreciate that our primary focus is on estimating the coe¢ cients
of the panel regression model. We do not wish to investigate the (co-)integration properties
of the unobserved factors. Rather, our focus is robustness to the properties of the unobserved
factors, for the estimation of the coe¢ cients of the observed regressors that vary over time as
well as over the cross section units. In this sense the extension provided by our work is of
great importance in empirical applications where the integration properties of the unobserved
common factors are typically unknown. In the CCE approach the nature of the factors does
not matter for inferential analysis of the coe¢ cients of the observed variables. The theoretical
�ndings of the paper are further supported for small samples via an extensive Monte Carlo
study. In particular, the results of the Monte Carlo study clearly show that the CCE estimator
is robust to a wide variety of data generation processes and has lower biases than all of the
alternative estimation methods considered in the paper.
The structure of the paper is as follows: Section 2 provides an overview of the method
suggested by Pesaran (2006) in the case of stationary factor processes. Section 3 provides the
theoretical framework of the analysis of nonstationarity. In this section the theoretical prop-
erties of the various estimators are presented. Section 4 presents an extensive Monte Carlo
study, and Section 5 concludes. The Appendix contains proofs of the theoretical results. Some
more technical results and proofs of Lemmas are relegated to a supplementary appendix that
is provided separately from the main paper.
Notations: K stands for a �nite positive constant, kAk = [Tr(AA0)]1=2 is the Frobenius
norm of the m� n matrix A, and A+ denotes the Moore-Penrose inverse of A. rk(A) denotes
the rank of A. supiWi is the supremum of Wi over i. an = O(bn) states the deterministic
sequence fang is at most of order bn, xn = Op(yn) states the vector of random variables, xn; is
at most of order yn in probability, and xn = op(yn) is of smaller order in probability than yn,q:m:!
denotes convergence in quadratic mean (or mean square error),p! convergence in probability,
d! convergence in distribution, and ds asymptotic equivalence of probability distributions. Allasymptotics are carried out under N !1, either with a �xed T , or jointly with T !1. Jointconvergence of N and T will be denoted by (N; T )
j!1. Restrictions (if any) on the relativerates of convergence of N and T will be speci�ed separately.
4
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
2 Panel Data Models with Observed and UnobservedCommon E¤ects
In this section we review the methodology introduced in Pesaran (2006). Let yit be the obser-
vation on the ith cross section unit at time t for i = 1; 2; :::; N ; t = 1; 2; :::; T; and suppose that
it is generated according to the following linear heterogeneous panel data model
yit = �0idt + �
0ixit +
0if t + "it; (1)
where dt is a n � 1 vector of observed common e¤ects, which is partitioned as dt = (d01t;d02t)0
where d1t is a n1�1 vector of deterministic components such as intercepts or seasonal dummiesand d2t is a n2 � 1 vector of unit root stochastic observed common e¤ects, with n = n1 + n2,
xit is a k � 1 vector of observed individual-speci�c regressors on the ith cross section unit attime t, f t is the m� 1 vector of unobserved common e¤ects, and "it are the individual-speci�c(idiosyncratic) errors assumed to be independently distributed of (dt;xit). The unobserved
factors, f t, could be correlated with (dt;xit), and to allow for such a possibility the following
speci�cation for the individual speci�c regressors will be considered
xit = A0idt + �
0if t + vit; (2)
whereAi and �i are n�k andm�k factor loading matrices with �xed and bounded components,vit = (vi1t; :::; vikt)
0 are the speci�c components of xit distributed independently of the common
e¤ects and across i; but assumed to follow general covariance stationary processes. In our setup,
"it is assumed to be stationary, which implies that in the case where ft and/or dt contain unit
root processes, then yit, xit;dt and ft must be cointegrated.1 Some of the implications of this
property are explored further in Remark 6.
Combining (1) and (2) we now have
zit(k+1)�1
=
�yitxit
�= B0
i(k+1)�n
dtn�1
+ C 0i
(k+1)�mf tm�1
+ uit(k+1)�1
; (3)
where
uit =
�"it + �
0ivit
vit
�=
�1 �0i0 Ik
��"itvit
�; (4)
Bi =��i Ai
�� 1 0�i Ik
�, Ci =
� i �i
�� 1 0�i Ik
�; (5)
1However, as will be shown later, our results on the estimators of � hold even if the factor loadings i and/or�i are zero (or weak in the sense of Chudik, Pesaran, and Tosetti (2010)), and it is not necessary that xit andft are cointegrated. What is required for our results is that conditional on dt and ft, the idiosyncratic errors "itand vit are stationary.
5
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Ik is an identity matrix of order k, and the rank of Ci is determined by the rank of the
m� (k + 1) matrix of the unobserved factor loadings
~�i =� i �i
�: (6)
As discussed in Pesaran (2006), the above set up is su¢ ciently general and renders a variety
of panel data models as special cases. In the panel literature with T small and N large, the
primary parameters of interest are the means of the individual speci�c slope coe¢ cients, �i,
i = 1; 2; :::; N . The common factor loadings, �i and i, are generally treated as nuisance
parameters. In cases where both N and T are large, it is also possible to consider consistent
estimation of the factor loadings, but this topic will not be pursued here. The presence of
unobserved factors in (1) implies that estimation of �i and its cross sectional mean cannot be
undertaken using standard methods. Pesaran (2006) has suggested using cross section averages
of yit and xit to deal with the e¤ects of proxies for the unobserved factors in (1). To see why
such an approach could work, consider simple cross section averages of the equations in (3)2
�zt = �B0dt + �C
0f t + �ut; (7)
where
�zt =1
N
NXi=1
zit, �ut =1
N
NXi=1
uit;
and
�B =1
N
NXi=1
Bi, �C =1
N
NXi=1
Ci. (8)
We distinguish between two important cases: when the rank condition
rk( �C) = m � k + 1, for all N; and as N !1; (9)
holds, and when it does not. Under the former, the analysis simpli�es considerably since it is
possible to proxy the unobserved factors by linear combinations of cross section averages, �ztand the observed common components, dt. But if the rank condition is not satis�ed this is not
possible, although as we shall see it is still possible to consistently estimate the mean of the
regression coe¢ cients, �, by the CCE procedure.
In the case where the rank condition is met we have
f t =��C �C
0��1
�C��zt � �B
0dt � �ut
�: (10)
2Pesaran (2006) considers cross section weighted averages that are more general. But to simplify the expo-sition we con�ne our discussion to simple averages throughout.
6
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
But since
�utq:m:! 0, as N !1, for each t; (11)
and�C
p! C = ~�
�1 0� Ik
�; as N !1; (12)
where~� = (E ( i) ; E (�i)) = ( ;�), (13)
it follows, assuming that Rank(~�) = m, that
f t � (CC 0)�1C��zt � �B
0dt
�q:m:! 0, as N !1:
This suggests that for su¢ ciently large N , it is valid to use �ht = (d0t; �z
0t)0 as observable proxies
for f t. This result holds irrespective of whether the unobserved factor loadings, i and �i, are
�xed or random.
When the rank condition is not satis�ed the use of cross section averages alone do not
allow consistent estimation of all of the unobserved factors and as a result the estimation of
the individual coe¢ cients �i by means of the cross section averages alone will not be possible.
But interestingly enough consistent estimates of the mean of the slope coe¢ cients, �, and their
asymptotic distribution can be obtained if it is further assumed that the factor loadings are
distributed independently of the factors and the individual-speci�c error processes.
2.1 The CCE Estimators
We now discuss the two estimators for the means of the individual speci�c slope coe¢ cients
proposed by Pesaran (2006). One is the Mean Group (MG) estimator proposed in Pesaran and
Smith (1995) and the other is a generalization of the �xed e¤ects estimator that allows for the
possibility of cross section dependence. The former is referred to as the �Common Correlated
E¤ects Mean Group�(CCEMG) estimator, and the latter as the �Common Correlated E¤ects
Pooled�(CCEP) estimator.
The CCEMG estimator is a simple average of the individual CCE estimators, bi of �i,
bMG = N�1NXi=1
bi; (14)
where
bi = (X0i�MX i)
�1X 0i�Myi; (15)
X i = (xi1;xi2; :::;xiT )0, yi = (yi1; yi2; :::; yiT )
0, �M is de�ned by
�M = IT � �H��H0 �H��1
�H0; (16)
7
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
�H = (D; �Z), D and �Z being, respectively, the T �n and T � (k+1) matrices of observationson dt and �zt. We also de�ne for later use
M g = IT �G (G0G)�1G0; (17)
and
M q = IT �Q (Q0Q)+Q0, with Q = G �P , (18)
where G = (D;F ), D = (d1;d2; :::;dT )0, F = (f 1;f 2; :::;fT )
0 are T � n and T � m data
matrices on observed and unobserved common factors, respectively, (A)+ denotes the Moore-
Penrose inverse of A, and
�P(n+m)�(n+k+1)
=
�In �B0 �C
�; �U
�= (0; �U ); (19)
where �U � has the same dimension as �H and �U = (�u1; �u2; :::; �uT )0 is a T � (k + 1) matrix of
observations on �ut. E¢ ciency gains from pooling of observations over the cross section units can
be achieved when the individual slope coe¢ cients, �i, are the same. Such a pooled estimator
of �, denoted by CCEP, is given by
bP =
NXi=1
X 0i�MX i
!�1 NXi=1
X 0i�Myi; (20)
which can also be viewed as a generalized �xed e¤ects (GFE) estimator, and reduces to the
standard FE estimator if �H = � T with � T being a T � 1 vector of ones.
3 Theoretical Properties of CCE Estimators in Nonsta-tionary Panel Data Models
The following assumptions will be used in the derivation of the asymptotic properties of the
CCE estimators.
Assumption 1 (non-stationary common e¤ects): The (n2+m)�1 vector of stochastic commone¤ects, gt = (d
02t;f
0t)0, follows the multivariate unit root process
gt = gt�1 + �gt
where �gt is a (n2 + m) � 1 vector of L2+�, � > 0, stationary near epoque dependent (NED)
processes of size 1/2, on some �-mixing process of size �(2+ �)=�, distributed independently ofthe individual-speci�c errors, "it0 and vit0 for all i, t and t0.
8
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Assumption 2 (individual-speci�c errors): (i) The individual speci�c errors "it and vjt are
distributed independently of each other, for all i; j and t. "it have uniformly bounded positive
variance, supi �2i < K; for some constant K, and uniformly bounded fourth-order cumulants.
vit have covariance matrices, �vi, which are nonsingular and satisfy supi k�vik < K < 1,autocovariance matrices, �iv (s), such that supi
P1s=�1 k�iv (s)k < K <1, and have uniformly
bounded fourth-order cumulants. (ii) For each i, ("it;v0it)0 is an (k+1)�1 vector of L2+�, � > 0,
stationary near epoque dependent (NED) processes of size 2�2��4 on some �-mixing process it
of size �(2 + �)=� which is partitioned conformably to ("it;v0it)0 as ( "it;
0vit)
0 where "it and
vjt are independent for all i and j.
Assumption 3 The coe¢ cient matrices, Bi and Ci are independently and identically distrib-
uted across i, and of the individual speci�c errors, "jt and vjt, the common factors, �gt, for
all i; j and t with �xed means B and C, and uniformly bounded second-order moments. In
particular,
vec(Bi) = vec(B) + �B;i, �B;i v IID (0;B�); for i = 1; 2; :::; N; (21)
and
vec(Ci) = vec(C) + �C;i, �C;i v IID (0;C�); for i = 1; 2; :::; N; (22)
where B� and C� are (k+ 1)n� (k+ 1)n and (k+ 1)m� (k+ 1)m symmetric non-negative
de�nite matrices, kBk < K, kCk < K, kB�k < K and kC�k < K; for some constant K.
Assumption 4 (random slope coe¢ cients): The slope coe¢ cients, �i, follow the random co-
e¢ cient model
�i = � + {i, {i v IID (0;{); for i = 1; 2; :::; N; (23)
where k�k < K, k{k < K, for some constant K, { is a k � k symmetric non-negative
de�nite matrix, and the random deviations, {i, are distributed independently of j;�j,"jt, vjt,and �gt for all i, j and t. {i has �nite fourth moments uniformly over i.
Assumption 5 (identi�cation of �i and �):�X0i�MXi
T
��1exists for all i and T , and limN!1
1N
PNi=1�vi is nonsingular.
Assumption 6�X0iMgXi
T
��1exists for all i and T , and supiE
X0i�MXi
T
2 < K <1.
Assumption 7 When the rank condition (9) is not satis�ed, (i) 1N
PNi=1
X0iMqXi
T 2and � =
limN;T!1
�1N
PNi=1�iT
�, where �iT = E (T�2X 0
iM qX i), are nonsingular. (ii) If m � 2k+1,
9
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
then�X0iMqXi
T 2
��1exists for all i and T and supiE
�X0iMqXi
T 2
��1 �X0iMqF
T 2
� 2 < 1. (iii) If
m < 2k + 1, then E F 0FT 2
2 <1 and E
�F 0FT 2 ��1 2 <1.Remark 1 Assumption 1 departs from the standard practice in the analysis of large panels
with common factors and speci�es that the factors are non-stationary. Assumption 2 concerns
the individual speci�c errors and relaxes the assumption that "it are serially uncorrelated, often
adopted in the literature (see, e.g., Pesaran (2006)). Assumptions 2-6 are standard in large
panels with random coe¢ cients. But some comments on Assumption 7 seems to be in order.
