Panel Econometrics V

ECONOMETRICS
Outline

Dynamic models

Introduction

The Anderson and Hsiao estimator

The Arellano and Bond estimator - The pure AR(1) case

ECONOMETRICS
Dynamics
Many economic relations are dynamic in nature:
I E.g. state dependence: today’s state yit depends on yesterday’s yi,t −1 .
I Then, one of the advantages of panel data is that they allow the researcher to
better understand the dynamics of adjustment.

Take a model with adjustment costs. Suppose the optimal quantity yit∗ is
yit∗ = xit γ + ṽit ,
but due to adjustment costs (governed by α), the realization is
yit = αyi,t −1 + (1 − α)yit∗ .
Putting the two equations together yields the estimable model:
yit = αyi,t −1 + xit (1 − α)γ + (1 − α)ṽit .
| {z } | {z }
β vit
ECONOMETRICS
The dynamic panel model

Suppose the structural model is

yit = αyi,t −1 + xit β + vit i = 1, . . . , N, t = 2, . . . , T

where α is a scalar, xit is 1 × K and β is K × 1.

We assume there is a one-way error component structure

vit = ci + uit .

We will be precise about the assumptions regarding ci and uit later.

This dynamic panel data model has two sources of persistence over time:
I state dependency (yit is a direct function of yi,t −1 ) and
I individual effects (like in static models).
ECONOMETRICS
Inconsistency of the OLS estimator

I Since yit is a function of ci , it immediately follows that yi,t −1 is also a function of

ci and the strict exogeneity assumption fails.
I Therefore, the right-hand side regressor yi,t −1 is correlated with the error term vit
through ci .
I Consequence: the OLS estimator is biased and inconsistent!
I Compare this to pure time series models: here the OLS estimator of the dynamic
model is consistent if the disturbance is white noise.
I In time series models, (a) there is no individual effect and (b) T is assumed to be
large (hence we use large T asymptotics).
I In panel models, (a) the individual effect introduces autocorrelation in the
disturbance and (b) N is assumed to be large (hence we use large N asymptotics).

ECONOMETRICS
Inconsistency of the within estimator
The FE estimator wipes out the ci by the within transformation:
T
1
ÿi,t −1 = yi,t −1 − ȳi,−1 = yi,t −1 − ∑
T − 1 t =2
yi,t −1 .

and
T
1
T − 1 t∑
üit = uit − ūi = uit − uit .
=2

I Hence, the regressor ÿi,t −1 is a function of yi,1 , . . . , yi,T −1 while the disturbance
üit is a function of ui,2 , . . . , ui,T , so there is obvious correlation between the two
(“regressor endogeneity”).
I This correlation makes the FE estimator biased (“Nickell bias”, see Nickell, 1981).
I This bias does not vanish as the number of individuals increases, so the FE
estimator is inconsistent for N large and T small.
I Only as T gets large the FE estimator becomes consistent.
ECONOMETRICS
Inconsistency of the RE estimator

The RE estimator applies quasi-demeaning to the regressor

ỹi,t −1 = yi,t −1 − (1 − φ)ȳi,−1

and the disturbance

ũit = uit − (1 − φ)ūi .
I As before, the regressor ỹi,t −1 is a function of yi,1 , . . . , yi,T −1 while the
disturbance ũit is a function of ui,2 , . . . , ui,T , so there is obvious correlation
between the two.
I Hence, the RE estimator is biased and inconsistent in a dynamic panel data model
as well.

ECONOMETRICS
Bias correction procedures

Several suggestions to correct for the bias of the popular FE estimator have been
proposed in the literature:
I Kiviet (1995): derives an approximation for the bias of the FE estimator in a
dynamic panel data model with serially uncorrelated disturbances and strictly
exogenous regressors. A bias corrected FE estimator then subtracts a consistent
estimator of this bias from the original FE estimator.
I Everaert and Pozzi (2007): bias correction for the FE estimator based on an
iterative bootstrap procedure.
I Bun and Carree (2006): derive the asymptotic bias of the FE estimator for finite
T and large N in the presence of both time-series and cross-section
heteroskedasticity; again, bias correction procedures.
Instead of such correction procedures we will follow the bulk of the literature and use
instrumental variables approaches to find consistent estimators (AH, AB, BB; others
exist, see textbooks).
ECONOMETRICS
Panel AR(1) model and assumptions
Since yi,t −1 is the problematic regressor, let us start with the AR(1) model:

yit = αyi,t −1 + vit , vit = ci + uit , t = 2, . . . , T .

Assumption AR.1 (sequential exogeneity/predeterminedness):

E(uit |yi,t −1 , . . . , yi1 , ci ) = 0 for all t = 2, . . . , T .

