Condition (C1) is standard for implementing the kernel estimation, in which the requirement of∞
−∞urK(u)duto be finite forr∈Nis satisfied by commonly used kernel functions listed in Wand & Jones (1995). Condition (C2) is regarded as the optimal bandwidth in the sense of Wand and Jones (1995, Sect. 2.5), and thus, we takehas of the rate ofn−1/5in the following development.
3 Feature Screening for Censored Data with Error-Prone
whereX¯i∗· =n1
i
ni
#
r=1
Xir∗.
Scenario III: Both functional form of f(·) and its associated parameters are unknown, but external validation data are available.
Suppose thatMis the subject set for the main study containing measurements Ti, Ci, δi, Xi∗
:i∈M
for nsubjects and that V is the subject set for the external validation study containing measurements
Xi, X∗i
:i∈V for m subjects, whereMandVdo not overlap. Assume that the main study and the validation study share the same measurement error model (7); this is the so-called transportabilityassumption (e.g., Yi et al.2015).
With the availability of external validation data,f(j )(·)forj =1,· · ·, pand
can be estimated. Fori∈Vandj =1,· · · , p, thejth component ofi is given byi(j ) =X∗i(j )−Xi(j ), which is known. Then adopting the estimator (6) with Xi(j )replaced byi(j )andnreplaced bymgives an estimate of the probability density functionfi(j )(·)ofi(j ):
f(j )(u)= 1 mh
i∈V
K
u−i(j ) h
.
Thus, the corresponding characteristic function φ(j )(u) is estimated by φ(j )(u) = ∞
−∞exp(iux)f(j )(x)dx. In addition, applying the least squares regression method gives the estimator of:
= 1 m−1
i∈V
(Xi∗−Xi)(Xi∗−Xi). (9)
3.2 Feature Screening with Measurement Error Effects Accommodated
In the presence of measurement error in covariates, the method in Sect. 2.3 cannot apply because the estimator (5) cannot be directly calculated due to the unavailability of the Xi. In this subsection, we derive an estimator (5) using the observed surrogateXi∗. First, we re-express the probability density functionfX(j )(x) by the inverse Fouriertransformation, given by
fX(j )(x)= 1 2π
∞
−∞exp(−iux) φX(j )(u)du, (10) whereφX(j )(u)is the characteristic function ofX(j ).
Forj =1,· · ·, p, letφX∗
(j )(u)andφ(j )(u)denote the characteristic functions of X(j )∗ and(j ), respectively, whereX∗(j )and(j )are thejth component ofX∗and, respectively; andX∗andfollow the same distribution asXi∗andi, respectively.
Then model (7) yields that φX∗
(j )(u)=φX(j )(u)φ(j )(u), and thus,φX(j )(u)= φφX(j )∗ (u)
(j)(u), assumingφ(j )(u)=0. Then (10) becomes fX(j )(x)= 1
2π ∞
−∞exp(−iux)φX∗ (j )(u)
φ(j )(u)du. (11) To emphasize that (11) is expressed in terms of the surrogateX∗(j ), we letfadj,j(x) to replacefX(j )(x)in the left-hand side of (11).
Next, to implement (11), we need to calculate φX∗
(j )(u) and φ(j )(u), where φ(j )(u)is derived from the distributionf(j )(·)of(j ), thejth marginal distribution derived fromf(·).
It now remains to calculateφX∗
(j )(u), which is given by φX∗
(j )(u)= ∞
−∞exp(iux) fX∗
(j )(x)dx, (12)
where fX∗
(j )(x) denotes the probability density function of X(j )∗ . Since X∗(j ) is observable, then the probability density function ofX∗(j )can be estimated by the kernel estimation, given by
fX∗
(j )(x)= 1
nh n i=1
K
$x−X∗i(j ) h
%
, (13)
wherehandK(·)are described for (6). In our numerical examination, we specify K(u)to be the normal kernel andhcan be estimated by the cross-validation method (e.g., Wand & Jones1995).
Consequently, withfX∗
(j )(x)in (12) replaced by fX∗
(j )(x),φX∗(u) can be esti- mated by
φX∗
(j )(u)= ∞
−∞exp(iux)fX∗ (j )(x)dx
= ∞
−∞exp(iux) 1 nh
n i=1
K
$x−X∗i(j ) h
% dx.