This Assumption is only used when the rank condition (9) is not satis�ed. It is made up of
three regularity conditions.3 The last two are of greater signi�cance and only relate to the
Mean Group estimator presented in the next Section. In e¤ect, these assumptions ensure that
the individual slope coe¢ cient estimators possess second-order moments asymptotically, which
seems plausible in most economic applications.
Remark 2 Note that Assumption 3 implies that i are independently and identically distributed
across i, and
i = + �i, �i v IID (0;�); for i = 1; 2; :::; N; (24)
where � is a m�m symmetric non-negative de�nite matrix, and k k < K, and k�k < K;
for some constant K.
For each i and t = 1; 2; :::; T , writing the model in matrix notation we have
yi =D�i +X i�i + F i + "i; (25)
where "i = ("i1; "i2; :::; "iT )0. Using (25) in (15) we have
bi � �i =�X 0
i�MX i
T
��1�X 0
i�MF
T
� i +
�X 0
i�MX i
T
��1�X 0
i�M"iT
�; (26)
which shows the direct dependence of bi on the unobserved factors through T�1X 0i�MF . To
examine the properties of this component, we �rst note that (2) and (7) can be written in
matrix notations as
X i = G�i + V i; (27)
and�H = (D;�Z) = (D;D�B+ F�C+ �U) = G �P + �U
�; (28)
3E T�2F 0F
2 <1, which is part of Assumption 7(iii), can be established under mild regularity conditions(see Lemma 4 of Phillips and Moon (1999)).
10
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
where �i = (A0i;�
0i)0, V i = (vi1;vi2; :::;viT )
0, G = (D;F), and �P and �U � are de�ned by (19).
Using Lemmas 3 and 4 in Appendix A and assuming that the rank condition (9) is satis�ed,
it follows thatX 0
i�MF
T= Op
�1pNT
�+Op
�1
N
�; uniformly over i; (29)
X 0i�MX i
T� X
0iM gX i
T= Op
�1pN
�; uniformly over i; (30)
andX 0
i�M"iT
� X0iM g"iT
= Op
�1pNT
�+Op
�1
N
�; uniformly over i: (31)
If the rank condition does not hold then by Lemma 6 in Appendix A it follows that
X 0i�MF
T� X
0iM qF
T= Op
�1pN
�; uniformly over i; (32)
X 0i�MX i
T� X
0iM qX i
T= Op
�1pN
�; uniformly over i; (33)
andX 0
i�M"iT
� X0iM q"iT
= Op
�1pNT
�+Op
�1
N
�; uniformly over i: (34)
In the next subsections we discuss our main theoretical results.
3.1 Results for Pooled Estimators
We now examine the asymptotic properties of the pooled estimators. Focusing �rst on the MG
estimator, and using (26) we have
pN�bMG � �
�=
1pN
NXi=1
{i +1
N
NXi=1
�1iT
pNX 0
i�MF
T
! i+
1
N
NXi=1
�1iT
pNX 0
i�M"i
T
!; (35)
where iT = T�1X 0i�MX i. In the case where the rank condition (9) is satis�ed, by (29) we
have pN�X 0
i�MF
�T
= Op
�1pT
�+Op
�1pN
�: (36)
Using this, we can formally show that
pN�bMG � �
�=
1pN
NXi=1
{i +Op
�1pT
�+Op
�1pN
�:
11
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Hence pN�bMG � �
�d! N(0;�MG); as (N; T )
j!1: (37)
The variance estimator for �MG suggested by Pesaran (2006) is given by
�MG =1
N � 1
NXi=1
�bi � bMG
��bi � bMG
�0, (38)
which can be used here as well. The following theorem summarises the results for the mean
group estimator. The result is proven in Appendix B.
Theorem 1 Consider the panel data model (1) and (2). Let assumptions 1-6 and 7(ii),(iii)
hold. Then, for the Common Correlated E¤ects Mean Group estimator, bMG, de�ned by (14),
we have, as (N; T )j!1, that
pN�bMG � �
�d! N(0;�MG);
where
�MG = { +�; (39)
� = limN;T!1
"1
N
NXi=1
�iqT
#: (40)
and �iqT is de�ned in (A28). �MG can be consistently estimated by (38).
Note that this theorem does not require that the rank condition, (9), holds for any number,
m, of unobserved factors so long as m is �xed. Also, it does not impose any restrictions on the
relative rates of expansion of N and T . The following Theorem summarizes the results for the
second pooled estimator, bP . The proof is provided in Appendix B.
Theorem 2 Consider the panel data model (1) and (2), and suppose that Assumptions 1-6 and
7(i) hold. Then, for the Common Correlated E¤ects Pooled estimator, bP , de�ned by (20), as
(N; T )j!1, we have that p
N�bP � �
�d! N(0;��
P );
where ��P is given by
��P = ��1 (�+�)��1 (41)
where
� = limN;T!1
1
N
NXi=1
�Ti
!; � = lim
N;T!1
1
N
NXi=1
�Ti
!; � = lim
N;T!1
1
N
NXi=1
�Ti
!
12
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
�Ti = V ar [T�2X 0iM qX i{i], and �Ti and �Ti are given by (A48) and (A45), respectively.
��P can be estimated consistently by
��P =
��1R���1
; (42)
where
�= N�1
NXi=1
X 0i�MX i
T; (43)
R�=
1
(N � 1)
NXi=1
�X 0
i�MX i
T
��bi � bMG
��bi � bMG
�0�X 0i�MX i
T
�: (44)
Overall we see that despite a number of di¤erences in the above analysis, especially in terms
of the results given in (29)-(34), compared to the results in Pesaran (2006), the conclusions are
remarkably similar when the factors are assumed to follow unit root processes.
Remark 3 The formal analysis in the Appendices focuses on the case where the factor is an
I(1) process and no cointegration is present among the factors. But, as shown by Johansen
(1995, pp. 40), when the factor process is cointegrated and there are l < m cointegrating
vectors, we have that F i = F 1�1i+F 2�2i where F 1 is an m� l-dimensional I(1) process withno cointegration whereas F 2 is an l-dimensional I(0) process. This implies that the cointegration
case is equivalent to a case where the model contains a mix of non-cointegrated I(1) and I(0)
factor processes. Since we know that the results of the paper hold for both non-cointegrated I(1)
and, by Pesaran (2006), I(0) factor processes, we conjecture that they hold for the cointegrated
case, as well. However, we feel that a formal proof of this statement is beyond the scope of the
present paper. We consider a case of cointegrated factors in the Monte Carlo study. The results
clearly support the above claim.
Remark 4 In the case of standard linear panel data models with strictly exogenous regressors
and homogeneous slopes, and without unobserved common factors, Pesaran, Smith, and Im
(1996) show that in general the �xed e¤ect estimator is asymptotically at least as e¢ cient as
the mean group estimator. It is reasonable to expect that this result also applies to the CCE type
estimators, namely that under �i = � for all i, the CCEP estimator would be at least as e¢ cient
as the CCEMG estimator. Although, a formal proof is beyond the scope of the present paper,
the Monte Carlo results reported below provide some evidence in favour of this conjecture.
As we noted above, the whole analysis does not depend on whether the rank condition holds
or not. But in the case where the rank condition is satis�ed, a number of simpli�cations arise.
In particular, the technical Assumption 7 is not needed, and Assumption 3 can be relaxed.
13
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Namely the factor loadings, i, need not follow the random coe¢ cient model. It would be
su¢ cient that they are bounded. Also the expressions for the theoretical covariance matrices
of the estimators change, although crucially the estimators of these covariance matrices do not.
For completeness, we present Corollaries on the theoretical properties of the pooled estimators
when the rank condition holds, below. Proofs are provided in Supplementary Appendix B.
Corollary 1 Consider the panel data model (1) and (2). Assume the rank condition, (9), is
met and suppose that Assumptions 1-6 hold. Then, for the Common Correlated E¤ects Mean
Group estimator, bMG, de�ned by (14), we have, as (N; T )j!1, that
pN�bMG � �
�d! N(0;�MG);
where �MG is given by {. �MG can be consistently estimated by (38).
Corollary 2 Consider the panel data model (1) and (2), and suppose that the rank condi-
tion, (9), is met and Assumptions 1-6 hold. Then, for the Common Correlated E¤ects Pooled
estimator, bP , de�ned by (20), as (N; T )j!1, we have that
pN�bP � �
�d! N(0;��
P );
where
��P = ��1R���1; (45)
R� = limN;T!1
"N�1
NXi=1
�viT
#; (46)
� = limN!1
N�1
NXi=1
�vi
!; (47)
and �viT denotes the variance ofX0iMgXi
T{i. ��
P can be estimated consistently by (42).
3.2 Estimation of Individual Slope Coe¢ cients
In panel data models where N is large the estimation of the individual slope coe¢ cients is likely
to be of secondary importance as compared to establishing the properties of pooled estimators.
However, it might still be of interest to consider conditions under which they can be consistently
estimated. In the case of our set up the following further assumption is needed.
Assumption 8 For each i, "it is a martingale di¤erence sequence. For each i, vit is an k � 1vector of L2+�, � > 0, stationary near epoque dependent (NED) process of size 1/2, on some
�-mixing process of size �(2 + �)=�.
14
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Then, we have the following result. The proof is provided in Supplementary Appendix C.
Theorem 3 Consider the panel data model (1) and (2) and suppose that Assumptions 1, 2(i)
and 3-8 hold. LetpT=N ! 0, as (N; T )
j! 1, and assume that the rank condition (9) issatis�ed. As (N; T )
j!1, bi, de�ned by (15), is a consistent estimator of �i. FurtherpT�bi � �i
�d! N(0;�bi): (48)
A consistent estimator of �bi is given by
�bi =��2i
�X 0
i�MX i
T
��1; (49)
where
��2i =
�yi �X ibi
�0�M�yi �X ibi
�T � (n+ 2k + 1) : (50)
Remark 5 Parts of the above result hold under weaker versions of Assumption 8. In partic-
ular we note that the central limit theorem in (B72), in Supplementary Appendix C, holds if
Assumption 2(ii) holds. However, in this case the asymptotic variance has a di¤erent form as
autocovariances of "itvit enter the asymptotic variance expression. If, then, a consistent esti-
mate of the asymptotic variance is required a Newey and West (1987) type correction needs to
be used. Consistency of this variance estimator requires more stringent assumptions than the
NED assumption 2(ii). It is su¢ cient to assume that ("it;v0it)0 is a strongly mixing process for
this consistency to hold.
Remark 6 It is worth noting that despite the fact that under our Assumptions f t, yit and xitare I(1) and cointegrated, implying that "it is an I(0) process, in the results of Theorem 3, the
rate of convergence of bi to �i as (N; T )j!1 is
pT and not T . It is helpful to develop some
intuition behind this result. Since for N su¢ ciently large f t can be well approximated by the
cross section averages, for pedagogic purposes we might as well consider the case where f t is
observed. Without loss of generality we also abstract from dt, and substitute (2) in (1) to obtain
yit = �0i (�
0if t + vit) +
0if t + "it = #
0if t + �it; (51)
where #i = �i�i+ i and �it = "it+�0ivit. First, it is clear that under our assumptions and for
all values of �i, �it is I(0) irrespective of whether f t is I(0) or I(1). But if f t is I(1), since
�it v I(0), then yit will also be I(1) and cointegrated with f t. Hence, it follows that #i can be
estimated superconsistently. However, the OLS estimator of �i need not be superconsistent. To
see this note that �i can be estimated equivalently by regressing the residuals from the regressions
of yit on f t on the residuals from the regressions of xit on f t. Both these sets of residuals are
stationary processes and the resulting estimator of �i will be at mostpT -consistent.
15
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Remark 7 An issue related to the above remark concerns the probability limit of the OLS
estimator of the coe¢ cients of xit in a regression of yit on xit alone. In general, such a
regression will be subject to the omitted variable problem and hence misspeci�ed. Also the
asymptotic properties of such OLS estimators can not be derived without further assumptions.
However, there is a special case which illustrates the utility of our method. Abstracting from dt,
assuming that k = m and that �i is invertible, and similarly to (51) write the model for yit as
yit = �0ixit +
0i�0�1i (xit � vit) + "it = %
0ixit + &it (52)
where %0i = �0i +
0i�0�1i and &it = "it � 0i�0�1i vit. Note that &it is, by construction, correlated
with vit. The question is whether estimating a regression of the form (52) provides a consistent
estimate of %i. For stationary processes this would not be case due to the correlation between
&it and vit. However, in the case of nonstationary data this is not clear and consistency would
depend on the exact speci�cation of the model. Under the assumptions we have made in this
remark, the estimator of %i would be consistent. However, even in this case it is clear that the
application of the least squares method to (52) can only lead to a consistent estimator of %i and
not of �i. To consistently estimate the latter we need to augment the regressions of yit on xitwith their cross-section averages.
