GMM Estimation from Incomplete and Rotating Panels · GMM Estimation from Incomplete and Rotating...
Transcript of GMM Estimation from Incomplete and Rotating Panels · GMM Estimation from Incomplete and Rotating...
GMM Estimation from Incomplete andRotating Panels
Pedro Albarran∗ and Manuel Arellano†
September 2008
1 Introduction
We consider the general problem of estimation and testing from a sequenceof overlapping moment conditions generated by incomplete or rotating paneldata. The crucial idea of our suggested method is to separate the problemof moment choice from that of estimation of optimal instruments. In thisway, we are able to form optimal combinations of all the moment conditionsgenerated by incomplete or rotating panels without experiencing an uncon-trolled increase in the number of first-stage coefficients. Our estimators areonly “GMM estimators” in the Sargan–Hansen sense of setting to zero linearcombinations of orthogonality conditions, but not in the sense of minimizinga cuadratic form in all the available moments. Rather, we form direct esti-mates of individual-specific optimal instruments pooling all the informationavailable in the sample.
2 Model and Estimator
Assumptions and Notation Consider a vector stochastic process wt∞t=−∞such that the joint distribution of a given time series wj =
(w(t0+1), ..., w(t0+T )
)satisfies rj moment conditions
Eψj(wj, θ
)= 0, (1)
∗Universidad Carlos III, Madrid†CEMFI, Madrid
1
where j is an index for the pair (t0, T ) and θ is a vector of unknown coefficientsof order k.1 Moreover, let V j = E
[ψj (wj, θ)ψj (wj, θ)
′], Dj = E [Υj (wj, θ)],
Υj (wj, θ) = ∂ψj (wj, θ) /∂θ′, and Πj = Dj′ (V j)−1
.The data consists of independent observations on N cross-sectional units
wj(1)1 , ..., w
j(N)N
where j (i) is the value of j for the i-th unit, which is
independent of wj(i)i . Thus, two units with the same value of j have identical
initial periods and time series length. The index j takes on values in the set1, 2, ..., J.
Let (t0i, Ti) be the pair that corresponds to j (i). The observed variablesfor individual i are therefore wi(t0i+1), ..., wi(t0i+Ti). Any wit with t ≤ t0i ort > t0i + Ti is well defined but regarded as a missing or latent variable.
Let ψj` be the `-th component of ψj (wj, θ) and let ιj`i be an indicator of
whether ψj` is observed for individual i (for given θ). Moreover, let Ij
i bea diagonal matrix of order rj whose `-th element is given by ιj`i. Note that
ψj (wj, θ) is observed for individual i when j = j (i) (i.e. Ij(i)i is an identity
matrix), but some of its elements may still be observable even if j 6= j (i).If ιj`i = 1 for j 6= j (i), then Eψj
` (wj, θ) = 0 is a redundant momentgiven those in Eψj(i)
(wj(i), θ
)= 0. For example, the entire vector ψj (wj, θ)
could just be a subset of ψj(i)(wj(i), θ
). This assumption is just a coherency
requirement, because in its absence the distribution of wj would satisfy moremoment conditions than those stated in (1).
Estimation We consider cross-sample (or multisample) estimators θthat solve
N∑i=1
Πj(i)ψj(i)(w
j(i)i , θ
)= 0 (2)
where Πj is an estimator of Πj based on a preliminary consistent estimate θas follows:
Πj = Dj′(V j)−1
Dj =
(N∑
i=1
Iji
)−1 N∑i=1
Iji Υj
i
1The vector of moments for a given j may effectively depend on only some of thecomponents of θ. We regard θ as the full parameter vector for all relevant j.
2
vec(V j)
=
(N∑
i=1
Iji ⊗ I
ji
)−1
vec
(N∑
i=1
Iji ψ
ji ψ
j′i I
ji
)
and ψji = ψj
(wj
i , θ)
and Υji = Υj
(wj
i , θ)
. Note that in these expressions j
need not coincide with j (i), so that some or all of the components in ψji or
Υji may be latent variables.
A computationally convenient form of extremum estimator for this prob-lem is
θ = arg minc∈Θ
N∑i=1
ψj(i)(w
j(i)i , c
)′Πj(i)′
[N∑
i=1
Πj(i)ψj(i)i ψ
j(i)′i Πj(i)′
]−1 N∑i=1
Πj(i)ψj(i)(w
j(i)i , c
).
Another possibility is given by
θ = arg minc∈Θ
N∑i=1
ψj(i)(w
j(i)i , c
)′Πj(i)′
[N∑
i=1
Dj(i)′(V j(i)
)−1
Dj(i)
]−1 N∑i=1
Πj(i)ψj(i)(w
j(i)i , c
).
Note that the weight matrices in these two expressions are different, butnevertheless the estimators coincide because the number of effective momentsis the same as the number of parameters.
Asymptotic Normality Taking a first-order expansion of (2) scaledby N−1/2 around the true value we have
0 =1√N
N∑i=1
Πj(i)ψj(i)(w
j(i)i , θ
)+
1
N
N∑i=1
Πj(i)∂ψj(i)
(w
j(i)i , θ
)∂c′
√N (θ − θ)+op (1) .
Moreover, under the assumption that for all j Πj p→ Πj as N →∞,
−E(Πj(i)Dj(i)
)√N(θ − θ
)=
1√N
N∑i=1
Πj(i)ψj(i)(w
j(i)i , θ
)+ op (1)
d→ N[0, E
(Πj(i)V j(i)Πj(i)′)] .
