2.8 Supplementary Appendix
2.8.2 Auxiliary Lemmas and Proofs
the influence functions directly. Let
ψp,x,1(a;θ0,G)≡E
(1−R)`(Z)Λx(W;β0) (1−Q0)fX|R(X|1) ·
(1−r)`(z)1(x≤X)
1−Q0 −FX|R(X|1)
, (2.8.25)
ψp,x,3(a;θ0,G)≡ −r−Q0 Q0
·E
(1−R)`(Z)Λx(W;β0) 1−Q0
· G(X) fX|R(X|1)
. (2.8.26)
Combining (2.8.25), (2.8.23), and (2.8.26), we deduce that
∆p,3=1 n
n
∑
i=1
ψp3,i=−1 n
n
∑
i=1
1 fY|R(qτ|1)
ψg,p,i+(1−Ri)`(Zi)Λx(Wi;β0)gp(Xi) 1−Q0
−dp(θ0,G)) +op(n−1/2). (2.8.27)
whereψg,pis defined in (2.7.10). Collecting the results in (2.8.20), (2.8.21), and (2.8.27), it follows that the asymptotic linear representation forU QE[pis given as in (2.7.3). This completes our proof.
verify two conditions: (i) the classF is Donsker, and (ii) the variance of∆m4(A;θ,qbτ,qτ)tends to zero.
To show (i), we let
F1≡ {y∈Y 7→1(y≤q)−1(y≤qτ): q∈Y},
F2≡ {z∈Z 7→L(k(z)0γ)/L(k(z)0γ+t(z)0λs): (γ,λs)∈Θγ,λs}, F3≡ {(r,z)∈ {0,1} ×Z 7→rej(z): j=1, ...,dβ}.
Hence,F⊆F1·F2·F3.
Here,F1is a collection of indicator functions overR. Under Assumptions 2.7(a)(ii) and (e)(i) ,F2is pointwise compact; for definition, see Example 19.8 of Van der Vaart (1998). It is well-known that the pointwise compact class of functions and collections of indicators of cells in Euclidean space are pointwise measurable; see the discussion in Section 2.3 of Van Der Vaart and Wellner (1996) for definition. The pointwise measurability ofF3follows by definition. Then, by Lemma 8.10 in Kosorok (2008),F is pointwise measurable. F1is a VC-subgraph class with VC-index equal to 2. In addition,F1is uniformly bounded byF1≡2. Theorem 2.6.7 in Van Der Vaart and Wellner (1996) implies that
sup
Q
N(εkF1kQ,2,F1,L2(Q))≤K1ε−2. (2.8.29)
Due to Assumption 2.7(e)(ii),
L(k(z)0γ1)
L(k(z)0γ1+t(z)0λs1)−L(k(z)L(k(z)0 0γ2)
γ2+t(z)0λs2)
≤K2· kk(z)k kγ1−γ2k+K3· kt(z)k kλs1−λs2k.
Combining this fact and Assumption 2.7(f),F2admits an integrable envelop function,F2≡K2γ¯kk(z)k+K3λ¯skt(z)k+
K4, where ¯γ ≡supγ∈Θ
γkγk and ¯λs≡supλ
s∈Θλskλsk are finite under Assumption 2.7(a)(ii). It then follows from Theorem 2.10.20 of Van Der Vaart and Wellner (1996) that for allδ >0,
Z δ
0
sup
Q
q
logN(εkF2kQ,2,F2,L2(Q))dε≤
∑
j∈{γ,λs} Z δ
0
q
logN(εj,¯Θj,k·k)dε
≤
∑
j∈{γ,λs}
pdj Z δ
0
s log
1+4diam(Θj) εj¯
dε<∞, (2.8.30)
wherediam(Θ)is the diameter ofΘ, andQis any finite discrete measure. By Assumption 2.7(f),E[F22]<∞, and therefore,F2is Donsker. Another application of Theorem 2.10.20 of Van Der Vaart and Wellner (1996) toF yields that
Z δ
0
sup
Q
q
logN(4εkFkQ,2,F,L2(Q))dε
≤
3
∑
j=1 Z δ
0
sup
Q
q
logN(ε Fj
Q,2,Fj,L2(Q))dε<∞, (2.8.31)
whereF3(z)≡ ke(z)k, is an envelop function forF3. The last inequality follows from (2.8.29), (2.8.30), and the fact thatF3is a finite set, and therefore, BUEI.
To show thatFis square integrable, note
E[|F1F2F3|2]≤4E[ke(Z)k2(K2γ¯kk(Z)k+K3λ¯skt(Z)k+K4)2]<∞,
where the second inequality follows by the Cauchy–Schwartz inequality and Assumption 2.7(f). By Theorem 2.10.20 and Theorem 2.10.1 in Van Der Vaart and Wellner (1996),F is Donsker.