4 Monte Carlo Design and Evidence
In this section we provide Monte Carlo evidence on the small sample properties of the CCEMG
and the CCEP estimators, which are de�ned by (14) and (20), respectively. We consider nine
alternative estimators. The �rst one is the CupBC estimator proposed by Bai, Kao, and Ng
(2009), which is a bias-corrected version of a continuously-updated estimator that estimates
both the slope parameters and the unobserved factors iteratively. The CupBC estimator, as
analysed by Bai, Kao, and Ng (2009), assumes the number of unobserved factors is known
and only considers the case where the slopes are homogeneous.4 In addition, we consider two
alternative principal component augmentation approaches discussed in Kapetanios and Pesaran
(2007). The �rst PC approach applies the Bai and Ng (2002) procedure to zit = (yit;x0it)0 to
obtain consistent estimates of the unobserved factors, and then uses the estimated factors
to augment the regression (1), and thus produces consistent estimates of �. We consider
both pooled and mean group versions of this estimator which we refer to as PC1POOL and
PC1MG. The second PC approach begins with extracting the principal component estimates
of the unobserved factors from yit and xit separately. In the second step yit and xit are regressed
4See Bai, Kao, and Ng (2009), for more details.
16
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
on their respective factor estimates, and in the third step the residuals from these regressions
are used to compute the standard pooled and mean group estimators, with no cross-sectional
dependence adjustments. We refer to the estimators based on this approach as PC2POOL and
PC2MG, respectively. On top of these principal component estimators, we consider two sets
of benchmark estimators. The �rst set consists of infeasible mean group and pooled estimators,
which are obtained assuming the factors are observable (i.e. �zt for the CCE estimators is
replaced by true factor f t). The other set consists of naive mean group and pooled estimators,
which ignore the factor structure. The naive estimators are expected to illustrate the extent of
bias and size distortions that can occur if the error cross section dependence that induced by
the factor structure is ignored.
We report summaries of the performance of the estimators in the Monte Carlo experiments
in terms of average biases, root mean square errors, and rejection probabilities of the t-test for
slope parameters under both the null hypothesis and an alternative hypothesis. For computing
the t-statistics, the standard errors of mean group and pooled CCE estimators are estimated
using (38) and (42), respectively. The standard errors of PC1, PC2, infeasible and naive
estimators are estimated similarly to those of the CCE estimators. The standard errors of the
CupBC estimator is computed following Bai, Kao, and Ng (2009).
4.1 Baseline Design
The experimental design of the Monte Carlo study closely follows the one used in Pesaran
(2006). Consider the data generating process (DGP):
yit = �i1d1t + �i1x1it + �i2x2it + i1f1t + i2f2t + "it, (53)
and
xijt = aij1d1t + aij2d2t + ij1f1t + ij3f3t + vijt, j = 1; 2; (54)
for i = 1; 2; :::; N , and t = 1; 2; :::; T . This DGP is a restricted version of the general linear
model considered in Pesaran (2006), and sets n = k = 2, and m = 3, with �0i = (�i1; 0);
�0i = (�i1; �i2), and 0i = ( i1; i2; 0); and
A0i =
�ai11 ai12ai21 ai22
�; �0i =
� i11 0 i13 i21 0 i23
�:
The observed common factors and the individual-speci�c errors of xit are generated as inde-
pendent stationary AR(1) processes with zero means and unit variances:
d1t = 1; d2t = �dd2;t�1 + vdt; t = �49; :::1; :::; T ,
vdt � IIDN(0; 1� �2d), �d = 0:5; d2;�50 = 0;
17
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
vijt = �vijvijt�1 + {ijt; t = �49; :::1; :::; T;
{ijt � IIDN�0; 1� �2vij
�, vji;�50 = 0;
and
�vij � IIDU [0:05; 0:95] ; for j = 1; 2:
But the unobserved common factors are generated as non-stationary processes:
fjt = fjt�1 + vfj;t, for j = 1; 2; 3, t = �49; ::; 0; ::; T; (55)
vfj;t � IIDN(0; 1); fj;�50 = 0, for j = 1; 2; 3:
The �rst 50 observations are discarded.
To illustrate the robustness of the CCE estimators and others to the dynamics of the
individual-speci�c errors of yit, these are generated as the (cross sectional) mixture of stationary
heterogeneous AR(1) and MA(1) errors. Namely,
"it = �i""i;t�1 + �i
q1� �2i"!it, i = 1; 2; :::; N1, t = �49; ::; 0; ::; T;
and
"it =�ip1 + �2i"
(!it + �i"!i;t�1) , i = N1 + 1; :::; N , t = �49; ::; 0; ::; T;
where N1 is the nearest integer to N=2,
!it � IIDN (0; 1) , �2i � IIDU [0:5; 1:5] , �i" � IIDU [0:05; 0:95] , �i" � IIDU [0; 1] .
�vij, �i", �i" and �i are not changed across replications. The �rst 49 observations are discarded.
The factor loadings of the observed common e¤ects, �i1 and vec(Ai) = (ai11; ai21; ai12; ai22)0 are
generated as IIDN(1; 1) and IIDN(0:5� 4; 0:5 I4) with � 4 = (1; 1; 1; 1)0, respectively, which
are not changed across replications. The parameters of the unobserved common e¤ects in the
xit equation are generated independently across replications as
�0i =
� i11 0 i13 i21 0 i23
�� IID
�N (0:5; 0:50) 0 N (0; 0:50)N (0; 0:50) 0 N (0:5; 0:50)
�.
For the parameters of the unobserved common e¤ects in the yit equation, i, we considered
two di¤erent sets that we denote by A and B. Under set A, i are drawn such that the rankcondition is satis�ed, namely
i1 � IIDN (1; 0:2) ; i2A � IIDN (1; 0:2) ; i3 = 0;
and
E�~�iA
�= (E ( iA) ; E (�i)) =
0@ 1 0:5 01 0 00 0 0:5
1A :
18
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Under set B i1 � IIDN (1; 0:2) ; i2B � IIDN (0; 1) ; i3 = 0;
so that
E�~�iB
�= (E ( iB) ; E (�i)) =
0@ 1 0:5 00 0 00 0 0:5
1A ;
and the rank condition is not satis�ed. For each set we conducted two di¤erent experiments:
� Experiment 1 examines the case of heterogeneous slopes with �ij = 1 + �ij; j = 1; 2,
and �ij � IIDN(0; 0:04), across replications.
� Experiment 2 considers the case of homogeneous slopes with �i = � = (1; 1)0.
The two versions of experiment 1 will be denoted by 1A and 1B, and those of experiment2 by 2A and 2B.Concerning the infeasible pooled estimator, it is important to note that although this esti-
mator is unbiased under all the four sets of experiments, it need not be e¢ cient since in these
experiments the slope coe¢ cients, �i, and/or error variances, �2i , di¤er across i. As a result
the CCE or PC augmented estimators may in fact dominate the infeasible estimator in terms
of RMSE, particularly in the case of experiments 1A and 1B where the slopes as well as theerror variances are allowed to vary across i.
Another important consideration worth bearing in mind when comparing the CCE and the
principal component type estimators is the fact that the computation of the CupBC, PC1 and
PC2 estimators assumes thatm = 3; namely that the number of unobserved factors is known.
In practice, m might be di¢ cult to estimate accurately particularly when N or T happen to
be smaller than 50. By contrast the CCE type estimators are valid for any �xed m and do not
require an a prior estimate for m.
Each experiment was replicated 2000 times for the (N; T ) pairs withN; T = 20; 30; 50; 100; 200.
In what follows we shall focus on �1 (the cross section mean of �i1) and the results for �2, which
are very similar to those for �1, will not be reported. The results for all the estimators consid-
ered are reported in Tables 1. Since the performance of CCE and CupBC estimators dominates
other feasible estimators in most of the designs considered, to save space we do not report the
results of these estimators for the remaining experiments.
4.2 Designs for Robustness Checks
In this subsection we consider a number of Monte Carlo experiment designs that aim to check
the robustness of the estimators to a variety of empirical settings.
19
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
4.2.1 The Number of Factors Exceeds k + 1
In order to show the e¤ect of a di¤erent type of violation of the rank condition from experiment
B, we consider the DGP 1A but an extra factor term i4f4t is added to the right hand side
of the y equation (53), where i4 � IIDN(0:5; 0:2), f4t = f4t�1 + vf4;t, vf4;t � IIDN(0; 1),
f4;�50 = 0. In this case, observe that
E( i;�i)0 =
0@ 1 1 0 0:50:5 0 0 00 0 0:5 0
1Awhose rank is k+1 = 3, which is less than the number of unobserved factors, m = 4. Under this
experiment the number of factors is treated as unknown and estimated, using the information
criterion �PCp2�which is proposed by (Bai and Ng, 2002, pp. 201).5 The information criterion
is applied to the �rst di¤erenced variables with the maximum number of factors set to six. The
results are reported in Table 5. However, recall that the CCE type estimators does not make
use of the number of the factors and is valid irrespective of whether k + 1 is more or less than
m.
4.2.2 Cointegrating Factors
In this design the unobserved common factors are generated as cointegrated non-stationary
processes. There are two underlying stochastic trends given by
f tjt = f tjt�1 + vtfj;t, for j = 1; 2, t = �49; ::; 0; ::; T; (56)
vtfj;t � IIDN(0; 1); f tj;�50 = 0, for j = 1; 2:
Then, this experiment uses the same design as 1A, but the I(1) factors in (53) and (54) arereplaced by
f1t = f t1t + 0:5ft2t + vf1;t, t = �49; ::; 0; ::; T;
f2t = 0:5ft1t + f t2t + vf2;t, t = �49; ::; 0; ::; T;
f3t = 0:75ft1t + 0:25f
t2t + vf3;t, t = �49; ::; 0; ::; T;
vfj;t � IIDN(0; 1); fj;�50 = 0, for j = 1; 2; 3:
The �rst 50 observations are discarded. The results are reported in Table 6.
5PCp2 is one of the information criteria which performed well in the �nite sample investigations reported inBai and Ng (2002).
20
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
4.2.3 Semi-Strong Factor Structure
Chudik, Pesaran, and Tosetti (2010) introduce the notions of weak, semi-strong and strong
factor structures and prove that these di¤erent factor structures do not a¤ect the consistency
of the CCE type estimators with I(0) factors. Here we consider the e¤ect of having semi-strong
factor structure when the factors are I(1). For this purpose, the same DGP of the experiment
1A is used, but all factor loadings in (53) and (54) are multiplied by N�1=2. The results are
reported in Table 7. It is easily seen that when the factors are weak or semi-strong they can
not be consistently estimated by the principal components and this could adversely impact the
estimators of � that rely on the PC�s as estimators of the unobserved factors.
4.2.4 A Structural Break in the Means of the Unobserved Factors
Finally, the results of recent research by Stock and Watson (2008) suggest that the possible
structural breaks in the means of the unobserved factors will not a¤ect the consistency of the
CCE type estimators, as well as the principal component type estimators. In view of this, we
considered another set of experiments, corresponding to the DGPs speci�ed as 1A, but now theunobserved factors are generated subject to mean shifts. Speci�cally, under these experiments
the unobserved factors are generated as fjt = 'jt for t < [2T=3] and fjt = 1+'jt for t � [2T=3]with [A] being the greatest integer less than or equal to A, where 'jt = 'j;t�1 + �jt, and
�jt � IIDN(0; 1), for j = 1; 2; 3: Results are reported in Table 8.
4.3 Results
Results of experiments 1A, 2A, 1B, 2B are summarized in Tables 1 to 4, respectively. We alsoprovide results for the naive estimator (that excludes the unobserved factors or their estimates)
and the infeasible estimator (that includes the unobserved factors as additional regressors) for
comparison purposes. But for the sake of brevity we include the simulation results for these
estimators only for experiment 1A.As can be seen from Table 1 the naive estimator is substantially biased, performs very poorly
and is subject to large size distortions; an outcome that continues to apply in the case of other
experiments (not reported here). In contrast, the feasible CCE estimators perform well, have
bias that are close to the bias of the infeasible estimators, show little size distortions even for
relatively small values of N and T , and their RMSE falls steadily with increases in N and/or
T . These results are quite similar to the results presented in Pesaran (2006), and illustrate
the robustness of the CCE estimators to the presence of unit roots in the unobserved common
factors. This is important since it obviates the need for pre-testing of unobserved common
21
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
factors for the possibility of non-stationary components.
The CCE estimators perform well, in both heterogeneous and homogeneous slope cases, and
irrespective of whether the rank condition is satis�ed, although the CCE estimators with rank
de�ciency have sightly higher RMSEs than those under the full rank condition. The RMSEs of
the CCE estimators of Tables 1 and 3 (heterogeneous case) are higher than those reported in
Tables 2 and 4 for the homogeneous case. The sizes of the t-test based on the CCE estimators
are very close to the nominal 5% level. In the case of full rank, the power of the tests for the
CCE estimators are much higher than in the rank de�cient case. Finally, not surprisingly the
power of the tests for the CCE estimators in the homogeneous case is higher than that in the
heterogeneous case.
It is also important to note that the small sample properties of the CCE estimator does not
seem to be much a¤ected by the residual serial correlation of the idiosyncratic errors, "it. The
robustness of the CCE estimator to the short run dynamics is particularly helpful in practice
where typically little is known about such dynamics. In fact a comparison of the results for
the CCEP estimator with the infeasible counterpart given in Table 1 shows that the former
can even be more e¢ cient (in the RMSE sense). For example the RMSE of the CCEP for
N = T = 50 is 3.97 whilst the RMSE of the infeasible pooled estimator is 4.31. This might
seem counter intuitive at �rst, but as indicated above the infeasible estimator does not take
account of the residual serial correlation of the idiosyncratic errors, but the CCE estimator does
allow for such possibilities indirectly through the use of the cross section averages that partly
embody the serial correlation properties of f t and "it�s.