Finally, since Πj = Dj′ (V j)−1
, we have
√N(θ − θ
)d→ N (0,W ) (3)
3
where
W =[E(Dj(i)′ (V j(i)
)−1Dj(i)
)]−1
, (4)
which can be consistently estimated as
W =
[1
N
N∑i=1
Dj(i)′(V j(i)
)−1
Dj(i)
]−1
. (5)
Note that an alternative, equivalent expression for W is
W =
[J∑
j=1
Dj′ (V j)−1
Dj Pr (j)
]−1
. (6)
3 Linear Models with Fixed Effects and Pre-
determined Variables
A leading situation in the panel context is one in which moments are obtainedas orthogonality conditions between a transformed disturbance and laggedvalues of a vector of conditioning variables. In a linear model, we have
yit = x′itθ + ηi + vit E (zisvit) = 0 (s ≤ t)
where ηi is a fixed effect and zis is a vector of predetermined instruments.Letting wt = (yt, x
′t, z′t)′, the time series wj =
(w(t0+1), ..., w(t0+T )
)implies
the moment conditions
E
z(t0+1)...
z(t0+t)
(y∗(t0+t) − x∗′(t0+t)θ)
= 0 (t = 1, ..., T )
where v∗(t0+t) denotes forward orthogonal deviations (Arellano and Bover,
1995):
v∗(t0+t) =
(T − t
T − t+ 1
)1/2 [v(t0+t) −
1
T − t(v(t0+t+1) + ...+ v(t0+T )
)].
4
In a more compact notation, we can write
ψj(wj, θ
)= Zj′ (yj∗ −Xj∗θ
)≡ Zj′vj∗
Dj = −E(Zj′Xj∗)
V j = E(Zj′vj∗vj∗′Zj
)Πj = −E
(Zj′Xj∗) [E (Zj′vj∗vj∗′Zj
)]−1
where yj∗ =(y∗(t0+1), ..., y
∗(t0+T−1)
)′, etc.
Since ψj (wj, θ) is linear in θ, the cross-sample estimator has a closed-formexpression given by
θ =
(N∑
i=1
Πj(i)Zj(i)′i X
j(i)∗i
)−1 N∑i=1
Πj(i)Zj(i)′i y
j(i)∗i
where Πj = Dj′(V j)−1
and
Dj =
(N∑
i=1
Iji
)−1 N∑i=1
Iji Z
j(i)′i X
j(i)∗i .
A one-step choice of V j is
vec(V j
I
)=
(N∑
i=1
Iji ⊗ I
ji
)−1
vec
(N∑
i=1
Iji Z
j(i)′i Z
j(i)i Ij
i
),
and a two-step choice
vec(V j
II
)=
(N∑
i=1
Iji ⊗ I
ji
)−1
vec
(N∑
i=1
Iji Z
j(i)′i v
j(i)∗i v
j(i)∗′i Z
j(i)i Ij
i
)
where vj∗i denotes one-step residuals.
4 Comparisons with Alternative Estimators
In this section we compare the previous cross-sample GMM estimator θ withtwo alternative estimators. The first one is a pooled GMM estimator based
5
on the union of the available sample moments. The second is an expandedGMM estimator that minimizes the sum of GMM criteria for each balancedsubpanel. We find that pooled (or stacked) GMM is generally inefficient rel-
ative to θ, and that expanded GMM, while asymptotically equivalent to θ, isbased on a much larger number of first-stage coefficients than θ. The impli-cation is that expanded GMM is less robust than θ to alternative asymptoticplans, and is likely to exhibit poor finite sample properties.
4.1 Nonredundant Moments
Let ψ (w, θ) be a vector of dimension r containing the total number of nonre-dundant moments spanned by the J different time series available:
ψ (w, θ) =⋃j∈J
ψj(wj, θ
).
Note that ψ (w, θ) need not correspond to the moment implications from thedistribution of any single time series (e.g. the moment implications froma rotating panel of overlapping time series of four periods each, coveringtwenty periods in total, will differ from those of a complete twenty year-period panel).
The construction of ψ (w, θ) can be approached as follows. Let j1 be an
index for(t10, T
1)
corresponding to the longest time series among those with
the earliest start, so that
t10 = min (t0i)
T1
= max(Ti | t0i = t
10
),
and let ψj1
(wj1 , θ
)be the moments associated with such time series. Next,
let t20 be the earliest start for a time series going beyond t
10 + T
1:
t20 = min
(t0i | t0i + Ti > t
10 + T
1)
andT
2= max
(Ti | t0i = t
20
).
6
Form j2 ≡(t20, T
2)
and ψj2
(wj2 , θ
), and consider the partition
ψj2
(wj2 , θ
)=
ψj2a
(wj2 , θ
)ψ
j2b
(wj2 , θ
) ,
such that ψj2a
(wj2 , θ
)is observable to j1 individuals but ψ
j2b
(wj2 , θ
)is not.
Then form
ψ[2] (w, θ) =
ψj1
(wj1 , θ
)ψ
j2b
(wj2 , θ
) .
