Econometría Avanzada
CAN Estimators
Henceforth, we consider a parameter space Θand a “true” parameter
θ0 2Θ.
The estimator θn of the r 1 parameter vectorθ0 is CAN if,
n1/2(θn θ0) d
!Nr(0,D),
and it is usually said that,
θn
asy Nr(θ0,AsyVar(θn)),
and
AsyVar(θn) =
1
nD
CAN Estimators
Henceforth, we consider a parameter space Θand a “true” parameter
θ0 2Θ.
The estimatorθn of the r 1 parameter vectorθ0 is CAN if,
n1/2(θn θ0)!d Nr(0,D),
and it is usually said that,
θn
asy Nr(θ0,AsyVar(θn)),
and
AsyVar(θn) =
1
nD
CAN Estimators
Henceforth, we consider a parameter space Θand a “true” parameter
θ0 2Θ.
The estimatorθn of the r 1 parameter vectorθ0 is CAN if,
n1/2(θn θ0) d
!Nr(0,D),
and it is usually said that,
θn
asy Nr(θ0,AsyVar(θn)),
and
AsyVar(θn) =
1
nD
CAN Estimators
Henceforth, we consider a parameter space Θand a “true” parameter
θ0 2Θ.
The estimatorθn of the r 1 parameter vectorθ0 is CAN if,
n1/2(θn θ0) d
!Nr(0,D),
and it is usually said that,
θn
asy Nr(θ0,AsyVar(θn)),
and
AsyVar(θn) = 1
nD
CAN Estimators
Henceforth, we consider a parameter space Θand a “true” parameter
θ0 2Θ.
The estimatorθn of the r 1 parameter vectorθ0 is CAN if,
n1/2(θn θ0) d
!Nr(0,D),
and it is usually said that,
θn
asy Nr(θ0,AsyVar(θn)),
and
Example:Given W1, ...,Wn scalar r.v. withE(W1) =µand
V(W1) =σ2,the sample mean µn =En(W)satis…es
n1/2(µn µ)!d N 0,σ2 .
Thus,
AsyVar(µn) = σ 2
Example:Given W1, ...,Wn scalar r.v. withE(W1) =µand
V(W1) =σ2,the sample mean µn =En(W)satis…es
n1/2(µn µ)!d N 0,σ2 .
Thus,
AsyVar(µn) = σ
2
The matrix D can usually be consistently estimated from the sample, byDn say, and the corresponding estimator of the asymptotic
variance of θn is,
\
AsyVar(θn) = 1
nDn.
Therefore,
\
AsyVar(θn) 1/2(θn θ0) d
The matrix D can usually be consistently estimated from the sample,
byDn say, and the corresponding estimator of the asymptotic
variance of θn is,
\
AsyVar(θn) =
1
nDn. Therefore,
\
Notation
We assume that W1, ...,Wn,with Wi = (Yi,Zi0)0 is a random sample
of W= (Y,Z0)0 (i.e. are independent copies of the random vector
W), Y is scalar, andZis a p 1 valued vector. We assume that:
Assumption A1: V(Z)exists and is non singular.
De…ningZ= 1
Z , A1 can alternatively be written as:
Notation
We assume that W1, ...,Wn,with Wi = (Yi,Zi0)0 is a random sample
of W= (Y,Z0)0 (i.e. are independent copies of the random vector
W), Y is scalar, andZis a p 1 valued vector. We assume that:
Assumption A1: V(Z)exists and is non singular.
De…ningZ= 1
Z , A1 can alternatively be written as:
Notation
We assume that W1, ...,Wn,with Wi = (Yi,Zi0)0 is a random sample
of W= (Y,Z0)0 (i.e. are independent copies of the random vector
W), Y is scalar, andZis a p 1 valued vector. We assume that:
Assumption A1: V(Z)exists and is non singular.
De…ningZ= 1
Z , A1 can alternatively be written as:
Notation
We assume that W1, ...,Wn,with Wi = (Yi,Zi0)0 is a random sample
of W= (Y,Z0)0 (i.e. are independent copies of the random vector
W), Y is scalar, andZis a p 1 valued vector. We assume that:
Assumption A1: V(Z)exists and is non singular.
De…ningZ= 1
Z , A1 can alternatively be written as:
The best linear predictor of Y given Zis the linear projector:
L(Yj1,Z) = β0+Z0β
= Z0β,
where,
β = β0
β =
E(Y) E(Z)0β
V(Z) 1C(Z,Y) =E ZZ
0 1
E(ZY).
Therefore,U =Y Z0βsatis…es that,E(U) =E(ZU) =0,or
more compactly E(ZU) =0.
TheOLS estimator of βis its sample analog:
βn =En ZZ0
1
The best linear predictor of Y given Zis the linear projector:
L(Yj1,Z) = β0+Z0β = Z0β,
where,
β = β0
β =
E(Y) E(Z)0β
V(Z) 1C(Z,Y) =E ZZ
0 1
E(ZY).
Therefore,U =Y Z0βsatis…es that,E(U) =E(ZU) =0,or
more compactly E(ZU) =0.
TheOLS estimator of βis its sample analog:
βn =En ZZ0
1
The best linear predictor of Y given Zis the linear projector:
L(Yj1,Z) = β0+Z0β = Z0β,
where,
β = β0
β =
E(Y) E(Z)0β
V(Z) 1C(Z,Y) =E ZZ
0 1
E(ZY).
Therefore,U =Y Z0βsatis…es that,E(U) =E(ZU) =0,or more compactly E(ZU) =0.
TheOLS estimator of βis its sample analog:
βn =En ZZ0
1
The best linear predictor of Y given Zis the linear projector:
L(Yj1,Z) = β0+Z0β = Z0β,
where,
β = β0
β =
E(Y) E(Z)0β
V(Z) 1C(Z,Y) =E ZZ
0 1
E(ZY).
Therefore,U =Y Z0βsatis…es that,E(U) =E(ZU) =0,or
more compactly E(ZU) =0.
TheOLS estimator of β is its sample analog:
Consistency
Theorem (Consistency)Under A1,theOLS estimator of βis consistent.