This implies dynamic completeness conditional on ci :

E(yit |yi,t −1 , . . . , yi1 , ci ) = αyi,t −1 + ci .

Discussion:
I Sequential exogeneity replaces strict exogeneity assumptions.
I Dynamic completeness means the dynamics of yit is fully specified.
I A consequence is that uit is white noise, E(uit uis ) = 0 ∀ s 6= t.
ECONOMETRICS
Backward substitution

For further use, rewrite the AR(1) model by recursive backward substitution:

yit = αyi,t −1 + vit

= α(αyi,t −2 + vi,t −1 ) + vit = α2 yi,t −2 + vit + αvi,t −1
..
.
= αt −1 yi1 + vit + αvi,t −1 + . . . + αt −2 vi2

Taking first differences yields (since ∆vit = ∆uit )

∆yit = αt −2 ∆yi2 + ∆uit + α∆ui,t −1 + . . . + αt −3 ∆ui3 .

ECONOMETRICS
Moment conditions
Apply the first difference (FD) transformation to get rid of ci :
yit − yi,t −1 = α(yi,t −1 − yi,t −2 ) + uit − ui,t −1 .

How to consistently estimate this equation?

I The strict exogeneity assumption fails. We can also argue directly: From the
previous slide we know that the regressor ∆yi,t −1 is a function of ui,t −1 and hence
is correlated with the error ∆uit .
I Hence, POLS is inconsistent.
I Anderson and Hsiao (1981) suggest to use an instrument for ∆yi,t −1 .
I Use either ∆yi,t −2 = yi,t −2 − yi,t −3 or yi,t −2 as instrument.
I These instruments will not be correlated with ∆uit = ui,t − ui,t −1 as long as the
uit is white noise. ECONOMETRICS
Anderson-Hsiao details
Note that assumption AR.1 ensures validity of the Anderson and Hsiao moment
conditions. Specifically, it implies E(yi,t −k uit ) = 0 for k ≥ 1.

Hence, for all t = 3, . . . , T ,

E(yi,t −2 ∆uit ) = E(yi,t −2 uit ) − E(yi,t −2 ui,t −1 ) = 0
and, for all t = 4, . . . , T ,
E(∆yi,t −2 ∆uit ) = E(yi,t −2 ∆uit ) − E(yi,t −3 ∆uit ) = 0.

Which instrument should we use?

I Using yi,t −2 leaves us one observation more per individual.
I Arellano (1989) finds that the estimator using ∆yi,t −2 as instruments has a
singularity point and very large variances over a significant range of parameter
values.
I In contrast, the estimator that uses instruments in levels, i.e. yi,t −2 , has no
singularities and much smaller variances. ECONOMETRICS
Matrix form ?
FD of structural model

yit − yi,t −1 = α(yi,t −1 − yi,t −2 ) + uit − ui,t −1 .

Collect all observations for individual i in the (T − 2) × 1 vectors
     
yi ,3 − yi ,2 yi ,2 − yi ,1 ui ,3 − ui ,2
∆yi =  .. .. ..
 , ∆yi ,−1 =   , ∆ui =  .
     
. . .
yi , T − yi , T − 1 yi ,T −1 − yi ,T −2 ui ,T − ui ,T −1

The equation becomes

∆yi = α∆yi,−1 + ∆ui , i = 1, . . . , N.

Stacking the observations of all individuals yields

∆y = α∆y−1 + ∆u,

where ∆y = (∆y10 , . . . , ∆yN

0 )0 , ∆y
−1 = ( ∆y1,−1 , . . . , ∆yN,−1 ) etc.
0 0 0
ECONOMETRICS
The Anderson and Hsiao instrumental variables estimator
Starting point: orthogonality (moment) conditions

E[yi,t −2 ∆uit ] = 0, t = 3, . . . , T .

Define the (T − 2) × 1 vector of instruments

 
yi,1
yi,−2 =  ... 
 

yi,T −2

Write the orthogonality condition in stacked form:

E[yi,0 −2 ∆ui ] = 0.
ECONOMETRICS
Sample moment conditions

Sample equivalent (normal equation):

N N
1 1
∑ yi,0 −2 ∆ûi = N ∑ yi,0 −2 (∆yi − α̂AH ∆yi,−1 ) = 0
!
N i =1 i =1

Solving for α̂AH yields the estimator:

i =1 yi,−2 ∆yi
0
0 ∆y
y− 2 ∑N ∑Ni =1 ∑t =3 yi,t −2 ∆yi,t
T
α̂AH = 0 ∆y = =
i =1 yi,−2 ∆yi,−1
∑N i =1 ∑t =3 yi,t −2 ∆yi,t −1
∑N
y− 0 T
2 −1

(which is not surprising, though).