Next, consider
t30 = min
(t0i | t0i + Ti > t
20 + T
2)
T3
= max(Ti | t0i = t
30
)get j3 ≡
(t30, T
3)
and form
ψ[3] (w, θ) =
ψj1
(wj1 , θ
)ψ
j2b
(wj2 , θ
)ψ
j3b
(wj3 , θ
)
where ψj3b
(wj3 , θ
)is the subset of ψj3
(wj3 , θ
)that is not observed by the
j1 or j2 individuals. Moments are accumulated in this way until we get a
ψ[`] (w, θ) such that t`0 + T
`= max (t0i + Ti), which then coincides with the
full vector of nonredundant moments ψ (w, θ).
4.2 Pooled GMM
We can form ψi (c) = ψ (wi, c) for each i, despite the fact that there couldbe no single individual in the sample for whom the entire vector ψi (c) isobservable. Define an r × r diagonal matrix Ii of indicators of observabilityof the components of ψ (w, θ) for individual i. A pooled GMM estimator is
7
given by
θp = arg minc∈Θ
[N∑
i=1
Iiψ (wi, c)
]′ [ N∑i=1
Iiψ(wi, θ
)ψ(wi, θ
)′Ii
]−1 [ N∑i=1
Iiψ (wi, c)
].
An example of this method is the unbalanced panel estimator for dynamiclinear models proposed in Arellano and Bond (1991).
Following standard GMM theory, the asymptotic variance matrix of the
estimation error√N(θp − θ
)is
V ar(θp
)=(D′V −1D
)−1
where
D = E
[Ii∂ψ (wi, θ)
∂c′
]= E (Ii)E
[∂ψ (wi, θ)
∂c′
]V = E
[Iiψ (wi, θ)ψ (wi, θ)
′ Ii].
4.3 Expanded GMM: Minimizing the sum of GMMcriteria for each balanced subpanel
On the other hand, letting dki = 1 [j (i) = k], we can consider GMM estima-tion based on the list of moments:
ψ† (wi, c) =
d1iψ1 (w1
i , θ)...
dJiψJ(wJ
i , θ)
which leads to the estimator
θs = arg minc∈Θ
J∑j=1
[
N∑i=1
djiψj(wj
i , c)]′ [ N∑
i=1
djiψj(wj
i , θ)ψj(wj
i , θ)′]−1 [ N∑
i=1
djiψj(wj
i , c)]
with first-order conditions
J∑j=1
N∑i=1
djiΠ (c)j ψj(wj
i , c)
= 0
8
where
˜Π (c)j =
[N∑
i=1
dji
∂ψj(wj
i , c)
∂c
]′ [ N∑i=1
djiψj(wj
i , θ)ψj(wj
i , θ)′]−1
orN∑
i=1
˜Π (c)j(i) ψj(i)
(w
j(i)i , c
)= 0 (7)
Note that (7) differs in two ways from (2). Firstly the estimate of Π in
(2) is kept fixed, but more importantly,˜Π (c)j is estimated using only obser-
vations with dji = 1, whereas the component matrices of Πj are estimatedelement-by-element using all the observations available in each case.
As long as plimN→∞N−1∑N
i=1 dji > 0 for all j, θs and θ are asymptoti-cally equivalent, although their finite sample properties may be very different,specially if J is large, some N−1
∑Ni=1 dji are small, but there is considerable
overlap among individual time series for different values of j.Let N j =
∑Ni=1 dji be the number of individuals for which we observe a
time series with the length and origin specified by j. Let N j`k =
∑Ni=1 ι
j`iι
jki be
the number of individuals for which moments ψj` and ψj
k are observable. Notethat N j
`k ≥ N j. Standard asymptotic analysis for (7) requires that for all jplimN→∞N
j/N > 0, whereas for (2) the requirement is the milder conditionplimN→∞N
j`k/N > 0.
Example 1 As a simple example, suppose that for j = 1, 2 we ob-serve w1
i = wi1, wi2, wi3 and w2i = wi2, wi3, respectively, with associated
moments
ψ1(w1
i , θ)
=
zi1vi1
zi1vi2
zi2vi2
zi1vi3
zi2vi3
zi3vi3
, ψ2(w2
i , θ)
=
zi2vi2
zi2vi3
zi3vi3
.
Moreover, suppose that plimN→∞N1/N > 0 but N2/N → 0, so that the
condition for (7) does not hold. However, since ψ2 (w2i , θ) is also observed for
individuals with j = 1 the requirement for (2) is still satisfied.
9
Asymptotic Efficiency Let us write the asymptotic variance of θs andθ as
V ar(θs
)=(D†′V †−1D†
)−1(8)
where D† = E[∂ψ† (wi, c) /∂c
′] and V † = E[ψ† (wi, c)ψ
† (wi, c)′]. Equation
(8) is just an alternative expression for (4) or (6). Let the dimension ofψ† (wi, c) be r† =
∑Jj=1 rj. We can write
Iiψ (wi, c) = Hψ† (wi, c)
where H is an r × r† selection matrix (r ≤ r†). Therefore, D = HD†,V = HV †H ′, and[
V ar(θs
)]−1
−[V ar
(θp
)]−1
= D†′V †−1D† −D†′H ′(HV †H ′
)−1HD†
= G′[I −H ′
(HH
′)−1
H
]G ≥ 0
where G = V †−1/2D† and H = HV †1/2. This shows that θp is dominated by
θs in terms of asymptotic efficiency.