Consider now the PC augmented estimators and recall that they are computed assuming
the true number of common factors is known. The results in Table 1 bear some resemblance to
those presented in Kapetanios and Pesaran (2007). The bias and RMSEs of the PC1POOL and
PC1MG estimators improve as both N and T increase, but the t-tests based on these estimators
substantially over-reject the null hypothesis. The PC2POOL and PC2MG estimators perform
even worse. The biases of the PC estimators are always larger in absolute value than the
respective biases of the CCE estimators. The size distortion of the PC augmented estimators is
particularly pronounced. Finally, it is worth noting that the performance of the PC estimators
actually gets worse when N is small and kept small but T rises. This may be related to the
fact that the accuracy of the factor estimates depends on the minimum of N and T .
Now consider the CupBC estimator and again recall that it is computed assuming the true
number of common factors is known. Let us begin with discussing results in the case in which
the rank condition is satis�ed, the results of which are reported in Tables 1 and 2. As is evident,
the average bias and RMSEs of CupBC estimator are comparable to those of CCE estimators.
22
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Because of this, the results of CCEMG, CCEP and CupBC estimators only are reported in
Table 2 onwards. In the case of heterogeneous slopes with the rank condition satis�ed, the
RMSEs of the CCE estimator are uniformly smaller than those of the CupBC estimator (as can
be seen from Table 1). This might be expected since the CupBC estimator is designed for the
model with homogeneous slopes. In the case of homogeneous slopes with the rank condition
satis�ed, as can be seen from Table 2, the RMSEs of CCEP estimator are smaller than those
of CupBC estimator when T is relatively small (T = 20 and 30). Turning our attention to the
performance of the t-test, it is apparent that the size of the test based on CupBC estimator is
far from the nominal level across all experiments. This is especially so for experiments where
the slopes are heterogeneous. In these cases, increases in N and T do not seem to help to
improve test performance. Even for homogeneous slope cases, the best rejection probability
result is 14.90% for T = N = 200 in Table 2. In contrast, the size of the t-test based on the
CCE estimators is close to 5% nominal level across all experiments. Tables 3 and 4 provide
the summary of experimental results in the rank de�cient case. For this design, even though
the size of the t-test based on the CupBC estimator is grossly oversized, the RMSEs of the
estimator are smaller than those of CCE estimators. However, note that in these experiments
the number of factors are treated as known, which is rarely expected in practical situation. We
return to this issue below.
Tables 5-8 report the results of the experiments carried out as robustness checks. 6 Table 5
reports the results of the experiments where the number of unobserved factors is four (m = 4)
which exceeds k + 1 = 3, in the case of heterogeneous slopes. In this experiment, CupBC
estimates are obtained supposing that m is unknown but estimated using the information
criterion PCP2, which is proposed by Bai and Ng (2002), applied to the �rst-di¤erences of
(yit; x1it; x2it). We set the maximum number of factors to six.7 Firstly, despite the number of
unobserved factors, m = 4; exceeding the number of regressors and regressand (k+ 1 = 3), the
RMSEs of CCE estimators decrease as N and T are increased, which con�rms the consistency
of the estimators in the rank de�cient case. Furthermore, the RMSEs of CCE estimators
dominate those of the CupBC estimator, except only when T is very large (� 100). We notethat, although not reported for brevity, the size of the t-test based on CCE estimators is very
close to the nominal 5% level, whilst the size distortion of the CupBC estimators is acute for
all cases considered. Tables 6, 7 and 8 report the results of experiments with the same DGP as
6For brevity the size and power of t-tests are not reported in Tables 5-8, since they are qualitatively similarto those in Tables 1-4. For similar reasons, the results for homogeneous slopes and/or rank de�cient cases (forTables 6-8) are not reported. A full set of results is available upon request from the authors.
7For small N and T the information criterion tends to over-estimate the number of the factors in the �rst-di¤erenced data (yit; x1it; x2it), and the estimates tend to four as N and T get larger.
23
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
in Table 1 but where the unobserved factors are cointegrated, factor structures are semi-strong,
and the unobserved factors are subject to mean shifts, respectively. In all of these designs the
CCE estimators uniformly dominate the CupBC estimator in terms of both RMSEs (and the
size of the t-test, which is not reported in the tables). These are consistent with the �ndings
of Chudik, Pesaran, and Tosetti (2010) and Stock and Watson (2008).
5 Conclusions
Recently, there has been increased interest in analysis of panel data models where the standard
assumption that the errors of the panel regressions are cross-sectionally uncorrelated is violated.
When the errors of a panel regression are cross-sectionally correlated then standard estimation
methods do not necessarily produce consistent estimates of the parameters of interest. An in-
�uential strand of the relevant literature provides a convenient parameterisation of the problem
in terms of a factor model for the error terms.
Pesaran (2006) adopts an error multifactor structure and suggests new estimators that
take into account cross-sectional dependence, making use of cross-sectional averages of the
dependent and explanatory variables. However, he focusses on the case of weakly stationary
factors that could be restrictive in some applications. This paper provides a formal extension
of the results of Pesaran (2006) to the case where the unobserved factors are allowed to follow
unit root processes. It is shown that the main results of Pesaran continue to hold in this
more general case. This is certainly of interest given the fact that usually there are major
di¤erences between results obtained for unit root and stationary processes. When we consider
the small sample properties of the new estimators, we observe that again the results accord
with the conclusions reached in the stationary case, lending further support to the use of the
CCE estimators irrespective of the order of integration of the data observed. The Monte Carlo
experiments also show that the CCE type estimators are robust to a number of important
departures from the theory developed in this paper, and in general have better small sample
properties than alternatives that are available in the literature. Most importantly the tests
based on CCE estimators have the correct size whilst the factor-based estimators (including
the one recently proposed by Bai, Kao, and Ng (2009)) show substantial size distortions even
in the case of relatively large samples.
24
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
ReferencesAndrews, D. W. (1987): �Asymptotic Results for Generalized Wald Tests,�Econometric Theory, 3,348�358.
Anselin, L. (2001): �Spatial Econometrics,� in A Companion to Theoretical Econometrics, ed. byB. Baltagi. Blackwell, Oxford.
Bai, J. (2003): �Inferential Theory for Factor Models of Large Dimensions,�Econometrica, 71, 135�173.
Bai, J., C. Kao, and S. Ng (2009): �Panel cointegration with global stochastic trends,�Journal ofEconometrics, 149, 82�99.
Bai, J., and S. Ng (2002): �Determining the Number of Factors in Approximate Factor Models,�Econometrica, 70, 191�221.
Chamberlain, G., and M. Rothschild (1983): �Arbitrage, Factor Structure and Mean-VarianceAnalysis in Large Asset Markets,�Econometrica, 51, 1305�1324.
Chudik, A., M. H. Pesaran, and E. Tosetti (2010): �Weak and Strong Cross Section Dependenceand Estimation of Large Panels,�Working Paper Series 1100, European Central Bank, forthcomingin The Econometrics Journal.
Coakley, J., A. Fuertes, and R. P. Smith (2002): �A Principal Components Approach to Cross-Section Dependence in Panels,�Mimeo, Birkbeck College, University of London.
Conley, T. G., and B. Dupor (2003): �A Spatial Analysis of Sectoral Complementarity,�Journalof Political Economy, 111, 311�352.
Conley, T. G., and G. Topa (2002): �Socio-economic Distance and Spatial Patterns in Unemploy-ment,�Journal of Applied Econometrics, 17, 303�327.
Connor, G., and R. Korajzcyk (1986): �Performance Measurement with the Arbitrage PricingTheory: A New Framework for Analysis,�Journal of Financial Economics, 15, 373�394.
(1988): �Risk and Return in an Equilibrium APT: Application to a new test methodology,�Journal of Financial Economics, 21, 255�289.
Davidson, J. (1994): Stochastic Limit Theory. Oxford University Press.
De Jong, R. M. (1997): �Central Limit Theorems for Dependent Heterogeneous Random Variables,�Econometric Theory, 13, 353�367.
Forni, M., M. Hallin, M. Lippi, and L. Reichlin (2000): �The Generalised Factor Model:Identi�cation and Estimation,�Review of Economics and Statistics, 82, 540�554.
Forni, M., and L. Reichlin (1998): �Let�s Get Real: A Factor Analytical Approach to Disaggre-gated Business Cycle Dynamics,�Review of Economic Studies, 65, 453�473.
Johansen, S. (1995): Likelihood-Based Inference in Cointegrated Vector Autoregressive Models. Ox-ford University Press, Oxford.
25
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Kapetanios, G., and M. H. Pesaran (2007): �Alternative Approaches to Estimation and In-ference in Large Multifactor Panels: Small Sample Results with an Application to Modelling ofAsset Returns,�in The Re�nement of Econometric Estimation and Test Procedures: Finite Sampleand Asymptotic Analysis, ed. by G. D. A. Phillips, and E. Tzavalis. Cambridge University Press,Cambridge.
Lee, K. C., and M. H. Pesaran (1993): �The Role of Sectoral Interactions in Wage Determinationin the UK Economy,�The Economic Journal, 103, 21�55.
Newey, W. K., and K. D. West (1987): �A Simple, Positive Semi-De�nite, Heteroskedasticity andAutocorrelation Consistent Covariance Matrix,�Econometrica, 55, 703�708.
Pesaran, M. H. (2006): �Estimation and Inference in Large Heterogeneous panels with a MultifactorError Structure,�Econometrica, 74, 967�1012.
Pesaran, M. H., T. Schuermann, and S. M. Weiner (2004): �Modeling Regional Interdepen-dencies using a Global Error-Correcting Macroeconomic Model,� Journal of Business Economicsand Statistics, (with Discussions and a Rejoinder), 22, 129�181.
Pesaran, M. H., R. Smith, and K. S. Im (1996): �Dynamic Linear Models for HeterogeneousPanels,� in The Econometrics of Panel Data, ed. by L. Matyas, and P. Sevestre, pp. 145�195.Kluwer.
Pesaran, M. H., and R. P. Smith (1995): �Estimating Long-Run Relationships from DynamicHeterogeneous Panels,�Journal of Econometrics, 68, 79�113.
Phillips, P., and H. R. Moon (1999): �Linear Regression Limit Theory for Nonstationary PanelData,�Econometrica, 67, 1057�1111.
Stock, J. H., and M. W. Watson (2002): �Macroeconomic Forecasting Using Di¤usion Indices,�Journal of Business and Economic Statistics, 20, 147�162.
Stock, J. H., and M. W. Watson (2008): �Forecasting in Dynamic Factor Models Subject toStructural Instability,�in The Methodology and Practice of Econometrics, A Festschrift in Honourof Professor David F. Hendry, ed. by J. Castle, and N. Shephard. Oxford University Press, Oxford.
Stone, R. (1947): �On the Interdependence of Blocks of Transactions,�Supplement of the Journalof the Royal Statistical Society, 9, 1�45.
26
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Appendix A
LemmasProofs of Lemmas are provided in Supplementary Appendix A.