ECONOMETRICS
Consistency
To see that the estimator is consistent substitute the model

i =1 yi,−2 ∆yi i =1 yi,−2 ∆ui

0 0
∑N ∑N
α̂AH = = α+ .
i =1 yi,−2 ∆yi,−1
∑N i =1 yi,−2 ∆yi,−1
∑N
0 0

Under standard conditions, as N → ∞,

N
1 p
N ∑ yi,0 −2 ∆yi,−1 → B ≡ E(yi,0 −2 ∆yi,−1 ) 6= 0.
i =1

By assumption,
N
1 p
N ∑ yi,0 −2 ∆ui → E[yi,0 −2 ∆ui ] = 0
i =1
(a CLT also typically holds). Taken together
p 0
α̂AH → α + = α.
B
ECONOMETRICS
Not all is fine

The instrumental variable (IV) estimation method leads to consistent but not
necessarily efficient estimates of the parameters in the model because
(a) it does not make use of all the available moment conditions and
(b) it does not take into account the differenced structure on the disturbances (∆uit )
which are NOT white noise.

To see (a), note that the stacked moment condition E[yi,0 −2 ∆ui ] = 0 means:

E[yi,1 ∆ui3 ] + . . . + E[yi,T −2 ∆uiT ] = 0

which is much weaker than what we originally assumed:

E[yi,1 ∆ui3 ] = . . . = E[yi,T −2 ∆uiT ] = 0.

ECONOMETRICS
Model

Start again from

yit = αyi,t −1 + vit , vit = ci + uit , t = 2, . . . , T .

Take first differences to eliminate the individual effect:

yit − yi,t −1 = α (yi,t −1 − yi,t −2 ) + (uit − ui,t −1 ) , t = 3, . . . , T

(not much new, so far).

ECONOMETRICS
Intuition
I Consider period t = 3, the first period we observe the relationship:

yi3 − yi2 = α (yi2 − yi1 ) + (ui3 − ui2 ) .

In this case, yi1 is a likely valid instrument: correlated with (yi2 − yi1 ), not
correlated with (ui3 − ui2 ) if uit is white noise.
I Consider period t = 4, the second period we observe the relation:

yi4 − yi3 = α (yi3 − yi2 ) + (ui4 − ui3 ) .

In this case, yi2 and yi1 are valid instruments for (yi3 − yi2 ).
I Consider period t = 5, the third period we observe the relation:

yi5 − yi4 = α (yi4 − yi3 ) + (ui5 − ui4 ) .

In this case, yi1 , yi2 and yi3 are valid instruments for (yi4 − yi3 ).
I And so on until period T , for which the set of valid instruments is yi1 , . . . , yi,T −2 .
ECONOMETRICS
Assumption and moment conditions
Assumption AR.1 (sequential exogeneity):

E(uit |yi,t −1 , . . . , yi1 , ci ) = 0 t = 2, . . . , T .

The assumption implies that

I uit is uncorrelated with yi,t −1 , . . . , yi1 ,
I ui,t −1 is uncorrelated with yi,t −2 , . . . , yi1 , and thus
I ∆uit is uncorrelated with yi,t −2 , . . . , yi1 .

The Arellano-Bond estimator uses all these moment conditions:

E[yi1 ∆uit ] = · · · = E[yi,t −2 ∆uit ] = 0, t = 3, . . . , T .

ECONOMETRICS
The number of moment conditions
To count the moment conditions, write them out:

E[yi1 ∆ui3 ] = 0, 1 condition for t = 3,

E[yi1 ∆ui4 ] = E[yi2 ∆ui4 ] = 0, 2 conditions for t = 4,
.. ..
. .
E[yi1 ∆uiT ] = . . . = E[yi,T −2 ∆uiT ] = 0 T-2 conditions for t = T .

Altogether, the number of moment conditions grows quadratically with T :

L = 1 + 2 + · · · + T − 2 = (T − 2)(T − 1)/2.

For T = 3 we have L = 1, for T = 4 we have L = 3, for T = 10 we have L = 36

moment conditions.

This implies that for T ≥ 4, we have overidentifying moment conditions. This is why
Arellano and Bond use GMM estimation. ECONOMETRICS
Instrument matrix
The instrument matrix for individual i thus is
 
[yi1 ] 0
 [ y , y
i1 i2 ] 
Wi = 
 
.. 
 . 
0 [yi1 , . . . , yi,T −2 ]
and the set of moment conditions is written (differently from PIV/P2SLS) as
E(Wi0 ∆ui ) = 0.

