Professional Documents
Culture Documents
in Dyadic Regression
1 / 16
Setting
N
1 XX
θ̂ = arg max `ij (θ) (1)
θ∈Θ N(N − 1)
i=1 j6=i
where
`ij (θ) = ln fYij |Xi ,Xj (Yij |Xi , Xj ; θ).
2 / 16
Example
[exp(Wij0 θ)]Yij
fYij |Xi ,Xj (Yij |Xi , Xj ; θ) = exp[− exp(Wij0 θ)]
Yij !
where
I Yij = exports from country i to country j,
I Wij = (ln GDPi , ln GDPj , ln distij )0 .
3 / 16
Limit Distribution
p
I Assume HN (θ̄) → Γ0 invertible, and that we have a consistent
estimator Γ̂ of Γ0 .
I Decompose the score:
(1st) Projec-
(1st) Projection tion Error
z }| { z}|{
SN (θ0 ) = U1N + U2N + VN . (2)
|{z} |{z}
(2nd) Hájek (2nd) Projec-
Projection tion Error
4 / 16
Details on Score Decomposition I
5 / 16
Details on Score Decomposition II
I Projection error:
−1 X
N (sij − s̄ij ) + (sji − s̄ji )
VN := .
2 2
i<j
N
2 X s̄ e (Xi , Ai ) + s̄ a (Xi , Ai )
U1N := ,
N 2
i=1
6 / 16
Details on Score Decomposition III
s̄ e (Xj , Aj ) + s̄ a (Xj , Aj )
− .
2
7 / 16
Variance of the Score I
I For p = 1, 2 let
s̄i1 ,i2 + s̄i2 ,i1 s̄j1 ,j2 + s̄j2 ,j1
Σp := Cov ,
2 2
√
I Estimate Ω := Var( NSN ) = 4Σ1 + 2/(N − 1)(Σ2 + Σ3 − 2Σ1 ) by
sample counterparts using:
−1 X ( 0
N 1 ŝij + ŝji ŝik + ŝki
Σ̂1 :=
3 3 2 2
i<j<k
0 0 )
ŝij + ŝji ŝjk + ŝkj ŝjk + ŝkj ŝik + ŝki
+ + .
2 2 2 2
with an index in common for each of them: {ij} and {ik}, {ij} and
{jk}, {jk} and {ik}. Hence, Σ̂1 is the sample covariance of
(ŝij + ŝji )/2 with other dyads sharing exactly 1 index in common.
10 / 16
Estimators of the Variance
1. Benchmark Analogue of (3) II
I Also: −1 X 0
N ŝij + ŝji ŝij + ŝji
Σ\
2 + Σ3 = .
2 2 2
i<j
(Γ̂0 Ω̂−1
A Γ̂)
−1
for
2
Ω̂A := 4Σ̂1 + 2 + Σ3 − 2Σ̂1 ).
(Σ\
N −1
11 / 16
Estimators of the Variance
2. Fafchamps and Gubert (2007) I
1 XXXX
Ω̂FG := 2
Ci1 ,i2 ,j1 ,j2 ŝi01 ,i2 ŝj1 ,j2 ,
N(N − 1)
i1 i2 6=i1 j1 j2 6=j1
where Ci1 ,i2 ,j1 ,j2 = 1 if {i1 , i2 } and {j1 , j2 } share at least one index in
common and 0 otherwise.
I Widely cited paper, but contained only an informal discussion of why
the proposed estimator leads to asymptotically correct inference.
12 / 16
Estimators of the Variance
2. Fafchamps and Gubert (2007) II
1 N N \
Ω̂FG = 8 × 3 Σ̂ 1 + 4 × Σ 2 + Σ 3
N(N − 1)2 3 2
2
= 4Σ̂1 + 2 + Σ3 − 2Σ̂1 ) = Ω̂A .
(Σ\
N −1
13 / 16
Estimators of the Variance
3. Snijders and Borgatti (1999) Jacknife I
N
N −2X
Ω̂SB = [SN,−i (θ̂) − SN (θ̂)][SN,−i (θ̂) − SN (θ̂)]0 ,
2
i=1
N−1
where SN,−i is the average of the dyadic scores over the 2 dyads
that do not include agent i.
I Graham shows the above not consistent, but can modify it slightly
to make it consistent.
14 / 16
Estimators of the Variance
3. Snijders and Borgatti (1999) Jacknife II
I Define
1 X ŝij + ŝji
ˆs̄1i :=
N −1 2
j6=i
I ˆs̄1i is the usual estimate of the i-th term of the Hájek projection, so
PN
a natural estimate of Σ1 is Σ̃1 := N1 i=1 ˆs̄1i ˆs̄1i0 . With some
algebraic manipulation:
2(N − 2) 4
Ω̂SB = 4Σ̃1 = 4Σ̂1 + 2 + Σ3 − Σ̂1 )
(Σ\
N N −1
which is consistent.
15 / 16
Estimators of the Variance
3. Snijders and Borgatti (1999) Jacknife II
2(N − 2) 2
Ω̂SB − (Σ\
2 + Σ3 ) = Ω̂FG = Ω̂A .
N N −1
16 / 16