Lemma 1 Under Assumptions 1-4,
�U0 �U
T= Op
�1
N
�(A1)
V 0i�U
T= Op
�1
N
�+Op
�1pNT
�,"0i�U
T= Op
�1
N
�+Op
�1pNT
�; uniformly over i (A2)
F 0 �U
T= Op
�1pN
�,D0 �U
T= Op
�1pN
�(A3)
X 0i�U
T= Op
�1pN
�; uniformly over i (A4)
Q0 �U
T= Op
�1pN
�(A5)
Q0Q
T 2= Op (1) (A6)
Q0Xi
T 2= Op (1) , uniformly over i (A7)
Q0G
T 2= Op (1) (A8)
�H0 �H
T 2= Op(1) (A9)
�H0G
T 2= Op(1) (A10)
�H0"iT
= Op(1), uniformly over i (A11)
�H0V i
T= Op(1), uniformly over i (A12)
�H0Xi
T 2= Op(1), uniformly over i (A13)
�H0 �U
T= Op
�1pN
�: (A14)
Lemma 2 Under assumptions 1-4,
V 0i�U
T= Op
�1pTN
�+Op
�1
N
�uniformly over i: (A15)
Lemma 3 Under Assumptions 1-4 and assuming that the rank condition (9) holds, then
X 0i�MXi
T� X
0iMgXi
T= Op
�1pN
�; uniformly over i; (A16)
X 0i�M"iT
� X0iMg"iT
= Op
�1pNT
�+Op
�1
N
�; uniformly over i: (A17)
27
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Lemma 4 Assume that the rank condition (9) holds. Then, under Assumptions 1-4
X 0i�MF
T= Op
�1
N
�+Op
�1pNT
�, uniformly over i: (A18)
Lemma 5 Under Assumptions 1-4,
X 0iMgXi
T��vi = Op
�1pT
�:
Lemma 6 Under Assumptions 1-4 and assuming that the rank condition (9) does not hold, then
X 0i�MXi
T� X
0iM qXi
T= Op
�1pN
�; uniformly over i; (A19)
X 0i�MF
T� X
0iM qF
T= Op
�1pN
�; uniformly over i; (A20)
X 0i�M"iT
� X0iM q"iT
= Op
�1pNT
�+Op
�1
N
�; uniformly over i: (A21)
Lemma 7 Under Assumptions 1-4 and assuming that the rank condition (9) does not hold,�X 0i�MF
T
��C = Op
�1
N
�+Op
�1pNT
�, uniformly over i: (A22)
Appendix B: Proofs of theorems for pooled estimators
Proof of Theorem 1We know that
�C =
� + ���+
1
N
NXi=1
�i{i; ��
!;
where �� = 1N
PNi=1 �i and � =
1N
PNi=1 i. Substituting this result in (A22) now yields�
X 0i�MF
T
� � + ���+
1
N
NXi=1
�i{i
!= Op
�1
N
�+Op
�1pNT
�, uniformly over i;�
X 0i�MF
T
��� = Op
�1
N
�+Op
�1pNT
�, uniformly over i;
which in turn yieldspNX 0
i�MF
T
� +
1
N
NXi=1
�i{i
!= Op
�1pN
�+Op
�1pT
�, uniformly over i:
But under Assumption 4, 1N
PNi=1 �i{i = Op
�N�1=2�, and therefore
pN�X 0i�MF
��
T= Op
�1pN
�+Op
�1pT
�, uniformly over i: (A23)
We next reconsider the second term on the RHS of (35), which is the only term a¤ected by the fact that rankcondition does not hold. The second term on the RHS in (35) can be written as
�NT �1
N
NXi=1
�X 0i�MXi
T 2
�+ pNX 0
i�MF
T 2
!(� + �i � ��) ; (A24)
28
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
where �� = 1N
PNi=1 �i. By (A19) and (A20) it follows that
�NT �1
N
NXi=1
�X 0iM qXi
T 2
�+ pNX 0
iM qF
T 2
!(� + �i � ��) +Op
�1pN
�: (A25)
Note that for the above two expressions, we have changed the normalisation from T to T 2. This is because inthe case where the rank condition does not hold, the use of cross-sectional averages is not su¢ cient to removethe e¤ect of the I(1) unobserved factors and so X 0
i�MXi, X
0i�MF , X 0
iM qXi and X0iM qF would involve
nonstationary components. Then, since by (A23),pN(X0
i�MF )�
T 2 = Op
�1
TpN
�+ Op
�1
T 3=2
�, uniformly over i; it
is the case that for N and T large
pN�bMG � �
�d� 1p
N
NXi=1
{i +1pN
NXi=1
�X 0iM qXi
T 2
�+�X 0iM qF
T 2
�(�i � ��) : (A26)
We next focus on analysing the RHS of (A26). The �rst term on the RHS of (A26) tends to a Normaldensity with mean zero and �nite variance. The second term needs further analysis. Letting
Q1iT =
�X 0iM qXi
T 2
�+�X 0iM qF
T 2
�and �Q1T =
1N
PNi=1Q1iT , we have that
1pN
NXi=1
Q1iT (�i � ��) =1pN
NXi=1
�Q1iT � �Q1T
��i: (A27)
We note that �i is i.i.d. with zero mean and �nite variance and independent of all other stochastic quantitiesin the second term of the RHS on (A27). We de�ne
Q1iT;�i =
�X 0iM q;�iXi
T 2
�+�X 0iM q;�iF
T 2
�
and �Q1T;�i =1N
PNi=1Q1iT;�i, whereM q;�i = IT�Q�i
�Q0�iQ�i
�+Q0�i,Q�i = G �P�i, �P�i =
�In �B�i0 �C�i
��B�i =
1N
PNj=1;j 6=iBj and �C�i =
1N
PNj=1;j 6=iCj . Then, it is straightforward that�
Q1iT � �Q1T
���Q1iT;�i � �Q1T;�i
�= Op
�1
N
�, uniformly over i;
and1pN
NXi=1
�Q1iT � �Q1T
��i �
1pN
NXi=1
�Q1iT;�i � �Q1T;�i
��i = Op
�1
N1=2
�:
Then, it is easy to show that if zTi = xiyTi, xi is an i.i.d. sequence with zero mean and �nite variance and yTiis a triangular array of random variables with �nite variance then zTi is a martingale di¤erence triangular arrayfor which a central limit theorem holds (see, e.g., Theorem 24.3 of Davidson (1994)). But this is the case here,for any ordering over i, setting yTi =
�Q1iT;�i � �Q1T;�i
�and xi = �i. Using this result, it follows that the
second term on the RHS of (A26) tends to a Normal density if�Q1iT � �Q1T
��i has variance with �nite norm,
uniformly over i, denoted by �iqT , i.e.
�iqT = V ar��Q1iT � �Q1T
��i�: (A28)
In order to establish the existence of second moments, it is su¢ cient to prove that �Q1iT � �Q1T
� , or equiv-alently
�Q1iT;�i � �Q1T;�i� , has �nite second moments. We carry out the analysis for �Q1iT � �Q1T
� . Forthis, we need to provide further analysis of X
0iMqXi
T 2 and X0iMqFT 2 . First, note that Xi can be written as
Xi = QBi1 + SBi2 + V i; (A29)
29
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
where S is the T �m� k � 1 dimensional complement of Q, i.e. Q and S are orthogonal and
F = QK1 + SK2; (A30)
where K1 and K2 are full row rank matrices of constants with bounded norm. Note that if m < 2k + 1, weassume, without loss of generality, that Bi2 has full row rank whereas if m � 2k+1, Bi2 has full column rank.Then,
X 0iM qXi =X
0iM q (QBi1 + SBi2 + V i) =X
0iM qSBi2 +X
0iM qV i =
B0i2S
0M qSBi2 + V0iM qV i +B
0i2S
0M qV i + V0iM qSBi2:
But, it easily follows thatV 0iM qV i
T 2= Op
�1
T
�, uniformly over i;
andB0i2S
0M qV i
T 2= Op
�1
T
�, uniformly over i:
Then,X 0iM qXi
T 2= B0
i2
S0S
T 2Bi2 +Op
�1
T
�, uniformly over i: (A31)
Similarly, using (A30),X 0iM qF
T 2= B0
i2
S0S
T 2K2 +Op
�1
T
�, uniformly over i:
Thus �X 0iM qXi
T 2
�+�X 0iM qF
T 2
�=
�B0i2
S0S
T 2Bi2
�+�B0i2
S0S
T 2K2
�+Op
�1
T
�, uniformly over i:
We need to distinguish between two cases. In the �rst case, m � 2k + 1. Then, it is easy to see that X0iMqXi
T 2
and B0i2S0ST 2 Bi2 have an inverse. Then, by Assumption 7(ii)
�Q1iT � �Q1T
� has �nite second moments. Thecase where m < 2k + 1 is more complicated. Denoting � = T�2S0S and ~Bi2 = �
1=2Bi2, we have
B0i2
S0S
T 2Bi2 = ~B
0i2~Bi2:
Then, noting that�~B0i2~Bi2
�+= ~B
+
i2~B0+
i2 and since in this case Bi2 has full row rank then
~B+
i2 = B0i2
�Bi2B
0i2
��1��1=2;
and we obtain �B0i2
S0S
T 2Bi2
�+= B0
i2
�Bi2B
0i2
��1�S0ST 2
��1 �Bi2B
0i2
��1Bi2: (A32)
Hence �X 0iM qXi
T 2
�+�X 0iM qF
T 2
�= B0
i2
�Bi2B
0i2
��1K2 +Op
�1
T
�, uniformly over i;
and the required result now follows by the boundedness assumption for Bi2 and K2. The assumption that Bi2
has full row rank if m < 2k+1 implies that the whole of S enters the equations for Xi. If that is not the casethen the argument above has to be modi�ed as follows: We have that
Xi = QBi1 + S1Bi2 + V i;
30
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
where S1 is a subset of S. Then,
X 0iM qXi
T 2= B0
i2
S01S1T 2
Bi2 +Op
�1
T
�, uniformly over i:
and the analysis proceeds as above until�X 0iM qXi
T 2
�+�X 0iM qF
T 2
�= B0
i2
�Bi2B
0i2
��1�S01S1T 2
��1�S01S
T 2
�K2 +Op
�1
T
�, uniformly over i:
Then, the required result follows by Assumption 7(iii) which implies that E
�S01S1T 2
��1 <1 and E S01ST 2 <
1, and the boundedness assumption for Bi2 and K2.Thus, in general we have that
pN�bMG � �
�d! N(0;�MG); as (N;T )
j!1;
where�MG = { +�; (A33)
and
� = limN;T!1
"1
N
NXi=1
�iqT
#: (A34)
To complete the proof we have to consider two further issues. First we note that in (A26), we disregard a term
involving�X0
iMqXi
T 2
�+ �X0
iMq"iT
�. In particular we have to prove that
�1
T
�1pN
NXi=1
�X 0iM qXi
T 2
�+�X 0iM q"iT
�= Op
�1
T
�: (A35)
For this, it is enough to show that 1pN
PNi=1
�X0
iMqXi
T 2
�+ �X0
iMq"iT
�follows a CLT. This follows if (i) for any
ordering of the cross-sectional units, X0iMq"iT is a martingale di¤erence and (ii)
�X0
iMqXi
T 2
�+ �X0
iMq"iT
�has
�nite second moments. (ii) follows easily from the argument made in other parts of the appendix about the
existence of moments of�X0
iMqXi
T 2
�+ �X0
iMqFT
�. Then, one has to simply prove (i). We need to show that for
any orderingE(Q�i jQ�i�1) = 0; (A36)
where Q�i =�X0
iMqXi
T 2
�+ �X0
iMq"iT
�: Denote Q��i =
�X0
iMqXi
T 2
�+: Then Q�i = Q
��i
�X0
iMq"iT
�. Now X0
iMq"iT =
1T
PTt=1 st"it where st is a unit root process (cf. the de�nition of S in (A29) above). Then, for (A36) to hold it
is su¢ cient to note that for all t,l, E(Q��i st"itjQ��i sl"i�1l) = 0. This completes the proof of (A35).Finally, we need to show that the variance estimator given by
�MG =1
N � 1
NXi=1
�bi � bMG
��bi � bMG
�0, (A37)
which is consistent. To see this �rst note that
bi � � = {i + hiT +Op�1pN
�+Op
�1pT
�, uniformly over i; (A38)
where
hiT =
�X 0i�MXi
T 2
�+X 0i�M [F (�i � ��) + "i]
T 2; (A39)
31
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
and so
bi � bMG = ({i � �{) +�hiT � �hT
�+Op
�1pN
�+Op
�1pT
�, uniformly over i; (A40)
where �hT = 1N
PNi=1 hiT . Since by assumption {i and hiT are independently distributed across i, then
1
N � 1
NXi=1
�bi � bMG
��bi � bMG
�0= �MG +Op
�1pN
�+Op
�1pT
�;
and the desired result follows.
Proof of Theorem 2As before the pooled estimator, bP , de�ned by (20), can be written as
pN�bP � �
�=
1
N
NXi=1
X 0i�MXi
T 2
!�1 "1pN
NXi=1
X 0i�M(Xi{i + "i)
T 2+ qNT
#; (A41)
where
qNT =1pN
NXi=1
�X 0i�MF
� i
T 2: (A42)
Assuming random coe¢ cients we note that i = � + �i � ��, where �� = 1N
PNi=1 �i. Hence
qNT =1
N
NXi=1
pNX 0
i�MF
T 2
!� +
1pN
NXi=1
�X 0i�MF
T 2
�(�i � ��) :
But by (A23), the �rst component of qNT is Op�
1TpN
�+ Op
�1
T 3=2
�. Substituting this result in (A41), and
making use of (33) and (34) we have
pN�bP � �
�=
1
N
NXi=1
X 0iM qXi
T 2
!�1 "1pN
NXi=1
X 0iM q(Xi{i + "i + F (�i � ��))
T 2
#+ (A43)
Op
�1
TpN
�+Op
�1
T 3=2
�:
Also by Assumption 7, when the rank condition is not satis�ed, 1N
PNi=1
X0iMqXi
T 2 is nonsingular. Further, by(A31),
1
N
NXi=1
X 0iM qXi
T 2=1
N
NXi=1
B0i2
S0S
T 2Bi2 +Op
�1
T
�:
We note that, by assumption 3, Bi2 is an i.i.d. sequence with �nite second moments. Further, by Assumption 7,
it follows that E S0ST 2 2 <1. Hence, T�2B0
i2S0SBi2 forms asymptotically a martingale di¤erence triangular
array with �nite mean and variance and, as a result, T�2B0i2S
0SBi2 obeys the martingale di¤erence triangulararray law of large numbers across i, (see, e.g., Theorem 19.7 of Davidson (1994)) and, therefore, its mean tendsto a nonstochastic limit which we denote by �, i.e.
