Econometric Analysis of Panel Data

• Panel Data Analysis

– Linear Model
• One-Way Effects
• Two-Way Effects
– Pooled Regression
• Classical Model
• Extensions
 Panel Data Analysis
• Linear Model Representation
yit  xit' β   it
 it  ui  vt  eit
t  1, 2,..., T (Ti ); i  1, 2,..., N ( N t )
 v1   u1 
 v  u 
(1) y i  Xi β  ui iTi  v i  ei vi   2
, ut   2 

   
or
   
(2) y t  Xt β  ut  vt i Nt  et vTi  u Nt 
  yi1   xi' 1   x1,i1 x2,i1  xK ,i1   1   ei1 
y   '  
 i2   xi 2   x1,i 2 x2,i 2  xK ,i 2    e 
 i2 
yi  ,X   ,β  2 
, ei 
  i            
   '       
 yiTi   xiTi   x1,iTi x2,iTi  xK ,iTi  K  eiTi 

 y1t   x1' t   x1,1t x2,1t  xK ,1t   e1t 

y   '  
 2t   x 2t   x1,2t x2,2t  xK ,2t  e 
, et   
2t
yt  , Xt  
            
   '     
 y Nt t   x Nt t   x1, Nt t x2, Nt t  xK , Nt t  eNt t 
 Linear Panel Data Model (1)
• One-Way (Individual) Effects
yit  xit' β  ui  eit  y i  Xi β  ui iTi  ei
(t  1, 2,..., Ti ; i  1, 2,..., N )

y  Xβ  u  e

 y1   X1   1   u1iT1   e1 
y  X      e 
u i
 2 T2 
y   2 , X   2 , β   2 , u   , e   2
         
         
y N  XN  K  u N iTN  e N 
 Linear Panel Data Model (1)
• One-Way (Time) Effects
yit  xit' β  vt  eit  y i  Xi β  v i  ei
(t  1, 2,..., Ti ; i  1, 2,..., N )

y  Xβ  v  e

 y1   X1   1   v1   e1 
y  X    v  e 
y   , X   , β   , v   , e   2 
2 2 2 2

         
         
y
 N X
 N 
 K v
 N e N 
 Linear Panel Data Model (1)
• Two-Way Effects
yit  xit' β  ui  vt  eit  y i  Xi β  ui iTi  v i  ei
(t  1, 2,..., Ti ; i  1, 2,..., N )

y  Xβ  u  v  e
 Linear Panel Data Model (2)
• One-Way (Individual) Effects
yit  xit' β  ui  eit  y t  Xt β  u t  et
(i  1, 2,..., N t ; t  1, 2,..., T )

y  Xβ  u  e

 y1   X1   1   u1   e1 
y  X    u  e 
y   2 , X   2 , β   2 , u   2, e   2
         
         
yT   XT  K  uT   eT 
 Linear Panel Data Model (2)
• One-Way (Time) Effects
yit  xit' β  vt  eit  y t  Xt β  vt i Nt  et
(i  1, 2,..., N t ; t  1, 2,..., T )

y  Xβ  v  e

 y1   X1   1   v1i N1   e1 
y  X      e 
v i
 2 N2 
y   2 , X   2 , β   2 , v   , e   2
         
         
y
 T X
 T 
 K vT i NT  eT 
 Linear Panel Data Model (2)
• Two-Way Effects
yit  xit' β  ui  vt  eit  y t  Xt β  u t  vt i Nt  et
(i  1, 2,..., N t ; t  1, 2,..., T )

y  Xβ  u  v  e
 Panel Data Analysis
• Between Estimator
yit  xit' β  ui  eit  yi  xi' β  ui  ei
1 1 1
 t 1 yit , x  T  t 1 x , ei  T 
Ti Ti Ti
yi  '
i
'
it e
t 1 it
Ti i i

1
 i 1 ui ,
N
• If ui  u 
N
then the pooled or population-averaged
model is more efficient.
 Panel Data Analysis
• Linear Pooled (Constant Effects) Model
yit  xit' β  ui  eit  yit  xit' β  u  eit  yit  w it δ  eit
β 
w it   x '
it 1 , δ   
u 
(t  1, 2,..., Ti ; i  1, 2,..., N ; NT   i 1Ti )
N


y  Wδ  e
 Pooled Regression Model
• Classical Assumptions
– Strict Exogeneity
E (eit | W )  0; Cov( w it , eit )  0
– Homoschedasticity
Var (eit | W)   e2
– No cross section and time series correlation
Var (e | W )   e2 I NT
 Pooled Regression Model
• Extensions
– Weak Exogeneity
E (eit | w i1 , w i 2 ,..., w iTi )  E (eit | Wi )  0
E (eit | w i1 , w i 2 ,..., w it )  0
E (eit | w it )  0
– Heteroschedasticity
Var (eit | Wi )   it2
Var (eit | Wi )   t2
Var (eit | Wi )   i2
 Pooled Regression Model
• Extensions
– Time Series Correlation (with cross section
independence for short panels)
Cov(eit , eis | w it , w is )   ts , t  s
Cov(eit , e js | w it , w js )  0, i  j
Var (eit | w it )   tt   t2  Var (ei | Wi )  i  Var (e | W)  Ω

