8 Spatial Econometric Analysis Using GAUSS Kuan-Pin Lin
Download
Report
Transcript 8 Spatial Econometric Analysis Using GAUSS Kuan-Pin Lin
Spatial Econometric Analysis
Using GAUSS
8
Kuan-Pin Lin
Portland State University
Panel Data Analysis
A Review
Model Representation
N-first or T-first representation
Pooled Model
Fixed Effects Model
Random Effects Model
Asymptotic Theory
N→∞, or T→∞
N→∞, T→∞
Panel-Robust Inference
Panel Data Analysis
A Review
The Model
yit xit' it
'
y
x
it
it ui vt eit
it ui vt eit
'
One-Way (Individual) Effects: yit x it ui eit
Unobserved Heterogeneity
Cross Section and Time Series Correlation
Cov(ui , u j ) 0, Cov(eit , e jt ) 0, i j
Cov(eit , ei ) 0, t
Panel Data Analysis
A Review
N-first Representation
T-first Representation
yit xit' β ui eit
yti xti' β ui eti
i 1, 2,..., N ; t 1, 2,..., T
t 1, 2,..., T ; i 1, 2,..., N
y i Xi β ui iT ei
y t X t β u et
y Xβ (I N iT )u e
y Xβ (iT I N )u e
Dummy Variables
Representation
y Xβ Du e
D I N iT or D iT I N
Panel Data Analysis
A Review
Notations
xi' 1 x1,i1
yi1
' x
y
x
1,i 2
y i i 2 , Xi i 2
'
y
xiT x1,iT
iT
xt' 1 x1,t1
yt1
' x
y
x
1,t 2
y t t 2 , Xt t 2
'
ytN
xtN x1,tN
x2,i1
x2,i 2
x2,iT
x2,t1
x2,t 2
x2,tN
xK ,i1
ei1
1
e
xK ,i 2
, ei i 2 , β 2
xK ,iT
e
K
iT
x K ,t 1
et1
u1
e
u
xK ,t 2
, et t 2 , u 2
xK ,tN
e
tN
u N
Pooled (Constant Effects) Model
yit xit' β ui eit (i 1, 2,..., N ; t 1, 2,..., T )
assuming u ui i
yit xit' β u eit
β
w it x 1 , δ
u
yit w it δ eit y Wδ e
'
it
E (e | X) 0, Var (e | X) e2I
Fixed Effects Model
yit xit' β ui eit (i 1, 2,..., N ; t 1, 2,..., T )
ui is fixed, independent of eit, and may be
correlated with xit.
Cov(ui , eit ) 0, Cov(ui , xit ) 0
y i Xi ui iT ei , i 1, 2,..., N
y t Xt u et , t 1, 2,..., T
Fixed Effects Model
Fixed Effects Model
Classical Assumptions
Strict Exogeneity: E (eit | u, X) 0
Homoschedasticity: Var (eit | u, X) e2
No cross section and time series correlation:
Var (e | u, X) e2 I NT
Extensions: Var (e | u, X)
Panel Robust Variance-Covariance Matrix
Random Effects Model
Error Components
yit xit' β it
it ui eit (i 1, 2,..., N ; t 1, 2,..., T )
ui is random, independent of eit and xit.
Cov(ui , eit ) 0, Cov(ui , xit ) 0, Cov(eit , xit ) 0
Define the error components as it = ui + eit
y i Xi (ui iT ei ), i 1, 2,..., N
y t Xt (u et ), t 1, 2,..., T
Random Effects Model
Random Effects Model
Classical Assumptions
Strict Exogeneity
E (eit | X) 0, E (ui | X) 0 E ( it | X) 0
X includes a constant term, otherwise E(ui|X)=u.
