Professional Documents
Culture Documents
1
T1
(y
i2
+. . . +y
iT
) and
it
=
it
1
T1
(
i2
+. . . +
iT
)
One can show that the correlation between y
i,t1
and
it
is negative, due to
cor(y
i,t1
,
1
T1
i,t1
) < 0 and cor(
1
T1
y
it
,
it
) < 0, Therefore the LSDV
estimator is inconsistent and biased in dynamic models (for N and xed
T). The bias of the LSDV estimator was rst estimated by Nickell (1981). His
bias approximation follows:
= plim
N
(
lsdv
) =
2
h(, T)
(1
2
x y1
)
2
y1
3
, where =
x y1
/
2
x
h(, T) =
(T1)T+
T
T(T1)(1)
2
and
x y1
=
x y1
/
x
y1
annot.: variables denoted as x and y are within-transformed
Because h(, T) is always positive (when > 0), the LSDV estimate is down-
ward biased. It can be seen that the bias for the autoregressive coecient is
especially severe, when
1. the autoregressive coecient is high
2. the number of time periods T is low
3. the ratio of
2
/
2
y1
is high
Notice that the bias of the exogenous variables is proportional to the bias of the
autoregressive coecient, where simply indicates the regression coecient of
y
i,t1
on x
it
.
1.2 Advantages of dynamic modelling
However, dynamic modelling includes several advantages. One not only takes
into account (temporal) autocorrelation in the residuals, but one is also able to
reduce the amount of potential spurious regression, which may lead to wrong
inferences and inconsistent estimation in static models. Especially in the context
of tourism demand models, static models may lead to an overestimation of the
eects of the exogenous variables.
1
Furthermore the coecient of the lagged
dependent variable itself may be of interest. In the tourism demand context, it
indicates habit formation due to mouth-to-mouth eects.
2 Consistent estimators
This section deals with consistent estimators for dynamic panel data models.
2.1 First Diererence IV (Anderson/Hsiao, 1981)
A specic case of a GMM estimator is the First-Dierence IV. Individual xed
eects
i
are eliminated by dierencing instead of within-transforming.
y
it
= y
i,t1
+x
it
+
i
+
it
y
it
= y
i,t1
+x
it
+
it
In matrix notation:
F
y
= F
y1
+FX +F
1
Spurious regression was also present in the application described in this paper. Because
when using a static specication of the winter tourism demand model, a negative coecient
for snow cover is obtained, probably resulting from a positive evolution of the overstay nights
together with a negative evolution of the snow cover variable. In dynamic specication, the
more intuitive positive coecient is obtained, meaning that higher snow cover leads to higher
winter tourism demand.
4
where F = I
N
F
T
and F
T
=
_
_
_
_
_
1 1 0 . . . 0
0 1 1 . . . 0
.
.
. 0
.
.
.
.
.
.
0
0 0 0 1 1
_
_
_
_
_
However y
i,t1
is correlated with the error term
i,t1
. Consistent estima-
tion is still possible by using the IV method with y
i,t2
as instrument for y
i,t1
, because
2
E(y
i,t2
it
) = 0
Even when all moments are employed, this means if all valid lagged internal
variables are included as instruments, estimation is inecient, because not all
information, e.g.
it
MA(1), is used. Only for the case of homoskedasticity
of the residuals it can be shown that Andersion/Hsiao (1981) is the most ecient
GMM estimator.
2.2 GMM estimators
2.2.1 Dierence GMM (Arellano and Bond, 1991)
Ecient estimates are obtained using a GMM framework. Following moments
are exploited:
E[y
i,ts
it
] = 0 and E[X
i,ts
it
] = 0 for s 2; t = 3, . . . T
X
i
=
2
6
6
6
4
y
i2
y
i1
x
i3
x
i2
y
i3
y
i2
x
i4
x
i3
.
.
.
.
.
.
y
i,T1
y
i,T2
x
iT
x
i,T1
3
7
7
7
5
Z
i
=
2
6
6
6
6
6
4
[y
i1,
x
i1
, x
i2
] 0 0
0
.