� = limN;T!1
1
N
NXi=1
�iT
!; (A44)
where�iT = E
�T�2B0
i2S0SBi2
�: (A45)
32
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
But, by similar arguments to those used for the mean group estimator in the case when the rank condition doesnot hold, we can show that
1pN
NXi=1
X 0iM qXi
T 2{i
d! N (0;�) ;
where
� = limN;T!1
1
N
NXi=1
�Ti
!; (A46)
and �Ti = V ar�T�2X 0
iM qXi{i�: Further, by independence of "i across i,
1pN
NXi=1
X 0iM q"iT 2
= Op
�1
T
�:
Further, letting Q2iT = T�2X 0
iM qF and �Q2T =1N
PNi=1Q2iT , we have
1pN
NXi=1
�X 0iM qF
T 2
�(�i � ��) =
1pN
NXi=1
�Q2iT � �Q2T
��i:
Then, similarly to the analysis used above for T�2X 0iM qXi, we have
1pN
NXi=1
�Q2iT � �Q2T
��i
d! N (0;�)
where
� = limN;T!1
1
N
NXi=1
�Ti
!(A47)
and�Ti = V ar
��Q2iT � �Q2T
��i�: (A48)
Thus, overall by the independence of {i and �i, it follows thatpN�bP � �
�d! N(0;��P ); as (N;T )
j!1; (A49)
where��P = �
�1 (�+�)��1 (A50)
proving the result for the pooled estimator. The result for the consistency of the variance estimator followsalong similar lines to that for the mean group estimator.
33
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Table
1:
Sm
all
Sam
ple
Propertie
sof
Com
mon
Correla
ted
E¤ects
Type
Estim
ato
rs
inth
eC
ase
of
Experim
ent
1A
(Hete
rogeneous
Slo
pes
+Full
Rank)
Bia
s(£
100)
Root
Mean
Square
Errors
(£
100)
Siz
e(5%
level,
0:
1=
10
0)
Pow
er
(5%
level,
1:
1=
09
5)
(N,T
)20
30
50
100
200
20
30
50
100
200
20
30
50
100
200
20
30
50
100
200
CC
ET
ype
Est
imato
rs
CC
EM
G20
0.0
5-0
.10
-0.0
30.0
6-0
.07
9.6
77.8
96.7
45.8
75.5
47.2
06.9
07.1
57.9
07.5
511.6
513.0
016.1
017.5
020.1
030
0.0
9-0
.01
-0.0
1-0
.13
0.1
07.6
96.0
95.1
14.5
44.2
26.9
55.3
05.9
06.2
56.3
511.4
014.2
518.0
522.0
526.8
550
-0.1
90.2
2-0
.11
0.1
4-0
.04
5.8
84.6
14.0
13.4
43.1
35.7
05.0
56.6
56.2
05.9
515.1
020.4
025.6
034.1
036.6
5100
0.0
00.0
40.0
40.0
30.0
44.2
53.4
62.8
92.3
32.2
75.7
55.8
55.2
54.9
06.2
023.3
534.3
044.4
056.0
063.2
5200
-0.0
5-0
.02
-0.0
30.0
50.0
03.0
72.4
92.0
11.7
21.5
14.4
05.1
54.9
05.6
05.1
035.5
552.6
568.7
083.6
590.5
0C
CE
P20
0.1
80.0
0-0
.05
-0.0
1-0
.13
8.7
57.6
76.8
56.3
26.2
17.7
08.1
07.3
08.0
57.1
512.7
513.5
016.0
516.8
018.3
030
-0.1
7-0
.12
0.0
9-0
.15
0.1
37.1
05.9
95.3
24.7
84.4
67.5
56.2
56.7
56.6
56.4
512.4
015.0
019.3
020.6
526.9
050
0.0
00.1
8-0
.07
0.1
2-0
.01
5.3
34.5
13.9
73.4
73.2
26.8
06.2
05.9
06.3
56.4
517.4
522.1
526.4
032.9
036.2
5100
0.0
00.0
90.0
30.0
00.0
23.7
83.2
52.8
52.3
42.2
85.7
05.6
55.6
05.1
56.2
528.1
537.4
044.8
055.2
061.7
5200
-0.0
7-0
.04
-0.0
50.0
50.0
02.7
12.2
91.9
51.7
01.5
35.1
04.3
55.0
54.7
04.7
544.7
556.8
070.3
083.5
589.7
5B
ai,
Kao
and
Ng
Prin
cip
al
Com
ponent
Estim
ato
rC
upB
C20
0.6
20.7
00.8
10.7
70.8
711.1
69.8
68.3
57.4
66.9
567.2
564.4
057.9
060.9
565.7
572.0
568.7
565.0
068.5
074.6
030
0.3
50.4
20.7
30.5
90.8
38.9
17.7
06.5
15.6
65.2
866.8
061.0
555.9
555.4
063.3
071.8
569.9
566.3
570.7
577.6
050
0.5
30.6
70.3
30.6
30.5
46.7
76.0
15.0
54.2
03.8
364.4
558.8
551.9
551.3
556.7
077.2
072.6
569.5
576.9
083.4
5100
0.2
10.3
40.3
50.2
80.3
34.8
34.1
53.3
92.7
62.5
564.6
056.4
047.8
543.3
552.6
580.7
080.3
582.5
087.9
092.5
0200
0.1
00.1
00.0
80.2
30.1
73.5
52.9
42.4
52.0
01.6
962.8
552.8
545.0
044.5
048.0
086.6
588.7
090.7
596.6
099.1
0In
feasib
leE
stim
ato
rs
(inclu
din
g1
and2 )
Infe
asib
leM
G20
0.0
1-0
.19
-0.0
80.1
5-0
.08
7.2
16.3
35.6
24.9
84.7
66.4
06.2
06.8
05.9
56.5
012.7
515.3
516.8
519.7
020.4
030
0.0
2-0
.14
0.0
1-0
.02
0.1
25.9
14.9
54.4
33.9
73.8
76.5
05.8
06.0
55.3
05.9
016.1
518.0
523.3
525.2
028.8
050
-0.1
00.0
7-0
.06
0.1
4-0
.04
4.4
83.7
53.3
93.0
92.9
46.4
55.2
55.9
05.2
55.2
021.7
027.3
531.4
538.4
540.2
5100
0.0
10.0
70.0
20.0
00.0
43.1
62.7
82.4
92.1
52.1
45.5
05.1
55.4
54.7
05.4
536.8
546.1
555.1
062.5
066.6
5200
-0.0
70.0
4-0
.07
0.0
60.0
12.2
21.9
31.6
91.5
71.4
44.8
55.0
05.0
05.6
04.7
059.1
572.8
582.2
590.4
092.7
5In
feasib
leP
oole
d20
0.1
5-0
.13
-0.1
5-0
.26
-0.2
17.3
06.9
66.9
27.1
17.4
06.4
06.8
06.6
07.0
05.1
013.7
013.7
514.5
514.1
012.6
530
-0.2
0-0
.15
0.2
2-0
.07
0.2
76.2
35.7
85.7
95.8
96.6
17.0
55.9
07.0
05.2
55.7
015.7
015.3
518.9
516.7
016.6
050
0.1
20.0
7-0
.08
0.2
10.0
24.6
14.4
04.3
14.7
15.0
25.7
05.8
05.5
06.2
55.0
022.2
022.5
523.6
525.5
021.0
0100
-0.0
50.0
70.0
90.0
60.0
03.3
03.2
63.1
23.3
03.5
25.2
55.6
05.2
05.2
05.3
033.4
538.2
038.8
536.7
532.3
0200
-0.0
80.0
6-0
.12
0.0
7-0
.02
2.3
52.2
22.2
02.4
52.4
94.9
54.7
04.5
05.8
54.7
056.1
562.1
059.5
059.0
552.2
0N
aïv
eE
stim
ators
(exclu
din
g1
and2 )
Naïv
eM
G20
22.1
823.1
326.8
229.9
632.6
231.7
632.9
737.3
741.4
947.0
432.0
532.9
534.8
535.4
531.5
041.0
042.6
543.5
041.9
538.0
530
22.2
325.0
628.3
631.3
334.0
130.5
133.3
137.8
741.4
645.3
240.4
544.1
046.6
543.8
539.4
551.0
053.9
557.4
552.2
047.1
550
22.2
123.9
125.6
529.6
133.6
429.7
531.1
232.7
537.7
342.6
655.8
059.3
058.0
059.2
554.7
568.3
070.8
570.3
069.2
065.0
5100
21.9
723.9
226.7
630.0
432.8
828.4
030.0
232.9
736.3
940.0
671.2
075.2
577.9
078.6
075.2
581.0
584.3
585.9
585.8
583.2
0200
22.1
524.0
927.4
930.0
933.2
327.8
729.4
432.8
035.7
139.3
481.8
586.0
087.8
588.0
587.9
588.7
591.9
592.3
092.9
092.0
5N
aïv
eP
oole
d20
25.2
526.6
031.2
733.5
934.8
435.3
037.0
142.6
645.4
247.6
742.1
543.6
547.7
545.2
044.5
052.5
052.6
555.9
553.4
051.9
530
25.7
629.3
932.4
535.3
735.4
635.4
839.1
342.7
045.9
746.8
151.5
556.7
057.6
559.5
556.2
061.0
566.6
066.5
567.7
564.5
550
26.5
428.7
530.3
934.0
135.8
835.6
137.3
939.0
544.0
445.9
364.7
567.1
569.2
570.3
569.3
573.5
576.2
578.2
578.6
577.4
5100
25.8
128.4
731.3
033.1
534.9
134.3
936.7
639.9
041.7
944.2
775.8
578.9
081.3
579.3
080.1
585.1
086.5
588.0
586.6
586.4
0200
25.9
528.3
231.8
933.6
534.1
134.2
036.2
139.6
342.3
942.6
883.4
586.2
587.7
087.4
087.2
089.9
591.9
093.5
592.2
092.2
0
34
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
(Table
1C
ontin
ued)
Bia
s(£
100)
Root
Mean
Square
Errors
(£
100)
Siz
e(5
%le
vel,
0:
1=
10
0)
Pow
er
(5%
level,
1:
1=
09
5)
(N,T
)20
30
50
100
200
20
30
50
100
200
20
30
50
100
200
20
30
50
100
200
Prin
cip
al
Com
ponent
Estim
ato
rs,
Augm
ente
dP
C1M
G20
-12.2
7-1
1.1
5-1
0.3
0-8
.87
-8.9
017.0
914.8
113.2
411.5
111.5
522.5
525.3
530.0
533.4
037.4
012.1
512.9
513.3
012.7
013.7
530
-9.2
5-7
.86
-6.4
6-5
.72
-5.2
513.5
510.8
48.9
87.8
07.1
520.6
020.9
021.6
524.7
524.7
010.7
58.2
57.3
57.4
06.7
550
-6.8
4-5
.05
-3.8
9-3
.01
-3.1
210.1
07.7
95.8
64.6
74.4
719.9
517.6
516.2
514.9
517.9
08.7
08.2
07.6
511.4
09.7
5100
-4.7
8-3
.21
-2.0
3-1
.57
-1.4
57.4
45.3
43.6
82.8
72.7
220.1
016.8
011.4
59.7
511.1
09.5
512.1
520.2
528.8
536.7
5200
-4.3
1-2
.54
-1.3
9-0
.81
-0.7
86.3
94.1
92.6
01.9
31.7
125.2
017.9
510.9
58.1
57.6
513.8
521.9
542.8
567.6
577.1
5P
C1P
OO
L20
-11.9
7-1
1.0
4-1
0.3
5-9
.09
-9.2
315.8
814.3
813.0
711.5
912.0
725.5
028.3
532.0
534.4
538.9
512.0
514.1
014.9
014.5
514.9
030
-8.8
6-7
.66
-6.3
4-5
.73
-5.3
712.4
810.4
58.8
97.8
07.3
421.4
523.7
522.0
524.7
025.5
011.0
08.8
07.5
57.9
56.3
550
-6.2
0-4
.86
-3.8
1-3
.07
-3.1
99.0
67.5
25.7
24.7
34.5
421.4
018.7
516.0
016.0
518.9
08.5
59.5
58.1
010.9
09.6
5100
-4.3
6-3
.00
-2.0
1-1
.60
-1.4
96.6
15.0
13.6
12.8
82.7
421.0
516.8
511.2
59.3
510.8
011.2
514.5
520.8
527.9
036.3
0200
-3.6
2-2
.32
-1.3
6-0
.81
-0.7
95.3
93.8
12.5
11.9
11.7
325.1
517.6
010.5
07.8
07.8
016.3
526.7
545.4
568.0
076.1
5P
rin
cip
al
Com
ponent
Estim
ato
rs,
Orthogonalis
ed
PC
2M
G20
-31.2
6-2
7.0
6-2
4.0
1-2
2.6
7-2
3.1
132.8
328.3
425.0
023.4
423.8
386.5
088.4
591.2
595.2
097.4
074.1
073.9
575.8
082.0
588.2
030
-25.5
0-2
1.2
1-1
8.2
7-1
6.6
9-1
6.3
326.8
222.2
519.1
317.3
516.9
286.8
587.1
089.1
093.3
595.9
570.1
567.8
066.1
069.2
574.7
050
-20.6
5-1
6.2
3-1
3.3
2-1
1.4
1-1
0.8
921.6
817.0
613.9
811.9
511.3
790.1
588.3
588.8
089.0
591.7
070.8
060.2
552.2
045.8
046.1
0100
-16.1
7-1
2.4
4-9
.69
-7.6
1-6
.60
16.8
712.9
710.1
87.9
97.0
293.6
593.3
089.7
587.5
083.3
072.3
556.2
037.6
019.3
013.6
0200
-14.6
1-1
0.7
8-8
.12
-5.7
9-4
.59
15.1
111.1
98.4
56.0
84.8
598.9
597.8
595.4
590.7
583.7
579.6
560.2
033.3
010.0
06.7
5P
C2P
OO
L20
-31.9
7-2
7.4
7-2
4.2
7-2
3.1
8-2
4.1
933.3
928.6
925.2
323.9
924.9
991.0
090.7
093.2
095.5
598.5
080.6
578.6
078.8
083.3
590.4
530
-26.3
2-2
1.5
1-1
8.2
4-1
6.8
3-1
6.7
527.5
322.4
819.1
317.5
117.3
791.3
590.4
089.7
093.3
596.1
578.5
071.8
066.6
570.6
576.9
050
-21.2
2-1
6.3
5-1
3.1
7-1
1.3
5-1
0.9
922.1
017.1
513.8
211.9
111.4
895.0
590.9
088.9
588.2
091.7
079.6
563.8
052.9
546.2
048.2
5100
-16.7
7-1
2.5
2-9
.62
-7.5
5-6
.60
17.4
313.0
610.1
17.9
57.0
397.9
595.0
590.5
086.4
582.3
080.9
060.8
038.1
018.3
014.2
5200
-15.1
6-1
0.9
1-8
.00
-5.6
6-4
.53
15.6
711.3
38.3
45.9
64.7
999.7
598.4
595.9
589.3
582.5
088.6
565.8
533.3
58.4
06.3
0
Note
s:T
he
DG
Pis
=11
+11
+22
+11
+22
+
with
=
¡1
+ (1
¡2)12 ,
=
12
[2],
and
= (1
+
2)¡
12(
+
¡1),
=
[2]+
1
,
»
(01),
2
»
[0515
],
»
[00
509
5],
»
[01].