Example 2 Suppose that for j = 1, 2 we observe w1i = wi1, wi2 and
w2i = wi2, wi3, respectively, with associated moments
ψ1(w1
i , θ)
=
zi1vi1
zi1vi2
zi2vi2
, ψ2(w2
i , θ)
=
zi2vi2
zi2vi3
zi3vi3
where vit = yit − x′itθ, xit is k × 1, zit is q × 1, and wit = (yit, x
′it, z
′it)′. Thus,
pooled GMM is based on
N∑i=1
Iiψ (wi, θ) =N∑
i=1
d1izi1vi1
d1izi1vi2
zi2vi2
d2izi2vi3
d2izi3vi3
10
with2
D′N =
[N∑
i=1
Ii∂ψ (wi, θ)
∂θ′
]′= −
N∑i=1
(d1ixi1z
′i1 d1ixi2z
′i1 xi2z
′i2 d2ixi3z
′i2 d2ixi3z
′i3
).
Let us consider a one-step pooled GMM estimator with weight matrix
AN =
N∑
i=1
d1izi1z
′i1 0 0 0 0
0 d1izi1z′i1 d1izi1z
′i2 0 0
0 d1izi2z′i1 zi2z
′i2 0 0
0 0 0 d2izi2z′i2 d2izi2z
′i3
0 0 0 d2izi3z′i2 d2izi3z
′i3
−1
so that
−D′NAN =
(Π1
...Π†2...Π3
)where
Π1 =N∑
i=1
d1ixi1z′i1
(N∑
i=1
d1izi1z′i1
)−1
Π†2 =
(Π†21
...Π†22
)=( ∑N
i=1 d1ixi2z′i1
∑Ni=1 xi2z
′i2
)( N∑i=1
d1izi1z′i1 d1izi1z
′i2
d1izi2z′i1 zi2z
′i2
)−1
Π3 =
(Π31
...Π32
)=( ∑N
i=1 d2ixi3z′i2
∑Ni=1 d2ixi3z
′i3
)( N∑i=1
d2izi2z′i2 d2izi2z
′i3
d2izi3z′i2 d2izi3z
′i3
)−1
.
Notice that Π1 is the regression coefficient of xi1 on zi1 in the d1i = 1subsample. As long as plimN−1
∑Ni=1 d1i > 0, it is a consistent estimate
2In terms of the notation used in Arellano and Bond (1991), we have
N∑i=1
d1izi1vi1
d1izi1vi2
zi2vi2
d2izi2vi3
d2izi3vi3
=N∑
i=1
d1i
zi1 00 zi1
0 zi2
0 00 0
(vi1
vi2
)+ d2i
0 00 0zi2 00 zi2
0 zi3
(vi2
vi3
)where (
vi1
vi2
)=(yi1
yi2
)−(x′i1x′i2
)θ.
11
of Π1 = E (xi1z′i1) [E (zi1z
′i1)]−1. Π†2 is the regression coefficient of xi2 on
(d1iz′i1, z
′i2) in the full sample. It is therefore a consistent estimate of
Π†2 =
(Π†21
...Π†22
)=(p1E (xi2z
′i1) E (xi2z
′i2))( p1E (zi1z
′i1) p1E (zi1z
′i2)
p1E (zi2z′i1) E (zi2z
′i2)
)−1
where p1 = E (d1i). Finally, Π3 is the regression coefficient of xi3 on (z′i2, z′i3)
in the d2i = 1 subsample.First-order conditions are
D′NAN
N∑i=1
d1i
zi1 00 zi1
0 zi2
0 00 0
(vi1
vi2
)+ d2i
0 00 0zi2 00 zi2
0 zi3
(vi2
vi3
) = 0
orN∑
i=1
[d1i
(xi1
...xi2
)(vi1
vi2
)+ d2i
(xi2
...xi3
)(vi2
vi3
)]= 0
where
xi1 = Π1zi1
xi2 = Π†21d1izi1 + Π†22zi2
xi3 = Π31zi2 + Π32zi3.
The estimator can be written in the general form
θ =
N∑
i=1
[d1i (xi1x′i1 + xi2x
′i2) + d2i (xi2x
′i2 + xi3x
′i3)]
−1
N∑i=1
[d1i (xi1yi1 + x′i2yi2) + d2i (xi2yi2 + xi3yi3)] (9)
Expanded GMM is based on
N∑i=1
ψ† (wi, c) =N∑
i=1
d1izi1vi1
d1izi1vi2
d1izi2vi2
d2izi2vi2
d2izi2vi3
d2izi3vi3
12
leading to an estimator of the same form as (9) but which uses:
xi1 = Π1zi1
xi2 = Π21d1izi1 + Π22d1izi2 + Π∗2d2izi2
xi3 = Π31zi2 + Π32zi3.
where
Π2 =
(Π21
...Π22
)=( ∑N
i=1 d1ixi2z′i1
∑Ni=1 d1ixi2z
′i2
)( N∑i=1
d1izi1z′i1 d1izi1z
′i2
d1izi2z′i1 d1izi2z
′i2
)−1
Π∗2 =N∑
i=1
d2ixi2z′i2
(N∑
i=1
d2izi2z′i2
)−1
Π2 and Π∗2 are, respectively, consistent estimators of
Π2 =
(Π21
...Π22
)=(E (xi2z
′i1) E (xi2z
′i2))( E (zi1z
′i1) E (zi1z
′i2)
E (zi2z′i1) E (zi2z
′i2)
)−1
andΠ∗2 = E (xi2z
′i2) [E (zi2z
′i2)]−1.
