0% found this document useful (0 votes)

70 views45 pages

GMM Estimation and Inference in Dynamic Panel Data

Uploaded by

RuthLis OM

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

70 views45 pages

GMM Estimation and Inference in Dynamic Panel Data

Uploaded by

RuthLis OM

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 45

GMM Estimation and Inference in Dynamic Panel Data Models with Persistent Data

Author(s): Hugo Kruiniger

Source: Econometric Theory, Vol. 25, No. 5 (Oct., 2009), pp. 1348-1391
Published by: Cambridge University Press
Stable URL: http://www.jstor.org/stable/40388591
Accessed: 12-07-2016 22:03 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact [email protected].

Cambridge University Press is collaborating with JSTOR to digitize, preserve and extend access to
Econometric Theory

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
Econometric Theory, 25, 2009, 1348-1391. Printed in the United States of America.
doi: 10. 1017/S026646660809053 1

GMM ESTIMATION AND INFERENCE

IN DYNAMIC PANEL DATA MODELS
WITH PERSISTENT DATA

Hugo Kruiniger
Queen Mary, University of London

In this paper we consider generalized method of moments-based (GMM-based)

estimation and inference for the panel AR(1) model when the data are persistent
and the time dimension of the panel is fixed. We find that the nature of the weak
instruments problem of the Arellano-Bond (Arellano and Bond, 1991, Review of
Economic Studies 58, 277-297) estimator depends on the distributional properties
of the initial observations. Subsequently, we derive local asymptotic approximations
to the finite-sample distributions of the Arellano-Bond estimator and the System
estimator, respectively, under a variety of distributional assumptions about the initial
observations and discuss the implications of the results we obtain for doing infer-
ence. We also propose two Lagrange multiplier-type (LM-type) panel unit root tests.

1. INTRODUCTION

In this paper we consider generalized method of moments-based (GMM-based)

estimation and inference for the panel AR(1) model yiit = pyi,t-' + (1 - p)ju¿ +
Si j , / = 1 , . . . , N and t = - S + 1 , . . . , 7' when the autoregression (AR) parameter
p is close to or equal to one. Throughout the paper we assume that S > -1, the
first observations occur at t = 1, the time dimension of the panel, T, is fixed,
and the cross-section dimension of the panel, N, is large. Among other things we
derive local asymptotic approximations to the finite-sample distributions of some
well-known linear GMM estimators for this model under a variety of assumptions
about the initial observations and discuss the implications of the results we obtain
for doing inference. We also propose two Lagrange multiplier-type (LM-type)
panel unit root tests.
Persistent data raise at least three issues in connection with GMM estimation
of the panel AR(1) model. First, some of the available moment conditions do
not identify the unit root. For instance, it is well known that the GMM estimator
of Arellano and Bond (1991) (the estimator is henceforth referred to as the AB

I thank Steve Bond and Frank Windmeijer for kindly making one of their computer programs available to me.
I also thank Liúdas Giraitis, three anonymous referees, a co-editor, and the editor for very helpful comments. This
research was funded by the ESRC under grant R000239 139. Address correspondence to Hugo Kruiniger, Department
of Economics, Queen Mary, University of London, Mile End Road, London El 4NS, United Kingdom; e-mail:
[email protected].

1 348 © 2009 Cambridge University Press 0266-4666/09 $ 1 5 .00

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
GMM ESTIMATION IN DYNAMIC PANEL DATA MODELS 1349

estimator) breaks down when p = 1 , because the instruments,which are lagged

levels of the data ()>/,j, s < t - 2), are no longer correlated with the first-
differences of the regressors (Ayijf_i): when p = 1, EÍ^Áy^-i) = 0. How-
ever, moment conditions such as E[AyiiS(yiit - py¡,t-')] = 0 with s < t - 1
still identify the unit root; see Arellano and Bover (1995) (henceforth Arbov).
In what follows we will refer to an estimator that only exploits such moment
conditions as an Arbov estimator. Second, most moment conditions are weak
in some sense when p is close to unity. The finite-sample distributions of the
corresponding estimators may therefore differ substantially from the first-order
fixed-parameter asymptotic distributions. The second-order bias approximation
for the AB estimator derived by Hahn, Hausman, and Kuersteiner (2001) also
becomes inaccurate when p is close to unity. Third, the first-order derivatives
with respect to p of some of the moment conditions are discontinuous at p = 1 : it
is easily verified for t > 2 that lim^i E(y¡j ay¡9Í) = - ^Varfo^) = - 'o2 when
the data are covariance stationary, whereas E( v/,i Av,,,) = 0 when p = '.
Recently, Han and Phillips (2007) proposed estimators for the panel AR(1)
model that avoid the weak moment conditions problem when p is near unity.
However the validity of their strong moment conditions relies on covariance sta-
tionarity of the initial conditions. Phillips and Han (2008) discuss a related least
squares estimator for the time series AR(1) model.
In the paper it is shown that the weakness of the moment conditions implied
by the panel AR(1) model can manifest itself in different ways depending on the
distributional properties of the initial observations. In some cases the first-order
derivatives of the moment conditions with respect to /?, i.e., the Jacobians, tend
to zero when p approaches one. In other cases the moment conditions are weak
because the standard deviations of the first-order derivatives of the underlying
moment functions with respect to p explode when p approaches one. In the first
type of situation the "signal" of the moment functions becomes weak, whereas in
the second type of situation the overwhelming "noise" of the moment functions
drowns their signal when p approaches one (cf. Han and Phillips, 2006). In both
situations doing a form of local asymptotics affords a better approximation to
the finite-sample distribution of the corresponding GMM estimator than doing
conventional first-order fixed-parameter asymptotics.
When S is fixed and p approaches one according to p = 1 - XN~X¡2 as N
grows large, where À > 0, both the means and the standard deviations of the
cross-sectional averages of the first-order derivatives of the AB moment func-
tions with respect to p vanish at a *J~Ñ rate. Thus when S is small one can
obtain local-to-nonidentification approximations to the finite-sample distribu-
tions of AB estimators by applying the local-to-zero asymptotics of Staiger
and Stock (1997). On the other hand, when the data are covariance stationary,
the standard deviations of the first-order derivatives of the AB moment functions
with respect to p explode when p approaches one, whereas their means tend
to a nonzero constant because of the fact that Var(y/f, - /¿¿) = o2/{' - p2). In
this case one can obtain local-to-nonidentification asymptotic approximations

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1350 HUGO KRUINIGER

to the finite-sample distributions of AB estimators by assuming that p ap-

proaches one according to p = I - ÀN~l as N grows large, where again X > 0.
Furthermore, we show that in this case the Arbov estimators and the two-step
"optimal" System estimator (cf. the paper by Arellano and Bover, 1995) have
nonnormal distributions also under such local asymptotics although they are still
consistent.

The cases of fixed S and covariance stationarity, which corresponds to 5 = oo,

represent two extreme cases for the local-to-nonidentification asymptotic approx-
imation to the finite-sample distributions of the AB estimator. Furthermore, under
covariance stationarity and p = 1 - ÀN~l, we have Var(j/5Í) = O(N), which im-
plies that the variance of the data is large when TV is large. This implication may
not be compatible with the data in hand. Therefore we also derive in this paper for
all ¿/ > 0 the local-to-nonidentification asymptotic distributions of the AB estima-
tor under the assumption that both S and N grow large with 5/ Nd -> c* > 0. We
show that ifp approaches one according to p = 1 - À N~8 as N grows large, where
X > 0 and 0 < g < 1, and if 0 < d < g, then Var(^ r) = O(Nd). Note that the vari-
ance of the data in hand can be used to choose a suitable value for d. Moreover, in
this setup E^A^r-i) = O(Nd~8) and [Var( >/,•,, Aj^-i)]1/2 = <9(W(1/2)i/),
where s < t - 2. This means that both the strength of the signal and the noise
of the AB moment functions increase with d, i.e., with S. The value of g for
which the local-to-nonidentification asymptotic distributions of AB estimators
are obtained also increases with d from a value of 1/2 for d = 0 to a maxi-
mum value of 1 for d > 1. Thus a larger value of S strengthens the identifi-
cation of p by the AB moment conditions. The value of g is chosen in such
a way that the means and the standard deviations of the cross-sectional aver-
ages of the first-order derivatives of the AB moment functions with respect to
p either do not vanish or vanish at the same rate when p approaches one ac-
cording to p = 1 - XN~g as N grows large, where k > 0. When d > 1 and
g = 1, both the Arbov and the two-step "optimal" System estimator have the
same nonnormal local asymptotic distributions as under covariance stationarity
when g = 1 .
In the paper we also show that under local (-to-nonidentification) asymptotics
(1) the "optimal" AB estimators are biased downward, (2) the estimators of the
optimal weight matrices for the Arbov estimator and the System estimator and
their asymptotic standard errors are no longer consistent when g = 1 and the
data are covariance stationary or d > 1, (3) the vector of cross-sectional aver-
ages of the Arbov moment functions and the vector of their first-order derivatives
with respect to p are uncorrelated, (4) the two-step "optimal" System estimator
can have a skewed distribution, and (5) the truly optimal Arbov and System es-
timators have the same distribution, which is symmetric. The first result explains
the downward biases of the AB estimator reported in a Monte Carlo study con-
tained in Blundell and Bond (1998). The other results help to explain the fact that
when the data are covariance stationary and persistent, Wald tests based on the
two-step "optimal" System estimator have incorrect size, whereas LM tests that

are based on System estimators that use a restricted estimator of the weight ma-
trix that is optimal under the null have correct size. This paper therefore offers
a theoretical justification for using LM tests in the context of panel AR models.
Finally the paper proposes two LM-type panel unit root tests and studies their
properties.
The paper is organized as follows. In Section 2 we review GMM estimation of
the panel AR(1) model. In Section 3 we derive local asymptotic approximations to
the finite-sample distributions of the AB, Arbov, and System estimators under var-
ious assumptions about the initial conditions, i.e., under various asymptotic plans
for S and N. Section 4 proposes two LM-type panel unit root tests and inves-
tigates their properties both analytically and through Monte Carlo experiments.
Section 5 concludes. Proofs are given in the Appendix.
A few words on notation. We use the symbol - > to signify convergence in dis-
tribution, the symbol plim to signify convergence in probability, and the symbol
q-%' to signify convergence in quadratic mean. To state multi-index asymptotic
results we make use of the following notation (for definitions of the underlying
concepts, see also Phillips and Moon, 1999): 'ims,N-^oo,seq N~] S/Li */,* = xt is

equivalent to limyv->och'iris->oo N~l X/Li */,/ = Xt whereas N~l/2^=ì Xiìt -»

Xt as (S,N -> oo)^ signifies that Af~1/2X/ii xi,t converges in distribution to
Xt sequentially by letting S pass to infinity first and letting N pass to infinity
subsequently. Sometimes we only write lim^^oo N~l Z/Li ^/,/ = Xt instead of
lims5J/v->oo,jé><7 N~x X/Li Xi,t - Xt when it is clear that S passes to infinity first,
e.g., when we have assumed covariance stationarity. Finally, plimyv-» oc, s/w->c
N~l X/Li Xi,t = Xt is an example of a so-called diagonal path probability limit.
In this example both S and N pass to infinity with S/N converging to some con-
stant c. We also make use of indicator functions. For instance, '{d = 0} = 1 if
d = 0 and 1 {d = 0} = 0 if d ^ 0. Finally, the abbreviation PD denotes positive
definite.

2. GMM ESTIMATION OF THE PANEL AR(1) MODEL

2.1. The Panel AR(1) Model

Consider the panel AR(1) model with random effects:

yi,t =pyi,t-'+m,t, (l)

a/,/ = ty +e,'/, where ni = 0 -p)Mi> (2)

for i = 1, . . . , AT and t = -5 + 1, . . . , T. The starting date of the {y,,,}, t = -5,

need not coincide with the date of the initial observations on y, t = 1 , i.e., - S < 1 ;
the number of "individuals," N, is large whereas the number of observations on y
per individual, 7' is fixed. Moreover - 1 < p < 1 .

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1352 HUGO KRUINIGER

We assume random sampling in the cross section.1 The composite error

terms, the v¿j, satisfy the following standard assumptions (cf. Ahn and Schmidt,
1995):2

EO*f)=O and E(ju}) = rf<oo, (3)

E(<?/t,)=0, Efo->fi7/)=0, and (4)

E(e/,,ylW) =0, for t = -S + 1,..., 7' (5)

and

Efo-,Je,-ff) = 0, forjar, s,t = -S + 1, ..., T. (6)

For convenience we also assume that the idiosyncratic errors, the £;,r, are ho-
moskedastic over time:

E(e?,) = <72>0, forf = -S+l,...,7' (7)

However, in the paper we do not consider any moment conditions based on this
assumption. The initial conditions are given by j/5_5 = //,- 4- (1 - p)f/,-s, for / =
1, . . . , N. Note that when p = 1, the initial conditions are equal to the individual
effects, the /¿,- . Finally, we assume that E(<f ®_5) < oo, E(juf) < oo, E(eft) = koa,
and E(e?r) < oo, for t = -5 + 1, . . . , T9 where k is the kurtosis parameter. We
also define a^s = E(yf_s).
Note that the model can be rewritten as yiyt - /// = p (yì,t-' - ju¿) + Siìt and that
the rji disappear (the /// drop out) when p = '. The parametrization rji = (1 -/))///
prevents the individual effects from turning into individual trends at p = 1 and
thereby avoids a discontinuity in the data generating process at p = 1 (cf. Han
and Phillips, 2007). Furthermore, when combined with the assumption a2 < oo,
these restrictions on the rji ensure that under covariance stationarity the means
of the data, i.e., rj¿/(l - p), i = 1, . . . , N, are drawn from a distribution with a
finite variance rather than a variance that tends to infinity when p approaches
one. Without these restrictions, some of our asymptotic results would be different
and/or become discontinuous at p = 1 .
In the paper we consider various versions of the panel AR(1) model that differ
with respect to the assumptions made about the initial observations. Among them
are the following two versions:
(CS) The {y¡it} have reached covariance stationarity at / = 1 when 'p' < 1.
(FS) Fixed 5: thus the {v/,, } have not reached stationarity at t = 1 when 'p' < 1.
Ahn and Schmidt (1997) have shown that given assumptions (3)-(7), the [yitt]
have reached covariance stationarity at t = 1 if and only if the initial observations
satisfy the following assumptions:

E(j/,i -///) =0, E[(tf,i -///)///] =0, and (8)

Var(v/,i -///) =

Note that the FS model does not rule out that mean-stationarity, i.e., assumption
(8), holds.

2.2. GMM Estimators for the Panel AR(1) Model

Arellano and Bond (1991) proposed a GMM estimator that exploits the following
m = (T - l)(T - 2)/2 linear moment conditions:

E[mAB9sAP)i = E[y,',(Ay/f/ -pAyi,t-')] = 0,

forj = l,...,f-2 and i = 3,...,7' (10)

where rriAB,sAP) = yiA&yu ~P&yi,t-') and Ay,-,r = y¡,t -yi,t-'- These mo-
ment conditions are implied by assumptions (4)-(6). Note that they do not identify
the unit root because E(yZ)i A;y/,,_i) = E( j/^e/^-i) = 0 when p = 1.
Arellano and Bover (1995) noted that if mean-stationarity, i.e., assumption (8),
holds also, one can add T - 2 linear moment conditions to those in (10):

E[(y«,r-W,/-i)Ay/fi-i]=0, for/ = 3,...,7' (11)

The latter moment conditions do identify the unit root because E( >>/,,_ i àyij-i)
= E(y,-,f-i £i,t-') = o1 when p = '.
A GMM estimator that exploits the moment conditions in both (10) and (1 1) is
known as a System (SYS) estimator.
The set of moment conditions in ( 1 0) and ( 1 1 ) is equivalent to a set that contains
T - 2 AB-type and m Arbov-type moment conditions:

E[y/fi(A#,,-pAtf,,-i)]=0, forf = 3,...,7' (12)

and

E[mArbov,s,t(p)] = E[(j/,/ -pyi,t-')àyi,s] =0,

fors = 2,...,í-l and / = 3,...,7' (13)

where m Ar bov,s, tip) = (yi,t - W,/-i)Ay/fJ. A GMM estimator that only exploits
the latter m moment conditions will be referred to as an Arbov estimator.
Lety/ = [y/,i ... yi%t' and let Z/ =diag(j/,...,y/r"2) bea (T - 2) x m block-
diagonal matrix. Then we can write the set of AB moment conditions in (10)
as E(ZjAi),-) = 0 where Av¡ = v¡ - o/,-i = [Ao/,3 ... Avij]'. Under our as-
sumptions, E(Au¿ Av¡)/a2 = //, where // = //r-2 is a (7 -2) band-diagonal
matrix with 2's on the main diagonal, - l's on the first sub- and superdiagonal,
and zeros elsewhere. It follows that the AB GMM estimator that uses Wm,ab' =
(N~l X/Li Z¡HZi)~l as weight matrix is an optimal one-step GMM estimator.
This estimator is denoted as pab'> In this paper {W^} denotes an arbitrary se-
quence of PD weight matrices with plimyv^oo^v = W, where W is PD. An AB
estimator based on Wn is simply denoted as pab-

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1354 HUGO KRUINIGER

There exist no feasible optimal one-step weight matrices for the Arbov
and System estimators, except when ajj = 0. Let Ayj = [Aj/,2 ... A >>/,,], let
Z' = y¡, 1 It -2, where It -2 is an (T - 2) identity matrix, let Z77 = diag( Ayf, . . . ,
Ayf~l) be a (T - 2) x m block-diagonal matrix, and let Zf = diag(Z/, Z/7) be
a 2(T - 2) x (r - 2 + ra) block-diagonal matrix. When a2 = 0, optimal one-step
weight matrices for the Arbov and System estimators are given by W^^rboui =
(AT^Z^'Z/7)-1 and Wn,Sysi = (N~l 1^, Z?'AZfy' respectively,
where A is given by

'H C
[C It-i'
where C = Ct-2 is a (T - 2) x (T - 2) matrix with ones on the main diagonal,
- l's on the first subdiagonal, and zeros elsewhere. Note that C = E(uz Av¡)'/a2.
The one-step GMM estimators based on W^^Arbov] and Wn^sysì will be re-
ferred to as the Arbov 1 estimator and the SYS1 estimator, respectively. Finally,
let WN9sYSib = (N-1 Iti Zi'ZfTX-
The "optimal" two-step System (S YS2) estimator is based on the weight matrix
Wn,sys2(Pi) = (N-] Iti Zi''A»'i S/]'[a5 v¡]Zf)-' with Ait = ¿¿¡(pi) =
Ay, - pi A y¡y-' , Si = y¡ - piy¡9-', and, unless stated otherwise, p¡ = p' , where
^1 is an initial V^-consistent estimator for p, i.e., 'ÍÑ(p' - p) - » /C # 0. We
define the "optimal" two-step AB (AB2) and Arbov (Arbov2) estimators and
their weight matrices analogously. The W^^YSk matrices (k = 1,2) can be parti-
tioned as

'WN,SYSk,'' WnjYSIcM
Wu,SYSk =
|_WyV,SrS*,21 WN,SYSk,22_

where the W^^YSkj 1 block corresponds to E(Z7/ Av¡) = 0.

