Chapter 1.

Panel Data

Joan Llull

Quantitative Statistical Methods II  Part II

Barcelona School of Economics
 Introduction

Chapter 1. Panel Data 2

 Panel data
The term panel data refers to data sets with repeated observations over time
(or other dimensions) for a given cross-section of units.

Units can be persons, households, rms, countries,...

It is dierent from repeated cross-sections.

Main advantages of panel data:

Permanent unobserved heterogeneity.

Dynamic responses and error components.

Chapter 1. Panel Data 3

 Cigarette Consumption
We will use the same example throughout the chapter.

Consider the estimation of the demand for cigarettes:

ln Cit = β0 + β1 ln Pit + β2 ln Yit + ηi + vit ,

where:

ln Cit is log consumption of cigarettes,

ln Pit is log price of the cigarettes,

ln Yit is log income of the individual,

ηi + vit is unobserved.

Chapter 1. Panel Data 4

 Static Models

Chapter 1. Panel Data 5

 General notation
We consider the following model:
yit = x0it β + (ηi + vit ).
where yit and xit are observed, and ηi + vit is unobserved.
Let {yit , xit }t=1,...,T
i=1,...,N be our sample. We dene:

x0i1
     
yi1 vi1
y i ≡  ...  , Xi =  ...  , η i = ηi ιT , and v i =  ...  ,
     

yiT x0iT viT

       
y1 X1 η1 v1
y ≡  ...  , X =  ...  , η =  ...  , and v =  ...  ,
       

yN XN ηN vN
where ιT is a size T vector of ones.
Hence, we can use compact notation: y i = Xi β+(η i +v i ), and y = Xβ+(η+v).
Chapter 1. Panel Data 6
 General assumptions for static models
For static models, we assume:

Fixed eects: E[xit ηi ] 6= 0 or random eects: E[xit ηi ] = 0.

Strict exogeneity: E[xit vis ] = 0 ∀s, t. This assumption rules out eects of past vis
on current xit (e.g. xit cannot include lagged dependent variables).

Error components: E[ηi ] = E[vit ] = E[ηi vit ] = 0.

Serially uncorrelated shocks: E[vit vis ] = 0 ∀s 6= t.

Homoskedasticity and i.i.d. errors: ηi ∼ iid(0, ση2 ) and vit ∼ iid(0, σv2 ), which
does not aect any crucial result, but simplies some derivations.

Chapter 1. Panel Data 7

 Pooled OLS
A simple approach: dene: u≡η+v and estimate β by OLS:

β̂ OLS = (X 0 X)−1 X 0 y.

The properties of β̂OLS depend on E[xit ηi ], as E[xit vit ] = 0:

If E[xit ηi ] = 0 ⇒ E[xj uj ] = 0 (random eects):

E[xit vit ]=0

 β̂ OLS is consistent as N → ∞, or T → ∞, or both.

 it is ecient only if ση2 = 0.

If E[xit ηi ] 6= 0 ⇒ E[xj uj ] 6= 0 (xed eects):

 β̂ OLS is inconsistent as N → ∞, or T → ∞, or both.

 cross-section results are also inconsistent (but panel helps in constructing a consistent
alternative).
Chapter 1. Panel Data 8
 Pooled OLS in Cigarette Demand
In our previous example, we would redene our regression as:
ln Cj = β0 + β1 ln Pj + β2 ln Yj + uj ,
where I used subindex j to emphasize that each observation it is considered as one inde-
pendent observation.

OLS
ln Prices (β1 ) -0.083 (0.015)
ln Income (β2 ) -0.032 (0.006)
Potential problems:
preferences,
education,
parental smoking behavior,...
Chapter 1. Panel Data 9
 The xed eects model.
Within groups estimation.

Chapter 1. Panel Data 10

 Within groups estimator
WritePthe model in deviations from individual means, ỹit ≡ yit − ȳi , where ȳi ≡
T −1 t=1 yit :
T

ỹit = (xit − x̄i )0 β + (ηi − η̄i ) + (vit − v̄i ) = x̃0it β + ṽit .

Given the previous assumptions:

E[x̃it ṽit ] = 0.

Therefore, OLS on the transformed model:

 −1
β̂ W G = X̃ 0 X̃ X̃ 0 ỹ,

is a consistent estimator either if E[xit ηi ] 6= 0 or E[xit ηi ] = 0.

Strict exogeneity is a crucial assumption (see next slide).