PROOF:
βn =β+En ZZ0
1
En(ZU)
so
βn = β+ En ZZ0
1
| {z }
p
!E ZZ0 1 LLN & Slutsky
En(ZU)
| {z }
p
!0
LLN
| {z }
Consistency
Theorem (Consistency)Under A1,theOLS estimator of βis
consistent.
PROOF:
βn =β+En ZZ0 1
En(ZU)
so
βn = β+ En ZZ0
1
| {z }
p
!E ZZ0 1 LLN & Slutsky
En(ZU)
| {z }
p
!0
LLN
| {z }
Consistency
Theorem (Consistency)Under A1,theOLS estimator of βis
consistent. PROOF:
βn =β+En ZZ0
1
En(ZU)
so
βn = β+ En ZZ0
1
| {z }
p
!E ZZ0 1 LLN & Slutsky
En(ZU)
| {z }
p
!0
LLN
| {z }
Consistency
Theorem (Consistency)Under A1,theOLS estimator of βis
consistent. PROOF:
βn =β+En ZZ0
1
En(ZU)
so
βn = β+ En ZZ0 1
| {z }
p
!E ZZ0 1 LLN & Slutsky
En(ZU)
| {z }
p
!0
LLN
Asymptotic Normality
Assumption A2 E ZZ0U2 exists.
Theorem (Asymptotic Normality)Under A1 and A2, β
n is CAN
with
AsyVar β
n =
1
nE ZZ
0 1 E ZZ0U2 E ZZ0 1.
proof:
n1/2 βn β =En ZZ0
1
Asymptotic Normality
Assumption A2 E ZZ0U2 exists.
Theorem (Asymptotic Normality)Under A1 and A2, β
n is CAN
with
AsyVar βn =
1
nE ZZ
0 1 E ZZ0U2 E ZZ0 1.
proof:
n1/2 βn β =En ZZ0
1
Asymptotic Normality
Assumption A2 E ZZ0U2 exists.
Theorem (Asymptotic Normality)Under A1 and A2, β
n is CAN
with
AsyVar βn =
1
nE ZZ
0 1 E ZZ0U2 E ZZ0 1.
proof:
n1/2 βn β =En ZZ0
1
n1/2 β
n β = En ZZ
0 1
| {z }
p
!E ZZ0 1
by LLN
n1/2En(ZU)
| {z }
d
!Np+1 0,E ZZ0U2
by the CLT
| {z }
d
!Np+1 0,E ZZ0 1
Assumption A2’(Homoskedasticity)
E ZZ0U2 =E ZZ0 E U2 with E U2 <∞.
Corollary: Under A1 and A2’, β
n is CAN with
AsyVar βn = 1
nE U
2 E ZZ0 1.
We shall call
Assumption A2’(Homoskedasticity)
E ZZ0U2 =E ZZ0 E U2 with E U2 <∞. Corollary: Under A1 and A2’, β
n is CAN with
AsyVar βn = 1
nE U
2 E ZZ0 1
.
We shall call
Assumption A2’(Homoskedasticity)
E ZZ0U2 =E ZZ0 E U2 with E U2 <∞.
Corollary: Under A1 and A2’, β
n is CAN with
AsyVar βn = 1
nE U
2 E ZZ0 1
.
We shall call
AsyVar estimates
Consider OLS residuals Uni =Yi Zi0βn.
IHomoskedasticity: AsyVar βn is estimated by:
^
AsyVar β
n =
1
nEn U 2
n En ZZ0
1
.
IHeteroskedasticity: AsyVar βn is estimated by:
\
AsyVar βn = 1
n En ZZ
0 1
En ZZ0Un2 En ZZ0
1
AsyVar estimates
Consider OLS residuals Uni =Yi Zi0βn.
IHomoskedasticity: AsyVar βn is estimated by:
^
AsyVar β
n =
1
nEn U
2
n En ZZ0 1
.
IHeteroskedasticity: AsyVar βn is estimated by:
\
AsyVar βn = 1
n En ZZ
0 1
En ZZ0Un2 En ZZ0
1
AsyVar estimates
Consider OLS residuals Uni =Yi Zi0βn.
IHomoskedasticity: AsyVar βn is estimated by:
^
AsyVar β
n =
1
nEn U 2
n En ZZ0
1
.
IHeteroskedasticity: AsyVar βn is estimated by:
\
AsyVar β = 1 En ZZ0
1
En ZZ0Un2 En ZZ0
1
Next, we provide a theorem that justi…es the consistency of AsyVar^ β
n
Theorem: Under A1, ifE U2 exists,
PROOF: Since
Un =U Z0(βn β)
En Un2 = En U2 + β β
n
0
En ZZ0 β β
n
+2 En(ZU)0 β β
n
By the LLN and consistency of βn
En Un2 = En U2
| {z } p
!E(U2)
+ β β
n
| {z }
=op(1)
0
En ZZ0
| {z } p
!E(ZZ0)
β β
n
| {z }
=op(1) +2 En(ZU)0
| {z }
p
!00
β β
n
| {z }
=op(1) = E U2
| {z } =σ2
PROOF: Since
Un =U Z0(βn β)
En Un2 = En U2 + β β
n
0
En ZZ0 β β
n
+2 En(ZU)0 β β
n
By the LLN and consistency of βn
En Un2 = En U2
| {z } p
!E(U2)
+ β β
n
| {z }
=op(1)
0
En ZZ0
| {z } p
!E(ZZ0)
β β
n
| {z }
=op(1) +2 En(ZU)0
| {z }
p
!00
β β
n
| {z }
=op(1) = E U2
| {z } =σ2
PROOF: Since
Un =U Z0(βn β)
En Un2 = En U2 + β β
n
0
En ZZ0 β β
n
+2 En(ZU)0 β β
n
By the LLN and consistency of βn
En Un2 = En U2
| {z } p
!E(U2)
+ β β
n
| {z }
=op(1)
0
En ZZ0
| {z } p
!E(ZZ0)
β β
n
| {z }
=op(1) +2 En(ZU)0
| {z }
p
!00
β β
n
| {z }
=op(1) = E U2
| {z } =σ2
PROOF: Since
Un =U Z0(βn β)
En Un2 = En U2 + β β
n
0
En ZZ0 β β
n
+2 En(ZU)0 β β
n
By the LLN and consistency of βn
En Un2 = En U2
| {z } p !E(U2)
+ β β
n
| {z }
=op(1) 0
En ZZ0
| {z } p !E(ZZ0)
β β
n
| {z }
=op(1)
+2 En(ZU)0
| {z }
p !00
β β
n
| {z }
=op(1)
Assumption A3 EkZk4 <∞ andEjUj4 <∞.