Cross-sample GMM uses the same form of instruments as expanded GMM,but different estimates of the first-stage coefficients:
xi1 = Π1zi1
xi2 = Π21d1izi1 + Π22d1izi2 + Π∗2d2izi2
xi3 = Π31zi2 + Π32zi3.
where
Π2 =
(Π21
...Π22
)=( P
i d1ixi2z′i1Pi d1i
Pi xi2z′i2N
)( Pi d1izi1z′i1P
i d1i
Pi d1izi1z′i2P
i d1iPi d1izi2z′i1P
i d1i
Pi zi2z′i2N
)−1
=N∑
i=1
(d1ixi2z
′i1 d1xi2z
′i2
)( N∑i=1
d1izi1z′i1 d1izi1z
′i2
d1izi2z′i1 d1zi2z
′i2
)−1
,
13
Π∗2 =N∑
i=1
xi2z′i2
(N∑
i=1
zi2z′i2
)−1
,
and d1 = N−1∑N
i=1 d1i.
Note that Π∗2 and Π∗2 are both consistent for Π∗2, but Π∗2 is obtainedfrom the whole sample whereas Π∗2 is only based in the d2i = 1 subsample.Similarly, Π2 and Π2 are both consistent for Π2, but Π2 only uses the d1i = 1subsample, whereas Π2 also uses the information from the d2i = 1 observa-tions when available. Thus, contrary to expanded GMM, cross-sample GMMimposes the cross-subsample restrictions on first-stage coefficients implied bythe model.
Pooled GMM can be regarded as imposing the restriction
Π22 = Π∗2
in its specification of the instruments. That is, it imposes the constraint thatthe simple regression coefficient of xi2 on zi2 (in the d2i = 1 sample) coincideswith the zi2 coefficient in the multiple regression of xi2 on zi1 and zi2 (in thed1i = 1 sample). Since this restriction will only hold in special cases (ifΠ21 = 0 or if E (zi1z
′i2) = 0), in general pooled GMM will be asymptotically
less efficient than expanded GMM or cross-sample GMM.
5 Monte Carlo Experiments
In this section, we present some experimental evidence on the finite sampleperformance of our proposed estimator, the Cross-Sample GMM, and thetwo other competing alternatives, the Pooled and the Expanded GMM.
5.1 Minimum Distance Estimation
If the moment conditions are linear, the estimation problem can be formu-lated as one of enforcing restrictions on a covariance matrix. Suppose thatwe have
E [zs (yt − x′tβ)] = 0 s ≤ t.
14
Let us define ωst = E (zsyt), Ωst = E (zsx′t), dit is an indicator of whether
period t variables are observed for individual i, and for∑N
i=1 disdit > 0:
ωst =1∑N
i=1 disdit
N∑i=1
disditzisyit
Ωst =1∑N
i=1 disdit
N∑i=1
disditzisx′it.
Next, form
bstN =
(ωst − Ωstβ
vecΩst − vecΩst
)≡(ωst − (I ⊗ β′) vecΩst
vecΩst − vecΩst
)and let bN be a vector containing the bstN for all s, t such that
∑Ni=1 disdit >
0, and let θ contain β and the corresponding vecΩst. A pooled minimumdistance estimator of θ is
θPMD = arg min b′N V−1bN
where V is a consistent estimator of the variance of bN . Moreover, under thetransformation(
I − (I ⊗ β′)0 I
)bstN =
(ωst − Ωstβ
vecΩst − vecΩst
),
the second block is seen to consist of unrestricted moments. Thus, letting b∗Nbe a vector containing all the available ωst − Ωstβ, from standard propertiesof minimum distance estimation it turns out that βPMD (which is part of the
θPMD vector) is asymptotically equivalent to
β = arg min b∗′N V∗−1b∗N
where V ∗ is a consistent estimator of the variance of b∗N . Since
ωst − Ωstβ =1∑N
i=1 disdit
N∑i=1
disditzis (yit − x′itβ) ,
it should be clear that β coincides with the pooled GMM estimator.
15
Similarly, an extended minimum distance estimator can be constructedas follows. Let (s, t) be an observable pair for the j-th subpanel. Form
bst[j]N =
(ωj
st − Ωstβ
vecΩjst − vecΩst
)where ωj
st and Ωjst are j-th subpanel sample averages. Form a vector b
[j]N
for all (s, t) that are observable for the j-th subpanel. Thus, letting b†N =(b
[1]′N , ..., b
[J ]′N
)′, an extended minimum distance estimator is
θEMD = arg min b†′N
(V †)−1
b†N
where V † is a consistent estimator of the variance of b†N . Using a similar
argument as before, βEMD can be seen to be asymptotically equivalent tothe extended GMM estimator of β.
Suppose an (s, t) pair that is observable in subpanels j and j′. Pooled MD
merges bst[j]N and b
st[j′]N into a single average, whereas extended MD treats them
as separate moments. Now consider another (s′, t′) pair that is observable
in j but not in j′, so that bs′t′[j]N is correlated to b
st[j]N but not to b
st[j′]N . The
efficiency of EMD relative to PMD comes from the fact that extended MDtakes into account these patterns of correlations across subpanels in imposingthe constraints. In contrast, pooled MD cannot allow for these differencesin correlations because subpanel-specific moments have been pooled into asingle aggregate moment.