Let the (T - 2) x (T - / - 2) matrix d, be given by d, = [0 IT-î-2Ï- In ad-
dition, let ZfB = [yLilT-2 ZP] = [Z{ Z?l where Z¡> = [d' Ayia d2Ay^ ...
dT-3 Ayij-2] is a (T - 2) x (m - (T - 2)) matrix. Then we can rewrite the set of
m AB moment conditions in E(Z¡Aü¡) = 0 as E(Zf Bf Avi) = 0. Thus there exists
a nonsingular constant matrix KAB such that Z¿ = KABZfBf.
An alternative transformation that can be used for removing individual effects
is the Helmert transformation, which amounts to taking forward orthogonal de-
viations. The Helmert transformation of (u/,2 v¡)' is 5/ = (S/,2 • • -S^r-i)', where
Si$/ = ((r-o/(^-r + i))1/2Kr-i/(r-oiL+i^]', / = 2,...,r-i.
Note that Vitt - 1/(7 - 0lJ=/+i Vi,s is equal to a linear combination of first dif-
ferences of the errors. An advantage of using the Helmert transformation rather
than taking first differences is that it preserves the orthogonality among the errors;
i.e., if E(e/ef0 = ^2h-2, then E(e{e[) = cr2/r_3.

Applying the Helmert transformation to y^ = pj/,/-i + (1 - p)m + £;,/,

t = 2,...,7' yields yitt = >9ji,-i,r-i +£/,*, t = 2,..., 7-1, where j/,-i,r-i =
«r - 0/(r - í + 1))1/2 b/,r-i - l/(r - 0 If",1 >>/,,] . Note that all lagged val-
ues of yij are valid instruments for the ah equation of the transformed system.
If we let z'it=y¡~~1, then the set of AB moment conditions in ( 1 0) is equivalent to
the following set of moment conditions: E(zíjVíj) = 0, t = 2, . . . , T - 1 . That is,
there is a nonsingular matrix KH such that Z' At)/ = KH Zfâ; see Arellano and
Bover (1995). An optimal weight matrix for E(Zj5i) = 0 is [AT1 j£=l (Z¡Zi)]-'
which is a block-diagonal matrix with typical diagonal block equal to [N~] X/Li
fc./Z/r)]"1. Note that Z¡HZ¡ = KHZ¡ZiKHf. It follows that the one-step
AB GMM estimator can be rewritten as a weighted average of T - 2 two-
stage least squares estimators: pab' = {Y^l2^y-',t-'zt{^t^t)~XZfty-'yt-'Yi~x
lj=2 ly-'j-'zt(z'tzt)~iz'tyt] where yt = [y',t ... W,r]', ¡V-u-i =[y',-',t-'
. . . yNt-u-'Y, and Z, = [z'j . • • znjY-

3. ASYMPTOTIC PROPERTIES OF THE ESTIMATORS WHEN THE

DATA ARE PERSISTENT

Here we investigate the asymptotic properties of various AB, Arbov, and System
estimators when p is close to unity under two different asymptotic schemes:
5, Af - » oo sequentially with S - > oo first (the CS model) and S, N -> oo
simultaneously with S/Nd - » c* > 0, where ¿/ > 0. The latter scheme includes
N -> oo, and 5 is fixed (the FS model) as special case for d = 0. Note that the
model assumes that T is fixed.
We will see that the estimators suffer from some kind of weak instruments
problem when p is close to unity, with the exact nature of the problem depend-
ing on the distributional properties of the initial conditions. Thus p is not only
the parameter of interest but also determines the "strength" of the instruments.
Conventional fixed parameter asymptotics yields a poor approximation to the
finite-sample distribution of the estimators when p is close to unity because of
the fact that the weak instruments problem vanishes when Af increases. To ensure
that the weak instruments problem is captured by the asymptotics, we let p tend
to unity as a function of Af when AT -> oo in such a way that asymptotically p
is local-to-nonidentified. That is, our asymptotics will make use of parameter se-
quences such as p = 1 - X N~g where X > 0 and '<g<'. The rate Af£ at which
p tends to unity depends on the distributional properties of the initial conditions.
Note that the parameter sequences p = 1 - 1 N~8 are introduced to obtain better
asymptotic approximations to the finite-sample distribution of the estimators for
p when the instruments are weak and not because of interest in local-to-unit-
root models. Thus the motivation for considering these parameter sequences is
different from that in the time series literature. Moon and Phillips (2000) have
considered consistent panel data estimators for the localizing parameter y < 0 in
p - exp(y / T) assuming that T grows large.

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1356 HUGO KRUINIGER

In the analysis the ratios Var(yift - ///)/cr2 play an important role. The panel
AR(1) model implies the following expression for Var( j/,i - /¿,-) when 'p' < 1:

Var(y/>i -Mi) = p2(1+5) Var(y/,-s - /¿/) + (1 - />2(1+5V 7 '(1 - A (14)

It is easily verified for any 5 and p > 0 that d[Var(;y/,i - jUi)]/dp > 0
and limpai Var(y/fi - Mi) = Var(yits -//;) + 0 + S)ff2. Moreover, lims_>oo
Var( vfji - /ii) = <t2/(1 - p2). It follows that when 5 is large and p tends to one,
the ratio Var(y/i - ¡ii)lo2 becomes large. Indeed, limîlimsôoVariy/,! -
m) = 00. However, when 5 is fixed limî Var( v/j - ju¡) < 00.

3.1. Local Asymptotics for the Covariance Stationary Case

We assume that the CS version of the model applies, i.e., S - > 00 and the
model has reached stationarity at t = 1 when 'p' < 1. In this case the finite-
sample distributions of the AB and System GMM estimators differ considerably
from their first-order fixed-parameter asymptotic distributions when p is close
to one; see, e.g., the Monte Carlo evidence reported in Blundell and Bond
(1998). To understand why this happens, we first examine the underlying moment
conditions.

When the CS model holds, limî Var(y/,r)/cr2 = limî [Var( y¡tt - jui)/a2 +
rf/a2] = limpTi[l/(l - p2)] + a2/(T2 = 00 for t = 1,...,7' Furthermore,
E[(A^_i)2] = 2<72/0 +p) and Ei-Âv^-i) = (1 - /Op'-ÊKy/,, -
Mi + Mi)(yi,s - /*/)]'= <720 - P)P1~2~' 7(1 - P2) when s < t - 2. The latter
equalities imply that 'imp<t'E(dmAB,sj/dp) = limp>piE(- v/Âv/^-i) = tr2/2
for s <t - 2 and limî E(dmArbou,s,t/dp) = limati E(- y,-f/_i Ay,-,,) = -cr2/2
when 5 < / - 1 .

Consider now the simple first-stage regression A yiit-i = tí v/)iS +cy; (5 < í - 2),
which corresponds to a typical AB moment condition. Then we find that
limpîplim^ôo^r = limîplim^ôoŒi^)"1!/^^/,/-! =0. Further-
more, if p = 1 - XN~' the first-stage F-statistic FAB = n2 N X£¿ y2s) / '£/
(Ay/f/-i -îy/tJ)2 = Op(l) and *- = [EÍ^W^Íy/,, Ay/,/-i) = O^"1). The
expression Fab - 1 provides an estimate of the so-called concentration param-
eter, which is a measure of the strength of the instrument(s); see, e.g., Stock,
Wright, and Yogo (2002) for details. The fact that FAb = Op(l) suggests
that there is a potential weak instruments problem. Note that if one would
choose the parameter sequence p = 1 - ÀN~Ì/2, then n = O(N~l/2) but
Fab = Op(N1/2). Thus when the CS version of the model applies and p is suf-
ficiently close to unity, the AB GMM estimator suffers from some sort of weak
instruments problem, albeit not from one of the Staiger and Stock (1997) type
where Fab = Op(') when n = O(N~1/2). In this case the weak instruments
problem arises because limîEiv2^ - > 00, whereas limîEiy/Ây/^-i) =
-<72/2 t¿ 0. Considering the multiple first-stage regressions Ay;,,_i = X^L2
Xkyi,k + (Oi for t = 3, . . . , T leads to the same conclusion because T is fixed.

Next consider the simple first-stage regression y¿yí-' = 7cAyiiS + a>i (s <
t - 1), which corresponds to a typical Arbov moment condition. Clearly, limî
plimw->oo7? = 1/2 7^0. Nonetheless, if p = 1 - kN~' the first-stage F-statistic
FArbov = ^2Î/(Aj/,,)2/I/(^/,/-i -*&yi,s)2 = Op(l), even though n =
0(1). The reason for this finding is that when 5 - > oo, E(yft_{) oc 1/(1 - p) and
hence plim;v->ooN~2Z/(}7,í-i -nàyi,s)2 = cr2/(2k) if p = 1 -kN~K However,
limî E( v/,í_i Av/,5) = <72/2 ^ 0.
The preceding analysis suggests that when the CS model holds and p ap-
proaches one the finite-sample distributions of the AB, Arbov, and System esti-
mators are affected by exploding higher order moments and that one should not
use the local-to-zero asymptotics of Staiger and Stock (1997) to obtain approxi-
mations to these distributions. Under covariance stationarity not only Variyî) a
1/(1 - p) but also Var(y/,iAv/,2) oc 1/(1-/9), VaríyÂy^) oc 1/(1 -p),
Var(y/jA£/,3) oc 1/(1 - p) and Cov(y/,i Ae/,3, y/,iAy/,2) oc 1/(1 - p); see
Lemma A.2 in the Appendix. In this case one can obtain local (-to-nonidenti-
fication) approximations to the distributions of the estimators by choosing a
parameter sequence such that the variances of the sample covariances of the
instruments and the differenced regressors become 0(1), i.e., by choosing
p = '-ÀN~l. Under this sequence Var(yzj) = O(N). A large value of Var(y;5i)
may or may not be an accurate description of the data.3
We have the following local (-to-nonidentification) asymptotic results for the
AB and Arbov estimators.

THEOREM 1. Let the CS model hold, T = 3,andp = '- AN~l with X > 0.
Then, as N -» oo,

ma.-,**;, ^re[g~«[-ivïp),|(_2, -')].

«oVÃW-,»4f, «*«» [^] ~ " [i (?) - (ó l)'
The distribution of the ratio of two, possibly correlated, normal variâtes does
not have finite moments; see, e.g., Fieller (1932). Furthermore, the density of such
a ratio may be unimodal or bimodal. Noting that such a ratio can be rewritten as
ci +C2(a + w)/(b + x) where a > 0, b > 0, c' and q are constants, and w and x
are independent standard normal variâtes, Marsaglia (1965) presents a necessary
and sufficient condition for bimodality in terms of the values of a and b; see
Figure 2 in his paper. For the Arbov estimator in Theorem l(ii) we have a = 0, and
this confirms that its limit density is unimodal. For the AB estimator in Theorem
1 (i) we obtain a=b = y/TJl, and its limit density is bimodal if and only if k > ¿o
with ko « 4. Even if the second mode exists, it is not of practical importance: as
pointed out by Woglom (2001), a second mode of practical importance requires
that the value of b is substantially less than the value of a. For instance, when

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1358 HUGO KRUINIGER

a =b = 2.25, the density at the second mode is only about 0.2% of the density
at the primary mode. A recent discussion of the possible bimodality of the (limit)
density of instrumental variables (IV) estimators can be found in the papers by
Forchini (2006), Hillier (2006), and Phillips (2006).
The local asymptotic results of Theorem 1 have been obtained under the
assumption of covariance stationarity. However, if we take k - 0 in the local
asymptotic distribution of the AB estimator, we obtain its distribution for p = 1 ,
despite the fact that we need to condition on the initial conditions in this case.4
Let p = 1 and T = 3. Then it can easily be shown that as (S, N - > oo)^,5

and

VS(^„-„4|, where[g]~«[(°),(¿ î)]. »O

The Arbov estimator remains consistent under local asymptotics, but its local
asymptotic distribution is nonnormal when p = 1 - XN~g with g > 1. When
g < 1, its local distribution is equal to its first-order fixed-parameter asymp-
totic distribution, whereas if g > 1, N^^8(pArbov -/>)-» Xi/X^g with X$g ~
N('{g = 1J0/2), 1/(22)) andCov(X3,X4g) =0.
The set of m Arbov moment conditions in E(Z¡Ifv¡) = 0 can be restated as
E(Z¡If[Av¡ v¡]') = 0, where Z'l = diag(Z/' Z') is a 2(7 - 2) x m matrix with
Z' = diag(Ay/52, . . . , Ay¡j-'). Let K11 be the nonsingular constant matrix such
that Z¡If[Av¡ v¡Y = KIIZJIfvi. Then we can generalize Theorem 1 to T > 3
using the following lemma.

LEMMA 1 . Let the CS model hold, T >3,andp = '- XN~l with I > 0. Let
N -> oo. Then

(0 (D N-^^Zl'AviÂXsu

N-WZ^zP'AviÂXsiand
N-l%tLiZ¡>'Ayi9-iA Xawith
X5 = (X'5l r52y - tf(0, ¿55), Xe = (X'6l X'62y^ W(¿6, 166), and
E(X5X¿) = ¿56 t¿ 0, where ¡L^y ¿55, I56, and ¿66 ore given in the
proof;
(2) whenT = 3, X5 = X5Ì =a2X'/VJand X6 = X6] =a2X2/'/X where
X' and X2 are defined in Theorem 1.

(ii) (1) N-{'21»=X Zfvi A X7, and

N-x Sf=1 Z/-V/.-1 4 ¿8i ~ A^((<t2//2), (<74/(2A))/) with

X7 = (X'52 X'7iy - W(0, ¿77), ¿77 = oAKllKllt, andEiX-jX'z) = 0,

where Xs = (0' X'slY;
(2) when T = 3,X1 = X71 = (J2X3 and X8 = Xgi = <72X4, w/zere X3 and
X4 are defined in Theorem 1.
In addition, we have the following results concerning the weight matrices:
(Hi) plimAôoAr2!^ (Z/'Z/) = (<T2/(2Ã))I,
plimôoJV-3/2!^, (ZJ'ZP) = 0, and
plimôoAT » I^=1 (Z,D/Z,D) = a2 1.
(iv) plimN^>ooW N,Arbovl -o~2I and

AT' X (¿/"[A? Sn'lA? SÍ1Z/7) 4 <74^"L2

1=1

^ tr4^77 /^/7/ + (a4/(2i))/C2diag(Om_(r-2), /r-2).

THEOREM 2. Leí í/*¿ C5 model hold, T >3,andp = '- ÀN~{ with X > 0.
Létf N -> 00. r/z^w

(i) (1) pABi -p-> X¿Z^X5/X¿S¿ X6.IfT> 3, paäi w asymptotically

biased downward.
(2) assuming that pab exploits E(ZfBfAvi) = 0 in lieu ofE(Z¡Aüi) = 0,

PAB-pl X'6ìWnXSÌ/X'6ìWuX6ì.

(ii) plimN-+oo(ßArbov - P) = 0, a«i/ assuming that pArbov exploits E(Z77/

x [A»; !>;]') = 0 in lieu of E(Z/7/ü,-) = 0,

VÑ(pArbov -P)^ X'iWXn/X'iWX*.

Parts (ii) of Lemma 1 and Theorem 2 imply that because X% is Gaussian and
hence Xj _L Xg, i.e., X7 and Xg are independent, the correct local asymptotic stan-
dard errors of Arbov estimators are given by (E(X¿ WÉ77 WXS/(X^WXS)2))0-5.
Thus when the data are CS the conventional asymptotic standard errors of Arbov
estimators are inconsistent under p = '-AN~l. Moreover, because the truly local
asymptotically optimal weight matrix is proportional to I^1 , the second result in
Lemma l(iv) implies that the conventional estimator of the optimal weight matrix
for the Arbov estimator is also inconsistent under these local asymptotics.6 How-
ever, the Arbov estimator that uses WM,Arbov2(p') as weight matrix is still consis-
tent, although when T > 3 its local asymptotic distribution is different from that of
the truly optimal Arbov estimator. Finally, if W 1 (Xf7 X's)' and p - 1 - kN~l 9
the local asymptotic distribution of the Arbov estimator is symmetric and, pro-
vided T > 3, the Arbov estimator is unbiased under these local asymptotics. In

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1360 HUGO KRUINIGER

contrast, when the data are CS and p = 1 - kN~' E(X$Xf6) / 0; i.e., the cross-
products of the instruments and the regressors in the AB estimator remain corre-
lated with the cross-products of the instruments and the dependent variables when
N - > oo, thereby causing the bias of this estimator.
Note that the set of m + T - 2 System moment conditions in E(Zf '[Av¡ v'Í) = 0
can be rewritten as E(Zf'[Aü¡ v¡]') = 0, where Zf = diag(Z^, Z[) is a
2(T - 2) x (m + T - 2) matrix. We have the following results for the System
estimator.

THEOREM 3. Let the CS model hold, let T > 3, and let p = 1 - ÁN~l with
A>0: Let N -> oo. Let Wsys* = diag(Or_2, Im), where Ot-2 is a (T -2) x
(T - 2) null matrix. Finally, let D = E[diag(e?, . . . , ej~x)fi A^J/a2. Then

(i) plim;v->ooWysrsi =0~2Wsys* and

plimN->ooWM,SYS'b =O~2WsYS*-

(ii) diagCAT1/^, ^-1/2/m)Il^1(zf/[Aüí S/]'[Ai>; ^]Zf)diag(^-1/r_2,

N-V2Im) A a4WsYlS2 with W¡'S2 = (¿)J/f W?YS2 = {W¡2YS2)f =
-(^ICK^D, and Wft^Wrfbau* ™here Orbavi ¿s defined in Lemma 1.
(Hi) plimM-tooipsYSk-p) = O,fork= l,lb,2, and

JÑ(psYSk-p) -4 X'strfXj/X'itrfX*. fork = 1,1ft,

^(psYS2-p)^X'nWSYS2Xi0/X'uWsYS2Xn

with Xio = (^5! X'-j)' and Xn = (0 X^)r, w^re ¿77, Í51, X7> flnrf X8 «^
defined in Lemma 1.
(iv) assuming that psYS exploits E(¿f'[Av¡ v¡]') = 0 in lieu of
E(Zf'[AD>;]') = 0,

PSYS-p^(X'6lWuX5ì+X'SÌW3lX5ì)/X'nWXì2

with X'2 = (X'6l 0 X'siy, where X51, À61, û«J Xgi «r^ defined in Lemma 1.

Part (iv) of Theorem 3 implies that when the data are CS, a System estimator
that uses a weight matrix estimator that has a PD probability limit is inconsis-
tent under p = 1 - ÀN~l. Moreover, the conventional estimator for the optimal
weight matrix for the System estimator, i.e., Wm,sys2ÌP'), and the conventional
asymptotic standard errors of System estimators are also inconsistent under
such asymptotics. Nevertheless, the System estimators that use Wn^ysi and
Wm,sys2ÌP'), respectively, as weight matrix are still consistent under such
asymptotics although the local asymptotic distribution of the two-step System
estimator is asymmetric because E(X5'XfSÌ) ^ 0. The truly optimal System es-
timator has the same local asymptotic distribution as the truly optimal Arbov

estimator, which is symmetric. Thus the local asymptotic distribution of the truly
optimal System estimator is not affected by the AB moment conditions.
The preceding results suggest that Wald tests based on the usual two-step Sys-
tem estimator will have incorrect size when the data are CS and p is close to one
as a consequence of its asymmetric distribution. On the other hand, LM tests that
are based on Arbov or System estimators that use a restricted estimator of the
weight matrix that is optimal under the null, e.g., Wu^rboviip) and W^^YSiip)^
respectively, have a standard normal asymptotic distribution under the null both
when first-order fixed-parameter asymptotics and when local asymptotics are em-
ployed because of the fact that X-j _L X%? Therefore, such LM tests will have
rejection probabilities very close to their nominal size.

3.2. Diagonal Path Local Asymptotics

The results in Section 3.1 are largely based on the fact that if p = 1 - XN~X , then
the sequential limits limA^ooh'ms->oo[Var(;y/5, -Mi)i/N = a1 ßX > 0, for t =
1 , 2, . . . , T. Now suppose that p = 1 - X N~ 1 and that N - » oo, S - > oo simultane-
ously with S/N -> c> 0, where c is a constant; then limyv-^cx),s/7V^c/?2(1+5) =
exp(- 2Xc) < 1 because limp_»oo0 + x/p)p = exp(jc). Using (14) it follows that
lim^^oo,5/^->c[Var(3j/f/ -^,-)]/N = o'' - exp(-2Xc)]/2X > 0. This result
suggests that when 5 is of the same order of magnitude as N and the data are
persistent, then the AB and Arbov moment conditions are still weak. Moreover,
in this case one can still obtain approximations to the distributions of the AB,
Arbov, and SYS estimators by using local asymptotics.
It turns out that we can derive diagonal path local asymptotic results for N,
S -> oo with (S/Nd) -> c* > 0 for any d > 0. Let <?(/l,c) = 1 -exp(-2>lc).
Then we have the following diagonal path local asymptotic results for the AB and
Arbov estimators when 7 = 3.