In the next step, we show (ii). That is,Rk∆m4(a;θ,qbτ,qτ)k2dFA(a)converges in probability to 0. Towards this end,
Z
||∆m4(a;θ,bqτ,qτ)||2dF(a)≤ c25
2c24· sup
y∈Y,z∈Z|fY|ZR(y|z,1)|
!2
E[ke(Z)k2](qbτ−qτ)2
≤K5(qbτ−qτ)2=Op(n−1s ), (2.8.32)
where the second inequality is by MVT and Assumption 2.7(b). The last equality follows by (2.8.7).
Combining (i) and (ii), Theorem 2.1 in Wellner and van der Vaart (2007) yields (2.8.28). To complete the proof, it suffices to show thatE[m4(A;θ,bqτ)−m4(A;θ,qτ)|Yn] =Op(n−1/2). By a first-order expansion,
E[km4(A;θ,qbτ)−m4(A;θ,qτ)k |Yn] =E[
R fY|ZR(eqτ|Z,1)e(Z)
|Yn]· |qbτ−qτ|
≤K6 sup
y∈Y,z∈Z|fY|ZR(y|Z,1)|E[ke(Z)k]·Op(n−1/2) =Op(1)·Op(n−1/2) =Op(n−1/2),
where the second line follows by MVT, Assumptions 2.7(b), and (f).
Lemma 2.5 Under the assumptions of Lemma 2.2,
H1≡ (
(r,z)∈ {0,1} ×Z 7→(r−L(k(z)0γ))L0(k(z)0γ)ek(z)
L(k(z)0γ)(1−L(k(z)0γ)) :ek(z)∈ {k1(z), ...,kdγ(z)},γ∈Θγ
o , H2≡
(r,z)∈ {0,1} ×Z 7→
r
L(k(z)0γ+t(z)0λs)−1
L(k(z)0γ)et(z): et(z)∈ {t1(z), ...,td
λ(z)},(γ,λs)∈Θγ,λs ,
H3≡
(r,z)∈ {0,1} ×Z 7→
1−r
1−L(k(z)0γ+t(z)0λa)−1
L(k(z)0γ)et(z): et(z)∈ {t1(z), ...,td
λ(z)},(γ,λa)∈Θγ,λa , H4≡
(r,z,x,y)∈ {0,1} ×Z X Y 7→
r·1(y≤q)
L(k(z)0γ+t(z)0λs)− (1−r)·Λ(w,β) 1−L(k(z)0γ+t(z)0λa)
·L(k(z)0γ)ee(z):
ee(z)∈ {e1(z), ...,ed
β(z)},q∈Y0,θ∈Θ o
,
are Glivenko-Cantelli (GC).
Proof of Lemma 2.5: Under Assumptions 2.7(e)(iii) and (f),H1,H2, andH3, admit integrable envelop functions, H1,H2, andH3, respectively, whereH1(r,z)≡K1kk(z)k,H2(r,z)≡K2kt(z)k, andH3(r,z)≡K3kt(z)k. Hence, by Example 19.8 in Van der Vaart (1998),Hj,j=1,2,3,are GC. To show thatH4is also GC, we define
H5≡ {y∈ {0,1} ×Y 7→1(y≤q),q∈Y0}, H6≡
(r,z,x)∈ {0,1} ×Z X Y 7→(1−r)Λ(w,β)L(k(z)0γ)e(z)e 1−L(k(z)0γ+t(z)0λa) , ee(z)∈ {e1(z), ...,ed
β(z)},q∈Y0,θ∈Θ o
.
Notice thatH4=F2F3H5−H6. Under Assumptions 2.7(c)(iii), (e)(iii), and (f),H6admits an integrable envelop, H6≡K4ke(z)k. By Example 19.8 in Van der Vaart (1998),H4is GC. It is straightforward to checkF2F3H5admits an integrable envelop function. From the proof of Lemma 2.4, we know thatF2andF3are Donsker. Then, by the Donskerness ofH5andH6, and Corollary 9.26 in Kosorok (2008),H4is also GC.
Lemma 2.6 Under the assumptions of Theorem 2.3,
dbq,n(θb,G)−dq(θ0,G) =op(1).