Next, we provide a theorem that justi…es the consistency of
\
AsyVar βn
Theorem. UnderA1,A2 and A3,
En ZZ0Un2 =E ZZ0U2 +op(1)
PROOF: Since,
En ZZ0Un2 = En ZZ0U2 +En ZZ0 β βn 0
ZZ0 β βn
+2 En ZZ0 β βn 0
Assumption A3 EkZk4 <∞ andEjUj4 <∞.
Next, we provide a theorem that justi…es the consistency of \
AsyVar βn
Theorem. UnderA1,A2 and A3,
En ZZ0Un2 =E ZZ0U2 +op(1)
PROOF: Since,
En ZZ0Un2 = En ZZ0U2 +En ZZ0 β βn 0
ZZ0 β βn
+2 En ZZ0 β βn 0
Assumption A3 EkZk4 <∞ andEjUj4 <∞.
Next, we provide a theorem that justi…es the consistency of
\
AsyVar βn
Theorem. UnderA1,A2 and A3,
En ZZ0Un2 =E ZZ0U2 +op(1)
PROOF: Since,
En ZZ0Un2 = En ZZ0U2 +En ZZ0 β βn 0
ZZ0 β βn
+2 En ZZ0 β βn 0
Assumption A3 EkZk4 <∞ andEjUj4 <∞.
Next, we provide a theorem that justi…es the consistency of
\
AsyVar βn
Theorem. UnderA1,A2 and A3,
En ZZ0Un2 =E ZZ0U2 +op(1)
PROOF: Since,
En ZZ0Un2 = En ZZ0U2 +En ZZ0 β βn
0
ZZ0 β βn
+2 En ZZ0 β βn 0
by the LLN and consistency of βn,
En ZZ0Un2 E ZZ0U2 En ZZ0U2 E ZZ0U2
+ β β
n 2
EnkZk4
+2 β βn En kZk3jUj
= op(1),
after noticing that, by Cauchy-Schwartz inequality,
E ZZ0U2 EkZk4 1/2 EjUj4 1/2 <∞,
and by Hölder’s inequality:
E kZk3jUj hE kZk4 i3/4 hEjUj4i1/4< ∞,
Observe that there is always a trade-o¤ between assumed moments of
by the LLN and consistency of βn,
En ZZ0Un2 E ZZ0U2 En ZZ0U2 E ZZ0U2
+ β β
n
2
EnkZk4
+2 β βn En kZk3jUj
= op(1),
after noticing that, by Cauchy-Schwartz inequality,
E ZZ0U2 EkZk4
1/2
EjUj4 1/2
< ∞,
and by Hölder’s inequality:
E kZk3jUj hE kZk4 i3/4 hEjUj4i1/4< ∞,
Observe that there is always a trade-o¤ between assumed moments of
by the LLN and consistency of βn,
En ZZ0Un2 E ZZ0U2 En ZZ0U2 E ZZ0U2
+ β β
n
2
EnkZk4
+2 β βn En kZk3jUj
= op(1),
after noticing that, by Cauchy-Schwartz inequality,
E ZZ0U2 EkZk4 1/2
EjUj4 1/2
< ∞,
and by Hölder’s inequality:
E kZk3jUj hE kZk4 i3/4 hEjUj4i1/4< ∞,
Observe that there is always a trade-o¤ between assumed moments of
by the LLN and consistency of βn,
En ZZ0Un2 E ZZ0U2 En ZZ0U2 E ZZ0U2
+ β β
n
2
EnkZk4
+2 β βn En kZk3jUj
= op(1),
after noticing that, by Cauchy-Schwartz inequality,
E ZZ0U2 EkZk4 1/2
EjUj4 1/2
< ∞,
and by Hölder’s inequality:
Detour: Recall the basic properties of a norm
kABk kAk kBk
kA+Bk kAk+kBk
Recall Hölder’s inequality:
E(kXYk) (EkXkp)1/p(EkYkq)1/q
with
1
p +
1
q =1
we have used 1/p=3/4 but alternative choices could be selected
Detour: Recall the basic properties of a norm
kABk kAk kBk
kA+Bk kAk+kBk
Recall Hölder’s inequality:
E(kXYk) (EkXkp)1/p(EkYkq)1/q
with
1
p +
1
q =1
we have used 1/p=3/4 but alternative choices could be selected
Detour: Recall the basic properties of a norm
kABk kAk kBk
kA+Bk kAk+kBk
Recall Hölder’s inequality:
E(kXYk) (EkXkp)1/p(EkYkq)1/q
with
1
p +
1
q =1
we have used 1/p=3/4 but alternative choices could be selected
Detour: Recall the basic properties of a norm
kABk kAk kBk
kA+Bk kAk+kBk
Recall Hölder’s inequality:
E(kXYk) (EkXkp)1/p(EkYkq)1/q
with
1
p +
1
q =1
we have used 1/p=3/4 but alternative choices could be selected
Detour: Recall the basic properties of a norm
kABk kAk kBk
kA+Bk kAk+kBk
Recall Hölder’s inequality:
E(kXYk) (EkXkp)1/p(EkYkq)1/q
with
1
p +
1
q =1
we have used 1/p=3/4 but alternative choices could be selected
Detour: Recall the basic properties of a norm
kABk kAk kBk
kA+Bk kAk+kBk
Recall Hölder’s inequality:
E(kXYk) (EkXkp)1/p(EkYkq)1/q
with
1
p +
1
q =1
we have used 1/p=3/4 but alternative choices could be selected
Why do not we use AsyVar\ βn in all occasions?
Because the con…dence interval will be usually wider and hypothesis
tests more ine¢ cient than those obtained with AsyVar^ β
n under
homoskedasticity.
Why do not we use AsyVar\ βn in all occasions?
Because the con…dence interval will be usually wider and hypothesis tests more ine¢ cient than those obtained with AsyVar^ β
n under
homoskedasticity.
Why do not we use AsyVar\ βn in all occasions?