THEOREM4. LetT = 3 andlet p = 1-XN~* with X > 0 andO < g < '.Let
0 < d < g and Agd = 1/2(1 - d) - (g - d). Furthermore, let N -> oo, S -> oo
simultaneously with S/N -> c > 0, S/N^ -> c+ > 0, and S/Nd -> c* > 0. Let
í(A,c*) = 2A(S+l) ifd = O, q(X,c*) = 2Xc* ifO <d <g, and q(X,c*) =
q(X,c*) if d = g. Finally, let crMy = E(y/,i///) and o] = E( y?j). Then

(i) NAsi(pAB-p)^N(0,2cT2tf/(X2(<Tfiy-tf)2)) if&gd>0 andd = 0t

NA*d(pAB-p)^N(0, 4/(Xq(X, c*))) ifAgd > 0 andd>0,

fiAB P 4 p + ii-, with L X - N ( 9 9 ) ,

fiAB P p + *2 with W - N ['X{a,y-a2y)/a2] 9 9 )

(y2/2)(Í, "/)] if^8d=0andd = 0f

?tBÍP+k, ^[^^[.^(^^(^ -•)]

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1362 HUGO KRUINIGER

ifAg¿ = 0 and d > 0,

»>+%■ -[;]-" [(S)- i(-i '>)] ifA"<o-

(ii)VÑ(pArbov-p)-^ £, with A

Note that Agd = 1/2(1 +d) -g and that Ag£/ < 1/2. When the values of d and
g are such that Agd = 0, p is weakly identified by the AB moment condition(s).
When Agd > 0, p is nearly weakly identified, and when A¿¿ < 0, p is nearly
nonidentified by the AB moment condition(s).
When 0 < d < 1 and Ag£/ = 0, there are two extreme cases: if d = 0, then one
obtains the local-to-nonidentification asymptotic distribution of pab for g = 1/2.
This case corresponds to local-to-zero asymptotics; cf. Staiger and Stock (1997).
On the other hand, if d = 1 , then one obtains the local-to-nonidentification asymp-
totic distribution of pab for g = 1. When Agd = 0 and the value of d increases
from zero to one, both the signal, E(y¡t' Ay/2) = O{Nd~g), and the noise of the
AB moment function, (Var(y/i Ay/,2))1^2 = O(N^^d), become stronger.
Some of the diagonal path local (-to-nonidentification) asymptotic results of
Theorem 4 are very similar to the sequential local asymptotic results of Theorem
1, which were obtained for g = 1. If d = 1 and Agd = 0 (so that g = 1), then
q(À, c*) = q(À, c) and the sequential local (-to-nonidentification) asymptotic dis-
tributions of pab and pArbov as (5, Af - » 00)^ could be obtained from Theorem 4
by letting c - > 00 so that q(À, c*) = q(À, c) = q(À, c+) = 1 . Furthermore, when
d > 1 diagonal path local (-to-nonidentification) asymptotic distributions of pab
and PArbov are still obtained for g = 1, and they can also be shown to be equal
to the sequential local (-to-nonidentification) asymptotic distributions of pab and
PArbov as (5, Af - > cx))5^ for g = 1. On the other hand, the first-order large Af
fixed S asymptotic distribution of pArbov for p = 1 can be obtained from Theo-
rem 4 by assuming that c+ = 0, i.e., d < g < 1, so that q(Á, c+) = q(À, c) = 0.
Finally, note that ôq(Á, c*)/(dc*) > 0.
If Ag¿ = 0 and d = 0, the limit density of pab given in Theorem 4 is bi-
modal if and only if Mcr^y - Gy'/{ay(r) > 2, whereas if Ag¿ = 0 and d > 0,
the limit density of pab is bimodal if and only if y/Xq(À9c*)/2 > 2. In both
cases the second mode, if it exists, is not of practical importance. When Agd <
0, the limit density of pab is unimodal. The limit density of PArbov is always
unimodal.

To see that the preceding diagonal path local asymptotic results can yield rea-
sonable approximations, consider the following example: let p = 0.95, Af = 100,
and S = 10. Choose d = g = 1. Then k = 5, c = 1/10, and q(k,c)N/(2k) =
(1 -exp(-l)) x 10 = 6.32, whereas (1 -p2(1+5>)/(l - p2) = 6.93.

In the remainder of this section it is assumed that plim^.^ ts/Nd->c* ^n = W

and a/Ñ(pi - p) - » K / 0 as N, S - > oo with S/Nd - > c*, and plimyv->oo is short
for plim^^oo s/Nd^>c*- The following lemma and two theorems extend some of
the diagonal path local (-to-nonidentification) asymptotic results of Theorem 4 to
T > 3 and to the System estimator.

LEMMA 2. Let T > 3, 0 < d < 1, and g = 1/2(1 +d). Furthermore, let p =
1 - i Af~~£ vv/i/z A > 0. Let N - > oo, 5 - » oo simultaneously with S/N - » c > 0 and
S/#¿ _> c* > o. Lei ^(A, c*) = 2/1(5 + 1) i/d = 0, q(l,c*) = 2Àc* ifO < d < 1,
andq(l9cm)=q(X,c*) ifd=l. Finally, let oj = o2(G2q(k,c*)/(2X)+ '[d =
0]als). Then

(i) (1) tf (-'/Wi-KO^ zj'Avi 4 X51,

N-]^=lZPfAviAX52,and

X5 = (¿m X'52Y - W(0, Í55), Xe = (X'6l X'62y~- N(Ji69 i6el and

E(X5X¿) = I56 # 0, where ¡i^ £55, £56, and £ô6 «^ given in the
proof;
(2) when T = 3, X5 = X51 = ít2Xi anJ X6 = X6' = o2Xly where X' and
X2 are defined in Theorem 4.

(ii) (1) aH/2^! Zf-'vi 4 X71 and

N-llLzt'yi,-i ■*> X*i ~N(<T2(l-q(X,c+)/2)l,
(<r4q(X,c)/(2X))I)with
X7 = (X'52 X'7iY ~ AT(O, Ê77), ¿77 = oAK"K"', andEiXjX'z) =0,
wAereX8 = (0'X^)';
C2; when T = 3, X7 = X71 = a2X3 a«d X8 = Xgi = (72X4, where X3 anrf
X4 are defined in Theorem 4.
In addition, we have the following results concerning the weight matrices:
(Hi) plim/ôoAr(1+rf) X,"=1 (Z!'Z!) = (o}/o2)I,
plimôo^-'/2)^) I,w=1 (Z''Zf) = 0, and
plimôoAT1 ifL^zP'Zf) = a2 1.
(iv) p'imNôoWNjArboü' -o~2l and

N-> £ (Z¡"[Á?, S¡]'m S¡ÍZ¡') S o*W-Albml

/=1

= oAKu K11' + (GA/(2k))q(X,c)K2àmg(Om-(T-2)jT-2).

Proof. See the complete version of the paper available at http://webspace.qmul.

ac.uk/hkruiniger/. B

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1364 HUGO KRUINIGER

THEOREM 5. Let T > 3, 0 < d < 1, and g = 1/2(1 + d). Furthermore, let
p = 1 -AN'8 with X > 0. Let N -> oo, S -> oo simultaneously with S/N -> c > 0
and S/Nd -> c* > 0. 77i£?/z

(i) (1) Pab' -pi X'6i^X5/X'6is5lX6. IfT > 3, A**, Í5 asymptotically

biased downward.

(2) assuming that pab exploits E(ZfBfAvi) = 0 in lieu ofE(Z¡ av¡) = 0,

PAB-pl X'hWXu/X'hWXh

with Xl3 = (X'5] '{d = 0}X'52Y and XH = {X'bX '[d = 0}X'62)'.

(ii) p'irriN-*oo(PArbov -p)=0, and assuming that pArbov exploits E(Z/7'
x [ Ai>; ufi) = 0 in lieu of E(Z/7/o/) = 0,

yíÑ (PArbov -p)l X'sWX7/X'iWX*.

Proof. See the complete version of the paper available at http://webspace.qmul.

ac.uk/hkruiniger/.
When 0 < d < 1 and g = 1/2(1 +d), both pabì and pab2 are inconsistent.
Moreover, plimN^0O9s/N<'-+c*(PAB2-PAB') ^ 0 because Avi(pAB') = Ai>/ -
(Pabi -p)Ay/,-i andpABi - P 1 Xf6±^X5/Xf6Í^ X6¿0. ■
THEOREM 6.LetT>3,0<d<',andg='/2(l+d). Furthermore, let p =
1 - À N~8 with À > 0. L^r yv - > oo, 5 - > oo simultaneously with S/N - > c > 0 fl«J
5/^ -> c* > 0. Letq(À9c*) = 2Ã(S+ 1) í/í/ = 0, q(À, c*) = 2Ác* ifO<d < 1,
andq(Á9c*) = q(X,c*) if d = 1. Let WSYS* = diag(l{¿ = 0}a-2H~!_2, (J-2Im).
Finally, let D = E[diag(e?, . . . , ej~x)fi ae¡]/a2. Then

(i) plimyv->ooWysrsi = WSys* and

plÍmw->ooWysKSl¿> = WsYS*'
(ii) diagíA^-1/2)^) /r.2, N-V2Im) I^=1(zf [Adì S/l' [Aof ?/] ¿f)
diagíyVÍ-1/2^1^)^^,^-1/2^) -4 (74^5-;52 W/ÍA (74^52 ^(72((72X
q(k,c*)/(2X)+ '{d = 0}o2_s)H, W2lYS2 = (Wls2S2)' = -(l/(2A))^(A,c)
x KK11 D, am/ VV2252 ^ Wrf^, where wrf^ is defined in Lemma 2.
(Hi) p'imN->oo(psYSk-p)=0>fork= 1, 1¿?,2, an¿/

y/Ñ(psYSk-p) -4 ^I^Xv/^i^Xg, /or* = 1, lfc,

VÑ(pS.ys2 -p)^ X' ! Wsmí 10/^í 1 ^S2*i !

wiíA Xio = (X'5X X'7y and Xn = (0 X¿)7, wA^ £77, X51, Xlt and X8 are
defined in Lemma 2.

(iv) assuming that psYS exploits E(Zf'[Av¡ v¡]) = 0 in lieu of E(Zf'

x[Av¡v¡])=0,

Nl/2V-dHpsYS-p)^ X[6WXÌ5/X'6WXÌ6

with Xl5 = (Xf5{ '{d = 0}X'7Y and Xl6 = ('{d = l}X'6l X'S)', where X5i,
X(,', Xj, and X% are defined in Lemma 2.

Proof. See the complete version of the paper available at http://webspace.qmul.

ac.uk/hkruiniger/. ■
When d > 1 the diagonal path local (-to-nonidentification) asymptotic proper-
ties of the AB, Arbov, and System estimators mentioned in Theorems 5 and 6 are
equal to their sequential local asymptotic properties given in Theorems 2 and 3.
Note that when d > 1 and g = 1 the estimators of the optimal weight matrices
for the Arbov estimator and the System estimator and the asymptotic standard
errors of the latter estimators are inconsistent and the local asymptotic distribution
of the two-step "optimal" System estimator is skewed.
We now consider the diagonal path asymptotic distributions of the AB and
Arbov estimators for p = 1 and T = 3.

THEOREM 7. Let T = 3 and let p = 1. Furthermore, let N -> oo, S -» oo

simultaneously with S/Nd -> c* > 0 and S/Nd -* c > 0 where d > 0 and d =
max(¿, 1). Finally, let '{d <'}='ifd<' and let '{d <'}=0ifd> 1. Then

wa.-,4fc. ^[ì;]-[(2).(_2, V)]-

w«»/(fc.-,>4£. •*»[!;] ~«[(lh,os,,).(¿ ")]■

When p = 1 and T = 3, the diagonal path asymptotic distribution of the AB
estimator is equal to both its first-order large N fixed S asymptotic distribution
and its sequential as (5, Af -> oo)^ asymptotic distribution (cf. (15)), regardless
of the value of d > 0. Furthermore, for d < 1 , the Arbov estimator has a normal
asymptotic distribution that is equal to its first-order large Af fixed S asymptotic
distribution. However, for d > 1, the Arbov estimator has a nonnormal asymp-
totic distribution. Moreover, for d > 1 , the diagonal path asymptotic distribution
of the Arbov estimator is equal to its sequential as (5, N -» oo)^ asymptotic
distribution, which is a Cauchy distribution; cf. (16).
The y/S rate (the faster than */Ñ rate) of convergence and the nonnormal shape
of the limiting distribution that are obtained for pArbov when p = 1 and 5, N - »
oo with S/Nd -» c* > 0 and d > 1 or (5, Af -> oo)^ are determined/caused

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1366 HUGO KRUINIGER

by the behavior of the sum that involves the initial conditions, which dominates
the sum(s) that only involve(s) the sample period innovations; see the proof of
Theorem 7 or note 5. Also when d = 1, the asymptotic distribution of pArbov is
affected by the behavior of the initial conditions. Phillips and Magdalinos (2008)
observe similar "tail wagging the dog" phenomena in the asymptotic behavior of
the LS estimator for a simple unit root AR model with a so-called infinite past
initialization and the related i-ratio. Note that although the initial conditions also
play a dominating role in the derivation of the asymptotic distribution of fi ab when
5, N -> oo with S/Nd -> c* > 0 and d > 1 or (S, N -> oo)seq, the distribution
obtained is the same as the one found for any d with 0 < d < 1.
Observe that ''mx'nq(k,c*) = 0 and limante (¿,<0/2ii) = lim^oiU-
exp(-2Ac))/(2A)) = linu¿o(2cexp(-2Ac)/2) = c. From these observations
and the results in Theorems 4 and 7 we conclude that the diagonal path local
asymptotic distributions of the Arbov and System estimators are left-continuous
atA=0(/? = l) provided that 0 < d < 1. A similar continuity result holds for
the local-to-nonidentification asymptotic distribution of the AB estimator for any
d>0.
The limit densities of fi ab and fi Arbov in Theorem 7 are both unimodal.
The following lemma and theorem extend the results in Theorem 7 to T > 3
and the System estimator.

LEMMA 3. Let T > 3 and let p = 1. Furthermore, let N -> oo, S -> oo
simultaneously with S/Nd -> c* > 0 and S/Nd -> c > 0 where d > 0 and d =
max(ú?, 1). Leto*r =o2(c*a2 + '{d = 0}(als-'-a2)). Finally, letfi'b bean initial
N (X/2W -consistent estimator for p, i.e., Nil/2)ã(pib - p) 4 Kb #0. Then

(i) (I) Nt-VW-WZ^Zl'Avi 4 *5i,

Ni-imiW^zl'Ay^lxn,
N-^Z^ZP'AviSxs^and
N-X/2^Z?fAyi,-x^X62with
X5 = (X'sl X'n)' - tf(0, S55), Xe = (X'6l X'62Y^ N{fo, S66), and
E{XsX'6) - £56 # 0, where ¡ùe, £55, £56, and £00 are given in the
proof;
(2) when T = 3, X5 = X5i = olrX' and X6 = X6i = a2rX2, where Xx
and X2 are defined in Theorem 7.

(ii) (Ì) N-V2ï»=i Zf-'vi S X7{ and

Ni-ißW+d)^ zYyi%-' 4 X81 - N('{d < '}a' ca4l) with
X7 = (X'52 X'7l)'^ N(0, £77) and E(X7 X's) = 0, where £77 is defined
in Lemma 2 and Xs = (0' X'^)';
(2) when 7 = 3, X7 = X7i = a2X3, and X8 = X81 = a2X4, where X3
and X4 are defined in Theorem 7.

In addition, we have the following results concerning the weight matrices:

070diag(^-|/2/m-(r-2),^(-l/2)¿/r-2)I,/L,(¿//'[A^(pi)";'(pi)]'

diag(^-1/2/m_(r-2),^-1/2/r-2)I,'Ll(Z//'tAt>;(p,ft)^(^iè)]'[At;;(yolfc)
V¡(pib)]Z;')diag(N-l'2Im-{T-2),N-V2lT-2) 4 ^^Ãrbovlb'
diasWl-WWlT-i.N-W^T-iy^-WlT-í)
xIìLl&f' [Aoí(pi) o!(fii)]'[Av'i(fit) Ô!(fiO]Zf)
xdiag(yv(-l/2)('+rf)/r-2, N-Wlm-(T-v, N^'^h-i) -4 a4^52>
diag(^(-'/2)O^)/r_2> A^-'/2/m_(r_2), Ar'/2/7--2)
x X,N=l (Zf [ Av'iipu,) v¡(p]b)]' [ A»í(pit) v¡(fiib)]Zf)

xdiagC^-'/2)*'^)/^, ^-|/2/m_(7-2),A^-1/2/r-2)^ «r4^-^,

THEOREM %.LetT>7> and let p = '. Furthermore, let N -> oo, 5 -» oo

simultaneously with S/Nd - > c* > 0 and S/W* - » c > 0 w/iere d > 0 and ¿ =
max(a' 1). Finally, let p'b be an initial N^/T)d -consistent estimator for p, i.e.,
N^Hpih-p) ÂiCh^O. Then

d) (D PAB'-P^K 1 5-5' x5/x'6 i;5l x6.

(2) assuming that pAB exploits E(ZfB'Aü¡) = 0 in lieu ofE(Z'¡ Ao,) = 0,

PAB-P^X'4WXÌ3/X'HWXH

with X,3 = (X'5l i{d = 0}X'52Y and XU = (X'6l '{d = 0}X'62)'.

(ii) p'imN-+oo(pArbov - P) = 0, and assuming that pArbov exploits E(Z¡"
x [Aw? v'¡]') = 0 in lieu ofE(Z¡"vi) = 0,

Nil/2)ã(PArboo -P) 4 X'SWX7/X'&WX8.

(Hi) p'imN->oo(pSYSk - P) = 0. for k = 1 , 'b, 2, 2b, and

N(l/2)3(psYSk-p) -^ Xíirfh/X'iÍrfXs, fork=', 'b,

V7J(pSYS2ÍPl)-p) -» ■^iiW'5K52-^10/An^SyS2-ïll,

N(]/2)ã(psYS2(pib)-p) 4 ¿il WsyS2*ÍK»/XÍ, WsKMtíll

w/fÄ Xio = (*5, Ì52 1W < l)^7i)'. Xiofr = (X^ X^)', am/ in = (0 *£)'.
f/vj (1) if d < ì, plimN->oo(psYS-p) = 0.

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1368 HUGO KRUINIGER

(2) assuming that pSYS exploits E(Z?f[Av¡ v¡]') = 0 in lieu of E(Zf

x[Av¡v¡])=0,

N(i/2)(d-d)(ßsYs_p) 4 x'6WXì5/X'6WXì6

with Xl5 = (X'5Ì '{d = 0}X'7y and X16 = ('{d> 1}X'6{ X's)f.