Proof of Lemma 2.6: Letη(θ) = (η1(·,θ),η2)0, f(a;θ,η(θ))≡ f1(a;θ,η(θ))·G−1(η1(x,θ)/η2) +f2(a;θ,η(θ)), where
f1(a;θ,η(θ))≡ 1−r Q0η2
· L(k(z)0γ+t(z)0λs)
1−L(k(z)0γ+t(z)0λa)·Λx(w;β), and f2(a;η(θ))≡ −x·f1(a;θ,η(θ)). With these quantities, we can write
dbq,n(bθ,G)−dq(θ0,G) =En[f(A;θ0,ηbn(θ))]−E[f(A;η0(θ0))],
where
ηbn(θ)≡ En
"
(1−R) Q0
L(k(z)0γb+t(z)0bλs) 1−L(k(z)0γb+t(z)0bλa)
1(X≤ ·)
# ,En[R]
Q0
!0
,
η0(θ)≡
E
(1−R) Q0
L(k(z)0γ+t(z)0λs)
1−L(k(z)0γ+t(z)0λa)1(X≤x)
,1 0
.
LetNθ0 denote a neighborhood ofθ0. We proceed by showing, (i) whenηbn(θ)is sufficiently close toη0(θ), we have supθ∈N
θ0
kf(a;θ,η(θ))−f(a;θ,η0(θ))k ≤c(a)·supx∈X,θ∈N
θ0
kη(θ)−η0(θ)k, for somec(·)satisfying, (ii)
E[c(A)] sup
x∈X,θ∈Nθ0kηbn(θ)−η0(θ)k→p 0, and (iii)E[supθ∈N
θ0,kη−η0k≤εkf(A;θ,η(θ))k]<∞.
By Assumptions 2.7(c) and (e),L(·)andΛx(·)are uniformly bounded. Under Assumption 2.8(d),G−1is Lipschitz with a bounded Lipschitz constant. Then, (i) follows immediately by
kf(a;θ,η(θ))−f(a;θ,η0(θ))k ≤K1sup
x∈Xkη1(x,θ)−η1,0(x,θ)k+K2kη2−η2,0k, and lettingc(·) =max{K1,K2}.
Next, we shall verify (ii). To this end, it suffices to show thatηb1,nconverges toη1,0uniformly. This follows from Gηbeing GC, whereGη≡Gη,1·Gη,2, and
Gη,1≡
(r,z)∈ {0,1} ×Z 7→ (1−r)L(k(z)0γ+t(z)0λs)
Q0(1−L(k(z)0γ+t(z)0λa)):θ∈Θ
, Gη,2≡ {x∈X 7→1(x≤q):q∈X}.
SinceGη,1is uniformly bounded, it follows from Example 19.2 in Van der Vaart (1998) that it is GC.Gη,2is VC subgraph with VC index equal to 2, and therefore, it is GC. Then, by Corollary 9.27 in Kosorok (2008),Gη is also uniformly bounded GC.
Lastly, (iii) holds under Assumption 2.1(d) and Assumption 2.7(e).
By (i), (ii), and the Markov inequality, supθ∈N
θ0kf(A;θ,ηb(θ))−f(A;θ,η0(θ))k→p 0. Then the desired result is
obtained by (iii) and Lemma 4.3 in Newey and McFadden (1994).
Lemma 2.7 Under the assumptions of Theorem 2.3, we have thatEn[∆φq(A;θ0,G)] =n−1∑ni=1ψg,q(Ai;θ0,G) +op (n−1/2),where∆φgandψg,qare defined in (2.8.17) and (2.7.9), respectively.
Proof of Lemma 2.7:Letf0(a;η)≡1−Q1−r
0`(z)Λx(w;β0)G−1(η1/η2),and hence,∆φq(a;θ0,G) =f0(a;ηbn)−f0(a;η0), whereηbn(·)≡
En
h(1−R)`(Z)
1−Q0 1(X≤ ·)i
,En[R]/Q00
,andη0(·)≡(FX|R(·|1),1)0. The first step of our proof is to show that
√n(En−E)(f0(a;ηbn)−f0(a;η0)) =op(1), (2.8.33)
by invoking Theorem 2.1 in Wellner and van der Vaart (2007).
Towards this end, we need to (i) define the functional space Hη such that P(ηbn∈Hη)→1, (ii) verify that Fη≡ {f0(·;η):η∈Hη}is Donsker, and (iii) show that
Z
(f0(a;η)b −f0(a;η0))2dFA(a)→p 0. (2.8.34)
For (i), letHη≡(Hη,1,Hη,2), where
Hη,1≡ {x∈X 7→ f(x):f non-decreasing, bounded between 0 and 1},
andHη,2≡[1−δη,1+δη], for someδη∈(0,1/2). GivenH, Condition (i) is implied byηb→p η0uniformly, which is shown in Lemma 2.6.