.
.
0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 [y
i1
, . . . , y
i,T2
, x
i1
, . . . , x
i,T1
]
3
7
7
7
7
7
5
X = (y
1
, X), Z = (Z
1
, Z
2
, . . . , Z
N
)
= (,
)
The idea of the GMM framework is the following: L instruments imply a set of
L moments, i.a. g
i
(
) = y
i,ts
it
, where exogeneity holds when E(g
i
()) = 0.
Each of the L moment equations corresponds to a sample moment g(
) =
1
n
n
i=1
g
i
(
) = 0. But here
this is not possible due to model overidentication. More equations are present
than unknown variables, or equivalently speaking, more excluded instruments
than endogenous variables. As a consequence one is not able to solve the equa-
tion exactly. Instead, the criterion J
N
is minimized.
J
N
(
) = g(
W
N
g(
)
or in our notation:
J
N
=
_
1
N
N
i=1
i
Z
i
_
W
N
_
1
N
N
i=1
Z
i
_
2
valid under the assumption, that errors do not depend on time
5
where
W
N
is an estimated weighting matrix. The weighting matrix is the weak
point of the Anderson/Hsiao estimator, which uses the identity matrix as weight-
ing matrix and therefore assumes conditional homoskedasticity. Such a matrix
however overweights moments with higher variances and moments which are
correlated with each other. Optimal weighting and therefore ecient estima-
tion is assessed by using the inverse of the moment covariance matrix:
W
N
= V ar(Z
)
1
= (Z
Z)
1
This procedure is similar to the GLS procedure, but where GLS minimizes the
weighted sum of the second moments of the residuals, GMM minimizes the
weighted sum of the covariance structure of the moments. Unless is known,
ecient GMM is not feasible. A two-step procedure is therefore assessed. First
replace with some simple G (here: assuming
it
i.i.d., equivalent to conditional
homoskedasticity)
W
1N
=
_
N
i=1
Z
i
G
T
Z
i
_
1
= (Z
GZ)
1
where G =
`
I
N
G
and G
T
= F
T
F
T
=
2
6
6
6
6
6
4
2 1 0 0
1 2
.
.
.
0
0
.
.
.
.
.
.
1
0 0 1 2
3
7
7
7
7
7
5
which delivers (consistent) rst-step estimates. Its residuals
1i
are used for
the two-step estimation of
W.
W =
_
N
i=1
Z
i
1i
1i
Z
i
_
1
Ecient estimates for the Di-GMM are then obtained with:
EGMM
=
_
X
Z
WZ
X
_
1
X
Z
WZ
y
One can show that under homoskedasticity one-step estimates are asymptoti-
cally equivalent to two-step estimates. However when there is a high persistency
of y
it
, instruments are weak. (Blundell and Bond, 1998, Kitazawa, 2001). Be-
cause the dierences of a time series which is persistent (near a unit root) are
near to innovations and therefore dicult to instrument. This is known as the
weak instrument problem.
Briey I also discuss the estimation of the variance matrix.
one-step procedure
replacing with a sandwich-type proxy
1
delivers consistent and robust vari-
ances.
d
V ar[
1
] =
Z
W
1
Z
1
X
Z
W
1
Z
1
Z
W
1
Z
Z
W
1
Z
1
two-step procedure
6
using the optimal weighting matrix W = (Z
Z)
1
, above formula reduces to
d
V ar[
2
] =
Z(Z
1
Z)
1
Z
1
however:
V ar[
2
] can be heavily downward biased. For many years econome-
tricians therefore printed out the t-statistics of the one-step procedure, when
estimating two-step GMMs. However nowadays, windmeyers correction (2005)
is mostly used for two-step t-statistics which have proven to perform very well
in Monte Carlo simulations.