Because the con…dence interval will be usually wider and hypothesis
tests more ine¢ cient than those obtained with AsyVar^ β
n under
homoskedasticity.
GLS
Assumption A4 E(YjZ) =Z0βand V(YjZ) =σ2(Z).
Assumption A5 E ZZ0/σ2(Z) is nonsingular.
The GLS estimator of βis:
βGLS
n = En
ZZ0
σ2(Z) 1
En
ZY
GLS
Assumption A4 E(YjZ) =Z0βand V(YjZ) =σ2(Z).
Assumption A5 E ZZ0/σ2(Z) is nonsingular.
The GLS estimator of βis:
βGLS
n = En
ZZ0
σ2(Z) 1
En
ZY
GLS
Assumption A4 E(YjZ) =Z0βand V(YjZ) =σ2(Z). Assumption A5 E ZZ0/σ2(Z) is nonsingular.
The GLS estimator of βis:
βGLS
n = En
ZZ0
σ2(Z) 1
En
ZY
Theorem. Under A4 and A5, the GLS estimator is CAN with
AsyVar βGLSn = 1
n E
ZZ0
σ2(Z) 1
.
PROOF:
βGLSn = β+ En
ZZ0
σ2(Z) 1
| {z }
=Op(1)
En
ZU
σ2(Z)
| {z }
=op(1)
whereU =Y Z0β,and,
n1/2 βGLS
n β = En
ZZ0
σ2(Z) 1
| {z }
p
!E σZZ02(Z) 1
n1/2En
ZU
σ2(Z)
| {z }
d
Theorem. Under A4 and A5, the GLS estimator is CAN with
AsyVar βGLSn = 1
n E
ZZ0
σ2(Z) 1
.
PROOF:
βGLSn =β+ En
ZZ0
σ2(Z) 1
| {z }
=Op(1)
En
ZU
σ2(Z)
| {z }
=op(1)
whereU =Y Z0β,and,
n1/2 βGLS
n β = En
ZZ0
σ2(Z) 1
| {z }
p
!E σZZ02(Z) 1
n1/2En
ZU
σ2(Z)
| {z }
d
Theorem. Under A4 and A5, the GLS estimator is CAN with
AsyVar βGLSn = 1
n E
ZZ0
σ2(Z) 1
.
PROOF:
βGLSn =β+ En
ZZ0
σ2(Z) 1
| {z }
=Op(1)
En
ZU
σ2(Z)
| {z }
=op(1)
whereU =Y Z0β,and,
n1/2 βGLS
n β = En
ZZ0
σ2(Z) 1
| {z }
n1/2En
ZU
σ2(Z)
TheAsyVar βGLSn is estimated by:
\
AsyVar βGLS
n =
1
n En
ZZ0
σ2(Z) 1
A serious problem consists of estimating σ2( ) without knowing its
TheAsyVar βGLSn is estimated by:
\
AsyVar βGLS
n =
1
n En
ZZ0
σ2(Z) 1
A serious problem consists of estimating σ2( ) without knowing its
Theorem. Under A4 and A5,
AsyVar βn AsyVar βGLSn is p.s.d.
Theorem. Under A4 and A5,
AsyVar βn AsyVar βGLSn is p.s.d.
Feasible GLS
In some occasions we know the functional form of σ2( ),i.e.
σ2(Z) = σ2γ(Z) with γ a given vector of parameters.
Notice that we can write,
U2 = σ2γ(Z) +error,
and we can estimate γ substitutingU2 by the OLS residuals. Let γn
Feasible GLS
In some occasions we know the functional form of σ2( ),i.e.
σ2(Z) = σ2γ(Z) with γ a given vector of parameters.
Notice that we can write,
U2 = σ2γ(Z) +error,
and we can estimate γ substitutingU2 by the OLS residuals. Let γn
Feasible GLS
In some occasions we know the functional form of σ2( ),i.e.
σ2(Z) = σ2γ(Z) with γ a given vector of parameters.
Notice that we can write,
U2 = σ2γ(Z) +error,
and we can estimate γsubstituting U2 by the OLS residuals. Let γn
The Feasible GLS is,
βFGLSn = "
En
ZZ0
σ2γ
n(Z)
!# 1
En
ZY
σ2γ
n(Z) !
.
Under suitable regularity conditions,
βFGLSn = βGLSn +op n 1/2 ,
The Feasible GLS is,
βFGLSn =
" En
ZZ0
σ2γ
n(Z)
!# 1 En
ZY
σ2γ
n(Z)
!
.
Under suitable regularity conditions,
βFGLSn = βGLSn +op n 1/2 ,
The Feasible GLS is,
βFGLSn =
" En
ZZ0
σ2γ
n(Z)
!# 1 En
ZY
σ2γ
n(Z)
!
.
Under suitable regularity conditions,
βFGLSn = βGLSn +op n 1/2 ,
Presentation of results: standard errors, con…dence
intervals and t ratios
Theorem. UnderA1,A2,and A3 and ifE ZZ0U2 is nonsingular,
\
AsyVar βn
1/2
βn β !d Np+1(0,Ip+1).
Under A1, A20 and ifE U2 >0,
^
AsyVar β
n
1/2
β
n β d
Presentation of results: standard errors, con…dence
intervals and t ratios
Theorem. UnderA1,A2,and A3 and ifE ZZ0U2 is nonsingular,
\
AsyVar βn
1/2
βn β !d Np+1(0,Ip+1).
Under A1, A20 and ifE U2 > 0,
^
AsyVar β
n 1/2
β
n β
d
We call standard error of the coe¢ cient βj to the estimate of
r
AsyVar βjn .
We have then the standard errors,
SE βjn =AsyVar^ βjn 1/2
under homoskedasticity.
SE βjn =AsyVar\ βjn 1/2
under heteroskedasticity.
We call standard error of the coe¢ cient βj to the estimate of r
AsyVar βjn .
We have then the standard errors,
SE βjn =AsyVar^ βjn
1/2
under homoskedasticity.
SE βjn =AsyVar\ βjn 1/2
under heteroskedasticity.
We call standard error of the coe¢ cient βj to the estimate of r
AsyVar βjn .
We have then the standard errors,
SE βjn =AsyVar^ βjn 1/2
under homoskedasticity.
SE βjn =AsyVar\ βjn
1/2
under heteroskedasticity.