Chapter 1. Panel Data 11

 The role of strict exogeneity
In the case where N → ∞ and T is xed, consistency depends on strict exogeneity.

To see it, recall that:

1 1
x̃it = xit − (xi1 + ... + xiT ) and ṽit = vit − (vi1 + ... + viT ).
T T

Therefore E[x̃it ṽit ] = 0 requires E[xit vis ] = 0 ∀ s, t unless T → ∞.

This has motivated the development of dynamic panel data models, to relax this
assumption.

Chapter 1. Panel Data 12

 Pros and cons of within groups estimator
Advantage: consistent either if E[xit ηi ] 6= 0 or E[xit ηi ] = 0.

Limitations:

Not ecient:
 When N → ∞ but T is xed, less ecient that e.g. β̂GLS if E[xit ηi ] = 0.

 Even if E[xit ηi ] 6= 0, dierencing introduces correlation in the errors.

It does not allow to identify coecients for time-invariant regressors, and

identication is through switchers.

Chapter 1. Panel Data 13

 Within Groups in Cigarette Demand

In our example:

OLS WG

ln Prices (β1 ) -0.083 (0.015) -0.292 (0.023)

ln Income (β2 ) -0.032 (0.006) 0.107 (0.019)

Chapter 1. Panel Data 14

 Least Squares Dummy Variables
The Within Groups estimator can also be obtained by including a set of N individual
dummy variables:
yit = x0it β + η1 D1i + ... + ηN DN i + vit ,
where Dhi = 1{h = i} (e.g. D1i takes the value of 1 for the observations on individual
1 and 0 for all other observations).

OLS estimation of this model gives numerically equivalent estimates to WG

(that's why β̂ W G is a.k.a. β̂ LSDV ).

Chapter 1. Panel Data 15

 LSDV in Cigarette Demand
We can generate individual (rm) dummies and estimate by OLS, to check that it
delivers the same results:

OLS WG LSDV

ln Prices (β1 ) -0.083 -0.292 -0.292

(0.015) (0.023) (0.023)

ln Income (β2 ) -0.032 0.107 0.107

(0.006) (0.019) (0.019)

Individual 1 (β0 + η1 ) 2.804

(0.288)

Individual 2 (β0 + η2 ) 3.455

(0.398)

Individual 3 (β0 + η3 ) 2.891

(0.416)

Individual 4 (β0 + η4 ) 2.908

(0.384)

Individual 5 (β0 + η5 ) 3.490

(0.433)

Individual 6 (β0 + η6 ) 2.092

Chapter 1. Panel Data 16
(0.325)
 First-Dierenced Least Squares
Another transformation we can consider is rst dierences:
∆yit = ∆x0it β + ∆vit , for i = 1, ..., N ; t = 2, ..., T

where ∆yit = yit − yit−1 .

Takes out time-invariant individual eects (∆ηi = ηi − ηi = 0), so OLS on the

consistent.
dierenced model is

Consistency requires E[∆xit ∆vit ] = 0 which is implied by but weaker than strict
exogeneity.

WG more ecient than FDLS under classical assumptions.

FDLS more ecient if vit random walk (∆vit = εit ∼ iid(0, σε2 )).
Chapter 1. Panel Data 17
 FDLS in Cigarette Demand
To get FDLS we generate st dierences and estimate by OLS:

OLS WG LSDV FDLS

ln Prices (β1 ) -0.083 -0.292 -0.292 -0.413
(0.015) (0.023) (0.023) (0.035)
ln Income (β2 ) -0.032 0.107 0.107 0.178
(0.006) (0.019) (0.019) (0.055)
Individual 1 (β0 + η1 ) 2.804
(0.288)
Individual 2 (β0 + η2 ) 3.455
(0.398)
Individual 3 (β0 + η3 ) 2.891
(0.416)
Individual 4 (β0 + η4 ) 2.908
(0.384)
... ...

Chapter 1. Panel Data 18

 The random eects model.
Error components.

Chapter 1. Panel Data 19

 Uncorrelated eects
Now we assume uncorrelated or random eects: E[xit ηi ] = 0.

In this case, OLS is consistent, but not ecient.

The ineciency is provided by the serial correlation induced by ηi :

E[uit uis ] = E[(ηi + vit )(ηi + vis )] = E[ηi2 ] = ση2 .

The variance of the unobservables (under classical assumptions) is:

E[u2it ] = E[ηi2 ] + E[vit
2
] = ση2 + σv2 .