Next, we establish the Donsker property ofFη. Note thatFη=1−Q1−r
0`(z)Λx(w;β0)·G−1(Hη1/Hη2).By Lemma 9.11 in Kosorok (2008),Hη,1is BUEI, relative to the envelopHη,1≡1.H2is a bounded convex set in the Euclidean space, and hence, trivially BUEI. Pointwise measurability is immediate from the definitions of the two sets. By Assumption 2.8(d), G−1 is a Lipschitz continuous function with a bounded Lipschitz constant. Given that H2 is bounded away from 0, we conclude from Lemma 9.14 and Theorem 9.15 in Kosorok (2008) thatG−1(Hη1/Hη2)is also BUEI and pointwise measurable relative to the envelopHη,3≡sup{x∈X}. By Theorem 2.10.1 in Van Der Vaart and Wellner (1996),G−1(Hη1/Hη2)is uniformly bounded Donsker. Finally, under Assumptions 2.7(c)(iii) and (e), Corollary 9.32 then implies thatFηis uniformly bounded Donsker.
Now, to show (2.8.34), note that Z
(f0(a;η)b −f0(a;η0))2dFA(a)≤K1sup
x∈X|ηb1,n(x)−η1,0(x)|2+K2|ηb2,n−η2,0|2=op(1),
where the first inequality follows by carefully bounding the coefficients associated with each term and by the fact that fourth moments ofkk(z)kandkt(z)kexist under Assumption 2.7(f). The last one follows becauseηbnconverges uniformly toη0.
Next, we prove that
√n
ηb1,n(·)
ηb2,n
−
η1,0(·)
η2,0
Gη1,η2(·),
whereGη1,η2(·)is a tight, two-dimensional mean zero Gaussian process with covariance functionΣ(x1,x2) =E[(ψη,1 (x1),ψη,2)(ψη,1(x2),ψη,2)0],ψη,1(·) =ψη,1(a,γ0;·)≡hη(r,z;γ0)1(x≤ ·)−FX|R(x|1),ψη,2≡(r−Q0)/Q0,andhη(r,z;
γ0)≡(1−r)`(z)1−Q
0 .
The weak convergence follows fromG0,η being Donsker, where
G0,η≡ {(r,x,z)∈ {0,1} ×X ×Z 7→hη(r,z;γ0)·1(x≤q):q∈X}.
Under Assumption 2.7(e),hηis uniformly bounded. Since the class of indicator functions is uniformly bounded Donsker, by Corollary 9.32 in Kosorok (2008), we conclude thatGηis also uniformly bounded Donsker.
In view of (2.8.33), we deduce that
En[∆φq(A;θ0,G)] =E[f0(A;ηbn)−f0(A;η0)] +op(n−1/2). (2.8.35)
In the next step, we derive the asymptotic linear representation of the first term on the right hand side of (2.8.35) . The proof continues by showing that the map, φG(η)≡R f(a;η)dFA(a),is Hadamard differentiable in η atη0. Observe that we can decomposeφGas follows
(η1,η2)7→η1
η2
7→G−1◦(η1/η2)7→hG◦(G−1◦(η1/η2)),
wherehG(g)≡R 1−Q1−r
0`(z)Λx(w;β0)gdFA(a).The first map is continuous and uniformly bounded onHη,2, and thus, Hadamard differentiable. The second and third maps are both composition maps. SinceG−1(·)is continuously differ- entiable with bounded first-order derivative under Assumption 2.8(d), by Lemma 3.9.25 in Van Der Vaart and Wellner (1996), it is Hadamard differentiable. Since integration is a linear functional,hGis also Hadamard differentiable. Now we invoke the chain rule, e.g. Theorem 20.9 in Van der Vaart (1998), and conclude thatφGis Hadamard differentiable, with the functional derivative ofφGatη0in the direction of(ψη,1,ψη,2)given as follows
(ψη,1,ψη,2)7→hg◦ 1
G0◦G−1◦ η1,0
η2,0
· ψη,1
η2,0−η1ψη,2
η2,02
!!
.
Given the Hadamard differentiability, we then apply the delta method (as in Theorem 3.9.4 in Van Der Vaart and
Wellner (1996)) to get the linear expansion,ψg,q,
ψg,q(a;θ0) =
Z 1−r
1−Q0· `(z)Λx(w;β0)
G0(G−1(η1,0/η2,0))· ψη,1
η2,0
−η1,0ψη,2
η2,02
! dFA(a)
=E
"
1−R
1−Q0· `(Z)Λx(W;β0) G0 G−1(FX|R(X|1))·
(1−r)`(z)1(x≤X)
1−Q0 −rFX|R(X|1) Q0
# .
This concludes our proof.