  11  12   1Ti  1 0  0 
  0   0 

 21    2Ti 
Ω 
22 2
i  
       
   
 Ti 1  Ti 2   TiTi   0 0   N
 Pooled Regression Model
• Extensions
– Cross Section Correlation (with time series
independence for long panels)
Cov(eit , e jt | w it , w jt )   ij , i  j
Cov(eit , e js | w it , w js )  0, t  s
Var (eit | w it )   i2
  12 I  12 I   1N I 
 
 I  2
I   I
Var (e | W)  Ω    IT   21 2 2N 
     
 2 
 N 1I  N 2 I   N I 
 Pooled Regression Model
• Extensions
– Cross Section and Time Series Correlation
Var (eit | w it )   i ,tt   i2
 1 12  1T 
Cov(eit , eis | w it , w is )   i ,ts   i ts , t  s  1   2T 
Cov(eit , e jt | w it , w jt )   ij , i  j R   21
    
 
Cov(eit , e js | w it , w js )   ij ts , t  s  T 1 T 2  1 

  12 R  12 R   1N R 
 
 R  2
R   R
Var (e | W)  Ω    R   21 2 2N 
     
 

 N 1 R  N2 R   2
N R 
 Alternative Pool Regression Models
• Between (Group Means) Estimator
yit  xit' β  u  eit  yi  xi' β  u  ei

• First-Difference Estimator
yit  yit 1  (xit'  xit' 1 )β  (eit  eit 1 )  yit  xit' β  eit

• Within (Group Mean Deviations) Estimator

yit  yi  (xit'  xi' )β  (eit  ei )
 Pooled Regression: OLS
• Classical Model Estimation (OLS)
1
δˆ OLS  ( W ' W ) 1 W ' y    i 1 Wi' Wi  
N N '
W yi
  i 1 i

1
ˆ (δˆ )  ˆ 2 ( W ' W) 1  ˆ 2   N W ' W 
Var OLS e e
 i 1 i i 
ˆ e2  eˆ ' eˆ / ( NT  K )
eˆ  y  Wδˆ

• Variance estimator Var

ˆ (δˆ )
is inconsistent OLS

because of heteroscedasticity and

autocorrelation.
 Pooled Regression: OLS
• Panel-Robust Variance-Covariance Matrix
– Adjusting general heteroscedasticity and serial correlation
within panel
Var (δˆ )  E[(δˆ  δ)(δˆ  δ) ']  ( W ' W) 1 W ' E (ee ') W( W ' W) 1
1 1
   i 1 Wi' Wi    i 1 Wi' E (ei ei' ) Wi    i 1 Wi' Wi 
N N N

    
1 1
ˆ (δˆ )    N W ' W    N W 'eˆ eˆ ' W    N W ' W 
Var
 i 1 i i   i 1 i i i i   i 1 i i 
1 1
   i 1  t i 1 w it w it'    i 1  t i 1  si1 w it w is' eˆit eˆis    i 1  t i 1 w it w it' 
N T N T T N T

    
eˆ i  y i  Wi δˆ , eˆit  yit  w it δˆ
 Pooled Regression: GLS
• The Model
  12  12   1N   1 12  1T 
y  Wδ  e   
  2
  1   2T 
E (e | W)  0  21 2 2 N  R   21
          
 2   
Var (e | W)  Ω    R  N 1  N 2   N   T 1 T 2  1 

• Generalized Least Squares (GLS)

1 1
δˆ GLS   W 'Ω
ˆ 1W  W 'Ω

ˆ 1y , Var
ˆ (δˆ )   W 'Ω
GLS 
ˆ 1W 


– If cross sections are independent (short panels)

1 1
δˆ   W  ˆ W   W ˆ (δˆ )   W 
ˆ y , Var ˆ W
N ' 1 N ' 1 N ' 1
GLS
 i 1 i
 i i

i 1 i
i i GLS i 1 i i i

ˆ
– where i is the consistent estimator of i
 Pooled Regression: GLS
• Heteroscedasticity
1 T 2
ˆ   t 1 eˆit
i
2

T
• Cross Section Correlation
1 T
ˆ ij 
T
 eˆ eˆ
t 1 it jt

• Time Series Correlation

ˆts
 Pooled Regression: GLS
• Examples of Time Series Correlation
– Equal-Correlation 1 if t  s
ts  
  if t  s

1 if t  s
– AR(1) ts   |t  s|
 if t  s

1 if t  s

– Stationary(1) ts    if | t  s | 1
 0 otherwise


– Nonstationary(1) 1 if t  s

ts   ts if | t  s | 1, ts   st
0 otherwise

 Model Extensions
• Time-invariant regressors
• Random regressors
• Lagged dependent variables
• Dynamic models
 Example: Investment Demand
• Grunfeld and Griliches [1960]
I it   i   Fit   Cit   it

– i = 10 firms: GM, CH, GE, WE, US, AF, DM, GY, UN,
IBM; t = 20 years: 1935-1954
– Iit = Gross investment
– Fit = Market value
– Cit = Value of the stock of plant and equipment
 Example: Investment Demand
• Pooled Model (Population-Averaged Model)
I it     Fit   Cit   it
• Classical OLS
• Panel-Robust OLS
• Feasible GLS
– Heteroscedastcity
– Autocorrelation