We call standard error of the coe¢ cient βj to the estimate of r
AsyVar βjn .
We have then the standard errors,
SE βjn =AsyVar^ βjn 1/2
under homoskedasticity.
SE βjn =AsyVar\ βjn 1/2
under heteroskedasticity.
SALIDA 1
Dependent Variable: Y Method: Least Squares Included observations: 899
White Heteroskedasticity-Consistent Standard Errors & Covariance
Variable Coe¢ cient Std. Error t-Statistic Prob.
C -17766.84 1424.957 -12.46833 0.0000
LOG(Z1) 2116.400 157.2934 13.45511 0.0000
Z2 311.4864 47.75287 6.522883 0.0000
Z3 292.3367 87.12519 3.355363 0.0008
Z4 9.276353 16.01485 0.579235 0.5626
Z5 7.441149 16.33873 0.455430 0.6489
Z6 -886.3335 220.5673 -4.018426 0.0001
These results are reported as follow,
b
Y = 17,766.8
(1424.9)
+2,116.4
(157.2)
log(Z1) +311.5
(47.8) Z2+292.3(87.1) Z3
+9.28
(16.0) Z4+(7.4416.3) Z5 (886.3220.6) Z6+(246.9119.2) Z6Z2
Therefore, an approximate con…dence interval at the 95% for βj is
given by,
βnj 1.96 h
\
AsyVar β
n
i1/2
jj
Under homoskedasticity, the con…dence interval is,
βnj 1.96 h
^
AsyVar β
n
i1/2
These results are reported as follow,
b
Y = 17,766.8
(1424.9)
+2,116.4
(157.2)
log(Z1) +311.5
(47.8) Z2+292(87.1.3) Z3
+9.28
(16.0) Z4+(716.44.3) Z5 (886220..63) Z6+(246119..29) Z6Z2
Therefore, an approximate con…dence interval at the 95% for βj is
given by,
βnj 1.96
h \
AsyVar β
n
i1/2
jj
Under homoskedasticity, the con…dence interval is,
βnj 1.96 h
^
AsyVar β
n
i1/2
These results are reported as follow,
b
Y = 17,766.8
(1424.9)
+2,116.4
(157.2)
log(Z1) +311.5
(47.8) Z2+292(87.1.3) Z3
+9.28
(16.0) Z4+(716.44.3) Z5 (886220..63) Z6+(246119..29) Z6Z2
Therefore, an approximate con…dence interval at the 95% for βj is
given by,
βnj 1.96 h
\
AsyVar β
n
i1/2
jj
Under homoskedasticity, the con…dence interval is,
Signi…cance testing of the parameter βj consists of testing:
H0 : βj =0 versus H1 :βj 6=0.
Under suitable conditions (theorems in this section),
t = βnj SE βnj
Signi…cance testing of the parameter βj consists of testing:
H0 : βj =0 versus H1 :βj 6=0.
Under suitable conditions (theorems in this section),
t = βnj
SE βnj
We call t ratio (or t test) for testing the hypothesis
H0 :b0β0+b1β1+....+bpβp =c
with b0,b1, ...,bp,c known constants, to the statistic,
t = b0βn0+b1βn1+....+bpβnp c SE b0βn0+b1βn1+....+bpβnp
!d N(0,1) under H0,
with
SE b0βn0+b1βn1+....+bkβnp
= 8 > > > > > < > > > > > : ∑p i=0∑
p j=0bibj
h
\
AsyVar βn
i
ij
1/2
or
∑k
i=0∑kj=0bibj
h
^
AsyVar βn
i
ij
1/2
We call t ratio (or t test) for testing the hypothesis
H0 :b0β0+b1β1+....+bpβp =c
with b0,b1, ...,bp,c known constants, to the statistic,
t = b0βn0+b1βn1+....+bpβnp c
SE b0βn0+b1βn1+....+bpβnp
!d N(0,1) under H0,
with
SE b0βn0+b1βn1+....+bkβnp
= 8 > > > > > < > > > > > : ∑p i=0∑
p j=0bibj
h
\
AsyVar βn
i
ij
1/2
or
∑k
i=0∑kj=0bibj
h
^
AsyVar βn
i
ij
1/2
We call t ratio (or t test) for testing the hypothesis
H0 :b0β0+b1β1+....+bpβp =c
with b0,b1, ...,bp,c known constants, to the statistic,
t = b0βn0+b1βn1+....+bpβnp c SE b0βn0+b1βn1+....+bpβnp
!d N(0,1) under H0,
with
SE b0βn0+b1βn1+....+bkβnp
= 8 > > > > > < > > > ∑p i=0∑
p j=0bibj
h \ AsyVar β n i ij 1/2
or
h i 1/2
Testing
notation and some things worth knowing
We focus on tests with large sample justi…cation.
These apply to a wide variety of models, often under general
conditions, and useful …nite-sample justi…cation can be given only in special circumstances.
Testing
notation and some things worth knowing
We focus on tests with large sample justi…cation.
These apply to a wide variety of models, often under general
conditions, and useful …nite-sample justi…cation can be given only in special circumstances.
Testing
notation and some things worth knowing
We focus on tests with large sample justi…cation.
These apply to a wide variety of models, often under general
conditions, and useful …nite-sample justi…cation can be given only in special circumstances.
Let
θ r 1=
θ1 θ2
q 1
s 1.
True value:
θ0 = θ01 θ02 .
Consider the null hypothesis
H0 : θ01 =0 Composite if q <r
Simple if q =r. .
(in the composite case θ02 are nuisance parameters)
Alternative hypothesis:
Let
θ
r 1= θ1 θ2
q 1
s 1.
True value:
θ0 = θ01 θ02 .
Consider the null hypothesis
H0 : θ01 =0 Composite if q <r
Simple if q =r. .
(in the composite case θ02 are nuisance parameters)
Alternative hypothesis:
Let
θ
r 1= θ1 θ2
q 1
s 1.
True value:
θ0 = θ01 θ02 .
Consider the null hypothesis
H0 : θ01 =0 Composite if q <r
Simple if q =r. .
(in the composite case θ02 are nuisance parameters)
Alternative hypothesis:
Let
θ
r 1= θ1 θ2
q 1
s 1.
True value:
θ0 = θ01 θ02 .