When p = 1 and T > 3, the diagonal path asymptotic distribution of the pab'
estimator is equal to both its first-order large N fixed S asymptotic distribution and
its sequential as (5, N -> oo)s^ asymptotic distribution for all d > 0. However,
in general, these equalities no longer hold simultaneously for any d > 0 when the
AB estimator uses an arbitrary sequence of PD weight matrices. Then one needs
to distinguish between two cases: the diagonal path asymptotic distribution of
the AB estimator is equal to its first-order large N fixed S asymptotic distribution
when d = 0, and equal to its sequential as (5, N -» oo)^ asymptotic distribution
when d > 0. Finally, when p = 1 and T > 3, the Arbov estimator that uses an
arbitrary sequence of PD weight matrices has a normal asymptotic distribution
that is equal to its first-order large N fixed S asymptotic distribution for d < 1 ,
whereas it has a nonnormal asymptotic distribution for d > 1, just as in the case
p = 1 and T = 3; cf. the discussion following Theorem 7. Moreover, for d > 1,
the diagonal path asymptotic distribution of the Arbov estimator is equal to its
sequential as (S, N -> oo)5^ asymptotic distribution.

4. GMM-BASED PANEL UNIT ROOT TESTS

4.1. LM Tests

In this section we propose two LM-type panel unit root (UR) test statistics that are
based on an Arbov estimator and a System estimator that use a weight matrix that
is optimal under the null, and on restricted conventional estimators of their first-
order fixed-parameter asymptotic standard errors. The critical values for these
LM tests can be taken from the standard normal distribution irrespective of the
assumptions made regarding the initial observations.
Let pArbov2,R and psysi^r be GMM estimators that use WNiArbou2(l) and
WNisYS2(l) as weight matrix, respectively. Let SE(pArbov2,R) = {N~lI,?=i
(^-iZ/Ot^^wil)]!^!^/7^,-!)}"172 and SE(psys2,r) = {AT1!^
[(Ay;., ^_1)Zf][^,5y52(l)]ir=i[zf/(A^-i^-i)/]}"1/2- Then we have
the following results.

THEOREM 9. Let T > 3.

(i) If p = 1 and if N -> oo, S -> oo simultaneously with S/Nd -> c* > Ofor
some d > 0, then

(PArbov2,R - ')/SE(pArbov2iR) ~> #(0, 1)

and

(PSYS2,R - l)/SE(psYS2,R) 4 tf (0, 1).

(ii) If p = '- XN~i/2 with X > 0 and */ S -> oo, tf -» oo sequentially, then

d f X ,- '
(pArbov2,R - l)/SE(pArboo2,R) "> N I--V/», ,- 1 1

and

(psys2,r - l)/5£(?5y52,Ä) 4 N (~Via9 1) .

(Hi) Ifp = l-XN~1/2 with X > 0 am/ // tf - > oo, S - > oo simultaneously with
S/Nl/2 -*c>0and S/Nd -> c* > 0 w/rcrc d > 0, /Ae/i

(PArbav2,R ~ 0/ SE(pArbov2,R) 4 JV(-A(1 -q(X9c)/2)y/iñ9 1)

(ÃKS2.A - l)/S£(psy52,Ä) 4 tf (-A(l -^a,c)/2)Vm, 1).

The results in Theorem 9(iii) corresponding to c = 0 are also valid under large
tf fixed 5 asymptotics. Madsen (2003) has derived some related local power
results.
Note that the local power is the lowest when the data are CS or d > 1/2.
Any Arbov estimator can be used to construct an LM-type panel unit root test,
but choosing the optimal Arbov estimator pArbov2,R yields a test statistic with
the highest local power within this class of LM tests. In particular, in the proof of
Theorem 9 it is shown that the weight matrix used by pArbov2,R is not only optimal
under the null but also optimal under local alternatives given by p = 1 - AN ~1/2
with X > 0, irrespective of the asymptotic plan for 5 and tf . On the other hand,
not every System estimator is suitable for the construction of an LM-type panel
unit root test that has correct size for any value of d and nontrivial power against
any local alternative. However, the optimal System estimator psYS2,R is v^V-
consistent under both the null and local alternatives and yields an LM test that has
correct size and the same local power properties as the LM test that is based on
the optimal Arbov estimator pArbou2,R, irrespective of the asymptotic plan for S
and tf.

In the last decade various other panel unit root tests have been proposed. For
instance, Breitung and Meyer (1994) proposed a test statistic that is based on an
ordinary least squares estimator for p in a model for deviations from the initial
observations. Harris and Tzavalis (1999) discussed an LM-type panel unit root
test that is based on the bias-corrected least squares dummy variables (LSDV)
estimator for p. Finally, Kruiniger (2008b) discussed a Wald-type panel unit root
test that is based on the first difference maximum likelihood estimator (FDMLE)
for p in the covariance stationary panel AR(l)/unit root model.

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1370 HUGO KRUINIGER

4.2. Monte Carlo Results

In this section we compare the finite-sample performance of our GMM-based

panel unit root tests with three other panel unit root tests, namely, the Wald test
that is based on the FDMLE, the LM test due to Harris and Tzavalis (1999) that is
based on the bias-corrected LSDV estimator, and a Wald test that is based on the
fixed effects maximum likelihood estimator (FEMLE) for p in the panel AR(1)
model (cf. Hsiao et al., 2002; Kruiniger, 2003, 2008b).8 We only consider size-
adjusted versions of the latter two tests.9 The implementation of the FEMLE-
based unit root test is further discussed in Kruiniger (2008b).
In most simulation experiments the errors have been drawn from normal distri-
butions: £iit ~ Af(0, 1) and ¡xl ■ ~ Af(O, 1). To assess how assumptions with respect
to y i 't i - ju i , i = 1 , . . . , N , affect the power of the tests, we have conducted four dif-
ferent kinds of experiments: in one set the initial observations are drawn from sta-
tionary distributions, i.e., (j/j - /¿/)|//z ~ N(0, 1/(1 - p2)), whereas in the other
three sets the initial observations are nonstationary. The three nonstationary cases
considered are (1) yiA - Mi = 0, (2) (yiti -///)|/// - N(0,2/(l -p2)), and (3)
()>/,i -2jUi)'jUi ~ N(0,p2/(i -p2)). Note that in all situations E(j/,r -yiit-') =
0 as is the case under the null hypothesis. In both case (1) and case (2) the vari-
ance of yz-, i - m is different from the variance under stationarity, whereas in case
(3) nonstationarity is due to the fact that E[ju¡ ( y¡9' - //,-)] ^ 0. Case (1) corre-
sponds to small 5. We have also considered experiments with o2 = 0, a2 = 100,
and su ~ (x2(l) - l)/'/2. Note that the size and the power of test statistics that
only exploit data in differences are not affected by changes in a2. Finally, in
the simulation experiments we have varied the dimensions of the panel data sets
also: (N, T) = (100, 10), (100,6), or (500,6). All simulation results are based
on 5,000 replications, and the (nominal) level of the tests is either 2.5% or 5%.
Tables 1 and 2 report the simulation results on the empirical size and power of
the panel unit root tests that were mentioned previously. Table 1 reports results
on power against stationary alternatives, whereas Table 2 reports results on power
against nonstationary alternatives. In the tables "W" denotes the Wald version of
a test, "LM" stands for the LM version of a test, and "SA" indicates that the test
has been size adjusted. When the (nominal) level of a test is 2.5% this is indicated
by a' otherwise the level of a test is 5%. Inspection of the results in Tables 1 and
2 leads to the following conclusions with respect to the GMM-based unit root
tests:

1 . In most cases considered the GMM-based tests have correct size. However
nonnormality of the errors affects the size of the tests.
2. The power of the test based on the System estimator is greater than or equal
to the power of the test based on the Arbov estimator. However, in many
cases the power of both tests is roughly the same and equal to the power of
the FDMLE.
3. When the variance of the y¿j - pa is larger than the value implied by co-
variance stationarity, the power of the test based on the System estimator

TABLE 1. Size and power against stationary alternatives

T N test size, 5 = -1 size, 5 = 49 p = 0.95 p = 0.9 p = 0.8

A. <72= 1 and<72 =0
6 100 ARBOV2-LMa 0.035 0.030 0.155 0.355 0.846
ARBOV2-LM 0.058 0.056 0.228 0.470 0.906
SYS2-LM0 0.029 0.030 0.143 0.370 0.881
SYS2-LM 0.050 0.054 0.215 0.477 0.930

B.<72 = 1 and<r2 = l
10 100 ARBOV2-LM* 0.030 0.029 0.261 0.663 0.987
ARBOV2-LM 0.059 0.054 0.370 0.756 0.993
SYS2-LM* 0.024 0.028 0.209 0.592 0.981
SYS2-LM 0.049 0.047 0.306 0.700 0.989
6 100 ARBOV2-LM* 0.035 0.030 0.145 0.339 0.844
ARBOV2-LM 0.058 0.056 0.217 0.459 0.901
SYS2-LM* 0.029 0.030 0.139 0.341 0.856
SYS2-LM 0.050 0.054 0.209 0.453 0.916
FDML-W 0.056 0.056 0.200 0.466 0.925
FEML-W-SA 0.049 0.049 0.118 0.162 0.319
LSDV-LM-SA 0.050 0.050 0.153 0.327 0.812
6 500 ARBOV2-LM* 0.029 0.029 0.446 0.939 1.000
ARBOV2-LM 0.055 0.053 0.553 0.968 1.000
SYS2-LM* 0.027 0.028 0.446 0.958 1.000
SYS2-LM 0.050 0.052 0.567 0.979 1.000

C. <72 = 1 and<72 = 100

6 100 ARBOV2-LM0 0.035 0.030 0.085 0.161 0.376
ARBOV2-LM 0.058 0.056 0.138 0.236 0.459
SYS2-LM* 0.029 0.030 0.077 0.147 0.351
SYS2-LM 0.050 0.054 0.127 0.215 0.435

D. (yiA -Mi)'fii - (20)- 1)^2(1 -P2) and a = 1

6 100 ARBOV2-LM* 0.052 0.047 0.163 0.311 0.619
ARBOV2-LM 0.100 0.091 0.256 0.420 0.732
SYS2-LM* 0.036 0.039 0.141 0.328 0.743
SYS2-LM 0.070 0.075 0.228 0.447 0.831

is greater than the power of the tests based on the Arbov estimator and the
FDMLE.

4. The power of the GMM-based tests decreases with an increase of the value
of 4
5. In the cases considered the GMM-based tests have greater power than either
the test that is based on the LSDV estimator or the test that is based on the
FEMLE.

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1372 HUGO KRUINIGER

TABLE 2. Size and power against nonstationary alternatives

T N test size, S = - 1 size, S = 49 p = 0.95 p = 0.9 /> = 0.8

A. )^ ' i = /// , cr - 1 , and <r^ = 1

6 ' 100 ARBOV2-LM0 0.035 0.030 0.308 0.799 0.999
ARBOV2-LM 0.058 0.056 0.416 0.872 0.999
SYS2-LM* 0.029 0.030 0.281 0.759 0.998
SYS2-LM 0.050 0.054 0.378 0.842 0.999
FDML-W 0.056 0.056 0.457 0.914 1.000
FEML-W-SA 0.049 0.049 0.138 0.182 0.290
LSDV-LM-SA 0.050 0.050 0.308 0.732 0.994

B. (y,M -Mi)'n ~ N(p,2/(1-P2)), °2 = 1, and aj = 1

6 100 ARBOV2-LM* 0.035 0.030 0.096 0.138 0.359
ARBOV2-LM 0.058 0.056 0.157 0.217 0.473
SYS2-LM* 0.029 0.030 0.089 0.189 0.609
SYS2-LM 0.050 0.054 0.151 0.278 0.721
FDML-W 0.056 0.056 0.068 0.116 0.351
FEML-W-SA 0.049 0.049 0.098 0.170 0.396
LSDV-LM-SA 0.050 0.050 0.060 0.101 0.378

C. (yÌ9Ì -2Mi)'Mi - N(0,p2/(' -P2)), ç2 = 1, and aj = 1

6 100 ARBOV2-LM0 0.035 0.030 0.138 0.351 0.799
ARBOV2-LM 0.058 0.056 0.216 0.459 0.870
SYS2-LM* 0.029 0.030 0.133 0.355 0.826
SYS2-LM 0.050 0.054 0.205 0.461 0.893
FDML-W 0.056 0.056 0.205 0.474 0.927
FEML-W-SA 0.049 0.049 0.118 0.166 0.325
LSDV-LM-SA 0.050 0.050 0.143 0.335 0.811

5. CONCLUDING REMARKS

In this paper we considered GMM-based estimation and inference for the panel
AR(1) model when the data are persistent and the time dimension of the panel is
fixed. We derived local asymptotic approximations to the finite-sample distribu-
tions of the AB, Arbov, and System estimators, respectively, under a variety of
distributional assumptions about the initial observations. Among other things we
found that the nature of the weak instruments problem of the AB estimator de-
pends on the distributional properties of the initial observations. Moreover, when
p = 1 - À/N and when either the data are co variance stationary or both S and Af
grow large with S/Nd -> c* > 0 and d > 1, then both the Arbov and the two-
step "optimal" System estimator have nonnormal local asymptotic distributions,
and the estimators of the optimal weight matrices for the Arbov estimator and the
System estimator and their asymptotic standard errors are no longer consistent.
We also argued that in these cases one should use LM tests and not Wald tests.

Two LM-type panel unit root tests that we proposed were found to have good size
and power properties.

NOTES

1 . We assume identical distributions across the individuals for convenience.

2. Note that E(e, , y¡tS ) = 0 for all t and all s < t and E(eitte^s ) = 0 for all t ^ s implies E(eijt rji ) =
0 for all t > -S + 2 but not E(siy-S+' li) = °-
3. In Section 3.2 we discuss a range of alternative local asymptotic distributions based on S/Nd - >
c* > 0 and p = 1 - XN~g with 0 < d < g < 1 so that Var( v/i) = O(Nd); cf. Lemma A.4 in the
Appendix. The value of d determines the order of magnitude of Var(^/j) in terms of N. One can
expect that if 5 or Var()>; j) is small relative to N, then a local asymptotics based on S/Nd - » c*
with d close to zero gives a better approximation to the distributions of the AB and Arbov estimators
compared to local asymptotics based on the CS model, which implies that Var(y/i) is large relative
to TV.

4. Note that E(dmAB,s,t/dP'P = 0 = ^(yi,s£i,t-l) = 0 when s < t -2 and that E(dmArbOVìSìt/

dp'p = ') = E(yiìt-'£iìS) = a2 when s < t - 1 . This implies that E(dmABiSj/dp) and E(dmArboVìSìt/
dp) are left-discontinuous at p = 1 .
5. When p = 1 and S - » oo, the y¡tt are no longer tight random variables. However, the re-
sults in (15) and (16) can still be shown to be correct upon scaling both the numerator and the
denominator in pab~P - Œ/Li ?/,l Ae/,3)/Œ/Li 37, 1*1,2) and the denominator in p Arbov ~P =
Œjii euetfVŒÎLi yi,2£i,2) bV ^/S•
6. EO/ÍJt^ig) < EiX'zWirjWXs/iX'xWXs)2) for any W.
7. Conditional on Xg, these LM test statistics have a standard normal asymptotic distribution. As
the latter distribution does not depend on Xg, it follows that the unconditional asymptotic distribution
of these LM test statistics is also equal to the standard normal distribution.
8. Bond, Nauges, and Windmeijer (2005) have also compared various panel unit root tests by
calculating their asymptotic local power and by conducting simulation experiments. The only GMM-
based test they consider is a Wald test based on the "optimal" System estimator. As expected this test
has poor size and power properties. They also find that the Breitung-Meyer test and the test based on
the FDMLE have very similar size and power properties.
9. Note that the FEMLE is different from the LSDV estimator.

REFERENCES

Ahn, S.C. & P. Schmidt (1995) Efficient estimation of models for dynamic panel data. Journal of
Econometrics 68, 5-28.
Ahn, S.C. & P. Schmidt (1997) Efficient estimation of dynamic panel data models: Alternative as-
sumptions and simplified estimation. Journal of Econometrics 76, 309-321.
Arellano, M. & S. Bond (1991) Some tests of specification for panel data: Monte Carlo evidence and
an application to employment equations. Review of Economic Studies 58, 277-297.
Arellano, M. & O. Bover (1995) Another look at the instrumental variable estimation of error-
components models. Journal of Econometrics 68, 29-51.
Blundell, R.W. & S. Bond (1998) Initial conditions and moment restrictions in dynamic panel data
models. Journal of Econometrics 87, 1 15-143.
Bond, S.R., C. Nauges, & F. Windmeijer (2005) Unit Roots: Identification and Testing in Micro Panels.
CEMMAP Working paper 07/05, Institute for Fiscal Studies, London.
Breitung, J. & W. Meyer (1994) Testing for unit roots using panel data: Are wages on different bar-
gaining levels cointegrated? Applied Economics 26, 353-361.
Fieller, E.C. (1932) The distribution of an index in a normal bivariate population. Biometrika 24,
428-440.

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1374 HUGO KRUINIGER

Forchini, G. (2006) On the bimodality of the exact distribution of the TSLS estimator. Econometric
Theory 22, 932-946.
Hahn, J., J. Hausman, & G. Kuersteiner (2001) Bias Corrected Instrumental Variables Estimation for
Dynamic Panel Models with Fixed Effects. Working paper, M.I.T.
Han, C. & P.C.B. Phillips (2006) GMM with many moment conditions. Econometrica 14, 147-192.
Han, C. & P.C.B. Phillips (2007) GMM Estimation for Dynamic Panels with Fixed Effects and Strong
Instruments at Unity. Cowles Foundation Discussion paper 1 599, Yale University.
Harris, R.D.F. & E. Tzavalis (1999) Inference for unit roots in dynamic panels where the time dimen-
sion is fixed. Journal of Econometrics 91, 201-226.
Hillier, G.H. (2006) Yet more on the exact properties of IV estimators. Econometric Theory 22, 913-
931.

Hsiao, C, M.H. Pesaran, & A.K. Tahmiscioglu (2002) Maximum likelihood estimation of fixed effects
dynamic panel models covering short time periods. Journal of Econometrics 109, 107-150.
Kruiniger, H. (2003) On the Estimation of Panel Regression Models with Fixed Effects. Revised ver-
sion of Working paper 450, Queen Mary, University of London.
Kruiniger, H. (2008a) GMM Estimation and Inference in Dynamic Panel Data Models with Persistent
Data. Revised version of Working paper 428, Queen Mary, University of London. Available at
http://webspace.qmul.ac.uk/hkruiniger/.
Kruiniger, H. (2008b) Maximum likelihood estimation and inference methods for the covariance sta-
tionary panel AR(l)/unit root model. Journal of Econometrics 144, 447-464.
Madsen, E. (2003) Using GMM when Testing for a Unit Root in Panels Where the Time-Dimension
is Fixed. CAM Working paper 2003- 1 1 , University of Copenhagen.
Marsaglia, G. (1965) Ratios of normal variables and ratios of sums of uniform variables. Journal of
the American Statistical Association 60, 193-204.
Moon, H.R. & P.C.B. Phillips (2000) Estimation of autoregressive roots near unity using panel data.
Econometric Theory 16, 927-997.
Phillips, P.C.B. (2006) A remark on bimodality and weak instrumentation in structural equation esti-
mation. Econometric Theory 22, 947-960.
Phillips, P.C.B. & C. Han (2008) Gaussian inference in AR(1) time series with or without a unit root.
Econometric Theory 24, 631-650.
Phillips, P.C.B. & T. Magdalinos (2008) Unit Root and Cointegrating Limit Theory when Initialization
is in the Infinite Past. Cowles Foundation Discussion paper 1655, Yale University.
Phillips, P.C.B. & H.R. Moon (1999) Linear regression limit theory for nonstationary panel data.
Econometrica 67, 1 057- 1 1 1 1 .
Staiger, D. & J.H. Stock (1997) Instrumental variables regression with weak instruments. Economet-
rica 65, 557-586.
Stock, J.H, J.H. Wright, & M. Yogo (2002) A survey of weak instruments and weak identification in
generalized method of moments. Journal of Business & Economic Statistics 20, 518-529.
Woglom, G. (2001) More results on the exact small sample properties of the instrumental variable
estimator. Econometrica 69, 1381-1389.