Regresso
rsare
gen
era
tedby
=11+22+11+33 +
v ,
=12,
for=
12
.1
=12
=052
¡1
+ ,
»
(01
¡05
2),2¡
50
=0;
=¡
1+ ,
»
(01),¡
50
=0,fo
r
=123;v
=v¡
1+ ,
»
(01
¡2),
v¡
50
=0
and
»
[00
509
5]fo
r
=12,fo
r=
¡49
with
the
…rst
50
observ
atio
ns
disca
rded
;1
»
(11);
»
(0505
)fo
r
=12,
=12;11
and23
»
(0505
0),13
and21
»
(005
0);
1
and2
»
(102
);
=1
+
with
»
(000
4)
for
=12.,,,
21,
for
=12,
=12
are
…xed
acro
ssre
plic
atio
ns.
CC
EM
G
and
CC
EP
are
de…ned
by
(14)
and
(20).
CupB
Cis
bia
s-corre
cted
iterate
dprin
cip
al
com
ponen
testim
ato
rof
Bai
etal.
(2009).
The
PC
1and
PC
2estim
ato
rsare
of
Kapeta
nio
sand
Pesa
ran
(2007).
The
varia
nce
estimato
rsof
all
mean
gro
up
and
pooled
estimato
rs(ex
cep
tth
at
of
CupB
C)
are
de…ned
by
(38)
and
(42),
respectiv
ely.
The
PC
type
estimato
rsare
com
puted
assu
min
gth
enum
ber
of
unobse
rved
facto
rs,
=3,
isknow
n.
All
experim
ents
are
base
don
2000
replic
atio
ns.
35
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Table
2:
Sm
all
Sam
ple
Propertie
sof
Com
mon
Correla
ted
E¤ects
Type
Estim
ators
inth
eC
ase
of
Experim
ent
2A
(Hom
ogeneous
Slo
pes
+Full
Rank)
Bia
s(£
100)
Root
Mean
Square
Errors
(£
100)
Siz
e(5
%le
vel,
0:
1=
10
0)
Pow
er
(5%
level,
1:
1=
09
5)
(N,T
)20
30
50
100
200
20
30
50
100
200
20
30
50
100
200
20
30
50
100
200
CC
EM
G20
0.0
5-0
.15
0.0
2-0
.15
0.0
98.4
56.2
95.1
03.7
83.1
47.1
56.4
06.8
06.7
56.8
511.7
013.8
021.7
531.2
547.9
030
-0.1
40.1
20.0
40.0
30.0
06.4
45.1
13.8
02.6
72.0
76.0
56.7
57.2
56.4
06.4
512.7
020.4
530.7
050.9
071.6
050
0.0
8-0
.06
0.0
20.0
50.0
35.0
83.7
92.8
01.9
41.3
96.1
05.9
04.8
55.4
05.3
518.0
026.9
044.4
575.6
595.0
0100
-0.0
4-0
.08
0.0
6-0
.04
-0.0
13.5
92.7
62.0
21.3
50.9
84.5
55.5
06.0
55.1
06.1
028.3
043.0
072.3
595.2
099.9
0200
0.0
6-0
.02
0.0
30.0
10.0
02.8
32.0
51.5
21.0
00.6
85.6
04.4
56.3
55.2
05.7
044.2
067.9
591.9
099.9
0100.0
0C
CE
P20
0.1
80.0
00.0
3-0
.14
0.0
86.9
55.5
64.9
43.9
83.7
46.6
06.7
57.3
06.7
56.8
014.2
516.2
525.2
533.7
046.2
530
-0.1
40.1
40.0
70.0
10.0
15.2
04.5
03.5
52.6
72.2
65.1
05.9
07.2
56.2
56.4
015.2
524.5
534.9
052.9
570.7
050
0.0
50.0
7-0
.02
0.0
40.0
34.0
83.2
92.5
61.8
41.3
95.4
05.4
05.4
56.2
05.3
024.6
034.3
551.7
078.6
595.0
0100
-0.0
2-0
.04
0.0
6-0
.04
-0.0
12.8
72.3
71.7
81.2
40.9
35.6
06.2
06.4
05.2
55.9
541.6
558.3
581.8
597.8
0100.0
0200
0.0
7-0
.03
0.0
10.0
20.0
02.1
71.6
31.3
20.9
20.6
55.6
03.9
55.7
05.6
05.3
565.2
584.4
096.9
5100.0
0100.0
0C
upB
C20
0.1
20.1
00.0
8-0
.01
0.0
18.2
56.1
34.1
42.3
21.2
964.0
052.4
038.2
025.1
518.8
570.4
066.6
565.9
084.7
598.3
530
0.0
40.0
80.0
70.0
2-0
.01
6.4
04.7
33.0
81.7
20.9
661.8
550.0
035.4
023.2
519.1
571.3
071.3
579.3
095.0
099.9
050
-0.0
40.2
2-0
.06
0.0
40.0
34.8
93.5
62.3
11.2
70.7
059.9
049.2
534.4
521.9
015.4
077.2
081.6
088.3
598.8
5100.0
0100
0.0
30.0
10.0
2-0
.05
0.0
13.2
72.4
31.6
60.8
60.4
860.3
048.4
034.4
020.2
517.1
587.1
591.6
597.4
0100.0
0100.0
0200
0.0
70.0
10.0
30.0
30.0
02.4
31.7
31.1
60.6
30.3
359.9
546.6
032.6
020.7
014.9
094.7
097.7
099.8
0100.0
0100.0
0N
otes:
The
DG
Pis
the
sam
eas
that
of
Table
1,
exce
pt
=1
for
all
and,=
12
,
=12.
See
notes
toT
able
1.
Table
3:
Sm
all
Sam
ple
Propertie
sof
Com
mon
Correla
ted
E¤ects
Type
Est
imato
rs
inthe
Case
of
Experim
ent
1B
(H
eterogeneous
Slo
pes
+R
ank
De…cie
nt)
Bia
s(£
100)
Root
Mean
Square
Errors
(£
100)
Siz
e(5
%le
vel,
0:
1=
10
0)
Pow
er
(5%
level,
1:
1=
09
5)
(N,T
)20
30
50
100
200
20
30
50
100
200
20
30
50
100
200
20
30
50
100
200
CC
EM
G20
0.3
3-0
.19
0.2
00.1
40.2
315.0
213.9
012.6
113.3
513.7
86.8
06.9
06.7
56.6
07.2
09.4
08.9
510.1
510.1
510.1
530
0.3
00.1
40.0
9-0
.17
0.3
512.9
112.0
310.7
010.0
710.5
95.5
06.8
05.2
56.1
54.8
08.4
010.0
59.4
510.3
511.6
550
-0.1
50.6
3-0
.20
-0.1
70.0
29.8
28.4
67.8
77.4
27.3
45.8
05.1
06.1
05.7
55.9
09.7
512.9
013.4
014.0
015.2
0100
0.2
50.1
30.2
70.0
00.0
67.0
16.5
55.8
55.2
55.0
15.7
55.9
55.4
55.4
56.1
014.5
017.7
521.6
522.6
527.3
0200
0.0
5-0
.11
-0.1
7-0
.07
-0.0
55.3
54.6
54.1
53.6
13.3
14.8
05.0
54.7
55.1
54.5
519.4
523.7
029.7
537.2
543.4
5C
CE
P20
0.4
80.0
6-0
.04
0.1
60.1
013.1
312.8
112.2
113.5
715.3
06.7
57.4
07.0
06.6
56.7
59.9
010.2
010.4
010.3
510.2
530
-0.2
3-0
.06
0.1
8-0
.25
0.4
311.4
810.7
010.3
99.9
511.0
46.1
06.9
05.7
06.0
05.5
09.0
59.9
510.5
510.2
510.6
050
0.0
00.4
8-0
.18
-0.1
7-0
.02
8.4
27.5
77.2
37.2
27.2
25.2
55.9
06.2
55.3
05.5
011.4
014.0
514.1
514.3
515.2
0100
0.1
10.1
80.2
4-0
.06
0.0
55.8
75.7
25.2
74.8
74.9
85.1
06.0
05.4
04.9
56.0
017.2
519.6
023.5
023.5
527.0
0200
0.0
4-0
.10
-0.1
6-0
.04
-0.0
34.3
53.9
93.7
53.3
03.1
55.4
04.7
05.2
54.1
03.9
525.7
528.5
034.5
041.1
046.0
5C
upB
C20
1.3
40.8
31.0
71.1
21.3
511.2
49.5
28.2
47.5
97.2
467.3
560.2
056.7
060.8
566.8
070.4
566.0
566.0
571.2
576.8
530
0.5
10.8
51.1
40.8
61.2
38.9
77.5
26.4
75.7
85.6
067.4
059.8
055.3
556.9
565.3
572.3
568.9
569.1
572.9
580.2
050
0.5
70.7
00.6
20.9
10.8
16.7
75.8
54.9
84.3
24.0
564.6
557.3
552.4
052.0
059.7
074.9
072.2
570.4
078.6
584.5
0100
0.3
00.4
40.4
50.4
20.4
64.8
64.2
03.4
42.7
62.6
166.4
056.5
548.2
044.0
053.0
079.3
580.4
083.1
089.4
093.6
0200
0.1
40.1
40.1
30.2
70.2
63.5
32.9
92.4
52.0
01.6
964.8
053.3
545.1
543.9
546.9
586.9
088.4
591.2
096.9
099.3
5N
otes:
The
DG
Pis
the
sam
eas
that
of
Table
1,
exce
pt2
»
(01),
soth
at
the
rank
conditio
nis
not
satis…
ed.
See
notes
toT
able
1.