16
References
[1] Arellano, M. and S. R. Bond (1991): “Some Tests of Specification forPanel Data: Monte Carlo Evidence and an Application to EmploymentEquations, Review of Economic Studies, 58, 277-297.
[2] Arellano, M. and O. Bover (1995): “Another Look at the Instrumental-Variable Estimation of Error-Components Models”, Journal of Econo-metrics, 68, 29-51.
[3] Hansen, L. P. (1982): “Large Sample Properties of Generalized Methodof Moments Estimators”, Econometrica, 50, 1029-1054.
[4] Sargan, J. D. (1958): “The Estimation of Economic Relationships UsingInstrumental Variables”, Econometrica, 26, 393-415.
17
Appendix
18
Tab
le1:
Mon
teC
arlo
Sim
ula
tion
Res
ult
s.P
aram
eter
valu
eα
=0.
2
T=
6T
=8
T=
10n=
100
n=
250
n=
100
n=
250
n=
100
n=
250
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
J=0(∗)
Med
ian
0.17
890.
1789
0.17
890.
1911
0.19
110.
1911
0.18
340.
1834
0.18
350.
1929
0.19
290.
1929
0.18
300.
1830
0.18
300.
1916
0.19
160.
1916
IQR
0.10
770.
1077
0.10
770.
0715
0.07
150.
0715
0.08
280.
0828
0.08
250.
0531
0.05
330.
0531
0.07
000.
0700
0.07
000.
0430
0.04
300.
0430
MA
E0.
0548
0.05
480.
0548
0.03
690.
0369
0.03
690.
0423
0.04
230.
0422
0.02
630.
0264
0.02
640.
0374
0.03
740.
0374
0.02
200.
0220
0.02
20J=
2M
edia
n0.
1632
0.15
490.
1706
0.18
610.
1814
0.18
950.
1752
0.16
540.
1814
0.18
970.
1854
0.19
250.
1723
0.16
450.
1783
0.19
040.
1865
0.19
25IQ
R0.
1334
0.12
920.
1395
0.08
750.
0848
0.08
940.
0915
0.08
920.
0977
0.06
290.
0615
0.06
140.
0769
0.07
360.
0778
0.04
740.
0456
0.04
64M
AE
0.06
950.
0725
0.07
070.
0446
0.04
410.
0465
0.04
900.
0503
0.04
900.
0332
0.03
320.
0336
0.04
150.
0446
0.04
140.
0235
0.02
410.
0239
J=4
Med
ian
0.15
770.
1199
0.17
480.
1792
0.16
190.
1843
0.16
570.
1308
0.17
470.
1865
0.16
930.
1901
0.16
860.
1326
0.17
930.
1870
0.17
050.
1914
IQR
0.15
130.
1517
0.15
460.
1041
0.09
580.
1037
0.10
050.
0962
0.10
460.
0682
0.06
700.
0697
0.07
530.
0705
0.07
790.
0479
0.04
570.
0490
MA
E0.
0836
0.09
470.
0801
0.05
250.
0570
0.05
300.
0565
0.07
320.
0538
0.03
620.
0392
0.03
590.
0445
0.06
770.
0414
0.02
580.
0327
0.02
55J=
6M
edia
n0.
1430
0.08
790.
1669
0.16
790.
1404
0.17
510.
1587
0.09
730.
1718
0.18
260.
1540
0.18
780.
1633
0.10
750.
1773
0.18
480.
1573
0.18
90IQ
R0.
1729
0.16
580.
1755
0.11
400.
1072
0.11
590.
1086
0.09
800.
1108
0.07
500.
0702
0.07
170.
0821
0.07
040.
0815
0.05
230.
0486
0.05
27M
AE
0.09
550.
1177
0.09
080.
0652
0.07
250.
0612
0.06
290.
1030
0.05
920.
0394
0.04
920.
0381
0.04
760.
0925
0.04
640.
0272
0.04
380.
0273
J=8
Med
ian
0.15
130.
0737
0.17
160.
1785
0.13
860.
1893
0.15
880.
0794
0.17
260.
1811
0.14
400.
1870
IQR
0.11
700.
1030
0.11
720.
0828
0.07
210.
0837
0.08
670.
0729
0.09
250.
0560
0.05
340.
0557
MA
E0.
0695
0.12
630.
0612
0.04
470.
0627
0.04
270.
0522
0.12
060.
0479
0.03
160.
0560
0.03
04J=
10M
edia
n0.
1375
0.04
730.
1689
0.17
560.
1275
0.18
590.
1480
0.05
690.
1720
0.17
790.
1294
0.18
58IQ
R0.
1356
0.11
410.
1374
0.08
350.
0776
0.08
780.
0903
0.07
450.
0964
0.05
990.
0537
0.05
97M
AE
0.07
960.
1527
0.07
090.
0457
0.07
370.
0451
0.05
970.
1431
0.05
250.
0354
0.07
060.
0324
J=12
Med
ian
0.13
970.
0357
0.16
580.
1716
0.11
660.
1835
IQR
0.09
230.
0766
0.10
020.
0611
0.05
450.
0626
MA
E0.
0672
0.16
430.
0560
0.03
780.
0834
0.03
46J=
14M
edia
n0.
1342
0.01
800.
1639
0.16
710.
1060
0.18
01IQ
R0.
1038
0.08
600.
1065
0.06
590.
0588
0.06
75M
AE
0.07
170.
1820
0.05
940.
0418
0.09
400.