APPENDIX: Proofs of the Results

LEMMA A. 1 . Let [y¡ , } be a stationary process and let u¡tl - y¡t - fi¡. Then

E(«,v_, A«,-,,) = -~, E[(A«,-,,)2] = ^-,

4 v d-p2)K + 6p2 4
E("i-')=(l-^)2(,+p2)-' 4 v 4

cr,A v4, (l-/>)2[*(l-p2)

('+p)20+p2) , ^ 1+p
cr,A E[(Aw, ,) v4, ]=

EK,_l(A«,v)3]
EK,_l(A«,v)3] ,. ,.,3l
,3l =
(l-p)[K('-p2) + 6p2] 4 3 4

or 2 ,A *2, («(l-/>2) + 6/>2) 4 , 1 4

or EK,-.(A«,v) 2 ,A *2, ]=(1+p)2(1+/>2)g +TT7ff , •

Proof of Lemma A.I. Note that u¡j = pult-' +e,-if and Am/?í = (p - ')uit-' +
£/,. Moreover {wzi} is a stationary process. Then verification of the first two claims is
straightforward. The other claims are proved as follows:

E(<,)=E(pV,í-l+6pM/-le2,r+eu)=/E(</) + ^Í+^4
n/ 4 ^ k<74 6gy (i-p2)*+6/>2 4
«b("'V> n/ ^ - ! _p4 + (1 .^d .^4) - (, _p2)2(i +p2)ff '

E[(A«;,,)4] =Wp- l)V,r-l +6(p-l)2M?/_,eff +£f,]

(l-/>)2[K(l-p2) + 6p2] 4 , 6(l-j>) 4 , 4
(l+p)2(l+p2) 1+p

E[tt,,,_,(A«M)3] =E[(p-l)3«4,_, +3(p- 1 )«?,_,£?,]

= _(l-p)[K(l-^) + 6^]g4_^_g4 and

(l+p)2(l+/)2) (1+/9)

LEMMA A.2. Let [y¡¿} be a stationary process. Then

a2 a2
E(>v,i Ay/,2) = ---- , E(y,-,2Ay/,2) = y- ,

v / a , 2a2l*2 + ei('-p2)]
v Var(,;,,A£,3)= / a , (i+p)0_p) .
2(T2[íJ2 + ít2(1-/9)2]
Var[fe,3 H- (1 -/>)///) Ay/,2] =

Variai A^i.2) = (1+p)2(1+p2) + (i+p)(i-p)

a2[2^(l+p)-t72]
+ d-fp)2

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1376 HUGO KRUINIGER

v , a * pHk('-p2) + 6p2) 4 , oA

jc(1 + />)2 - 1 -6/>(l + p) + 2(<t¿At2)(1 + />) 4

CovUi A.,-,3, J/,i Ayw) =

2(7^2(1 -/>)
Cov[(e/,3 + (l -/J^iOAy/^.y/^Ay/^] =

Proof of Lemma A.2. Noting that >>;}/ = utj + //,- , application of Lemma A. 1 yields
-?
<7

E(y/,iAy/f2) = E[(M/ii+|i/)Ai</i2] = -j- ,

2 *2
E( y/,2Ay/,2)=E[y/jAy/f2 + ( Allori 2 = yt-.

Var(y/,iAe/,3) = E[(y/fiAe/i3)2] = E[((ii/,i+/i/)Ae/i3)2]

2^4 2 2 2^2[a2+^(l-/>2)]

Var[fo,3 + (1 -p)jui)àyi,2] =E[(fo,3 + (l -p)/ii)Auit 2)2]

2c72[(72 + ^2(l-p)2]

r ^2 2I
Variai A^2) = E (m/,i Am/,2 + /ìiAiìi,2 + - - )
L ^ J
, 2a4 2<72<r2
= E[KlA""2) , ÕW+TT7 2a4 2<72<r2
, ^4 1
(i+p)2J

_ (K(l-p2) + 6p2)g4 g4
(l+p)2(l+p2) +(1+/>)(1-P)
(T2[2g2(l+p)-ff2]
+ (l+/>)2

Var(y,,2 Ay¡,2) = E («,,i A«,-,2 + (A«/,2)2 + n¡ Am,,2 - -^- )2

= E[(«M A«;,2)2 + (A«i)2)4 + (/i,- Am,,2)2

+ (T^)2+2"'.'(A"'.2)3

lo1 lo2 ,
- 77- (w/,1 A11/.2) - - - (Am/,2) , ]

= Wl-/>2) + 6/)2) " 4

(l+/>)2(l+p2) "

<74 (1 -p)2[ic(l -p2) + 6p2] 4

X-P2 (l+/>)2(l+/>2) "

i+/> '+p (i+p)2

2(1 -/>)[«:(! -p2) + 6p2] 4

(l+p)2(l+/,2) "
6 4 2<r4 4g4
~(l+p)a +(l+p)2~(l+p)2
_^2(>c(l-/)2) + 6p2) a 4
(l+p)2(l+p2) a
a4
+ (l+p)(l-p)
*(1 +p)2 - 1 -6/>(l +p) + 2(<72/a2)(l +/») 4
+ (TT^P a'
Covij/j Ae/,3, y/,i Aj/,2) = Cov(uit' Ae/,3, m/,i Aii/,2)

+ Cov(/i/Ae/f3,/i/A«/f2)

= E(M?1Aei,3AM|i2) + <T^E(Ae/>3Aii|f2)

= -E(«?tlCi2f2)-^E(c22)
¿72[<72+<72(l-p2)]
(l-p)(l+/>) ' *"

Cov[(e/ï3 + (l -^///JA^^^A^J = E[(1 -/))(//, AM/, 2)2] =

A.7. Proo/ 0/ Theorem 1. When 7 = 3, Ä4fi = p + (A^"1 I^j yiAAeii3)/(N-1

ÏÎLi W,i Ay/,2). and pAr6m, = p + (A^"1 XiLi fe,3 + d -P)î) A^VÍ^"1 I/Li y/,2
Aj>,'2).
Let us define X! = N-^^yÂe^ X2 = M^^yÂy^ X3 = N~1/2
liLl fe',3 + d - P)Mi) Ay/,2. and X4 = iV"1 I^j y/>2 Ayif2.
Let the CS model hold. Then using the results in Lemma A.2 we obtain as Af -> 00 for

the parameter sequence p = l-AA^"1 thatXi A N(0,a4/k), X2 4 N(-a2/2,a4/(2X)),

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1378 HUGO KRUINIGER

X3 4 N(0,<74), and X4 4 N(a2/2,(T4/(2Á)). It is also easily verified that lim^oo

Cov(X!, X2) = - cr4/(2/l) and lim^_^0OCov(X3, X4) = 0. It follows that

»'4 - [î;]~"[-t^(î)-t(-2. -.')]• -

LEMMA A.3. Lei {^r} be a stationary process, let wz?i = yit - ///, let /, p, oazú? g
be positive integers, and let Kq = E(ef t)/aq < oo. Then assuming that the sum of the
powers ofsik, ui,s> ui,t> and ^ui,t I<J less tnan or equal to 8, we have

(i) 'impV('-p)PE(ulPt)<oo,
(ii) limpT1(l-y9^|E(^+1)| <oo,
(Hi) 'imp<tiE[(AuiJ)P] = Kp(jP,

(iv) 'impV(' -p)P E[«g(Aii/fi)«]| <oo,

(v) 'impn(' -p)P E[M^+1(AM/,r)^]| <oo,
(vi) limpîl(l-p)^ E[elikulPs(Aui^]' < oo, and
(vii) 'imprì(' -p)P E[^^+1(AW/>/)^]| <oo.
Proof of Lemma A.3. Note that {u¿it } is a stationary process, Uit =pu¿it-'-'- £¡j , and
A«/,r = (p- l)«/,/-i +e/,i. Moreover, E(i<?f) = E[(/9iilV_1 +e/,f)p] = E[X£=0 £(/?,*)
x(P«i,/-l)p~*eff/], where Ä(p,*) = [k'(p-k)']-{ p' .
We prove (i) and (ii) together. The proof proceeds by induction:
First consider p = 0: E(u® t) = 1 and E(uj t) = 0, and hence lim^j Efw^) < oo and
limpT1 E(ujt) <oo.
Now let p > 0 and suppose that lim^i(l - p)qE(u^qt) < oo and lim^i((l - p)Q
x |E(Mf2J+1)|) < oo for q < p - 1. Note that limpT1(l - p2p)~x{' - p) = limpT1
Œ^ô1^)"1 = 1/(2p) and limpîl(l - p2P+x)-x{' - p) = 1/(2/7+1). It follows
that limpid -p)PE{u2ipt) = 'impV(' - p2P)-l(l - p)pB(2p,2)E(ulp-2)E(elt) =
(T2(2p)-lB(2p,2) x limpT1(l - p)P~xE{u2ip~2) < oo. It also follows that limpT1

(1-P)^|e(M^+1)| < limptl(l->92^1)-1(l-p)^Ä(2/7+l92)|E(W/2J-1)|E(e/?/)+

'impV(ì-p2P^rx('-p)PB(2p^'i3)E(ulP~2)'E(elt)'=a2(2p
1,2) x limpîl(l - p)/7-1 JEÍwfj"1)! + 'K3'cT3(2p + l)"1^/? + 1,3) x liny^
(l-^-1E(Wfj)-2)<oo.
The proofs of (iii)-(vii) are now straightforward:

(iii) limn, E[(Aii/if )^] = limpîl E[((p - 1 )«/,/_! +^,i)/?] = E[(e/,f )^] = /c^îj/7,

(iv) Hmpîl(l-p)*|E[i«^

(v) limpT1(l -p)P |E[«g+1(AiilV)*] | =limpTl (1 - pY |e[«^+1(^, 1)(/?- 1)

XM^.jef^+e^)]! <oo,
(vi) limpîl(l -p)P'E[slkulPs(àui^]' = 'imprì(' - p)P'E[slikulPsclt]' < oo,
and finally

(vii) lim^T1(l - p)P'E[slkulPs+ì(Aui^]' = limpîl(l - P^E^^f"1

x (B(q9 '){p - ')uitt-XeqiJX+ <??,,)] | < oo. ■
A.2. Proofs of Lemma I and Theorem 2

Proof of Lemma 1. Part (i): The proof of the first two results is similar to the proof of
the first part of Theorem 1 :
For/? = 1 -XN~X 'imN->ooE(yiìiAyis-ì) = 'imN->OoPt-3E(yiAAyiy2) = ~'o2
and limyv->oo AT^Varíî Aü/ff) = lim^ôo N~lE[(y¡9' A^,)2] = lim/v-ôo^"1
xE[(j/JA^3)2]) = lim^ôo2^2N-lE(>;?1) = lim^ôo2a2^-1E(W?1) = ^^.
Note that Var(y/fl Ayiyt-') = E[(uiA AiilV_i +^/Aii/ff_i + y^p'~3)2], M,-,/ =
puij-'+£itt, and A«/if = (p-l)w/ir_i +£/,/• Then it is easily seen that for/? = l-AN~l
limn-too N-âriyijAyij-i) = 'imNôoN~ìE[(uiìiAuiftî)2] = lim^ôo^"1
E{[(p - l)w?! +K/,ie/f,-i]2} = limyvôoyV-^tÍM/,!^-!)2] = a4/(2A); that if
s <t- 1 and/? = l-AA^~1limyv->oo^~1Cov(y/?1A^jJ,y/iAî_i) = lim^ôo^"1
EíiiÂii/Âii^-i) = lim^ôo^'^ÍMÂiiiÂii/^+i) = . lim^ôo^"^^?!
[(/? - 1)m5_i + es][(p - 1)(pmj-i + ej)]} = 0; lim^ôo^'ôviî Au/,/,
yÌ9'Ayu-l) = YimNôoN-ÌEiyfÂeijAuij-O = -tr4/(2A); and finally that
'imNÔo^~ìC0'(yiiiAvij9yiAAyij) = 'imN^0ON-ìE(ylìA£iìiAuij) = (74/(2A).
Thus for/? = 1-/1^- * limyvôo^-1E(Z//At;/Ay;z/) = lim^ôoi727V-1E(>'?1)
xf/ = aV(2A)//JimNôo^-1E(Z//Ay/í_1A>;;_lZ/) = lim^ôoíT2^-1E(^1)/
= í74/(2^)/, and lim;v->oo W-'EÍZ/'A^-! Aü¡Z¡) = -MmNôo^NÊiyf^C =
-g4/(2à)C, and hence

r at'iM'a», i, r*5.i r/ o ' /^w« -cy

Note that lys,, i = E(X51 X¿,) = -ít4/(2/1)C' # 0.

Consider now the scaled sums AT1/2!,^, Z^'Ao,- and W~I/2Z/L, Z^'Ay,-,.).
For ¿> = 1 -A//"1 we obtain limA,^00W1/2E(Z/:)'A}';)_1) = lim/vôo Nxl2E(zf>'
x Am,,-!) = lim/vôo/v'ÊIzP'ti/j - 1)«,,_2 + £¿,-i]) = 0. In addition, limôo
E(Z,P;Ao, Ay;Z,P) = UmNôo<r2E(ZP'HZP), Mm^ÊizP' Ay^ Ay¡ _,Z,P) =
limA,ôo^2E(zP'zP), and lim/vôoEízP'Ay,,-! Auízf) = -limJv^wff2E(Z/>/
x CZP). Hence for p = 1 - AW"1

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1380 HUGO KRUINIGER

|V' V-diag(Cr-3,Cr-4,...,C1) / )'

Note that Ë56>22 = E(X52Xf62) = -a4dmg(CT.3, Cr_4, . . . , Q)' ^ 0.
Next consider the off-diagonal blocks ¿55,12» ¿66,12» ¿56,12» anc* ¿56,21- Because
limôoEÍZ/'Ao/AoJz/^limtf^
limyVôoíT2//£(>;/íi[úf1A^52 d2pAyit2 ••• ¿7-3P7~4Ay/,2]) = limyv-xx)-^4
x //[J^2 ... rfr-3]»we have ¿55,12 = E(Í51Í£2) = 'imNôoa2N-l/2E(Z¡fHZP)
= 0. Similarly, ¿66>12 = E(X61X^2) = limAr_>00(T2^-1/2E(Z//zP) = 0, ¿56)12 =
E(X51X^2) = -'imN_ooa2N-l/2E(Z¡fCfZP) = 0, and ¿56>2i = E(X52X'6l) =
-'imNôo<T2N-i/2E(zP'C'Z{) = 0.
Finally, it is easily seen that if T = 3, then X5 = X51 = o2X'/y/J and X^ = X6' =
(T2X2/'/X, where Xj and X2 are defined in Theorem 1.
Part (ii): The proof of the first two results is similar to the proof of the last part of
Theorem 1 and the proof of part (i) of that theorem:
Stationarity implies that for p = 1 - XN~^ lim^ôo E(y^t Ay¡tt) = limjvôo E(yl ;¿
x Ay/>2) = o2ß and limyv _> 00 A^Var ()>/,, Ayitt) = lim^ôoA^-WariÂy/^) =
(74/(2À). Moreover, if s < t, then for/? = 1 -AN~l limN^00N~1E(y;JA)>; ^/j Ay;,)
= lim^_>cx)^"1E(ii/>JAi<J-iJii/tf Aii,->f) = lim^-xjo^'^ÍM/^-iAii/îi^-iAMj^) =
lim^ôoA^~1E(M/>5_iAM/5jM/î5_1AM/>J+1) = limyvôo^~1E{w^_i[(p-l)w5-l+^]
x[(p-l)(piíJ.i+eJ)]} = 0.
Thus for p = 1 - /lArMimAôoEizfV;^) = KmNÔ0o2E(Z'''Z'') = a4/,
limôo^-îZ^-.^^jZf) •= (74/(2A)/, and lim^^ AT^z^. _iy;
xZ^) = 0, and hence

Note that E(X71X^) = 0. Because limyv^oo A^~1/2E(Z/)/Ayzy/' _xz[) = 0, we also

have E(X52Xg1) = 0. We conclude that E(X7Xg) = 0. Furthermore it is easily verified
that X7 - N(0, ¿77) with ¿77 = limyvôo KIIE(Z¡Ifüiü¡Z¡I)KIIf = a4KnKIIf.
Finally, it is easily seen that if T = 3, then X7 = X71 = a2X3 and X8 = X81 = a2X4,
where X3 and X4 are defined in Theorem 1 .
Part (iii): Because limyvôo^"2^^^^) < 00, lim^ôo N~lWaT(yii Ayiit) < 00,
and limyv-4ooVar([Ay/r]2) < 00, it is easily seen that for p = 1 - AAf"1, N"2^^
(Z/'Z/) **• lim/ôo AT » E(Z/'Z/) = (<72/(2A))/) ¿V"3/2 Sf=1 (Z/'Z,P) *5" lim^^«,
N-WE(Z{'zf>) = 0, and N-^=¿zf>'Zf>) "^ limôoE(Z,D'Z/> ) = ,r2/. The
results are obtained by noting that convergence in quadratic mean implies convergence in
probability.

Part (iv): Because limw_>ooVar([A;yz?r]2) < oo Vf, it is easily seen that for p =
1 - AN'1, N~l ZfL^zV'Zf1) q^' limôo E(Z/7/z/7) = a2 1 and hence plim^^
[AT-1Z^1(Z///Z//)] = ^2/.
To establish the second claim of part (iv), note that v¡ = y; - pìy/,-1 =«/+(/?- p')
xy/,-i and Avi = Ay¡ -p' AyZì_i = Ao/ + (/>-pi)Ayif_i.
Lemma A.3 implies that lim;v->ocE(Aw/,/)8 < oo, lim^-Ôo^~1E[M?J(AM/>f)6] <
cxDîm^ôoA^ÊlM^ÍAM^)4] < oo, limAôoEÂw;,,)4] < oo, limôo |E[n/fJ
x (Am/^)3] < oo, and ìim^ôQ N~Ê[uj s(Auiìt)2] < oo Vs,r. It follows from the
first result, our model assumptions, and the Cauchy-Schwarz inequality that lim^ôo
Var([uiìSAyiìt]2) < oo Vs,t and hence N'^fL^Zf1'^ v¡][Av¡ oJ]Z/7) q™'
'imNôo KÊiZf^Viü'iZf^K11' = o*KuKIlf. Similarly it follows that AT2
x SÍLiíZpyí.-i^.!^) ^ limôoiV-^ÍZ^-i^.^/«) = ^4/(2A)/,
Ar2L*Li (Z/>' AWfli A^_, z/> 0 q5' 0, and ^-2^=1 (z^/'-i ^l-x*?') '$' 0.
Lemma A.3 and our model assumptions also imply that lim;v->oo ^[vi,kui,s(ûi,t)2] <

oo and limyv^oo N~lE[vf ku2 s(Auij)4] < oo Vk,s,t. It follows from these results, our
model assumptions, and the Cauchy-Schwarz inequality that N~3/2 X-^j (Z^'d/^ _{ z/O

Í5- O, tf"3/2 It^=1 (Zpo¿ A>if _! Z/>) ^- 0, AT3/2 1^=1 (Zpy/,-1 AdJ z/>) q4' 0, and
N-3/2S^=1 (z/> > A^_! Ay;Zzö) ^- 0.
We conclude from the preceding results that AT"1 I^j ¿/7/[AÏ- SJ]'[Ao- ^]Z/7 4
a*KllKw + o*/(2k)K2diag{Om-{T-1)9 ¡T-2). ■
Proof of Theorem 2. The proof relies on results in (the proof of) Lemma 1 .
Part (i): Recall that Z' = KABzfB' where ZfB = [Z/ ZfP] and rank(KAB) = m.
Therefore pab' is equal to a GMM estimator that exploits E(Z^'Aü;) = 0 and uses the
weight matrix (W1 X/î zfB'HZfB)~l. Consider the local-to-nonidentification limit-
ing behavior of this weight matrix. Because limwôo N~2Var(j2j) < oo, lim^ôo ^-1
Variai Ayij) < oo, and limyv-ôo VarilA^^]2) < oo, it is easily seen that for p = 1 -
XN-' o2N-2^{Z''HZ{)q™' ¿55,11 andÂ^X^Z/'/ZZ^
owcva2N-lJäfLi(zPfHZP)q^' t55a2- Therefore p'imNôoc72N-2^fLi(Z¡fHZIi)
= £55f „, plimôoa2^"1 IJL^ZP'HZP) = ¿55,22, and plimôoa^-3/2!^
(ZJ'HZP) = ¿55,12 = 0. Furthermore ¿55 is PD. It follows from the preceding results

and (the proof of) Lemma 1 that pAB' - p -* X'6 ¿^ X5/X¿ ¿^ X6. Assuming that pAB

exploits E(Zf Bf Avi ) = 0, we also obtain pAB~P^ *61 wl 1 ^5 1 /¿¿I W" *61 •
Finally we show that pAB' is asymptotically biased downward when T > 3:
Note that ¿As! = p + {I^I^j^jZ^ZÍZO-^íy.!,,.!]}-1!^1!^!^^,
x(Z;Z0-1Z;Dí].BecauseI^21[^1'í_1Z,(Z;Z0-1z;yí] = (I^1A);;j_
{Zf'HZf^^Zf'Avi ^à'^[yf_u_xZt{ZftZt)-xZfty.u^] = <£*LX
A^ _j Zf 5)[I^! (ZfB'HZ?B)rl ZfLi ZfBfAyi^h it follows from part (i) of the proof
of Lemma 1 that both the numerator and denominator of pAB' - p converge in distribution.