2.2.2 System-GMM (Blundell/Bond, 1998)
It was invented to tackle the weak instrument problem. It builds up a system of
two equations, the level equation and the dierence equation. Endogenous
variables in the level equation are instrumented by lagged dierences. Summa-
rizing, where Arellano-Bond instruments dierences [. . . ] with levels, Blundell-
Bond instruments levels with dierences. [...] For random walklike variables,
past changes may indeed be more predictive of current levels than past levels
are of current changes (Roodman, 2006). Following additional moments are
explored:
E[y
i,t1
(
i
+
it
)] = 0 and E[X
i,t1
(
i
+
it
)] = 0, for t = 3, . . . , T
In matrix notation:
X
i
=
_
_
y
i2
y
i1
x
i3
x
i2
y
i3
y
i2
x
i4
x
i3
.
.
.
.
.
.
y
i,T1
y
i,T2
x
iT
x
i,T1
y
i2
x
i2
.
.
.
.
.
.
y
i,T1
x
iT
_
_
Z
i
=
_
Z
D
i
0
0 Z
L
i
_
Z
L
i
=
_
_
[y
i2
, x
i2
, x
i3
] 0 0
0
.
.
.
0
.
.
.
.
.
.
.
.
.
0
0 0 [y
i2
, . . . , y
i,T2
, x
i,2
, . . . , x
i,T
]
_
_
2.2.3 Validity of the instruments
Performance depends strongly on the validity of the instruments. A valid in-
strumental variable z requires:
1. E[ | z] = 0 (exogeneity)
2. cov(z, x) = 0 (relevance)
These assumptions can be tested, using procedures like the overidentifying re-
strictions test which is also used in the context of IV estimation.
Overidentifying restrictions test (Sargan/Hansen Test)
The assumption of exogeneity of all instruments is tested using the Hansen-
Statistic J(
EGMM
)
2
LK
or the Sargan-Statistic, which, as opposed
7
to Hansen, assumes conditional homoskedasticity. The main problem with
the Hansen-statistic is that it is seriously weakened by the use of a large
amount of instruments. By denition, if the number of instruments is
above the number of (cross-section) observations, it will deliver a p-value
of one, which is meaningless. The Sargan-statistic is known as relatively
robust to the use of a large number of instruments, but may be inconsistent
under heteroskedasticity or at least, tests based on Sargan-statistic have
little power. One could also test subset of instruments, which is especially
helpful in analyzing the new instruments employed for the System-GMM.
This can be done by the Dierence-in-Sargan statistic: DS = S
u
S
r
2
Arellano and Bond - Autocorrelation test
It examines the dierenced residuals to nd autocorrelation of the resid-
uals in levels, which would be an indicator of weak, not exogenous instru-
ments.
It is dicult to detect how many instruments should be used in GMM esti-
mation. Typically one employs more instruments to increase eciency of the
estimation. However it was shown that more instruments also increase the -
nite sample bias (Bun/Kiviet, 2002). Thus it exists a trade-o between small
sample bias and eciency. It is suggested by Hahn that a restricted subset of
instruments should be used, using a number of instruments at least below the
number of observations.
2.3 Bias corrected LSDV estimator
Consistent estimation is obtained by additive bias correction. One could es-
timate LSDV coecients and correct these by subtracting their (estimated)
biases. The nowadays relatively large amount of bias corrected LSDV estima-
tors may be distinguished according to their use of external information. Bias
corrected LSDV estimators, which
use a preliminary consistent estimator
Kiviet (1995)
Hansen (2001)
Hahn and Kuersteiner (2002) - not for short T
Bruno (2005) - for unbalanced panels and short T
do not need a preliminary consistent estimator
Bun/Carree (2005)
2.3.1 Bruno (2005)
This paper will explain briey the procedure suggested by Bruno (2005). Bias
approximations emerge with an increasing level of accuracy:
B
1
= c
1
(
T
1
), B
2
= B
1
+c
2
(N
1
T
1
) and B
3
= B
2
+c
3
(N
1
T
2
)
where c
1
, c
2
and c
3
depend i.a. on
2