0383
Note
:W
eru
n1,0
00
rep
lica
tion
for
each
sam
ple
size
,u
nb
ala
nce
dn
ess
patt
ern
an
des
tim
ato
r.(∗
)T
he
thre
ees
tim
ato
rsare
equ
ivale
nt
for
bala
nce
dp
an
els
(save
for
min
or
nu
mer
ical
ap
pro
xim
ati
on
an
dro
un
din
gis
sues
).
19
Tab
le2:
Mon
teC
arlo
Sim
ula
tion
Res
ult
s.P
aram
eter
valu
eα
=0.
5
T=
6T
=8
T=
10n=
100
n=
250
n=
100
n=
250
n=
100
n=
250
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
J=0
Med
ian
0.46
990.
4699
0.46
990.
4871
0.48
710.
4871
0.47
210.
4721
0.47
220.
4890
0.48
900.
4890
0.47
310.
4731
0.47
310.
4885
0.48
850.
4885
IQR
0.12
970.
1297
0.12
970.
0851
0.08
510.
0851
0.09
500.
0950
0.09
490.
0575
0.05
750.
0575
0.07
400.
0740
0.07
400.
0462
0.04
620.
0462
MA
E0.
0678
0.06
780.
0678
0.04
340.
0434
0.04
340.
0502
0.05
020.
0501
0.03
030.
0303
0.03
030.
0398
0.03
980.
0398
0.02
440.
0244
0.02
44J=
2M
edia
n0.
4466
0.42
990.
4489
0.47
810.
4735
0.48
170.
4642
0.44
750.
4703
0.48
250.
4760
0.48
600.
4607
0.44
490.
4709
0.48
430.
4773
0.48
67IQ
R0.
1693
0.16
480.
1764
0.10
800.
1059
0.11
350.
1093
0.10
230.
1092
0.07
210.
0698
0.07
450.
0804
0.07
770.
0831
0.05
070.
0492
0.05
24M
AE
0.09
000.
0964
0.09
010.
0560
0.05
620.
0565
0.05
640.
0630
0.05
840.
0388
0.03
960.
0383
0.04
800.
0570
0.04
590.
0277
0.02
920.
0265
J=4
Med
ian
0.43
490.
3823
0.45
820.
4663
0.43
670.
4749
0.44
840.
3903
0.46
430.
4768
0.45
150.
4833
0.45
490.
4037
0.46
720.
4809
0.45
540.
4854
IQR
0.19
380.
1841
0.20
160.
1248
0.11
950.
1265
0.12
030.
1071
0.12
440.
0776
0.07
220.
0797
0.08
310.
0769
0.08
540.
0539
0.04
890.
0542
MA
E0.
1052
0.12
690.
0989
0.06
550.
0771
0.06
480.
0695
0.11
050.
0662
0.04
340.
0547
0.04
110.
0533
0.09
630.
0474
0.02
990.
0452
0.02
92J=
6M
edia
n0.
4019
0.32
400.
4392
0.44
680.
4064
0.46
560.
4329
0.34
730.
4544
0.47
050.
4282
0.48
080.
4454
0.36
500.
4611
0.47
710.
4347
0.48
20IQ
R0.
2057
0.19
460.
2213
0.14
460.
1310
0.14
330.
1303
0.10
630.
1261
0.08
400.
0789
0.08
260.
0894
0.07
670.
0940
0.05
970.
0528
0.05
89M
AE
0.12
640.
1769
0.12
050.
0822
0.10
580.
0786
0.08
050.
1527
0.07
300.
0465
0.07
270.
0446
0.06
090.
1350
0.05
340.
0342
0.06
540.
0330
J=8
Med
ian
0.41
940.
3137
0.45
340.
4656
0.40
650.
4785
0.43
580.
3300
0.45
700.
4711
0.41
530.
4807
IQR
0.13
910.
1256
0.13
530.
0933
0.08
050.
0929
0.09
830.
0808
0.10
480.
0625
0.05
630.
0634
MA
E0.
0912
0.18
630.
0785
0.05
280.
0941
0.04
720.
0691
0.17
000.
0568
0.03
760.
0847
0.03
48J=
10M
edia
n0.
3994
0.27
790.
4453
0.45
680.
3847
0.47
540.
4226
0.29
680.
4531
0.46
570.
3956
0.47
59IQ
R0.
1572
0.12
920.
1617
0.09
840.
0858
0.09
540.
1056
0.08
720.
1047
0.06
760.
0573
0.06
96M
AE
0.10
900.
2221
0.08
680.
0596
0.11
550.
0509
0.08
100.
2032
0.06
260.
0441
0.10
440.
0384
J=12
Med
ian
0.41
090.
2745
0.44
490.
4577
0.37
760.
4722
IQR
0.11
390.
0880
0.10
830.
0704
0.05
890.
0710
MA
E0.
0926
0.22
550.
0697
0.04
840.
1224
0.04
24J=
14M
edia
n0.
3985
0.25
490.
4378
0.45
040.
3627
0.46
98IQ
R0.
1177
0.09
450.
1220
0.07
800.
0658
0.07
60M
AE
0.10
440.
2451
0.07
570.
0551
0.13
730.
0469
Note
:W
eru
n1,0
00
rep
lica
tion
for
each
sam
ple
size
,u
nb
ala
nce
dn
ess
patt
ern
an
des
tim
ato
r.(∗
)T
he
thre
ees
tim
ato
rsare
equ
ivale
nt
for
bala
nce
dp
an
els
(save
for
min
or
nu
mer
ical
ap
pro
xim
ati
on
an
dro
un
din
gis
sues
).