Chapter 1. Panel Data 20

 Error structure
Therefore, the variance-covariance matrix of the unobservables is:

 2
ση + σv2 ση2 ση2

...
 σ2 ση2 + σv2 . . . ση2 
0 η
E[ui ui ] =   = Ωi ,
 
. . .. .
. . . .
 . . . 
ση2 ση2 2
. . . ση + σv2

whose dimensions are T × T, and E[ui u0h ] = 0 ∀ i 6= h, or:

 
Ω1 0 . . . 0
 0 Ω2 . . . 0 
E[uu0 ] =  . .  = Ω,
 
. ..
 .. .
. . . 
.
0 0 ... ΩN

which is block-diagonal with dimension NT × NT.

Chapter 1. Panel Data 21
 Generalized Least Squares
Under the classical assumptions, GLS (Balestra-Nerlove) estimator is consistent
and ecient if E[xit ηi ] = 0:

−1
β̂ GLS = X 0 Ω−1 X X 0 Ω−1 y.

If E[xit ηi ] 6= 0 GLS is inconsistent as N → ∞ and T is xed.

This estimator is unfeasible because we do not know ση2 and σv2 .

Chapter 1. Panel Data 22

 Theta-dierencing
β̂ GLS is equivalent to OLS on the theta-dierenced model:
∗
yit = x∗it 0 β + u∗it ,

where:
∗
yit = yit − (1 − θ)ȳi ,
and:
σv2
θ2 = .
σv2 + T ση2

Consistency relies on E[xit ηi ] = 0 (as ηi not eliminated).

Two special cases:

If ση2 = 0 (i.e. no individual eect), OLS is ecient.

If T → ∞, then θ → 0, and yit∗ → ỹit = yit − ȳi : WG is ecient.

Chapter 1. Panel Data 23
 Feasible GLS
β̂ GLS is unfeasible because we do not know ση2 and σv2 .

A consistent estimator of σv2 is provided by the WG residuals:

ṽˆit ≡ ỹit − x̃0it β̂ W G

ˆ0 ṽ
ṽ ˆ
σ̂v2 =
N (T − 1) − K
Then, a consistent estimator of ση2 is given by the BG residuals:
ȳi = x̄0i β + η̄i + v̄i , i = 1, ..., N ⇒ β̂ BG

ˆi ≡ ȳi − x̄0i β̂ BG
ū

ˆ0 ū
ˆ
 \ 
1 ū 1 2
σ̂ū2 = ση2 + σv2 = ⇒ σ̂η2 = σ̂ū2 − σ̂ .
T N −K T v
Chapter 1. Panel Data 24
 Feasible GLS in Cigarette Demand

In our example, if we now estimate β̂ F GLS , we get:

OLS WG FGLS

ln Prices (β1 ) -0.083 -0.292 -0.122

(0.015) (0.023) (0.014)

ln Income (β2 ) -0.032 0.107 -0.012

(0.006) (0.019) (0.004)

Chapter 1. Panel Data 25

 Testing for correlated individual eects.

Chapter 1. Panel Data 26

 Testing for correlated eects (Hausman test)
β̂ W G is consistent regardless of E[xit ηi ] being equal to zero or not.

β̂ F GLS is consistent only if E[xit ηi ] = 0.

⇒ we can test whether both estimates are similar!

The Hausman test does exactly this comparison:

a
h = q̂ 0 [avar(q̂)]−1 q̂ ∼ χ2 (K),
under the null hypothesis E[xit ηi ] = 0, where:
q̂ = β̂ W G − β̂ F GLS ,
and:    
avar(q̂) = avar β̂ W G − avar β̂ F GLS .

Requires classical assumptions (FGLS to be more ecient than WG).

Chapter 1. Panel Data 27
 Hausman test in Cigarette Demand

Output from software packages often includes the Hausman test. In our example:

Statistic P-value

Hausman test 24.661 0.000

Chapter 1. Panel Data 28

 Dynamic Models

Chapter 1. Panel Data 29

 Autoregressive models with individual eects

Chapter 1. Panel Data 30

 Autorregressive panel data model
We consider the following model:

yit = αyit−1 + ηi + vit |α| < 1.

Other regressors can be included, but main results unaected.

We assume:

Error components: E[ηi ] = E[vit ] = E[ηi vit ] = 0.

Serially uncorrelated shocks: E[vit vis ] = 0 ∀ s 6= t.