Stacking all observations i = 1, . . . , N, the matrix of instruments is

 
W1
W =  ...  .
 

WN
ECONOMETRICS
GMM estimation
Since for T ≥ 4, the conditions E(Wi0 ∆ui ) = 0 are overidentifying, it is not possible to
find an estimator of α that equates the sample equivalent
N
N −1 ∑ Wi0 ∆ûi
i =1

exactly to zero. We need a symmetric, positive definite L × L weighting matrix Ξ

(ideally the GMM optimal weighting scheme).

An asymptotically optimal weighting matrix satisfies

Ξ = Λ −1 , Λ = E(Wi0 ∆ui ∆ui0 Wi ).

While we can estimate this matrix in a second step based on first-step residuals ∆ûi ,
we need a good first-step choice.

Arellano and Bond suggest to use one that is based on homoskedasticity assumptions. ECONOMETRICS
First step weighting matrix
We make the conditional homoskedasticity assumption

E(∆ui ∆ui0 |Wi ) = E(∆ui ∆ui0 ) = Σ∆u .

By the LIE, it implies that Ξ−1 simplifies to

Ξ1−1 = Λ1 = E(Wi0 ∆ui ∆ui0 Wi ) = E(Wi0 Σ∆u Wi ).

So we need to find the (unconditional) variance matrix Σ∆u of the differenced error
term ∆ui .

Since by assumption uit is white noise, ∆uit is MA(1) with unit coefficient and zero
mean.

In addition, let us assume that second moments are time-invariant (homoskedastic over
time, stationary). ECONOMETRICS
Variance matrix of ∆ui under homoskedasticity ?
Under the white noise and homoskedasticity assumptions, we have:
E[(∆uit )2 ] = E[(uit − ui,t −1 )2 ] = E(uit2 ) + E(ui,t
2 2
−1 ) = 2σu ,

E[∆uit ∆ui,t −1 ] = E[(uit − ui,t −1 )(ui,t −1 − ui,t −2 )] = −E(ui,t

2 2
−1 ) = − σu ,

E[∆uit ∆ui,t −k ] = 0 ∀k ≥ 2.

Hence,
 
2 −1 0 ... 0 0
 −1 2 −1 ... 0 0 
0
 0 −1 2 ... 0 0 
Σ∆u = E ∆ui ∆ui = σu2   = σu2 G

 
.. .. .. .. .. ..
 . . . . . . 
... −1
 
0 0 0 2
0 0 0 ... −1 2

with the (T − 2) × (T − 2) matrix G defined implicitly. ECONOMETRICS

One-step GMM estimator ?

Under the white noise and homoskedasticity assumptions, we thus find

E(Wi0 ∆ui ∆ui0 Wi ) = σu2 E(Wi0 GWi ).

Since σu2 is a scalar that does not change the relative weighting, we may use

Ξ1−1 = Λ1 = E(Wi0 GWi ),

which can be consistently estimated as

N
−1
Ξ̂1 = Λ1 = N −1 ∑ Wi0 GWi = N −1 W0 (IN ⊗ G)W.
i =1

Note that all we need for this estimator is the data W.

ECONOMETRICS
This leads to... ?

The one-step GMM estimator minimizes the objective function

" #0 " # −1 " #
N N N
QN (α) = N −1
∑ Wi0 ∆ui ∑ Wi0 GWi N −1
∑ Wi0 ∆ui
i =1 i =1 i =1
 −1
= ∆u0 W W0 (IN ⊗ G)W W0 ∆u


This yields the Arellano and Bond (1991) one-step estimator

h
−1 0 i −1
0 0
α̂AB,1 = ∆y− 1 W W ( IN ⊗ G ) W W ∆y −1
h i
0 0 −1 0
× ∆y− ∆y


1 W W ( IN ⊗ G ) W W

ECONOMETRICS
Two-step GMM estimator ?
The homoskedasticity assumptions are often deemed too restrictive. Therefore,
Arellano and Bond (1991) suggest to use a nonparametric (robust) estimator of the
optimal weighting matrix in a second step.

Based on first-step residuals this is straightforward:

" # −1
N
Ξ̂2 = Λ2−1 = N −1
∑ Wi0 ∆ûi ∆ûi0 Wi .
i =1

The resulting estimator is the two-step GMM estimator:

 0 0
 −1  0
α̂AB,2 = ∆y− 1 W Ξ̂2 W ∆y−1 ∆y−1 WΞ̂2 W0 ∆y .


A consistent estimator of the asymptotic variance is given by

h√ i   −1
0 0
N (α̂AB,2 − α) = ∆y− 1 W Ξ̂2 W ∆y−1 .
ECONOMETRICS