Consider the null hypothesis
H0 : θ01 =0 Composite if q <r
Simple if q =r. .
(in the composite case θ02 are nuisance parameters)
Alternative hypothesis:
Let
θ
r 1= θ1 θ2
q 1
s 1.
True value:
θ0 = θ01 θ02 .
Consider the null hypothesis
H0 : θ01 =0 Composite if q <r
Simple if q =r. .
(in the composite case θ02 are nuisance parameters)
Alternative hypothesis:
Example: when testing on the mean, the variance is typically a nuisance parameter.
There is no real loss of generality in H0,since we can reparametrize
our problem,θ01 could be Mβ m,say.
De…nition: For a test statisticτˆn suppose that we reject H0 when
ˆ
τn >c.
Then
Πc
(θ01) = Pr(τˆn >cjθ01)
Example: when testing on the mean, the variance is typically a nuisance parameter.
There is no real loss of generality inH0,since we can reparametrize
our problem,θ01 could be Mβ m,say.
De…nition: For a test statisticτˆn suppose that we reject H0 when
ˆ
τn >c.
Then
Πc
(θ01) = Pr(τˆn >cjθ01)
Example: when testing on the mean, the variance is typically a nuisance parameter.
There is no real loss of generality inH0,since we can reparametrize
our problem,θ01 could be Mβ m,say.
De…nition: For a test statisticτnˆ suppose that we reject H0 when
ˆ
τn >c.
Then
Πc
(θ01) = Pr(τˆn >cjθ01)
Example: when testing on the mean, the variance is typically a nuisance parameter.
There is no real loss of generality inH0,since we can reparametrize
our problem,θ01 could be Mβ m,say.
De…nition: For a test statisticτˆn suppose that we reject H0 when
ˆ
τn >c.
Then
Πc
(θ01) = Pr(τnˆ >cjθ01)
Example: when testing on the mean, the variance is typically a nuisance parameter.
There is no real loss of generality inH0,since we can reparametrize
our problem,θ01 could be Mβ m,say.
De…nition: For a test statisticτˆn suppose that we reject H0 when
ˆ
τn >c.
Then
Πc
(θ01) = Pr(τˆn >cjθ01)
De…nition: (Consistency) The test in the previous de…nition is consistent i¤
Πc(
θ01) ! 1 as n!∞, 8θ01 6=0, 8c >0.
Unfortunately, in a given problem there are any number of statistics which have similar null distributions (whenH0 is true) and give
consistent tests.
How do we choose between them?
Consider a sequence of local alternatives (“Pitman” alternatives):
H1n :θ01 =δn 1/2 (1)
for a …xed q 1 vectorδ.The choice of δ determines the direction
of departure from H0,e.g.
δ= (1,0,. . .,0)0.
De…nition: (Consistency) The test in the previous de…nition is consistent i¤
Πc(θ01)
! 1 as n!∞, 8θ01 6=0, 8c >0.
Unfortunately, in a given problem there are any number of statistics which have similar null distributions (whenH0 is true) and give
consistent tests.
How do we choose between them?
Consider a sequence of local alternatives (“Pitman” alternatives):
H1n :θ01 =δn 1/2 (1)
for a …xed q 1 vectorδ.The choice of δ determines the direction
of departure from H0,e.g.
δ= (1,0,. . .,0)0.
De…nition: (Consistency) The test in the previous de…nition is consistent i¤
Πc(θ01)
! 1 as n!∞, 8θ01 6=0, 8c >0.
Unfortunately, in a given problem there are any number of statistics which have similar null distributions (whenH0 is true) and give
consistent tests.
How do we choose between them?
Consider a sequence of local alternatives (“Pitman” alternatives):
H1n :θ01 =δn 1/2 (1)
for a …xed q 1 vectorδ.The choice of δ determines the direction
of departure from H0,e.g.
δ= (1,0,. . .,0)0.
De…nition: (Consistency) The test in the previous de…nition is consistent i¤
Πc(θ01)
! 1 as n!∞, 8θ01 6=0, 8c >0.
Unfortunately, in a given problem there are any number of statistics which have similar null distributions (whenH0 is true) and give
consistent tests.
How do we choose between them?
Consider a sequence of local alternatives (“Pitman” alternatives):
H1n :θ01 =δn 1/2 (1)
for a …xed q 1 vectorδ.The choice of δ determines the direction
of departure from H0,e.g.
δ= (1,0,. . .,0)0.
De…nition: (Consistency) The test in the previous de…nition is consistent i¤
Πc(θ01)
! 1 as n!∞, 8θ01 6=0, 8c >0.
Unfortunately, in a given problem there are any number of statistics which have similar null distributions (whenH0 is true) and give
consistent tests.
How do we choose between them?
Consider a sequence of local alternatives (“Pitman” alternatives):
H1n :θ01 =δn 1/2 (1)
for a …xed q 1 vectorδ.The choice of δ determines the direction of departure from H0,e.g.
De…nition: (Consistency) The test in the previous de…nition is consistent i¤
Πc(θ01)
! 1 as n!∞, 8θ01 6=0, 8c >0.
Unfortunately, in a given problem there are any number of statistics which have similar null distributions (whenH0 is true) and give
consistent tests.
How do we choose between them?
Consider a sequence of local alternatives (“Pitman” alternatives):
H1n :θ01 =δn 1/2 (1)
for a …xed q 1 vectorδ.The choice of δ determines the direction
of departure from H0,e.g.
De…nition: Consider two statistics τn1ˆ , τn2ˆ which reject H0 when
ˆ
τni >c and where for some r.v. X
ˆ
τni !d X under H0 i =1,2.
If under H1n
Πc
1(θ01) =nlim
!∞Pr(τˆn1 >cjθ01)>Π c
2(θ01) =nlim
!∞Pr(τˆn2>cjθ01)
for all c >0 and for all δ,then we say thatτˆn1 is (asymptotically)
moree¢ cient than τˆn2.
IfΠc1(θ01) =Πc
De…nition: Consider two statistics τˆn1, τˆn2 which reject H0 when
ˆ
τni >c and where for some r.v. X
ˆ
τni d
!X under H0 i =1,2.