Predetermined initial cond.: E[yi0 vit ] = 0 ∀ t = 1, ..., T .

Chapter 1. Panel Data 31

 Properties of pooled OLS and WG estimators
Even assuming E[yit−1 vit ] = 0, still OLS delivers:
plim α̂OLS > α,
N →∞
 
because E[yit−1 ηi ] = ση2 1−αt−1
1−α
+ αt−1 E[yi0 ηi ] > 0.

Doing the within groups transformation we see that:

plim α̂W G < α,
N →∞

(1−α) 1+T 1−αt−1 −αT −1−t +αT 1−αT −1

    !

because E[ỹit−1 ṽit ] = −Aσv2 < 0. A= 2

T (1−α) 2

WG bias vanishes as T → ∞ (bias not small even with T = 15).

Supposedly consistent estimators with α̂ >> αOLS or α̂ << α̂W G should be seen with suspicion.

Chapter 1. Panel Data 32

 Anderson and Hsiao
Consider the model in rst dierences:

∆yit = α∆yit−1 + ∆vit .

OLS in rst dierences is inconsistent: E[∆yit−1 ∆vit ] = −σv2 < 0.

However, yit−2 or ∆yit−2 are valid instruments for ∆yit−1 :

Relevance: E[∆yit−2 ∆yit−1 ] 6= 0, E[yit−2 ∆yit−1 ] 6= 0.

Orthogonality: E[∆yit−2 ∆vit ] = E[yit−2 ∆vit ] = 0.

Anderson and Hsiao (1981) proposed this 2SLS estimators:

 −1
\ 0 \ \ 0
α̂AH = ∆y ∆y
−1 −1∆y ∆y, −1

where:
−1
0
∆y
\ −1 = Z Z Z Z 0 ∆y −1 ,
where Z can be y −2 or ∆y −2 .

Requires min. three periods (T = 2 and yi0 ). Only ecient if T = 2.

Chapter 1. Panel Data 33
 AR(1) Cigarette Demand (No Covariates)
In our example, we redene the model to be an AR(1) process (for now without
regressors):
ln Cit = α ln Cit−1 + ηi + vit .

The Anderson-Hsiao results (together with OLS and WG) are:

OLS WG Anderson-
Hsiao

Lagged cons. (ln Cit−1 ) 0.982 0.884 1.395

(0.003) (0.061) (0.090)

Chapter 1. Panel Data 34

 Dierenced GMM estimation.

Chapter 1. Panel Data 35

 GMM in 3 slides (I): the setup
GMM nds parameter estimates that come as close as possible to satisfy orthogo-
nality conditions (or moment conditions) in the sample.

E.g., consider K regressors xi and L instruments zi:

ui = yi − f (xi , β) z i = g(xi ).

The model species L moment conditions: E[z i ui ] = 0.

Sample analogue:
N
1 X
bN (β) = z i ui (β).
N
i=1

Chapter 1. Panel Data 36

 GMM in 3 slides (II): the estimation
Two possible cases cases:
L = K (just identied): bN (β̂ GM M ) = 0.

L > K (overidentied): β̂ GM M = arg minβ bN (β)0 WN bN (β).

WN is a positive denite weighting matrix.

−1
Optimal GMM (ecient) uses
 PN
1
E[z u u0 z0 ] , the inverse of the variance-
N i=1 i i i i
covariance matrix, as weighting matrix.

Chapter 1. Panel Data 37

 GMM in 3 slides (III): asymptotic properties
β̂ GM M is a consistent estimator of β.
It is asymptotically normal, with the following variance:

avar(β̂ GM M ) = (D0 W D)−1 D0 W S0 W D(D0 W D)−1 ,

where:

∂bN (β)
D ≡ plim ,
N →∞ ∂β 0

W ≡ plim WN ,
N →∞

N
1 X
S0 ≡ E[z i ui u0i z 0i ].
N
i=1

Chapter 1. Panel Data 38

 Moment conditions
Given previous assumptions, several moment conditions:
Equation Instruments Orthogonality cond.

∆yi2 = α∆yi1 + ∆vi2 yi0 E[∆vi2 yi0 ] = 0

  
yi0
∆yi3 = α∆yi2 + ∆vi3 yi0 , yi1 E ∆vi3 =0
yi1
  
yi0
∆yi4 = α∆yi3 + ∆vi4 yi0 , yi1 , yi2 E ∆vi4 yi1  = 0
yi2
.. .. ..
. . .
  
yi0



 yi1 

∆yiT = α∆yiT −1 + ∆viT yi0 , yi1 , yi2 , ..., yiT −2 E ∆viT
  yi2  = 0

..