36
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Table
4:
Sm
all
Sam
ple
Propertie
sof
Com
mon
Correla
ted
E¤ects
Type
Estim
ato
rs
inth
eC
ase
of
Experim
ent
2B
(Hom
ogeneous
Slo
pes
+R
ank
De…cie
nt)
Bia
s(£
100)
Root
Mean
Square
Errors
(£
100)
Siz
e(5
%le
vel,
0:
1=
10
0)
Pow
er
(5%
level,
1:
1=
09
5)
(N,T
)20
30
50
100
200
20
30
50
100
200
20
30
50
100
200
20
30
50
100
200
CC
EM
G20
-0.2
8-0
.26
0.4
1-0
.31
0.7
314.4
512.8
512.0
212.0
713.4
77.3
55.4
56.4
06.7
06.0
09.3
59.1
510.9
511.5
510.9
030
-0.1
10.0
70.0
90.4
5-0
.05
11.9
910.7
89.8
29.5
210.3
35.2
05.9
05.9
56.5
06.5
57.8
510.5
012.4
014.3
514.9
050
0.0
00.2
3-0
.07
-0.0
20.0
09.0
17.9
77.6
26.7
96.7
25.0
54.8
05.0
05.4
54.9
59.4
012.2
015.7
517.6
021.1
5100
0.1
4-0
.08
-0.1
2-0
.03
0.0
66.6
65.9
25.1
64.7
84.5
64.6
55.4
05.6
04.6
06.3
515.1
018.1
523.9
528.5
034.8
5200
0.1
40.1
10.0
1-0
.17
-0.0
75.1
34.4
53.8
83.2
73.3
45.4
55.1
05.4
54.6
55.1
522.3
528.8
036.6
044.7
556.7
0C
CE
P20
-0.1
2-0
.19
0.3
5-0
.26
0.6
612.6
611.5
311.5
612.1
215.0
77.4
57.0
07.5
56.3
56.5
09.8
510.0
012.6
012.6
511.5
030
-0.0
90.0
50.0
60.3
90.0
310.0
09.5
79.2
69.3
611.0
55.5
55.7
56.8
06.7
06.7
59.9
011.7
013.3
015.2
014.5
050
-0.1
40.3
9-0
.08
0.0
10.0
37.2
96.9
26.8
46.5
86.7
94.9
55.2
55.4
55.6
04.8
511.2
515.6
016.6
519.9
520.4
0100
0.2
0-0
.13
-0.1
1-0
.05
0.0
45.4
44.9
74.5
54.4
54.3
94.8
05.3
55.4
04.9
56.0
520.6
022.6
528.3
531.4
036.8
0200
0.1
90.1
1-0
.08
-0.1
3-0
.07
3.9
73.7
13.3
52.9
63.0
95.2
55.1
55.0
55.0
05.6
031.9
538.4
544.3
050.7
060.4
0C
upB
C20
0.4
40.3
30.2
90.2
60.2
08.1
16.0
24.1
12.4
11.3
359.6
548.2
534.4
026.1
519.0
069.4
565.9
068.1
585.8
599.3
030
0.1
80.2
20.2
30.1
40.0
96.3
34.6
43.0
41.7
21.0
060.0
548.7
533.8
521.6
020.0
071.4
572.1
579.2
095.1
0100.0
050
0.1
20.3
60.0
30.1
30.0
74.9
03.6
22.3
21.2
90.7
060.9
047.1
032.8
520.0
014.7
577.0
082.2
588.3
598.9
5100.0
0100
0.1
80.0
20.0
9-0
.01
0.0
43.2
32.4
81.6
50.8
60.4
859.6
548.7
033.2
019.8
016.9
087.8
591.1
097.8
5100.0
0100.0
0200
0.1
00.0
30.0
60.0
50.0
22.3
91.7
21.1
70.6
30.3
359.5
045.6
532.1
521.5
015.8
095.0
598.5
099.6
5100.0
0100.0
0N
otes:
The
DG
Pis
the
sam
eas
that
of
Table
1,
exce
pt2
»
(01),
soth
at
the
rank
conditio
nis
not
satis…
ed,
and
=1
for
all
and,=
12
,
=12.
See
note
sto
Table
1.
37
ACC
EPTE
DM
ANU
SCR
IPT
ACCEPTED MANUSCRIPT
Table
5:
Sm
all
Sam
ple
Propertie
sof
Com
mon
Correla
ted
E¤ects
Type
Esti
mators,
The
Num
ber
of
Facto
rs
=4
Exceeds
+
1=
3,
Inth
ecase
of
Hete
rogeneous
Slo
pes
Bia
s(£
100)
Root
Mean
Square
Errors
(£
100)
(N,T
)20
30
50
100
200
20
30
50
100
200
CC
EM
G20
0.2
30.2
90.0
6-0
.23
-0.1
610.9
79.5
98.2
97.6
17.7
030
0.2
00.0
8-0
.07
0.1
4-0
.03
8.9
87.6
56.8
46.4
26.2
950
-0.0
40.0
0-0
.16
-0.1
90.1
46.8
16.0
35.1
24.7
14.6
7100
0.1
2-0
.06
0.0
1-0
.01
0.1
24.8
14.2
53.6
93.5
33.4
6200
0.0
1-0
.04
0.0
3-0
.04
-0.1
03.7
83.0
82.8
42.6
12.5
3C
CE
P20
0.0
90.5
0-0
.02
-0.2
2-0
.11
9.5
78.9
48.0
77.7
07.8
330
0.0
3-0
.05
-0.0
80.0
4-0
.09
7.9
67.2
16.6
06.3
66.2
550
-0.0
4-0
.05
-0.1
3-0
.14
0.1
36.0
65.5
94.8
54.5
44.4
9100
0.0
6-0
.07
-0.0
10.0
10.1
14.2
13.8
53.5
13.3
73.3
8200
-0.0
4-0
.05
0.0
0-0
.03
-0.1
03.1
32.7
42.6
22.4
22.3
7C
upB
C20
0.4
90.3
20.0
60.1
10.1
111.5
610.2
68.9
47.0
96.3
030
0.0
10.1
20.1
20.2
10.0
79.3
87.9
86.6
85.5
84.6
250
-0.1
10.2
5-0
.08
-0.0
20.2
17.0
76.2
95.0
34.0
43.5
4100
0.0
60.0
40.1
00.0
4-0
.04
4.8
14.3
23.5
82.8
22.5
4200
0.0
5-0
.11
0.0
00.0
0-0
.03
3.5
53.1
42.6
12.0
01.6
7N
otes:
Th
eD
GP
isth
esa
me
as
th
at
of
Ta
ble
1,
ex
cep
ta
nex
tra
term
44
isa
dd
ed
to
th
e
eq
ua
tio
n,
wh
ere4
»
(0502
),4
=4¡
1+4
,4
»
(01),4¡
50
=0
.
Fo
rC
up
BC
estim
ato
r,
th
en
um
ber
of
un
ob
serv
ed
facto
rs
istrea
ted
un
kn
ow
nb
ut
estim
ated
by
th
ein
form
atio
ncrit
erio
n
2,
wh
ich
isp
ro
po
sed
by
Ba
ia
nd
NG
(2
00
2).
We
set
th
em
ax
imu
mn
um
ber
of
facto
rs
to
six
.S
ee
als
oth
en
otes
to
Ta
ble
1.
Table
6:
Sm
all
Sam
ple
Properti
es
of
Com
mon
Correla
ted
E¤ects
Type
Esti
mato
rs,
Hete
rogeneous
Slo
pes
and
Full
Rank,
Coin
tegrated
Facto
rs,
inth
eC
ase
of
Experim
ent
1A
(Hete
rogeneous
Slo
pes
+Full
Rank)
Bia
s(£
100)
Root
Mean
Square
Errors
(£
100)
(N,T
)20
30
50
100
200
20
30
50
100
200
CC
EM
G20
0.0
5-0
.05
-0.2
20.0
80.0
09.2
67.8
76.5
85.6
95.2
930
-0.1
40.0
90.0
3-0
.02
0.0
27.3
56.0
25.1
84.5
44.1
650
-0.0
30.1
4-0
.05
0.1
10.1
15.8
54.7
04.0
63.4
93.1
4100
-0.0
5-0
.01
0.0
3-0
.05
0.0
04.1
53.4
02.8
72.4
92.1
9200
-0.0
50.1
40.0
30.0
4-0
.04
3.0
82.4
62.0
21.7
21.5
9C
CE
P20
-0.0
6-0
.01
-0.2
30.0
6-0
.01
8.5
27.5
46.6
55.9
55.6
830
-0.0
6-0
.07
-0.0
7-0
.02
0.0
16.7
85.9
05.2
54.7
04.2
950
-0.0
30.1
4-0
.09
0.1
20.1
35.3
54.5
44.0
53.5
53.1
9100
-0.0
20.0
30.0
6-0
.03
-0.0
23.7
73.1
82.8
42.5
02.2
2200
-0.0
40.1
0-0
.01
0.0
5-0
.04
2.7
02.3
31.9
91.7
21.6
0C
upB
C20
0.5
40.8
50.6
10.6
80.7
811.0
19.5
88.0
16.9
46.3
230
0.6
90.5
20.5
40.5
00.6
88.6
57.4
86.2
65.3
94.9
150
0.4
90.5
40.5
00.5
30.5
86.8
25.6
94.9
94.2
43.7
0100
0.3
30.3
70.3
80.2
20.2
64.6
13.8
63.4
32.8
42.5
2200
0.1
30.3
10.1
30.2
50.0
93.3
92.8
82.4
12.0
31.8
2N
otes:
Th
eD
GP
of
th
esa
me
as
th
at
of
Ta
ble
1,
ex
cep
tth
efa
cto
rs
are
gen
era
ted
as
co
integ
ra
ted
no
n-s
ta
tio
na
ry
pro
cesses:1
= 1
+05 2
+1
,
2
=05 1
+ 2
+2
,3
=07
5 1
+02
5 2
+3
,w
ith
»
(01)
¡
50
=0
,fo
r
=123
,w
here
= ¡
1+
wit
h
»
(01)
for
=12
,
=¡
490
See
als
oth
en
otes
to
Ta
ble
1.
Table
7:
Sm
all
Sam
ple
Propertie
sof
Com
mon
Correla
ted
E¤ects
Type
Esti
mato
rs,
Sem
i-Str
ong
Factors,
inthe
Case
of
Experim
ent
1A
(Hete
rogeneous
Slo
pes
+Full
Rank)
Bia
s(£
100)
Root
Mean
Square
Errors
(£
100)
(N,T
)20
30
50
100
200
20
30
50
100
200
CC
EM
G20
-0.0
9-0
.22
-0.0
70.0
9-0
.09
9.9
28.0
16.5
75.6
35.1
730
0.0
20.0
10.0
1-0
.11
0.1
07.7
46.2
15.1
44.4
34.1
050
-0.1
20.1
6-0
.11
0.1
4-0
.04
5.9
64.5
73.9
93.4
23.1
0100
0.0
10.0
30.0
50.0
20.0
44.2
33.5
12.8
72.3
32.2
6200
-0.0
60.0
1-0
.01
0.0
50.0
03.0
62.4
62.0
01.7
21.5
1C
CE
P20
0.0
9-0
.07
-0.0
60.0
4-0
.12
8.6
47.4
96.3
45.6
55.3
430
-0.1
9-0
.10
0.0
9-0
.08
0.1
37.1
25.9
05.1
24.4
94.2
150
0.0
10.1
3-0
.05
0.1
3-0
.02
5.2
74.4
63.9
33.4
33.1
6100
0.0
40.0
80.0
20.0
00.0
33.7
73.2
82.8
42.3
52.2
8200
-0.0
7-0
.03
-0.0
40.0
50.0
02.6
82.3
01.9
61.7
01.5
3C
upB
C20
0.2
30.4
60.1
70.4
30.4
512.2
910.5
58.0
96.7
55.8
030
-0.2
00.0
90.3
80.2
00.4
99.5
38.0
36.3
95.1
44.5
850
0.3
90.3
70.0
60.2
00.1
57.3
46.0
85.0
73.9
93.4
0100
0.1
80.1
80.0
60.0
50.0
94.9
94.4
03.6
12.6
92.4
5200
0.0
00.0
30.0
30.0
90.0
13.7
73.0
32.5
51.9
81.6
4N
otes:
Th
eD
GP
of
th
esa
me
as
th
at
of
Ta
ble
1,
ex
cep
tth
efa
cto
rlo
ad
ing
s
ma
trix
¡0
ism
ult
ipli
ed
by
¡12
for
all.
See
als
oth
en
otes
to
Ta
ble
1.
38
Table
8:
Sm
all
Sam
ple
Properti
es
of
Com
mon
Correla
ted
E¤ects
Type
Esti
mators,
One
Break
inthe
Means
of
Unobse
rved
Facto
rs,
inthe
Case
of
Experim
ent
1A
(H
eterogeneous
Slo
pes
+Full
Rank)
Bia
s(£
100)
Root
Mean
Square
Errors
(£
100)
(N,T
)20
30
50
100
200
20
30
50
100
200
CC
EM
G20
0.0
1-0
.10
-0.0
20.0
6-0
.07
9.6
67.8
26.7
45.8
75.5
430
0.1
4-0
.03
-0.0
2-0
.13
0.1
07.6
86.0
85.1
14.5
44.2
250
-0.2
10.2
0-0
.11
0.1
4-0
.04
5.9
14.6
44.0
13.4
33.1
3100
0.0
20.0
30.0
50.0
30.0
44.2
63.4
82.8
82.3
32.2
6200
-0.0
8-0
.02
-0.0
20.0
60.0
03.0
82.4
92.0
11.7
21.5
1C
CE
P20
0.1
70.0
0-0
.05
0.0
0-0
.13
8.7
37.6
16.8
66.3
06.2
130
-0.1
5-0
.13
0.0
7-0
.14
0.1
47.1
05.9
85.3
14.7
84.4
650
-0.0
30.1
8-0
.06
0.1
1-0
.01
5.3
04.5
33.9
73.4
73.2
1100
0.0
50.0
90.0
40.0
10.0
23.8
03.2
62.8
52.3
42.2
8200
-0.0
6-0
.04
-0.0
50.0
50.0
02.7
22.2
91.9
51.7
11.5
3C
upB
C20
0.5
20.7
70.7
90.8
00.8
911.1
89.8
78.3
97.5
26.9
730
0.3
20.5
80.7
70.5
80.8
48.9
17.8
06.5
55.6
85.2
750
0.5
80.7
50.3
80.6
10.5
46.7
86.0
15.0
34.1
83.8
2100
0.2
80.3
50.3
80.2
90.3
24.8
54.2
23.4
12.7
52.5
5200
0.1
00.0
80.0
80.2
30.1
73.5
72.9
32.4
42.0
11.6
9N
otes:
Th
eD
GP
isth
esa
me
as
th
at
of
Ta
ble
1,
ex
cep
tth
at
=
for
b23
ca
nd
=1
+
for
¸b23
cw
ith
b
cb
ein
gth
eg
rea
test
integ
er
pa
rt
of
,w
here
=
¡1
+
,
»
(01),
=
123
See
als
oth
en
otes
to
Ta
ble
1.