If under H1n
Πc
1(θ01) =nlim
!∞Pr(τn1ˆ >cjθ01)>Π
c
2(θ01) =nlim
!∞Pr(τn2ˆ >cjθ01)
for all c >0 and for all δ,then we say thatτˆn1 is (asymptotically)
moree¢ cient than τˆn2.
IfΠc1(θ01) =Πc
De…nition: Consider two statistics τˆn1, τˆn2 which reject H0 when
ˆ
τni >c and where for some r.v. X
ˆ
τni d
!X under H0 i =1,2.
If under H1n
Πc
1(θ01) =nlim
!∞Pr(τˆn1 >cjθ01)>Π c
2(θ01) =nlim
!∞Pr(τˆn2>cjθ01)
for all c >0 and for all δ,then we say thatτn1ˆ is (asymptotically)
moree¢ cient than τn2ˆ .
IfΠc1(θ01) =Πc
De…nition: Consider two statistics τˆn1, τˆn2 which reject H0 when
ˆ
τni >c and where for some r.v. X
ˆ
τni d
!X under H0 i =1,2.
If under H1n
Πc
1(θ01) =nlim
!∞Pr(τˆn1 >cjθ01)>Π c
2(θ01) =nlim
!∞Pr(τˆn2>cjθ01)
for all c >0 and for all δ,then we say thatτˆn1 is (asymptotically)
moree¢ cient than τˆn2.
De…nition: (Noncentral χ2q) X has a non-central χ2q distribution,
with noncentrality parameter
Λ=
q
∑
1 λ2j
and we write
X χ2q(Λ)
if
X =
q
∑
1
(uj +λj)2, uj NID(0, 1)
that is
X =
q
∑
1
De…nition: (Noncentral χ2q) X has a non-central χ2q distribution,
with noncentrality parameter
Λ=
q
∑
1 λ2j
and we write
X χ2q(Λ)
if
X =
q
∑
1
(uj +λj)2, uj NID(0, 1)
that is
X =
q
∑
1
De…nition: (Noncentral χ2q) X has a non-central χ2q distribution,
with noncentrality parameter
Λ=
q
∑
1 λ2j
and we write
X χ2q(Λ)
if
X =
q
∑
1
(uj +λj)2, uj NID(0, 1)
that is
X =
q
∑
1
De…nition: (Noncentral χ2q) X has a non-central χ2q distribution,
with noncentrality parameter
Λ=
q
∑
1 λ2j
and we write
X χ2q(Λ)
if
X =
q
∑
1
(uj +λj)2, uj NID(0, 1)
that is
X =
q
∑
1
Note that if you have a vector
X Np(µ,Σ)
since
Σ 1/2
(X µ) Np(0,I)
or
Σ 1/2X N
p(Σ 1/2µ,I)
then
X0Σ 1X χ2p(Λ)
with noncentrality parameter
Note that if you have a vector
X Np(µ,Σ)
since
Σ 1/2
(X µ) Np(0,I)
or
Σ 1/2X N
p(Σ 1/2µ,I)
then
X0Σ 1X χ2p(Λ)
with noncentrality parameter
Note that if you have a vector
X Np(µ,Σ)
since
Σ 1/2
(X µ) Np(0,I)
or
Σ 1/2X N
p(Σ 1/2µ,I)
then
X0Σ 1X χ2p(Λ)
with noncentrality parameter
Note that if you have a vector
X Np(µ,Σ)
since
Σ 1/2
(X µ) Np(0,I)
or
Σ 1/2X N
p(Σ 1/2µ,I)
then
X0Σ 1X χ2p(Λ)
with noncentrality parameter
Note that if you have a vector
X Np(µ,Σ)
since
Σ 1/2
(X µ) Np(0,I)
or
Σ 1/2X N
p(Σ 1/2µ,I)
then
X0Σ 1X χ2p(Λ)
with noncentrality parameter
Testing linear restrictions
Let Mbe a q (p+1)matrix withq p+1 such that
rank(M) =q, and a q 1 vectorm. The hypothesis to be tested is:
H0 :Mβ=mversus H1 :Mβ6=m,
which can also be written as:
H0 :θ1 =0 versusH1 :θ16=0,
Testing linear restrictions
Let Mbe a q (p+1)matrix withq p+1 such that
rank(M) =q, and a q 1 vectorm. The hypothesis to be tested is:
H0 :Mβ=mversus H1 :Mβ6=m,
which can also be written as:
H0 :θ1 =0 versusH1 :θ16=0,
Testing linear restrictions
Let Mbe a q (p+1)matrix withq p+1 such that
rank(M) =q, and a q 1 vectorm. The hypothesis to be tested is:
H0 :Mβ=mversus H1 :Mβ6=m,
which can also be written as:
H0 :θ1 =0 versusH1 :θ16=0,
REMARK: In what follows, for the sake of notational convenience, we remove the subindex “0” from θ1.Then, θ1 refers to the “true”
parameters, and generic values are denoted by Latin letters (b,t,etc).
We can always rearrange the components of β such that,
Mβ = M1β(1)+M2β(2) where M1 is a nonsingular q q matrix.
β
(p+1) 1 =
0 B B @
β(1)
q 1
β(2)
(p+1 q) 1 1 C C
REMARK: In what follows, for the sake of notational convenience, we remove the subindex “0” from θ1.Then, θ1 refers to the “true”
parameters, and generic values are denoted by Latin letters (b,t,etc).
We can always rearrange the components of β such that,
Mβ = M1β(1)+M2β(2) where M1 is a nonsingular q q matrix.
β
(p+1) 1 =
0 B B @
β(1)
q 1
β(2)
(p+1 q) 1 1 C C
REMARK: In what follows, for the sake of notational convenience, we remove the subindex “0” from θ1.Then, θ1 refers to the “true”
parameters, and generic values are denoted by Latin letters (b,t,etc).
We can always rearrange the components of β such that,
Mβ = M1β(1)+M2β(2) where M1 is a nonsingular q q matrix.