.
  
  
yiT −2

We end up with (T − 1)T /2 moment conditions.

Chapter 1. Panel Data 39
 Moment conditions in matrix form
We can write these moment conditions as E[Zi0 ∆v i ] = 0, where:
   
yi0 0 0 0 ... 0 0 ... 0 ∆vi2
0 yi0 yi1 0 ... 0 0 ... 0   ∆vi3 
Zi =  .

.. .. .. .. .. .. .. 

and ∆v i = 
 . ,

 .. . . . . . . ... .   .
. 
0 0 0 0 ... yi0 yi1 ... yiT −2 ∆viT

and the sample analogue is:

N
1 X 0
bN (α) = Zi ∆v i (α).
N i=1
Hence, the GMM estimator (proposed by Arellano and Bond, 1991) is:
N
! N
!
1 X 1 X 0
α̂GM M = arg min ∆v 0i (α)Zi WN Zi ∆v i (α) =
α N i=1 N i=1

= (∆y 0−1 ZWN Z 0 ∆y −1 )−1 ∆y 0−1 ZWN Z 0 ∆y.

Chapter 1. Panel Data 40
 Optimal weighting matrix
The optimal weighting matrix (ecient GMM) is:
N
!−1
1 X
WN = E[Zi0 ∆v i ∆v 0i Zi ] .
N i=1
The sample analogue is obtained in a two-step procedure:
N
!−1
1 X 0 \ \
WN = [Zi ∆v i (α̂)∆v 0i (α̂)Zi ] .
N i=1
Windeijer (2005) proposes a nite sample correction of the variance that accounts for α being
estimated.

The most common one-step (and rst-step) matrix uses the structure of E[∆v i ∆v 0i ]:
−1
 
2 0 0 ... 0
−1 2 −1 0 ... 0
E[∆v i ∆v 0i ] = σv2  . .. .. .. .. .
 
 .. . . . . 
0 0 0 0 ... 2
Chapter 1. Panel Data 41
 GMM in Cigarette Demand
GMM results in our example are:

Coecient Standard Error

Least Squares (OLS) 0.982 (0.003)
Within Groups (WG) 0.884 (0.061)
Anderson-Hsiao 1.395 (0.090)
One-step GMM 1.023 (0.104)
Two-step GMM 0.994 (0.040)
Two-step GMM small sample 0.994 (0.121)

Chapter 1. Panel Data 42

 Potential limitations of Arellano-Bond
Weak instruments:

When α → 1, relevance of the instrument decreases.

Instruments are still valid, but have poor small sample properties.

Monte Carlo evidence shows that with α > 0.8, estimator behaves poorly unless huge
samples available.

There are alternatives in the literature.

Overtting:

Too many instruments if T relative to N is relatively large.

We might want to restrict the number of instruments used.

It is always good practice to check robustness to dierent combinations of instruments.

Chapter 1. Panel Data 43
 Dynamic linear model
Once we include regressors, the model is:
yit = αyit−1 + x0it β + ηi + vit |α| < 1.

We maintain the previous assumptions: error components, serially uncorrelated

shocks, and predetermined initial conditions.

Therefore, moment conditions of the kind:

E[yit−s ∆vit ] = 0 t = 2, ..., T, s ≥ 2,

are still valid.
Chapter 1. Panel Data 44
 Assumptions on regressors
Dierent assumptions regarding xit will enable dierent additional orthogonality
conditions:
xit can be correlated or uncorrelated with ηi .

xit can endogenous, predetermined, or strictly exogenous with respect to vit .

For instance, if assumptions are analogous to those for yit−1 , we may use xit−1 (and
longer lags) as instruments:

x0i0 x0i1 00 00
 
yi0 ... 0 ... 0 ...
 . .. .. .. .. .. .. .. .. 
Zi =  .. . . . . ... . . . . .
0 00 00 ... yi0 ... yiT −2 x0i0 ... x0iT −1

Chapter 1. Panel Data 45

 Dynamic Cigarette Demand with Regressors
In our example, we rewrite the employment equations as follows:

ln Cit = α ln Cit−1 + β1 ln Pit + β2 ln Yit + ηi + vit .