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1382 HUGO KRUINIGER

We first show that 'imNôoE[y'_] ,_, Zt (Z'tZt)~x Z'tvt] < 0 for t = 2, . . . , T - 1. Note
that limyvôcEI^j^jZ^ZÍZr)-^;^] = lim^Ê^ [?,•_!,,_!;; t(Z'tZtyx
x Zißij]} = limyv-^cx)I/ii [E(z'Ì3l(Z'tZt)-xZi,t) x E(5vs_i^_ie/s/)].
Now, Eiy/.-^-je/âr -0/(7-^

o/a-o) x[=,+1 */,,)] =(i/a-r+i»^^

SÍi1'íA?J=-^2(V(71-í + l))O/(7'-O)[l^ It fol-
lows that lim/v->ooE(J/,_U_i£/,r) = üm^Tl E(?/,-l,r-l^/,r) = -^2- We also have
lim^_>ooE(I-Li z?Í9t(Z'tZt)-lZi9,) = E[tr(/,_1)] = /- 1 > 0.
We conclude that limyv^c^X^Ei?-! t-'Zt(Z'tZt)-lZ'tot) < 0.
To complete the proof of the asymptotic biasedness of p&ß' , we write the numerator
of pab' -P, i.e., I^KS-Li ^-Ur-l^r)^^)"1^! z/./C/.i)], as the sum of two
terms:

(T-' ' N / ■ i '

{ia,5,í;-'(z;z'rií-(f^T) (T-' ' N / ■ i '

x - I A?, + ^7 S I AM (A.1)

/ r-i-r i r-i r-i-5 '1 ì

x' -¿=0I Pkd, + T- y I I P% • (A.2)
y ~r5=r+i it=0 / J J

The first term, (A.I), converges in probability to lim^^ooX^1 E|jL] t_xZt{Z'tZt)~x

xZ¡et]. Because the numerator of p^ß '-p, S^1 y'_ ' t-'Zt{ZftZt)~x Z't£t , converges in
distribution to a random variable with mean lim^^oo X,7^1 E[yf_i t_'Zt(ZftZt)~x Zfc]*
the second term, (A.2), converges in distribution to a random variable with mean zero. For
convenience we will assume that the s;, 's are symmetrically distributed around zero. Do-
ing so does not entail a loss of generality because imposing this assumption does not affect
the asymptotic distribution of p^ß' . Noting that the second term does not involve higher
. powers of e¡j (higher than one) and using that the £;,'s are symmetrically distributed
around zero, it follows that the second term is symmetrically distributed around zero and
also asymptotically uncorrelated with the denominator of $ab' ~ P> "¿[=2$-] /-l^r
x (ZftZt)~x Z'ty-'t-' ]. We conclude that the expectation of the second term divided by
the denominator converges to zero as TV - > oo. The ratio of the first term and the denomi-
nator converges to a negative constant divided by a positive random variable and gives rise
to the negative bias of p^ß' .
Part (ii): It follows in a straightforward manner from the results (in the proof of)

Lemma 1 thatplimA^«,^^ -p) = 0and >/Ñ{pArbov -p) 4 X^WX-j/X^WXsM

A3. Proof of Theorem 3.

Part (i): WN SYS' = (N-^^zf'AZfy1 where zf = diag(Z/,z/7) and

M? U-
Let MU = N-izfLi&f'HZf), Mn = M'2X = W"1 ^(z/'CZ/7), and M22 =
N"1 S^j (Z/7/Z/7). From parts (i) and (iv) of the proofs of Theorem 2 and Lemma 1
we have pUm^-ôQ N~^ M'' = tf~2 ¿55,11 = (a2/(2À))H and plimyv->oo^22 = ^ 2/-
Moreover, because limjv->ooE(;y/,i àyij) = -a2/2 and lim^ôo TV"1 Var(>>/ ' Ay/,,) =
cr4/(2A), wehaveplim^ôo^-1/2M12 = plim^_,ooN-1/2M^1 =0.

Note that WN,SYSi = {MM» ^]"'=[^! ^] . where M» =N-(Mn/

^-Ml2M2~21M2l/^)"^M12 = -M11M12M2"21,M21=-M2~21M2lM11,andM22 =
M^1 + M~2{ M2'MU MnM-2l .
The first claim follows now straightforwardly by Slutsky's theorem. The proof of the
second claim is very similar.

Part (ii): Note that zff[Kv' SJ]' = ((Z/'A^)' (Z/7/[AÌ- SJ]')')'.

From part (iv) of Lemma 1 wehave AT ^^z/7'^ S/]'[Ao] v^zf1 Âa4W^boo2.
Lemma A.3, our model assumptions, and the Cauchy-Schwarz inequality imply that
limyv-ôoNÊKAD/ s)2ujt] < 00 and limyv->oo^~2E[(At>/5)4M^r] < 00 V5,r and
hence A^Î^^Z/'Aü/Ay/z/j^'lim^ôo^-^ÍZ/'Aü/A^z/). In part (i) of the
proof of Lemma 1 we showed that lim^ôo N~lE(ZJf av¡ Av^ZJ) = limyv-xx)^2
N-lE(Z¡fHZ¡) = (fj)//. It then follows along the lines of part (iv) of the proof of
Lemma 1 thatplimyv->oo ^Î^^Z/'A^/ A^Z/) = (§£)#.
Similar arguments show that plimyvôo A^~3/2I^1(A:7/z///yz At>-Z/) =0 and
plimôo^-2!^!^7^^
= (a2/2A)KIIEldi3Lg(EJ,...,eJ-iyiAe'i] = {aA/2X)KnD. Note that Kllz'l%
= zfIf[Ao¡ o/]. It follows that A^"3/2Z/1i(¿/7/[Aü; S/]AoJz/)4 -((J4/2A)JCKI¡D.
Noting that iv]{,52 = (1/(22))//, M>|{,52 = (W]252)r = -(1/(2/1))A:a://D, and
*|yS2 = WÃÍbooV we conclude that diag(A^-1/r_2,^-1/2/m)I^1(Zf/[A^ {?/]'
x [A^y/lZ^diagi^-1/^^,^-1/2^)-^^4^-^.
Part (iii): From part (i) of Lemma 1 we have N"1 1^, z/'Au/ 4 Xs!,^"1
xxjL, z/'a*,-! 4 x61, ^-1/2Iz/v=i Zp/Ao. 4 x52j and^-i/2SN i Zp/A% _t 4
X62, with X5 = (X^ X'52Y - N(0, ¿55) and X6 = (X'6X X'62)f - tf (¿6, ¿66).
From part (ii) of Lemma 1 we have N~^2^=l z''i>i -^ Í71 ~ #(0, ¿77,22) and
AT1 X/Li zt'y¡,-' ^ *8l ^ ^V((^2'/2), (<T4/(2/l))/.
Recall that Z/7/[AdJ ^]' = KnZ¡!'»i. Let /:5 = diag(/r_2, AT77). It follows that
zf'iA«; 0;i' = «z/'Ai// (z/7/[a^ v¡íyy = ksz?Uv¡ v¡].
Considernow KsW-{SYSlKs' = N-lZfLi Kszf' A zf KSt and KSW~ ^YSÌb K S/ =
N-x^f=xKszffzfK^. From part (i) we obtain plimNôo N-^fL^Z1/ H Z{) =

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1384 HUGO KRUINIGER

(<72/(2/l))tf, p'imN^ooN-{^=[(KIIZIiífZ¡IKIIf) = c72KIIKII/ = a-2t17i and

plimyv^oo N-y2^=i(Z}fCZ¡ l K1If) = 0. We also have plim^^ AT2
„S?N (7ll7l'

We conclude from the preceding results that p'imN->oo(PSYSk ~ P) - 0 and

SÑ{psYSk-p) 4 X^X7/X'st^lX%9 for* = 1, 'b.
It immediately follows from part (ii) and the preceding results that plim^^ oo(PSYS2 ~

p) = 0 and VÑ(psyS2-P) 4 Ííi^ys2io/i 1^52*11 with Xl0 =

(X'5lX'52X'7iyandXn=(0 0X'Siy.
Part (iv): From the results mentioned in part (iii) we conclude that psYS ~ P -*
(X¿1W11X5i+í¿1W31*5l)/ÍÍ2^*12WÍthí12 = (í¿10^1)/. ■

LEMMA A.4. Let g, d, c*, and À be constants such that 0 < g < l,0<d < g,0 <c* <
oo, andO < I < oo. Furthermore, let p = 1 -ÀN~8. Then lim^^^ s/Nd^>c* Ng~d(' -
p2V+s)) = q(l9c*) <oo, when$(l,c*) = 2X(S+l)ifd = 0;q(l,c*) = 2Xc* ifO<
d < g; andq(Á,c*)=q(Á,c*) if d = g.

Proof of Lemma A.4. Note that (1 - p2^+s^)/{' -p) = (1 + PX+S)Y,Sk=Ç)Pk and

recall that limp_>oo(l +x/p)p = exp(jc) and q(X,c) = 1 - exp(- Ike). There are three
cases: d = 0, 0 < d < g, and d = g.
First assume that d = 0. Then 5 = c* is fixed, and lim^^^ s/Nd^c*^8~d
(1 -p2O+«)) = lim^oof 5^*^(1 -p2(1+5))/(l -p) = 1^^00,5^ WO +P1+5)
ZLoP*) = 2A(S+l).
Next assume that 0 < ¿/ < g. Then we have lim^^QQ s/Nd^>c* ^~^^k=0^ = ^»
because 1 > lim^^^ tS/Nd^c* S~l lSk=opk > K™N->oo,S/Nd^c* PS = L II follows
thatlim^œ5/^^c,^-^(l-p2(1+5)) = lirn^œ5/^^c,ayV-^(l-p2(1+5))/
x (1 -p)) = Hm^^^^^ 2(1 +p1+5)c*5"1 lf=op^ = 2Xc'
Finally assume that d = g. Then we have lim^^^ s,Nd_+c* N^~d(' -p2^1+5)) =
lim^_>OOf5/ArÄ^c.(l -p2(1+5>) = 1 -exp(-2/Lc*) = ^(A,c*) < oo. ■

A.4. Proof of Theorem 4. Note that limî Var(y/ _5 - ///) = 0. Then we obtain for
the parameter sequence p = 1 - ÀN~8 that lim^^^ ^S/Nd->c* N~dVax(yi9' - Hi) =
limôo,5/tf'-*c**2*"<'0 -p2(1+5))/0 -P2) = Hmôo^/yv^c^2^"^1 -
xp2(H-5))/(2/i) = i72â,c*)/(2/l) by Lemma AA It follows that Hm^^^^^,*
N-dVar(yÌ9ÌAyÌ92) = K™N->oo9S/N<'-+c<> N~d^(yiA£i,l) = cj2{a2q{X,c*)/{2X) +
1{J = 0}a2s). Furthermore, if 0 < d < g, we have HmAÔOíS/A,</_>c.[Af«""í/
x E(y/flA^2)l = -Wm^ôo.í/^^c^^^V^íyiM ~ /*/) = -^2Û,c*)/2.
Finally, if d = 0, we obtain that lim^^ iS/Nd^c< N8-dE(yiAAyia) = A(trMy - a2).
Note that when d = 0,v2 = (T2q(Ã,c*)/(2À) + (T2s.
When T = 3, pAB = p + (X^ w>1 Ae/)3)/(Xf=1 w>1 A^2) and pArbov = p +
ŒJL^c/.a + O-pîOA^VŒiL^.Â^).

Define X! =^"1/2)(1+^I^1 3^U A^, X3 = A^-1/2I^1((^ + (1 -p)//¿)A^2),

and X4 = AT1 X^ yi%2 Ay/,2- Moreover, let X2 = N'VW+di ^=i yiy' Ay/,2 if *gd < 0
and X2 = N-^-d^ y.A àyia if Agd > 0.
Recall that q(k,c) = 0 if c = 0. Then using results similar to those in Lemma A.2
we obtain as N, S -> oo with S/N -> c > 0, S/W« -> c+ > 0, and S/Nd -> c* > 0,
where 0 < g < 1 and 0 < d < g, for the parameter sequence p = 1 - ÀN~8 that Xj - >
AK0,2ff2((72$(A,c*)/(2A) + l{d = 0}(j15)), X3 S N(0,a4l and X4 4 W(<72(1 -
q(k, c+)/2), íT4^(/l, c)/(2A)). In addition, we obtain that

X2^N{X{afiy-a^),a2{a2q{)iic*)/{2X)^als)) if Ag¿/=0 and ¿ = 0,

X2-Íyv(-íT2^(/l,c)/2,í74^(/l,c)/(2/l)) ifA^=0 and ¿ > 0,

X2S N(0,(T2(<j2q(2.,c*)/(2l) + '{d = 0}cT2s)) if Ag¿ < 0,

X2 4/l(íT^ -a2) ifAg¿>0 and J = 0, and

X24-iT24(/l,c*)/2 ifAg£/>0 and d > 0,

and we obtain that limAf_^oo?5/^_^c*Cov(X1,X2) = -(72{o2q{k,c*)/{2X) +
l{d = 0}(j25) if Agd < 0, limN_>OOfS/^^c*Cov(X1,X2) = 0 if bgd > 0, and
limA(^oo,5/^^c* Cov(X3, X4) = 0.
Parts (i) and (ii) of Theorem 4 follow now straightforwardly from the preceding
results.

A.5. Proof of Theorem 7. When T = 3 and p = 1 , pAB = /> + (2/Li >7, 1 Ae/,3)A£/!l
^ iei)2) and pArbov = P + (I/Li e/,3e/f2)/Œ/Li 3>i,2*i'2)-
Let us define X, = ^-1/2(1+£/)I^i ^/,1 A^,3, X2 = iV1/2(l+^I^l ^ l£/j 2,
X3 = A^"1/2!^! C/>3Cif2. and X4 = N~^1^ I^=1 y/,2^,2. Let <r4r = .t2(cV2 +
'{d = 0}(ít15 +(J2)). Then it is easily seen that as N, S -> oo with S/Nd -> c* > 0 and
5/N^ -> c > 0 where ¿i > 0 and d = max(d, 1), one obtains that X' -> A^(0, 2o*r), X2 ->

N(0,a¿r), X3 4 A^(0,iJ4), and X4 4 A^(1{J < '}ct2,ccj4). In addition, one obtains that
limN_>OOíS/^_>c*Cov(X1,X2) = -í74r and lim;v_>oo5/yVi/_>c*Cov(X3,X4) = a
Parts (i) and (ii) of Theorem 7 now follow straightforwardly from the preceding results. ■

LEMMA A.5. Let p = 1. Furthermore, let N -^ oo, S -> oo simultaneously with

S/Nd -> c* > 0, H>/ié?ré? d > 0. Leí w;5, = y¡tt - ///, /e/ /, p, and q be positive integers, and

let Kq = E{eqit)laq < oo. Then assuming that the sum of the powers of "e,-^, m/,5, m;,,,
flttd Aujj is less than or equal to 8, we have

(i) limN^s/N^c N-dPE{u]Pt) < oo,

(ii) limAMoo,s/^_>itf-*'|E(«g+1)| <oo,
f/iO limAf_>00)S/Aíd_»^E[(Aií/,í)''] = KpoP,

(iv) WmN-,oo,siN'>^cN-dp'^ïPs^ui, ,)"]'< oo,

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1386 HUGO KRUINIGER

(v) lim^TO>5/JV</^^-*|E[iig+1(AiilV)^| <oo,

(vi) lim^TO>s/^^^-*|E[^iig(Aii/f/)^| <oo, am/
(vii) lim^^^^^-^lE^^fVA^)^! <oo.
Proof of Lemma A.5. Note that uut - w/,r_i + £/,> = Z¿=_s+1 ^k and Aii/^ = £;,/.
Note also that m;,_s = 0.
When d = 0 the proofs for (i)-(vii) are trivial because in this case the moments can be
written as sums of a finite number of bounded terms.
Therefore from now on we assume that d > 0.

Let limyv^oo be short for limyy^^ 9s/Nd^>c* •

We prove (i) and (ii) together.
First consider p = 0: E(w?,) = 1 and E(ujt) = 0 and hence limjvôoEiw?,) < oo and
Hindoo ECw/,) < oo. Now let p > 0: limôo NÊlu]?) = 'imNôo[N-dP

x IU,....,=-s+i E(n-i a,)] = c*Pa2/7 < °° and lim^°° N~dp IE(W^+1) =

The proofs of (iii) - (vii) are now straightforward.

(iii) Hindoo E[(Auitt)P] = E[(siit)P] = KpaP,

(iv) limyvôo^-^ EIiigíAii^^l^lim^ôo^-^lEIiige^ll < oo,
(v) Hindoo N-'P E[M^+'AM/,/)^]|=limôoN-^M^+1e^]<oo,
(vi) Hindoo N-dP E[e¡9kulP(Auu)«]'='imN^N-dP^
finally

(vii) limA^ootf-^lEIe^^
■

A.á. Proo/s of Lemma 3 and Theorem 8. In these proofs (p)limjv->oo is snort f°r
(pJlim^^QQ iS/Nd-*c*- Note that when P = ^ °/,i = ei,r» Ayi,t = e/,f » A^/ = e|» and
Z/^diagic?,...,^-1). ■
Proof of Lemma 3.
Part (i): The proof of the first two results is similar to the proof of the first part of
Theorem 7. As N, S -> oo with S/Nd -> c* > 0 and S/Nd -> c > 0 where d > 0 and
¿ = max(¿/, 1), we obtain that

r Nl-1/2W+d) jA Lz,,AD n p51 ^ ,„ _c/^

I" N-V^zf'Lvi 1 , FX521

~n'(°' ff4/diag(//r-3^7-_4,...,//i) -diag(Cr_3,C7-_4,...,C,)''l

.[W'ff V-diag(Cr_3(Cr_4,...,C,) / JJ'

Note that £56fll = E(X5lX'6l) = -o*rC # 0 and £55,22 = E(XS2X'62) = -a4

xdiag(Cr_3,Cr_4,..,C'1)/#0.
Next consider the off-diagonal blocks £55,12» ¿66,12» £56,12» and £56,21- We nave
£55,12 = E(^51^52) = limôoaW-VWEÍZ/'WzP) = 0. Similarly, £66,12 =
E(*6l ^62) = 'imNôo^^-lÊ(Z¡'zP) = 0, £56, 12 = E(X51 X'62) = - lim^«,
<72Af(-1/2)i/E(Z/'C"zP) = 0, and £56,2i = E(X52X'6l) = -limôo!*2^"1'2**
xE(Z/>/C/z/)] = 0.
Finally, it is easily seen that if T = 3, then X5 = X51 = o2rX' and X§ = X^' = o2rX.2,
where X' and X2 are defined in Theorem 7.
Part (ii): The proof of the first two results is similar to the proof of the last part of
Theorem 7. As N9 S -> 00 with S/Nd -> c* > 0 and S/N** -> c > 0 where J > 0 and
¿ = max(¿/, 1), we obtain that

r «-^-^ l^rH^ff » Vf Ml.