β
(p+1) 1
= 0 B B @
β(1)
q 1
β(2)
(p+1 q) 1
1 C C
Thus, underH0,
M1β(1)+M2β(2)=m=)
β(1) =M11
h
m M2β(2)
i
Thus, underH0,
M1β(1)+M2β(2)=m=)
β(1) =M11 h
m M2β(2) i
Example: suppose q =2:
2β1+β2 = 1 β1+β2+2β3 = 0
so,
M= 2 1 0
1 1 2 andm=
1 0 so
Mβ=M1β(1)+M2β(2)
where
M1= 2 1
1 1 , β
(1)
= β1
β2 , M2 =
0
2 , β
(2) =β3
so,
β(1)=M11
h
m M2β(2)
i
is
β1 β2 =
2 1 1 1 1 1 0 0
2 β3 =
1+2β3
Example: suppose q =2:
2β1+β2 = 1
β1+β2+2β3 = 0
so,
M= 2 1 0
1 1 2 and m= 1 0
so
Mβ=M1β(1)+M2β(2)
where
M1= 2 1
1 1 , β
(1)
= β1
β2 , M2 =
0
2 , β
(2) =β3
so,
β(1)=M11
h
m M2β(2)
i
is
β1 β2 =
2 1 1 1 1 1 0 0
2 β3 =
1+2β3
Example: suppose q =2:
2β1+β2 = 1
β1+β2+2β3 = 0
so,
M= 2 1 0
1 1 2 and m=
1 0
so
Mβ=M1β(1)+M2β(2)
where
M1= 2 1
1 1 , β
(1)
= β1
β2 , M2 =
0
2 , β
(2) =β3
so,
β(1)=M11
h
m M2β(2)
i
is
β1 β2 =
2 1 1 1 1 1 0 0
2 β3 =
1+2β3
Example: suppose q =2:
2β1+β2 = 1
β1+β2+2β3 = 0
so,
M= 2 1 0
1 1 2 and m=
1 0 so
Mβ=M1β(1)+M2β(2)
where
M1=
2 1 1 1 , β
(1)
= β1
β2 , M2 =
0 2 , β
(2)
=β3
so,
β(1)=M11
h
m M2β(2)
i
is
β1 β2 =
2 1 1 1 1 1 0 0
2 β3 =
1+2β3
Example: suppose q =2:
2β1+β2 = 1
β1+β2+2β3 = 0
so,
M= 2 1 0
1 1 2 and m=
1 0 so
Mβ=M1β(1)+M2β(2)
where
M1=
2 1
1 1 , β
(1)
= β1
β2 , M2 =
0
2 , β
(2) =β3
so,
β(1)=M11
h
m M2β(2) i
is
β1 β2 =
2 1 1 1 1 1 0 0
2 β3 =
1+2β3
Example: suppose q =2:
2β1+β2 = 1
β1+β2+2β3 = 0
so,
M= 2 1 0
1 1 2 and m=
1 0 so
Mβ=M1β(1)+M2β(2)
where
M1=
2 1
1 1 , β
(1)
= β1
β2 , M2 =
0
2 , β
(2) =β3
so,
β(1)=M11 h
m M2β(2)
i
IRestricted OLS:
(restricted means we imposeH0)
~β
n =arg min
b2Rp+1E
n
h
Y Z0b 2i
s.t.Mb=m.
Arranging the Z0 correspondingly
Z0 = Z(1)0 Z(2)0 ,
the model is written
Y =Z0β+U =Z(1)0β(1)+Z(2)0β(2)+U
and replacing (2), the restricted OLS estimator of β(2),~β
(2)
n ,is the
OLS estimator in the linear model:
h
Y Z(1)0M11mi= hZ(2)0 Z(1)0M11M2iβ(2)+U,
and the restricted estimator of β(1) is
~β(1)
n =M
1 1
h
m M2~β
(2)
n
i
IRestricted OLS:
(restricted means we imposeH0)
~β
n =arg min b2Rp+1E
n
h
Y Z0b 2
i
s.t.Mb=m.
Arranging the Z0 correspondingly
Z0 = Z(1)0 Z(2)0 ,
the model is written
Y =Z0β+U =Z(1)0β(1)+Z(2)0β(2)+U
and replacing (2), the restricted OLS estimator of β(2),~β
(2)
n ,is the
OLS estimator in the linear model:
h
Y Z(1)0M11mi= hZ(2)0 Z(1)0M11M2iβ(2)+U,
and the restricted estimator of β(1) is
~β(1)
n =M
1 1
h
m M2~β
(2)
n
i
IRestricted OLS:
(restricted means we imposeH0)
~β
n =arg min
b2Rp+1E
n
h
Y Z0b 2 i
s.t.Mb=m.
Arranging the Z0 correspondingly
Z0 = Z(1)0 Z(2)0 ,
the model is written
Y =Z0β+U =Z(1)0β(1)+Z(2)0β(2)+U
and replacing (2), the restricted OLS estimator of β(2),~β
(2)
n ,is the
OLS estimator in the linear model:
h
Y Z(1)0M11mi= hZ(2)0 Z(1)0M11M2iβ(2)+U,
and the restricted estimator of β(1) is
~β(1)
n =M
1 1
h
m M2~β
(2)
n
i
IRestricted OLS:
(restricted means we imposeH0)
~β
n =arg min
b2Rp+1E
n
h
Y Z0b 2 i
s.t.Mb=m.
Arranging the Z0 correspondingly
Z0 = Z(1)0 Z(2)0 ,
the model is written
Y =Z0β+U =Z(1)0β(1)+Z(2)0β(2)+U
and replacing (2), the restricted OLS estimator of β(2),~β
(2)
n ,is the
OLS estimator in the linear model:
h
Y Z(1)0M11mi= hZ(2)0 Z(1)0M11M2iβ(2)+U,
and the restricted estimator of β(1) is
~β(1)
n =M
1 1
h
m M2~β
(2)
n
i
IRestricted OLS:
(restricted means we imposeH0)
~β
n =arg min
b2Rp+1E
n
h
Y Z0b 2 i
s.t.Mb=m.
Arranging the Z0 correspondingly
Z0 = Z(1)0 Z(2)0 ,
the model is written
Y =Z0β+U =Z(1)0β(1)+Z(2)0β(2)+U
and replacing (2), the restricted OLS estimator of β(2),~β
(2)
n ,is the
OLS estimator in the linear model: h
Y Z(1)0M11mi= hZ(2)0 Z(1)0M11M2
i
β(2)+U,
and the restricted estimator of β(1) is
~β(1)
n =M
1 1
h
m M2~β
(2)
n
i