Homoschedasticity
Var (eit | X) e2 , Var (ui | X) u2 , Cov(ui , eit ) 0
Var ( it | X) 2 e2 u2
Constant Auto-covariance (within panels)
Var (εi | X) e2 IT u2iT iT'
Random Effects Model
Random Effects Model
Classical Assumptions (Continued)
Cross Section Independence
Var (εi | X) e2IT u2iT iT'
Var (ε | X) Ω I N
Extensions:
Panel Robust Variance-Covariance Matrix
Fixed Effects Model
Estimation
Within Model Representation
yit xit' β ui eit
yi xi' β ui ei
yit yi (xit' xi' )β (eit ei )
yit xit' β eit
y i X i β ei
or
Qy i QXi β Qei
1 '
where Q IT iT iT , (QiT 0, Q ' Q Q)
T
Fixed Effects Model
Estimation
Model Assumptions
E (eit | xit ) 0
Var (eit | xit ) (1 1/ T ) e2
Cov(eit , eis | xit , xis ) ( 1/ T ) e2 0, t s
1 '
Var (ei | Xi ) Q (IT iT iT )
T
Var (e | X) Ω I N
2
e
2
e
Fixed Effects Model
Estimation: OLS
Within Estimator: OLS
y i Xi β ei
y Xβ e
1
ˆβ ( X' X) 1 X' y N X' X
OLS
i 1 i i
ˆ (βˆ ) ( X' X) 1 X'ΩX( X' X) 1
Var
'
X
i1 i y i
N
OLS
1
ˆ i 1 X Xi i 1 X QXi i 1 X Xi
2
e
N
'
i
1
N
ˆ i 1 X Xi
ˆ e2 eˆ ' eˆ / ( NT N K ), eˆ y Xβˆ
2
e
N
'
i
'
i
N
'
i
1
Fixed Effects Model
Estimation: ML
Normality Assumption
yit xit' β ui eit
y i Xi β ui iTi ei
(t 1, 2,..., T )
(i 1, 2,..., N )
ei ~ normal iid (0, e2 IT )
y i Xi β ei with y i Qy i , Xi QXi , ei Qei ,
1 '
iT iT
T
ei ~ normal (0, ), where e2QQ ' e2Q
Q IT
Fixed Effects Model
Estimation: ML
Log-Likelihood Function
T
1
1 ' 1
lli (β, | y i , Xi ) ln 2 ln ei ei
2
2
2
T
T
1
1 ' 1
2
ln 2 ln( e ) ln Q 2 ei Q ei
2
2
2
2 e
2
e
Since Q is singular and |Q|=0, we maximize
T
T
1 '
2
lli (β, | y i , Xi ) ln 2 ln( e ) 2 ei ei
2
2
2 e
2
e
Fixed Effects Model
Estimation: ML
ML Estimator
N
2
ˆ
(β, e ) ML arg max i 1 lli (β, e2 | y i , Xi )
ˆ ' eˆ
e
i 1 i i
N
1 2 ˆ
ˆ
1 ˆ e , ei y i Xi βˆ
NT
T
eˆ ' eˆ
T 2
2
ˆ e
ˆ e
N (T 1)
T 1
2
e
Fixed Effects Model
Hypothesis Testing
Pool or Not Pool
F-Test based on dummy
variable model: constant or
zero coefficients for D w.r.t
F(N-1,NT-N-K)
F-test based on fixed
effects (unrestricted) model
vs. pooled (restricted)
model
yit xit' β ui eit
vs. (ui u, i )
yit xit' β u eit
( RSS R RSSUR ) / N 1
F
~ F ( N 1, NT N K )
RSSUR / ( NT N K )
RSSUR eˆ 'FE eˆ FE , RSS R eˆ 'POeˆ PO
Fixed Effects Model
Hypothesis Testing
Based on estimated residuals of the fixed
'ˆ
e
y
x
effects model: i
i
i β, i 1,..., N
Heteroscedasticity
Breusch and Pagan (1980)
Autocorrelation: AR(1)
Breusch and Godfrey (1981)
2
NT e ' e 1
2
LM
~
(1)
T 1 e 'e
2
Random Effects Model
Estimation: GLS
The Model
y i Xi β ε i , ε i ui iT ei
E (ε i | Xi ) 0
Var (ε i | Xi ) e2IT u2iT iT'
2
2
T
2
e
u
e Q
IT Q
2
e
1 '
1 '
where Q IT iT iT , IT Q iT iT
T
T
Random Effects Model
Estimation: GLS
GLS
βˆ GLS ( XΩ X) XΩ y i 1 Xi Xi
1
1
1
N
1
and
1
2
1
N
1
e2
1
I Q
Q
2
2 T
e
e T u
1
X
i 1 i y i
N
1
Var (βˆ GLS ) ( XΩ X) i 1 Xi Xi
2
1
1
1
'
u
where 2 IT 2
i i 2
2 T T
e
e T u
e
1
1
e2
I Q
Q 2
2 T
e T u
Random Effects Model
Estimation: GLS
Feasible GLS