Results are:

OLS WG GMM

Lagged dep (α) 0.947 0.528 0.495

(0.011) (0.064) (0.127)

ln Prices (β1 ) 0.010 -0.501 -0.607

(0.006) (0.098) (0.143)

ln Income (β2 ) 0.049 0.369 0.338

(0.011) (0.044) (0.051)

Chapter 1. Panel Data 46

 System GMM estimation

Chapter 1. Panel Data 47

 Additional orthogonality conditions

Recall our (T − 1)T /2 moment conditions:

E[yit−s ∆vit ] = 0 t = 2, ..., T ; s ≥ 2.

System GMM (Arellano and Bover, 1995) uses the assumption E[yi0 |ηi ] = 1−α
ηi
:

E[∆yit ηi ] = 0, ∀t

or, alternatively:

E[∆yiT −s uiT ] = 0, uiT = ηi + viT , s = 1, ..., T − 1.

Chapter 1. Panel Data 48

 The System GMM estimator

Analogously to the rst-dierenced GMM, the estimator is given by E[(Z ∗ )0 u∗i ] = 0:

−1
α̂Sys−GM M = (X ∗ )0 Z ∗ WN (Z ∗ )0 X ∗ X ∗ Z ∗ WN (Z ∗ )0 y ∗ ,

where:
       
Zi 0 ... 0 ∆v i ∆y −1i ∆y i
Zi∗ = , u∗i = , Xi∗ = and y ∗i = ,
00 ∆yi1 ... ∆yiT −1 ηi + viT yiT −1 yiT

This estimator is more ecient, as it uses additional moment conditions.

It reduces small sample bias, especially when α → 1.

Chapter 1. Panel Data 49

 System-GMM in Cigarette Demand
GMM results in our example are:

Coecient Standard Error

Least Squares (OLS) 0.982 (0.003)
Within Groups (WG) 0.884 (0.061)
Anderson-Hsiao 1.395 (0.090)
One-step GMM 1.023 (0.104)
Two-step GMM 0.994 (0.040)
Two-step GMM small sample 0.994 (0.121)
One-step System-GMM 0.926 (0.023)
Two-step System-GMM small 0.911 (0.032)

Chapter 1. Panel Data 50

 Specication tests

Chapter 1. Panel Data 51

 Specication tests
There are several relevant aspects for the validity of the estimation that can be
tested formally.

Orthogonality conditions: are they small enough to not reject that they are
zero (overidentifying restrictions).

Validity of a subset of more restrictive assumptions (dierence Sargan test,

Hausman test).

Serial correlation in the data: vital for the validity of the instruments
(Arellano-Bond test).

Chapter 1. Panel Data 52

 Sargan/Hansen overidentifying restrictions test
The null hypothesis is whether the orthogonality conditions are satised (i.e. moments
are equal to zero).

The test can only be implemented if the problem is overidentied (otherwise the sample
moments are exactly zero by construction).

The test is:


N N
!−1 N

1 X ˆ0 1 X 0 1 X 0ˆ 
S = N JN (β) = N  ûi Zi Zi ûi û0i Zi Zi ûi ,
N N N
i=1 i=1 i=1

where û are those predicted residuals from the rst step and û
ˆ are predicted from the second
stage, and:
a
S ∼ χ2 (L − K).

Chapter 1. Panel Data 53

 Testing stronger assumptions
Some of the assumptions that we make are stronger than others.
If the problem is overidentied, we can test whether results change if we include
or exclude the orthogonality conditions generated by them.
If they are true, increase eciency, but if not, inconsistent!
Two ways of testing it:
Overidentifying restrictions (dierence in Sargan should be close to zero:
a
∆S = S − S A ∼ χ2 (L − LA )).

Dierences in coecients (Hausman test: A

q̂ = δ̂ GM M − δ̂ GM M ).

Chapter 1. Panel Data 54

 Direct test for serial correlation
The test was proposed by Arellano-Bond (1991).

Tests for the presence of second order autocorrelation in the rst-dierenced

residuals.

If dierences in residuals are second-order correlated, some instruments would

not be valid!
The test is:
0
∆v
\ −2 ∆v ∗ a
d
m2 = ∼ N (0, 1),
se
where ∆v −2 is the second lagged residual in dierences, and ∆v ∗ is the part of
the vector of contemporaneous rst dierences for the periods that overlap with the
second lagged vector.

Values close to zero do not reject the hypothesis of no serial correlation.

Chapter 1. Panel Data 55