Letz= x−Xh∗i(j ); then applying the change of variables yields φX∗
(j )(u)= ∞
−∞
1 n
n i=1
exp
iuX∗(j )+iuhz
K (z) dz
=
& ∞
−∞exp(iuhz) K(z)dz '
× (
1 n
n i=1
exp
iuXi(j )∗ )
. (14)
Combining (11) and (14) gives an estimator of (11):
fadj,j(x)= 1 2π
∞
−∞exp(−iux)φX∗ (j )(u)
φ(j )(u)du, (15) and thus, an adjusted estimator of the cumulative distribution functionFX(j )(x)in terms ofX(j )∗ is
Fadj,j(x)= x
−∞
fadj,j(u)du. (16)
Therefore, the functional distance correlation (3) can be estimated using the observed surrogateX∗(j )together with the outcomeY, given by
ωj dcorr{ Fadj,j(X(j )∗ ),F (Y ) }
= dcov{Fadj,j(X(j )∗ ),F (Y ) }
dcov{ Fadj,j(X∗(j )),Fadj,j(X∗(j ))}dcov∗{F (Y ), F (Y ) }
, (17)
where dcov{Fadj,j(X∗(j )),F (Y ) } is determined by (4) with FX(j )(x) replaced by (16).
Remark The development here extends the discussion of Chen (2019) who assumed thatf(·)is the probability density function of a normal distribution under Scenarios I, II, and III. With the jth noise term (j ) assuming a normal distribution with mean zero and varianceσ2
(j ), we have that the characteristic function is given by φ(j )(u)=exp
−12u2σ2
(j )
, and thus, (15) becomes fadj,j(x)= 1
2π ∞
−∞exp
−iux+1 2uσ2
(j )
φX∗ (j )(u)du.
In contrast, if(j )follows at distribution with degrees of freedomv > 1, then the corresponding characteristic function is given by Dreiera & Kotzb (2002):
φ(j )(u)=2vvv/2 (v)
∞
0
exp
−v1/2(2x+ |u|)
× {x(x+ |u|)}(v−1)/2dx, (18)
and substituting (18) into (15) yieldsfadj,j(x).
3.3 Asymptotic Results
To establish theoretical results of the proposed method, we impose the following additional conditions:
(C3) There exists a positive constantw0such that for all 0< w≤2w0, sup
p
1max≤j≤pE
exp
w X(j ) 21
<∞ and E
exp
w Y 2q
<∞. (C4) The minimum of the functional distance correlations for the active covariates
satisfies
minj∈Iωj≥2cn−ζ for some constantsc >0 and 0≤ζ <1/2.
(C5) There exists a positive constantv0such that lim
p→∞
minj∈Iωj−max
j∈Icωj
>
v0, assuming the limits exists.
(C6) The covariatesXi∗fori=1,· · ·, nare bounded.
Condition (C3) is used to examine the boundness of the difference ωj −ωj between (3) and its estimator (17). Condition (C4) says that the marginal DC of active covariates cannot be too small, which is similar to Condition 3 of Fan & Lv (2008). Condition (C5) basically requires the signal carried by the active covariates to be stronger than that displayed by inactive covariates for at least a fixed amount if the dimensionpgoes to infinty. This condition was also imposed by other authors (e.g., Cui et al.2015). Condition (C6) indicates the finite boundness of surrogate measurements of the covariates.
Theorem 1 Under regularity conditions (C3) and (C5) and the assumptions of Lemmas1 and2in Appendix A, we have that forcandζ described in Condition (C4), there exists a constantD >0such that
P
j=max1,···,pωj −ωj≥cn−ζ
=O
pexp
−Dn1−2ζ
. (19)
Moreover,
P
maxj∈Icωj≥min
j∈Iωj
=O
&
exp
−1 4Dnv02
'
, (20)
wherev0is the constant described in Condition (C5).
Equation (19) in Theorem 1 indicates ωj is close to its estimate with a large probability. Similar to the discussion in Li et al. (2012) and Chen et al. (2018), (19) shows that the proposed method is able to handle the non-polynomial (NP) dimensionality of order logp = o(n1−2ζ) for some constant 0 ≤ ζ < 1/2.
Equation (20) in Theorem1ensures that the proposed estimator (17) has the ranking consistency property, similar to that discussed by Cui et al. (2015) and Hao et al.
(2019).
Theorem 2 Suppose that Conditions (C3)–(C4) and the assumptions of Lemmas1 and2in Appendix A hold. Let
I=
j :ωj≥cn−ζ forj =1,· · ·, p
(21) forcandζ described in Condition (C4). Then for a sufficiently largen,Ihas the sure screening property:
P I⊆I
≥1−O
qexp
−Dn1−2ζ
, whereDandζ are the constants described in Theorem1.
The sure screening property in Theorem2 shows that with a large probability, the true active set is included in the estimated active set. This property is important which is commonly required for any sensible screening procedure (e.g., Fan & Lv 2008; Li et al.2012; Chen et al.2018).
While (21) allows us to establish the sure screening property of the procedure, it does not tell us exactly about the choice of a suitable threshold value becausecand ζ are unknown. In the actual implementation, we often rank the covariates by the values of theωj for j = 1,· · ·, pand then retain, say,q covariates with the first q largestωj. A common choice ofqisq =*
n logn
+
, where·stands for the floor function (e.g., Li et al.2012; Cui et al.2015; Yan et al.2017; Chen et al.2018; Chen 2019).