^(-i/ZKi^^z^-iJ UsiJ [VlW<l}^2^ Vo cI).
Note that E(X71X^) = 0. Because lim^_>00iV"1/2áE(Z/)/Ao/^_1ZÍ') = 0, we also
have E(X52Xg,) = 0. We conclude that E{X-]X%) = 0. Furthermore it is easily verified
that X7 - N(0, £77) with £77 = limAôo KllE(Z'l'^'z'l)Kllt = oAKlIKIlf.
Finally, it is easily seen that if T = 3, then X7 = X71 = <t2X3 and X8 = Xgj = a2X4,
where X3 and X4 are defined in Theorem 7.
Part (iii): Note that S/(pi) = .v/ -/?iy/,-i = »/ + (p -Pi)y/,-i and AS¡ (p') = Ay¡ -
P'Ayit-' = At>/ + (p-pi)Ay/t_i.
Lemma A.5 implies that lim^->ooE(AMijr)8< 00, lim^-ôo^~£/E[M?s(AM/i)6]< 00,
lim^ôo^-ÊtwfÂw/^)4] < 00, liniAôoEKAtt/^)4] < 00, lim^-xx, |E[iiífJ
x (Am/;)3] < cxD, and lim/v-»oo N~dE[uf^(àuitt)2] < 00 Vj,í. It follows from the
first result, our model assumptions, and the Cauchy-Schwarz inequality that
limyvôoVaraÂ^]2) < 00 Vm and hence N-l^=i(Z¡If[Av¡ v¡][Av¡ ofiz/7)
^' 'imNôoKIIE(Z¡nüiv¡Z¡I)Kin = oAKnKIlf. Similarly it follows that
^-(l+J)llLl(^/y/,-i<.1Z/') ^>- limôo^ÊÍZ^-l^-iZ/-) = cff4/,
^V-2 ZJLi VP'AyiÂyi^zl») q5' 0 and AT* I^=1 (zfyÂy^Z?') q3' 0.
Lemma A.5 and our model assumptions also imply that lim/v->oo |E[t>; ¿ui,s
x (Aii/ff)2] < 00 and limyvôo ^"ÊIü^m^íAii,-^)4] < 00 V*,j,í. It follows from
these results, our model assumptions, and the Cauchy-Schwarz inequality that Af~3/2 Y,fL'

x A0;z,P) «-51 0, and ^~3/2 5^, (Z/» Ay,- _, Aoizf) '-i" 0.

We conclude from the preceding results that áiag(N~^2Im^f-2)^~^2d ^T-2) x

I/l,(z/l/[^;(?,)^(?i)]'[^(pi)S/(pl)]Z/l)diag(JV-1/2/m_(7-_2),W-1/2j/7--2)
4 oAKUK>" + c<r4/C2diag(Om_(7-_2),/r-2); ^"' I/li (¿/"tAÏ^Plft) ?/(piè)]'

x [ÁSlípi^S/ÍPitWZ/') 4 ff4K"K'" + ca4iC2diag(OmHT.2)JT-2)-

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1388 HUGO KRUINIGER

Note that zff[Av' o/]' = ((Z/'A^)' (z/7/[AÌ- S/]')')'.

Lemma A.5, our model assumptions, and the Cauchy-Schwarz inequality imply that
limAôoAf-ÊKAu^)2«?,] < oo and limôo N'ÊliAv^fuf t] < oo Vj,/, and
hence AT^+^Z-^ÌZ/'Ao,- Aojz/) *-T iims^^NÊiZ1/ AüíAv¡Z¡). From part
(i) we have that iim^jôoNÊiZP AvíAv¡Z¡) = ''mNôoa2N-dE(ZIi' HZ¡) =
<7¿r//. It follows that plimyVôo^"(1+£/)I-li(Z//A¿/A¡;|z/) = ^r//.
Similar arguments show that plim^^ N^MQ+^JifL^K1 ! zV'oiAolZJ) = 0,
plimôo tf<- 1/2X3+1/) Z^1(zP/A>;/-_1Ay;z/)=0,andplimôo(N(-1/2)(2+^^x
ZJI^Z/^-iAdÍZ/^
c(T2KIIE[dmg(£fi...,ef-])'iAe'i] = c(j4KIID. Note that KIIzJI% = ¿/"[AÏ- S/].
It follows that ^^^^^^^ZîíZ/^í^/GoOS/ípOJAüíípOZ/) A-cgaKKuD
andA^"1/2)^^^/7'^^
Weconcludethatdiag(7V(-1/2)(1+í/)/r_2,^-1/2/m_(r_2),^-1/2^/r-2)Z/l1 (¿f
x [ASÍ (polipi)]' [A^z-(y91)^(^1)]Zf)diag(^(-1/2)(l+^)/r_2î ^-l/2/m_(:r_2)5
N-1WIt_2) 4 tr4W5-;52 and that diagí^í-1/2«1^)/^^^"172/«^^)^"172
/r^Z^^ft^í^^^^^H^^^^^^lzfdiagí^
/m_(7_2), N~l/2IT_2) -> i74^5rS2Z?' wnere ^5^52 and ^5r52¿) are PD mat"ces. ■

Proof of Theorem 8. The proof relies on results in (the proof of) Lemma 3.
Part (i): Recall that Z' = KABzfB' where ZfB = 'z' ZP] and rank(KAB) = m.
Therefore /5¡4£i is equal to a GMM estimator that exploits E(ZfBfAv¡) = 0 and uses the
weight matrix (N"1 2/Li ZfBfHZfB)~l. Consider the diagonal path limiting behavior
of this weight matrix. Since lim^-x» N~2dVar(yf { ) < oo by Lemma A.5, limyv->oo N~d
Var(^ij Aj/r) < oo, and limyv->cxD Var([AjZ)/]2) < oo, it is easily seen that ¿j2A^~^1+^

Z^! [z¡'HZ¡) q™- ¿55,11 and a2N^2)^d)^=x{z'f HZf>) q™' 0. Moreover,

^N^lîLi&P'nzft^'ism Therefore plim^00[a2^-(1+rf)ZJL1(Z///ÍZ/)
= Í55J1,plim^^ooa2^-1I^1(zP///zP) = Í55522,andplim^^ooa2^(-1/2)(2+í/)
ifLliZf'HZP) = £55,12 = 0. Furthermore I55 is PD. It follows from the preceding

results and the proof of Lemma 3 that pABi - p -> X'6 t~5l X5/X'6 t^5l X6. Assuming that

PAB exploits E(ZfB'Aüi) = 0, it also follows that pAB~P^ ^14^13/^4^14 with

X,3 = (^1 lW = 0}^2)/andX14 = (X^1 '{d = 0}X'62)'.
Part (ii): It follows in a straightforward manner from the results in the proof of Lemma 3

that plimN^ooipArbov -P) = O and N^^d(pArbov - p) 4 XfsWX7/XfsWXs.

Part (iii): From part (i) of Lemma 3 we have ^-l/2^l^^=l z/'Au; 4 X5',
A,(-l/2)(l^)I^iZ/,A^_i 4 Í6if N-W^zp,^. S X52, and N~^2^=l
zP'Ay^ 4 X62, with X5 = (X^ X'S2y - A^(0,¿55) and X6 = (X^X^Y -
^V(//6, ¿66).

From part (iii) of Lemma 3 we have AT1/2!^ z'%i 4 X71 - N(0, 1 77,22) and
A^(-l/2)(l+J)Iî z'-'yi-x ^ ¿81 ~ N{'[d < '}a2i,ca4l).
Recall that zf'[Av' ü¡] = ((Z/'Ai/)' (Z/7'[A¿- S?]')')' = KsZ?'[Av¡ v¡]' where
z/7/[Ay; »<]' = KnZ¡Ifüi and AT5 =diag(/r_2, tf77).
Considernow * 5 W~sr51 AT5' = AT1 X^ Kszf'AZfKSf and KsW~XSYS{bKSf =
N~x I^j Kszf'zfK*f. It is easily verified that plimôo aH1+¿) ^ (Z/'//z/) =
KV*2)//, Plim^00^-Il^1(^//Z//'Z//^770 = a2^7^7^ = <r-2£77, and
plimNôoNÍ-WV+VzíLiiZfCZl1 K11') = 0. We also have plimyvôJAH1*^
xI^1(Z//z/)] = K4r/i72)/.
We conclude from the preceding results that plim/v-ôcíÂsrs* - p) = 0 and
W(1/2)<Ws*-/>)4 X'iirfXj/X'tirfXt, fork= 1, 1^.
It immediately follows from part (iii) of Lemma 3 and the preceding results that

plimyy -+oo(psYS2k - p) = 0 for k = 2, 2b, and >/Ñ(psYS2(Pl) ~ p) ->

*i1WSy52*10/ííiWsy52*ll withX10 = (X^ X^2 1 W < lJX^/ and Xn =(0 jfy',
and N^2^(psYS2(P'b) - P) ^ ^11 ^5^52^10^/^11 ^5^52^11 with ¿io*, =
(*51 Ä7);-
Part (iv): From the results mentioned in part (iii) we conclude that A^(1/2)(^-^)

(PSYS-P) ^ X'ì6WXÌS/X'l6WXì6 with Xl5 = (X'5{ '{d = 0}X'7Y and Xi6 =

('{d> '}X'6{ X^y,andthatif¿/ < 1 piimu^ooipsYS- P) = 0. ■
A.7. Proof of Theorem 9.
Part (i): Recall that z'l = diag(zP, zf) and that zf = diag(zfB, z'' where zfB =
[Z/ ZP]. Also recall that KIIZ¡I/vi = Z¡If[Av¡ v¡] and from the proof of Theorem
3 that Kszff[Av' v¡] = Z?'[Av¡ v¡] when Ks = diag(/r_2, AT77). Similarly, we
have KllZ'l'yU-X = ¿/"[Ay/., y¡ _,] and /:5Zf [A^_, y¡ _,] = zf [Ay¡ _,

Note that when /)= 1, t>(),=e()í, A^,=e,)f, A>>? =s?, and Z¡' =diag(e?,...,e(^ ').
Recall the distributional results from parts (i) and (ii) of the proof of Lemma 3 and in
particular that Xj ± Xg.
Note that WN<Arbov2(l) = (AT'S,!, z/^oíz/7)-1 and W^.sy^d) =
(A^-'ljLjZf'IAi»; v'iÚAv'i v'¡]Zf)-]. Consider now ^"^'^^(l)^7" and
KsW^lSYS2(l)KSl. Recall that o*KIlK11' = £77. Then it is easy to verify that

1- Plimôo.^ê.iV-C^SJt^/'AoiAoiZ/ =ff*./#,
2. plimôcs/^. AH If=1 K"z!"v,v',z!'K'" = £77, and
3- Pltaôo.^-.c^*-1^2^^, K"Z¡»o,Av¡Zl =0.
We conclude from the preceding results that P^^Môo,s/Nd-^c* (PArbov2,R - 1) = 0,

and N^/2^{pArbovliR - 1) -^ Xgl^X-y/Xgi^Xg. Furthermore, we can conclude

that plim^^^ s/N<^>c(PSYS2,R - 1) = 0 and N^^d{psYS2,R - 0 ~> X'^l^X-j/

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms
1390 HUGO KRUINIGER

Note that SE(ßArboo2,p) = I^"1 X^ii^-i^/7*77')^77')"1 W^rtoirtO)

(tf77)"1 xíf^í^ZÍVl)]"1^ and that SE(&r52,J?) = [tf-'ljLjWA^,
>^;_1)zf/:^)(^5/)-1vv^,5KS2(i)(^5)~1S/L1(^52:f/(A^;_l y;^!/)]-172-
It follows from the preceding results that [S E(pArbov2R)]~2 / Nd -> Xg£^LX8
and hence (pArbov2,R - l)/SE{pArbov2,R) 4 {X'%t^ X%)'1 12 X^1 Xn . Because
X7 -1 *g, (XfsÍ-7l Xsrl/2XfsÍ-7l X7'XS - 7V(0, 1), and therefore (pArbov2,R - 1)/
SE(pArboü2,R)-ÍN(0, l).Similarly we obtain that [5£(y95y52,/?)]"2/^-^ ^8^77^8
and (ASKS2,/? - 1)/^(P5KS2,ä) ^ ^(0, 1).
Part (ii): We first consider the weight matrix and some other fourth-order moments.
Note that p = ' -AN'1/2. Let Uij =yiìt-M Vi. Thenlim^yvôo^ÊIXAM;^)8] < oo,
Kms^N-tooêq E[(Auij)2] = <r2, and provided s ¿ i, 'imSìN->oo,seq E( Am/,s AwZii) = 0.
It follows from these results, our model assumptions, and the Cauchy-Schwarz inequality
that lims^ôoÊCZ/û/^z/7) = oAI and ''mSyN-xx>tseq Var([ü/>5A^f/]2) < oo
Vj,í. Consequently we obtain that N"1 J^, z/ô/üJz/7 ^?' ít4/ as (5, N -> oo)jeá7
and hence plim5jv->oo,j^ ^-1 S/Li ^/7/y/y/^/7 =o*I- Similar arguments show that
pKms9Nôo,SeqN-lÌìLlZ!!'*yi*ylz!i = °4l> &™S,Nôo,seq[N-V2i?=x{Z!'
x (Ay/ - Ay/,_i)(Atf - Ay/.-i/z/)] = plim5,ôo,^ N^2^ ZJ'Avi Av¡Z¡ =
(f^Jf.andplimsôo,,^^^
a^-5/4IÍL1(z//aü/ü;z//) = o.
Next we consider second moments. By Lemma A. 2 we have that lim^jv-ôo,^^
E(uiitAujj) = a2 ß Vi, and Lemma A.3 yields Iini5jv-»oo,5^ ^~1E(wfí(^M/,í)2) = 0
Vi. It follows that A^-^^jZ/7^/.! ^' (<72/2)/ as (5,,/V -» oo)je^ and hence
plim^yvôo^^N-^^jZ/7^/.! = X8 = (í72/2)¿. Similarly, we obtain that
plimsôo,^ N~x S^j Z/'Ay/.-i = -(a2/^.
The preceding results imply that both p'imSiN-ôoyseqlSE(pArbov2yP)]~2/N =
X^Xs/a4 = (l/2)2m and p'ims,N->oo,seq[SE(pSYS29R)r2/N = X'sXs/a4.
Because N^4^ z'' Av>i 4 X5l - N(09 (£)H) and A^"1/2!^! z/7/ü/ 4 X7 -
N(0,a4I) as (5,7V ^ oo)^, we also obtain that (pArbov2,R - P)/SE(PArbov2,R) ->
a"2^^)"1/2^^ and (psYS29R'P)/SE(pSYs2,R) ^ a'Hx^y^X^X-j as
(5,7V ->oo)^.
Note that X7 1 X%. Therefore o~2 (X^Xg)"1/2^^ |X8 - 7V(0, 1). We conclude that
(PArbov2,R - l)/SE(pArboü2tR) -> N(-(Ã/2)Jm, 1) and (psYS2,R - 1)/SE(PSYS2,r)
Â N(-(À/2)y/n9').
Part (iii): We first consider the weight matrix and some other fourth-order moments.
Note that p = 1 - ÀN~1/2 and d > 0. Let u¡j = y/^ - /i/ Vi. Then
limA^->oo,s/^-:>c*E[(Al|/,i)8l < °°» limyvôo,5/^->c*E[(AM/,í)2] = íj2' and Pro-
vided s ^ i, limyy^^ s/Nd-*c* E(Aw/,5Aw/,r) = 0. It follows from these results,
our model assumptions,' and the Cauchy-Schwarz inequality that lim^^^ S/Nd^>c*
E(Z/7V,'Z/7) = o4l and lim^^ >S/Nd_>c* Var([Â>v,,]2) < oo Vj,V Conse-
quently we obtain that N'^fL^Z^'vi x v>'z'l )q^' a4 1 and hence plimyy^^ s/Nd-^c*

N~l £A^ z/^ü/üJz/7 = a4 1. Similar arguments show that plim^^^ yS/Nd_>c* N~x
I^/'^A^Z/'^4/.
Because lims^-x»,«? N~lE(eflkuf s(Auijt)2cl) < oo Vk9s9t and V/,<? > 0 by
Lemma A.3, we have a fortiori 'imN^œS/Nd_+c* N~lVar([yijAyij]2) < oo and
'imN^oo s^Nd^c* N~lWar([yi'Aüit]2) < oo Vi when ¿/ > 0. Moreover, when
d = 0, we have lim^^ tS/Nd^c* Varfly/j Ayiit]2) < oo and lim^^^ iS/Nd-+c*
Var([y,M ¿K,]2) < oo Vf! It follows that plim^TOfS/^c, ATÍ^S^Z/'

= (<t2$(¿,O/(2¿) +'[d = V'ols)G2H and pHm^oo S/tf^c^"0072^

x I^Z/'KA* - AyÂyft1 = p'imN^s/N^c.N-W™ ^X{Z¡>
x (Ay/)y-Z/7) = 0, where d = min(¿, 1/2) and q(À,c*)/(2Ã) = Hm^^^^^^^
N"^(l -p2(1+5))/(l -p2) (cf. $(A,c*)/(2A) in the proof of Theorem 4). Note that
q(À,c*)/(2Á) = 5+1 if d = 0; $(¿,c*)/(2A) = c* if 0 < d < 1/2; q(X,c*)/
(2/1) = ? (/l,c*)/(2/l) if d = 1/2; and §(A, c*)/(2A) = 1/(2/1) if d > 1/2.
Next we consider second moments. When 0 < d < 1/2, we have lim^^^ ^/Nd^>c*
E((l -p)ii? j) = Hmôo^/jv^^ o-2(l -p2(1+5))/2 = 0 by part (ii) of Lemma A.4.
When d = 0, we obtain lim^^^ ys/Nd->c* E(^ "" P)uì ') - ° also- FmaHy» wnen d >
1/2, we have lim^^s/^^* E((l -p)«? j) = ümNôcS/NW^cS/N'^c* ^^ ~
pW+S)y2] = t,2q(X9c)/2.
Recall that if c = 0, then #(A,c) = 0. Then it is easily verified that for d > 0
Hm^ôo^/î/^ÊÍii/Âii/^) = o2 + lim^^Qo^/^^ÊÍM/^-iAii,^) = <72
+ lim^ôo^/î^^^^/^^c^íÍP - l)«?,i) = *2 - tf2?(A,c)/2. Moreover,
because 'imSìN->oo,seq N~lE(u}t(Auift)2) = 0 by Lemma A.3, we have a
fortiori lim^^QQ^/^-^ ^"^(«^(Am/^)2) = 0 when í/ > 0. It follows that
Ar^jZ/'X-l ^>' ^2(1 - q(l,c)/2)i and hence plim^^^ tS/Nd^c.
N~lyLiL'z!"yi,-' = ^8 = ^2(! - q(k,c)/2)i. Similarly, we obtain that
P^N-,oo,S/Nd^ N-W^tíLi z!fAy^{ =0 whenJ > 0.
Finally, note that AT1/2*1^!^, Z/'A«/ Â X5l - A^(0, (a2tf(/l,c*)/(2/l)4- '{d =
0}a2s) g2H) and N^^^Z^vi 4x7- yV(0,ir4/) as A^,5-> oo with S/NX'2 ->
c > 0 and S/Nd -> c* > 0, where d > 0.
The preceding results imply that plim^^^ ìS/Nd->c* ^SEÌPArbov2ìR)'~2/N = XgX8/
GA = ('-q(Á,c)/2)2m and plim^^^ iS/NdÂSE(PSYS2,R)r2/N = Xf%X%/a'
Moreover, as N,S -> oo with S/Nl/2 -^ c > 0 and S/Nd -> c* > 0, where d > 0,
(PArbov2,R - P)/SE(pArboü2,R) 4 tf-îg)-1/2^*; and (p5K52,/? " p)/
5£(P5y52,/?) ^ a-2(í¿¿8)-1/2ÍJ¿7.
Note that X7 1 X8 Therefore (j-2(XfsX^)-^2XfsX7'Xs ~ N(0, 1). We conclude that