Based on estimated residuals of fixed effects
model ˆ 2 eˆ ' eˆ / N (T 1)
e
1 T ˆ
ˆ
ˆ
ˆ
ˆ T ˆ ˆ T e ' e / N , where ei t 1 eit
T
ˆ 1X)1 XΩ
ˆ 1y
βˆ GLS ( XΩ
ˆ 1X) 1
Var (βˆ ) ( XΩ
2
1
2
u
2
e
GLS
1
2
1
1
ˆ
where 2 Q 2 IT Q , ˆ1 ˆ e2 Tˆ u2
ˆ1
ˆ e
Random Effects Model
Estimation: ML
Log-Likelihood Function
yit xit' β (ui eit ) xit' β it
y i Xi β ε i
(t 1, 2,..., T )
(i 1, 2,..., N )
εi ~ normal iid (0, )
lli (β, e2 , u2 | y i , Xi )
T
1
1
ln 2 ln εi 1εi
2
2
2
Random Effects Model
Estimation: ML
where
e2 T u2
I i i Q
(I T Q )
2
e
u2
e2
1
1
1
'
2 IT
i i 2 Q
(I T Q )
2
2 T T
2
2
e
T u e
T u e
e
2
e T
2
'
u T T
2
2
T
2 T
'
2 T
u
u
| | ( e ) IT 2 iT iT ( e ) 1 2
e
e
Random Effects Model
Estimation: ML
ML Estimator
(βˆ , ˆ e2 , ˆ u2 ) ML arg max i 1 lli (β, e2 , u2 | y i , Xi )
N
where
T
1
1
lli (β, , | y i , Xi ) ln 2 ln εi 1εi
2
2
2
2
2
T
T
1
2
e
u
ln 2 e ln
2
2 e2
2
2
1 T
T
'
2
'
u
( yit xit β)
2 t 1 ( yit xit β) 2
2
e T u t 1
2 e
2
e
2
u
Random Effects Model
Hypothesis Testing
Pool or Not Pool
Test for Var(ui) = 0, that is
Cov( it , is ) Cov(ui eit ,ui eis ) Cov(eit ,eis )
For balanced panel data, the Lagrange-multiplier
test statistic (Breusch-Pagan, 1980) is:
Random Effects Model
Hypothesis Testing
Pool or Not Pool (Cont.)
NT eˆ '(J T I N )eˆ
2
LM
1
~
(1)
2 T 1
eˆ ' eˆ
ˆ
e
NT i 1 t 1 it
1
N
T
2
2 T 1 eˆit
i 1
t 1
βˆ
'
where eˆit yit xit 1
uˆ Pooled
N
T
2
2
Random Effects Model
Hypothesis Testing
Fixed Effects vs. Random Effects
H 0 : Cov(ui , xit' ) 0 (random effects)
H1 : Cov(ui , xit' ) 0 ( fixed effects)
Estimator
Random Effects
E(ui|Xi) = 0
Fixed Effects
E(ui|Xi) =/= 0
GLS or RE-OLS Consistent and
(Random Effects) Efficient
Inconsistent
LSDV or FE-OLS Consistent
(Fixed Effects)
Inefficient
Consistent
Possibly Efficient
Random Effects Model
Hypothesis Testing
Fixed effects estimator is consistent under H0
and H1; Random effects estimator is efficient
under H0, but it is inconsistent under H1.
Hausman Test Statistic
'
1
H βˆ RE βˆ FE Var (βˆ RE ) Var (βˆ FE ) βˆ RE βˆ FE
~ 2 (# βˆ FE ), provided # βˆ FE # βˆ RE (no intercept )
Random Effects Model
Hypothesis Testing
Alternative Hausman Test
Estimate the random effects model
( yit yi ) (xit' xi' )β (xit' xi' ) γ eit
F Test that g = 0
H 0 : γ 0 H 0 : Cov(ui , xit ) 0
Random Effects Model
Hypothesis Testing
Heteroscedasticity
yit xit' β it
it ui eit
H0: θ2=0 | θ1=0
H0: θ1=0 | θ2=0
H0: θ2=0, θ1=0
eit ~ (0, e2it )
e2 e2 h(z it' 1 ), i 1,..., N , t 1,..., T
it
or e2it e2 h(hi'1 ), i 1,..., N , t
ui ~ (0, u2i )
u2 u2 h(fi' 2 ), i 1,..., N
i
Random Effects Model
Hypothesis Testing
Heteroscedasticity (Cont.)
Based on random effects model with
homoscedasticity:
eˆ y X βˆ , i 1,..., N ; ˆ 2 , ˆ 2 , ˆ 2 Tˆ 2 ˆ 2
i
i
LM 2 0|1
i
u
e
1
u
1
1
2
S
'
F
(
F
'
F
)
F
'
S
~
(# F )
4
2ˆ1
F [fi' , i 1,...N ], F (I N i N i 'N / N ) F
S [ Si , i 1,..., N ], Si eˆ i' (iT iT' / T )eˆ i
e
Random Effects Model
Hypothesis Testing
Heteroscedasticity (Cont.)