20
Tab
le3:
Mon
teC
arlo
Sim
ula
tion
Res
ult
s.P
aram
eter
valu
eα
=0.
8
T=
6T
=8
T=
10n=
100
n=
250
n=
100
n=
250
n=
100
n=
250
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
Pool
.Expd.
CSm
p.
J=0(∗)
Med
ian
0.74
210.
7421
0.74
210.
7721
0.77
210.
7721
0.74
710.
7471
0.74
710.
7780
0.77
800.
7780
0.75
380.
7538
0.75
380.
7803
0.78
020.
7802
IQR
0.16
380.
1638
0.16
380.
1042
0.10
420.
1042
0.11
520.
1152
0.11
480.
0758
0.07
580.
0758
0.07
580.
0758
0.07
580.
0509
0.05
100.
0509
MA
E0.
0908
0.09
080.
0908
0.05
580.
0558
0.05
580.
0660
0.06
600.
0660
0.03
810.
0381
0.03
810.
0528
0.05
280.
0528
0.03
010.
0303
0.03
03J=
2M
edia
n0.
6900
0.66
840.
7170
0.75
190.
7415
0.75
840.
7275
0.69
850.
7396
0.76
840.
7549
0.77
720.
7369
0.70
710.
7478
0.77
300.
7609
0.77
60IQ
R0.
2150
0.20
420.
2284
0.15
170.
1449
0.15
580.
1233
0.11
470.
1304
0.08
810.
0820
0.08
680.
0888
0.08
800.
0931
0.05
800.
0580
0.06
10M
AE
0.13
370.
1461
0.12
970.
0835
0.08
550.
0846
0.08
340.
1046
0.07
660.
0488
0.05
380.
0477
0.06
700.
0930
0.06
090.
0358
0.04
190.
0354
J=4
Med
ian
0.66
870.
5779
0.70
370.
7331
0.68
230.
7522
0.70
670.
6204
0.73
330.
7607
0.70
990.
7725
0.72
430.
6416
0.74
480.
7661
0.72
220.
7737
IQR
0.24
760.
2182
0.25
700.
1640
0.15
840.
1746
0.14
030.
1262
0.14
630.
0905
0.08
650.
0958
0.09
490.
0858
0.10
090.
0633
0.05
840.
0677
MA
E0.
1592
0.22
350.
1488
0.09
970.
1226
0.09
630.
1011
0.17
960.
0904
0.05
430.
0904
0.05
290.
0775
0.15
840.
0655
0.04
220.
0778
0.03
92J=
6M
edia
n0.
6121
0.49
140.
6729
0.70
400.
6308
0.73
800.
6806
0.55
690.
7147
0.74
420.
6692
0.76
010.
7082
0.59
340.
7383
0.75
860.
6917
0.77
09IQ
R0.
2817
0.24
690.
3105
0.20
060.
1685
0.19
820.
1559
0.13
300.
1596
0.10
410.
0882
0.10
200.
1066
0.08
890.
1122
0.06
760.
0597
0.06
99M
AE
0.20
370.
3086
0.18
340.
1244
0.17
110.
1149
0.12
490.
2431
0.10
750.
0672
0.13
080.
0597
0.09
270.
2066
0.07
330.
0485
0.10
830.
0421
J=8
Med
ian
0.65
600.
5048
0.71
290.
7346
0.63
560.
7577
0.68
910.
5500
0.72
920.
7521
0.66
050.
7675
IQR
0.16
780.
1477
0.17
670.
1172
0.09
530.
1094
0.11
470.
0946
0.12
240.
0708
0.06
250.
0752
MA
E0.
1468
0.29
520.
1098
0.07
710.
1644
0.06
410.
1113
0.25
000.
0826
0.05
240.
1395
0.04
39J=
10M
edia
n0.
6209
0.47
060.
7042
0.71
540.
6006
0.74
940.
6687
0.51
140.
7150
0.73
990.
6352
0.76
21IQ
R0.
1808
0.15
600.
2039
0.12
980.
1056
0.12
750.
1251
0.09
790.
1324
0.08
370.
0691
0.08
37M
AE
0.18
000.
3294
0.12
050.
0938
0.19
940.
0743
0.13
130.
2886
0.09
310.
0642
0.16
480.
0485
J=12
Med
ian
0.65
170.
4853
0.70
240.
7284
0.60
720.
7571
IQR
0.14
320.
1000
0.14
040.
0854
0.07
630.
0853
MA
E0.
1490
0.31
470.
1047
0.07
400.
1928
0.05
49J=
14M
edia
n0.
6325
0.46
510.
6945
0.71
640.
5864
0.74
74IQ
R0.
1426
0.10
180.
1550
0.09
350.
0796
0.09
45M
AE
0.16
750.
3349
0.11
340.
0845
0.21
360.
0622
Note
:W
eru
n1,0
00
rep
lica
tion
for
each
sam
ple
size
,u
nb
ala
nce
dn
ess
patt
ern
an
des
tim
ato
r.(∗
)T
he
thre
ees
tim
ato
rsare
equ
ivale
nt
for
bala
nce
dp
an
els
(save
for
min
or
nu
mer
ical
ap
pro
xim
ati
on
an
dro
un
din
gis
sues
).
21