(PArbov2,R - l)/S£(PArW,/?) 4 #(-¿(1 - q(l9c)/2)Jn, 1) and (p5K52,/? " 0/

SE(psYS2,R) ^ A^M(1 -^U,c)/2)Vm, 1). ■

This content downloaded from 129.252.86.83 on Tue, 12 Jul 2016 22:03:31 UTC
All use subject to http://about.jstor.org/terms

Blundell Bond 1998
No ratings yet
Blundell Bond 1998
29 pages
pdynmc-introLong (1)
No ratings yet
pdynmc-introLong (1)
25 pages
SSRN Id2896087 PDF
No ratings yet
SSRN Id2896087 PDF
49 pages
Drukker XTDPD
No ratings yet
Drukker XTDPD
34 pages
Week 4_8
No ratings yet
Week 4_8
31 pages
Panel VAR
No ratings yet
Panel VAR
29 pages
SSRN Id276592
No ratings yet
SSRN Id276592
64 pages
Generalized Method of Moments Estimation: Econometrics C Lecture Note 8 Heino Bohn Nielsen November 20, 2013
No ratings yet
Generalized Method of Moments Estimation: Econometrics C Lecture Note 8 Heino Bohn Nielsen November 20, 2013
31 pages
Bond(2007 Lecture)Dynamic PD first difference GMM
No ratings yet
Bond(2007 Lecture)Dynamic PD first difference GMM
50 pages
GMM Var Stata
No ratings yet
GMM Var Stata
27 pages
Carrasco-GeneralizationGMMContinuum-2000
No ratings yet
Carrasco-GeneralizationGMMContinuum-2000
39 pages
Dynamic Panel Data Models. Nickell's Bias. Anderson-Hsiao Estimator. Arellano-Bond Estimator. System GMM Estimator or Blundell-Bond
No ratings yet
Dynamic Panel Data Models. Nickell's Bias. Anderson-Hsiao Estimator. Arellano-Bond Estimator. System GMM Estimator or Blundell-Bond
22 pages
Dynamic Panel Estimator: GMM: Dr. Elya Nabila Abdul Bahri
No ratings yet
Dynamic Panel Estimator: GMM: Dr. Elya Nabila Abdul Bahri
18 pages
Makro Ekonomi
No ratings yet
Makro Ekonomi
22 pages
Pvar Stata Modul
No ratings yet
Pvar Stata Modul
29 pages
Dynamic Panel Data CH 2. Dynamic Linear Panel Models: Pr. Philippe Polomé, Université Lumière Lyon 2
No ratings yet
Dynamic Panel Data CH 2. Dynamic Linear Panel Models: Pr. Philippe Polomé, Université Lumière Lyon 2
72 pages
How2Do_xtabond2
No ratings yet
How2Do_xtabond2
30 pages
GMM Estimation with persistent panel data an application to production functions
No ratings yet
GMM Estimation with persistent panel data an application to production functions
21 pages
Dynamic Panel Data Model - 2011
No ratings yet
Dynamic Panel Data Model - 2011
1 page
Clase Panel
No ratings yet
Clase Panel
70 pages
Estimation_of_dynamic_panel_data_models_2
No ratings yet
Estimation_of_dynamic_panel_data_models_2
37 pages
Books Bedi, Rajinder Singh Lambi Larki
100% (3)
Books Bedi, Rajinder Singh Lambi Larki
169 pages
Dynamic Panel Data
No ratings yet
Dynamic Panel Data
51 pages
Wooldridge-InstrumentalVariablesEstimation-2005
No ratings yet
Wooldridge-InstrumentalVariablesEstimation-2005
6 pages
hansen1982
No ratings yet
hansen1982
27 pages
On Bias, Inconsistency, and Efficiency of Various Estimators in Dynamic Panel Data Models
No ratings yet
On Bias, Inconsistency, and Efficiency of Various Estimators in Dynamic Panel Data Models
26 pages
Blundell Bond ER
No ratings yet
Blundell Bond ER
20 pages
Wooldridge InstrumentalVariablesEstimation 2005
No ratings yet
Wooldridge InstrumentalVariablesEstimation 2005
6 pages
Proposed Solutions (A) Introduce Exogenous Variables: (I) Anderson-Hsiao
No ratings yet
Proposed Solutions (A) Introduce Exogenous Variables: (I) Anderson-Hsiao
11 pages
On The Remarkable Success of The Arellano-Bond Estimator
No ratings yet
On The Remarkable Success of The Arellano-Bond Estimator
6 pages
GMM, MLE and Tests For Non-Linear Restrictions: 1 Generalized Method of Moments (GMM)
No ratings yet
GMM, MLE and Tests For Non-Linear Restrictions: 1 Generalized Method of Moments (GMM)
15 pages
Estimating Regression Models of Finite But Unknown Order
No ratings yet
Estimating Regression Models of Finite But Unknown Order
17 pages
Chapter 2. Dynamic Panel Data Models
No ratings yet
Chapter 2. Dynamic Panel Data Models
209 pages
MIT14 382S17 Lec10
No ratings yet
MIT14 382S17 Lec10
10 pages
Panel Data V
No ratings yet
Panel Data V
28 pages
Estimation of Spatial
No ratings yet
Estimation of Spatial
13 pages
Which Moments To Match
No ratings yet
Which Moments To Match
25 pages
AdvEcx Chp2 Full 1005
No ratings yet
AdvEcx Chp2 Full 1005
105 pages
Estimating Dynamic Panel Data Models: A Guide For Macroeconomists
No ratings yet
Estimating Dynamic Panel Data Models: A Guide For Macroeconomists
7 pages
Some Estimation Methods For Dynamic Panel Data Models: July 2014
No ratings yet
Some Estimation Methods For Dynamic Panel Data Models: July 2014
9 pages
Arellano Bond
No ratings yet
Arellano Bond
22 pages
GDFM
No ratings yet
GDFM
22 pages
NeurIPS 2020 Minimax Estimation of Conditional Moment Models Paper
No ratings yet
NeurIPS 2020 Minimax Estimation of Conditional Moment Models Paper
15 pages
Generalized Method of Moments
No ratings yet
Generalized Method of Moments
24 pages
GMM by Wooldridge
No ratings yet
GMM by Wooldridge
15 pages
Anderson and Hsiao (1981)
No ratings yet
Anderson and Hsiao (1981)
10 pages
Anderson Darling Minimisation
No ratings yet
Anderson Darling Minimisation
16 pages
Blundell 2000
No ratings yet
Blundell 2000
21 pages
Panel Data Models: Dynamic Panels and Unit Roots
No ratings yet
Panel Data Models: Dynamic Panels and Unit Roots
20 pages
Cambridge University Press Econometric Theory
No ratings yet
Cambridge University Press Econometric Theory
6 pages
Autoregressivemodelsformatrix-valuedtimeseries
No ratings yet
Autoregressivemodelsformatrix-valuedtimeseries
22 pages
Dynamic Econometric Models Time Series Econometrics For Microeconometricians 2011
No ratings yet
Dynamic Econometric Models Time Series Econometrics For Microeconometricians 2011
51 pages
BBBBBBBB
No ratings yet
BBBBBBBB
22 pages
Arellano y Bover - 1995
No ratings yet
Arellano y Bover - 1995
23 pages
Dynamic Factor Models M Watson
No ratings yet
Dynamic Factor Models M Watson
43 pages
Chapter 4 - Lecture Notes
No ratings yet
Chapter 4 - Lecture Notes
39 pages
Bollerslev y Wooldridge1992
No ratings yet
Bollerslev y Wooldridge1992
31 pages
Digital Signal Processing (DSP) with Python Programming
From Everand
Digital Signal Processing (DSP) with Python Programming
Maurice Charbit
No ratings yet
Understanding Vector Calculus: Practical Development and Solved Problems
From Everand
Understanding Vector Calculus: Practical Development and Solved Problems
Jerrold Franklin
No ratings yet
Stationary and Related Stochastic Processes: Sample Function Properties and Their Applications
From Everand
Stationary and Related Stochastic Processes: Sample Function Properties and Their Applications
Harald Cramér
4/5 (2)
2022 Tutorial Tests
No ratings yet
2022 Tutorial Tests
7 pages
Pipeline Response Buried Explosives2
No ratings yet
Pipeline Response Buried Explosives2
265 pages
Math Project
100% (3)
Math Project
30 pages
Parabola
No ratings yet
Parabola
9 pages
Test certificate 4.0 Clear -Inv. 5154
No ratings yet
Test certificate 4.0 Clear -Inv. 5154
1 page
How Does A Theodolite Works?: Equipment and Materials
No ratings yet
How Does A Theodolite Works?: Equipment and Materials
4 pages
EASY TO SCORE MANUAL 2025
No ratings yet
EASY TO SCORE MANUAL 2025
58 pages
Chapter 48 Applications of Matrices and Determinants: EXERCISE 201 Page 545
No ratings yet
Chapter 48 Applications of Matrices and Determinants: EXERCISE 201 Page 545
32 pages
DOC-20250407-WA0015.
No ratings yet
DOC-20250407-WA0015.
8 pages
Manual-Labor-pH-Meter 766-En-Knick PDF
No ratings yet
Manual-Labor-pH-Meter 766-En-Knick PDF
52 pages
313313-Theory of Machines
No ratings yet
313313-Theory of Machines
9 pages
PREDICTION TERRAIN - Goldman Et Al. (1971) - Predicting Metaboolic Energy Cost
No ratings yet
PREDICTION TERRAIN - Goldman Et Al. (1971) - Predicting Metaboolic Energy Cost
5 pages
2. Preference and Utility II Ver1.01
No ratings yet
2. Preference and Utility II Ver1.01
36 pages
Doreen MAssey, Space Place Gender Alıntı
No ratings yet
Doreen MAssey, Space Place Gender Alıntı
4 pages
Beta Binomial Distribution
No ratings yet
Beta Binomial Distribution
4 pages
Gozinta® Force Transducer Gozinta Force Transducer: Model GZ-10
No ratings yet
Gozinta® Force Transducer Gozinta Force Transducer: Model GZ-10
3 pages
Applied Measurement Techniques Riassunto
No ratings yet
Applied Measurement Techniques Riassunto
203 pages
Example-Isolated Footing
No ratings yet
Example-Isolated Footing
10 pages
PSPM SP025 CH 5
No ratings yet
PSPM SP025 CH 5
14 pages
Physics 5 - Hooke's Law
No ratings yet
Physics 5 - Hooke's Law
17 pages
Sarukura Lab - Utralviolet Optical Properties
No ratings yet
Sarukura Lab - Utralviolet Optical Properties
5 pages
Two Higher Order Nonlinear Evolution Equations
No ratings yet
Two Higher Order Nonlinear Evolution Equations
20 pages
Quantum Mechanics
No ratings yet
Quantum Mechanics
100 pages
ATD-ATDO1 Model For Color Appearances, Color Differences and Chromatic Adaptation, Guth (2002)
No ratings yet
ATD-ATDO1 Model For Color Appearances, Color Differences and Chromatic Adaptation, Guth (2002)
4 pages
2022 Maharashtra State Board Important Question
No ratings yet
2022 Maharashtra State Board Important Question
17 pages
Complete Download Manufacturing Engineering: Select Proceedings of CPIE 2019 Vishal S. Sharma PDF All Chapters
100% (3)
Complete Download Manufacturing Engineering: Select Proceedings of CPIE 2019 Vishal S. Sharma PDF All Chapters
52 pages
Astm D7830 D7830M 14
No ratings yet
Astm D7830 D7830M 14
6 pages
Center of Gravity (INTRODUCTION)
No ratings yet
Center of Gravity (INTRODUCTION)
28 pages
tau lepton polarization measurement in L3
No ratings yet
tau lepton polarization measurement in L3
179 pages
Physical_Sciences_Gr_12_Assessent_Frame_June_Paper_2_2025
No ratings yet
Physical_Sciences_Gr_12_Assessent_Frame_June_Paper_2_2025
2 pages

GMM Estimation and Inference in Dynamic Panel Data

Uploaded by

GMM Estimation and Inference in Dynamic Panel Data

Uploaded by

GMM Estimation and Inference in Dynamic Panel Data Models with Persistent Data

Author(s): Hugo Kruiniger

GMM ESTIMATION AND INFERENCE

In this paper we consider generalized method of moments-based (GMM-based)

In this paper we consider generalized method of moments-based (GMM-based)

1 348 © 2009 Cambridge University Press 0266-4666/09 $ 1 5 .00

estimator) breaks down when p = 1 , because the instruments,which are lagged

to the finite-sample distributions of AB estimators by assuming that p ap-

The cases of fixed S and covariance stationarity, which corresponds to 5 = oo,

equivalent to limyv->och'iris->oo N~l X/Li */,/ = Xt whereas N~l/2^=ì Xiìt -»

2. GMM ESTIMATION OF THE PANEL AR(1) MODEL

Consider the panel AR(1) model with random effects:

yi,t =pyi,t-'+m,t, (l)

a/,/ = ty +e,'/, where ni = 0 -p)Mi> (2)

for i = 1, . . . , AT and t = -5 + 1, . . . , T. The starting date of the {y,,,}, t = -5,

We assume random sampling in the cross section.1 The composite error

EO*f)=O and E(ju}) = rf<oo, (3)

E(e/,,ylW) =0, for t = -S + 1,..., 7' (5)

Efo-,Je,-ff) = 0, forjar, s,t = -S + 1, ..., T. (6)

E(e?,) = <72>0, forf = -S+l,...,7' (7)

E(j/,i -///) =0, E[(tf,i -///)///] =0, and (8)

2.2. GMM Estimators for the Panel AR(1) Model

E[mAB9sAP)i = E[y,',(Ay/f/ -pAyi,t-')] = 0,

forj = l,...,f-2 and i = 3,...,7' (10)

E[(y«,r-W,/-i)Ay/fi-i]=0, for/ = 3,...,7' (11)

E[y/fi(A#,,-pAtf,,-i)]=0, forf = 3,...,7' (12)

E[mArbov,s,t(p)] = E[(j/,/ -pyi,t-')àyi,s] =0,

fors = 2,...,í-l and / = 3,...,7' (13)

where the W^^YSkj 1 block corresponds to E(Z7/ Av¡) = 0.

Applying the Helmert transformation to y^ = pj/,/-i + (1 - p)m + £;,/,

3. ASYMPTOTIC PROPERTIES OF THE ESTIMATORS WHEN THE

Var(y/>i -Mi) = p2(1+5) Var(y/,-s - /¿/) + (1 - />2(1+5V 7 '(1 - A (14)

3.1. Local Asymptotics for the Covariance Stationary Case

ma.-,**;, ^re[g~«[-ivïp),|(_2, -')].

VS(^„-„4|, where[g]~«[(°),(¿ î)]. »O

(ii) (1) N-{'21»=X Zfvi A X7, and

X7 = (X'52 X'7iy - W(0, ¿77), ¿77 = oAKllKllt, andEiX-jX'z) = 0,

AT' X (¿/"[A? Sn'lA? SÍ1Z/7) 4 <74^"L2

^ tr4^77 /^/7/ + (a4/(2i))/C2diag(Om_(r-2), /r-2).

(i) (1) pABi -p-> X¿Z^X5/X¿S¿ X6.IfT> 3, paäi w asymptotically

(ii) plimN-+oo(ßArbov - P) = 0, a«i/ assuming that pArbov exploits E(Z77/

VÑ(pArbov -P)^ X'iWXn/X'iWX*.

(i) plim;v->ooWysrsi =0~2Wsys* and

(ii) diagCAT1/^, ^-1/2/m)Il^1(zf/[Aüí S/]'[Ai>; ^]Zf)diag(^-1/r_2,

JÑ(psYSk-p) -4 X'strfXj/X'itrfX*. fork = 1,1ft,

3.2. Diagonal Path Local Asymptotics

(i) NAsi(pAB-p)^N(0,2cT2tf/(X2(<Tfiy-tf)2)) if&gd>0 andd = 0t

fiAB P 4 p + ii-, with L X - N ( 9 9 ) ,

(*y2/*2)(Í, "/)] if^8d=0andd = 0f

?tBÍP+k, ^[^^[.^(^^(^ -•)]

ifAg¿ = 0 and d > 0,

*»*>+%■ -*[*;]-" [(S)- i(-i '>)] ifA"<o-

In the remainder of this section it is assumed that plim^.^ ts/Nd->c* ^n = W

(i) (1) tf (-'/Wi-KO^ zj'Avi 4 X51,

X5 = (¿m X'52Y - W(0, Í55), Xe = (X'6l X'62y~- N(Ji69 i6el and

(ii) (1) aH/2^! Zf-'vi 4 X71 and

N-> £ (Z¡"[Á?, S¡]'m S¡ÍZ¡') S o*W-Albml

= oAKu K11' + (GA/(2k))q(X,c)K2àmg(Om-(T-2)jT-2).

Proof. See the complete version of the paper available at http://webspace.qmul.

(i) (1) Pab' -pi X'6i^X5/X'6is5lX6. IfT > 3, A**, Í5 asymptotically

(2) assuming that pab exploits E(ZfBfAvi) = 0 in lieu ofE(Z¡ av¡) = 0,

with Xl3 = (X'5] '{d = 0}X'52Y and XH = {X'bX '[d = 0}X'62)'.

yíÑ (PArbov -p)l X'sWX7/X'iWX*.

Proof. See the complete version of the paper available at http://webspace.qmul.

(i) plimyv->ooWysrsi = WSys* and

y/Ñ(psYSk-p) -4 ^I^Xv/^i^Xg, /or* = 1, lfc,

(iv) assuming that psYS exploits E(Zf'[Av¡ v¡]) = 0 in lieu of E(Zf'

Proof. See the complete version of the paper available at http://webspace.qmul.

THEOREM 7. Let T = 3 and let p = 1. Furthermore, let N -> oo, S -» oo

wa.-,4fc. ^[ì;]-[(2).(_2, V)]-

w«»/*(fc*.-,>4£. •*»[!;] ~«[(lh,os,,).(¿ ")]■

(i) (I) Nt-VW-WZ^Zl'Avi 4 *5i,

(ii) (Ì) N-V2ï»=i Zf-'vi S X7{ and

In addition, we have the following results concerning the weight matrices:

xdiagC^-'/2)*'^)/^, ^-|/2/m_(7-2),A^-1/2/r-2)^ «r4^-^,

THEOREM %.LetT>7> and let p = '. Furthermore, let N -> oo, 5 -» oo

(y2/2)(Í, "/)] if^8d=0andd = 0f

»>+%■ -[;]-" [(S)- i(-i '>)] ifA"<o-

w«»/(fc.-,>4£. •*»[!;] ~«[(lh,os,,).(¿ ")]■

D. (yiA -Mi)'fii - (20)- 1)^2(1 -P2) and a = 1

»'4 - [î;]~"[-t^(î)-t(-2. -.')]• -