1
LM 1 0|2
S ' H ( H ' H ) 1 H ' S ~ 2 (# H )
2a
H [hi' , i 1,...N ], H (I N i N i 'N / N ) H
1
ˆ e4 ˆ14 (T 1)
1
a
,S 4 S 4
4 4
ˆ1 ˆ e
ˆ e
ˆ1
S
S [ Si , i 1,..., N ], Si eˆ i' (iT iT' / T )eˆ i
S [ Si , i 1,..., N ], Si eˆ i' (IT iT iT' / T )eˆ i
Random Effects Model
Hypothesis Testing
Heteroscedasticity (Cont.)
LM1 0,2 0 LM2 0 LM1 0 ~ 2 (# F # H )
Baltagi, B., Bresson, G., Pirotte, A. (2006) Joint
LM test for homoscedasticity in a one-way error
component model. Journal of Econometrics, 134,
401-417.
Random Effects Model
Hypothesis Testing
Autocorrelation: AR(1)
Based on random effects model with no
autocorrelation: eˆ i y i Xi βˆ , i 1,..., N
ˆ u2 , ˆ e2 , ˆ12 T ˆ u2 ˆ e2
LM test statistic is tedious, see
Baltagi, B., Li, Q. (1995) Testing AR(1) against
MA(1) disturbances in an error component model.
Journal of Econometrics, 68, 133-151.
Random Effects Model
Hypothesis Testing
Joint Test for AR(1) and Random Effects
Based on OLS residuals: εˆ y - Xβˆ
LM 0, 2 0
u
NT 2
A2 4 AB 2TB 2 ~ 2 (2)
2(T 1)(T 2)
εˆ '(I N iT iT' )εˆ
εˆ ' εˆ 1
A
1, B
εˆ ' εˆ
εˆ ' εˆ
Marginal Test for AR(1) & Random Effects
LM 2 0
u
2
NT
NT
A2 ~ 2 (1); LM 0
B 2 ~ 2 (1)
2(T 1)
T 1
Random Effects Model
Hypothesis Testing
Robust LM Tests for AR(1) and Random
Effects
Because
LM 0, 2 0 LM * 0 LM 2 0 LM * 2 0 LM 0
u
u
u
LM 2 0
NT
(2 B A) 2 ~ 2 (1)
2(T 1)(1 2 / T )
LM * 0
NT 2
( B A / T ) 2 ~ 2 (1)
(T 1)(1 2 / T )
*
u
Panel Data Analysis
An Example: U. S. Productivity
The Model (Munnell [1988]):
ln( gspit ) 0 1 ln( publicit ) 2 ln( privateit )
3 ln(empit ) 4 (unempit ) ui eit
publicit hwyit waterit utilit
Panel Data Analysis
An Example: U. S. Productivity
Productivity Data
48 Continental U.S. States, 17 Years:1970-1986
STATE = State name,
ST_ABB = State abbreviation,
YR = Year, 1970, . . . ,1986,
PCAP = Public capital,
HWY = Highway capital,
WATER = Water utility capital,
UTIL = Utility capital,
PC = Private capital,
GSP = Gross state product,
EMP = Employment,
UNEMP = Unemployment rate
U. S. Productivity
Baltagi (2008) [munnell.1, munnell.2]
Panel Data Model
ln(GSP) = 0 + 1 ln(Public) + 2ln(Private) + 3ln(Labor)
+ 4(Unemp) +
Fixed
Effects
s.e
Random
Effects
s.e
1
-0.026
0.029
0.003
0.024
2
0.292
0.025
0.310
0.020
3
0.768
0.030
0.731
0.026
4
-0.005
0.001
-0.006
0.001
2.144
0.137
0
F(47,764) =75.82
Hausman LM(4) = 905.1
LM(1) = 4135
Panel Data Analysis
Another Example: China Provincial Productivity
Cobb-Douglass Production Function
ln(GDP) = a + ln(L) + g ln(K) +
Fixed
Effects
s.e.
Random
Effects
s.e
0.30204
0.078
0.4925
0.078
g
0.04236
0.0178
0.0121
0.0176
2.6714
0.6254
a
F(29,298) = 158.81
Hausman LM(2) = 48.4
LM(1) = 771.45
References
B. H. Baltagi, Econometric Analysis of Panel Data, 4th
ed., John Wiley, New York, 2008.
W. H. Greene, Econometric Analysis, 6th ed., Chapter 9:
Models for Panel Data, Prentice Hall, 2008.
C. Hsiao, Analysis of Panel Data, 2nd ed., Cambridge
University Press, 2003.
J. M. Wooldridge, Econometric Analysis of Cross Section
and Panel Data, The MIT Press, 2002.