Journal of Econometrics 198 (2017) 146–164

Contents lists available at ScienceDirect

Journal of Econometrics
journal homepage: [Link]/locate/jeconom

Measurement errors in quantile regression models✩

Sergio Firpo a , Antonio F. Galvao b, *, Suyong Song c
a
Insper, Rua Quata 300, Sao Paulo, SP 04546-042, Brazil
b
Department of Economics, University of Iowa, W284 Pappajohn Business Building, 21 E. Market Street, Iowa City, IA 52242, United States
c
Department of Economics, University of Iowa, W360 Pappajohn Business Building, 21 E. Market Street, Iowa City, IA 52242, United States

article info a b s t r a c t
Article history: This paper develops estimation and inference for quantile regression models with measurement errors.
Received 1 July 2015 We propose an easily-implementable semiparametric two-step estimator when repeated measures
Received in revised form 24 September for the covariates are available. Building on recent theory on Z -estimation with infinite-dimensional
2016
parameters, consistency and asymptotic normality of the proposed estimator are established. We also
Accepted 6 February 2017
develop statistical inference procedures and show the validity of a bootstrap approach to implement
Available online 21 February 2017
the methods in practice. Monte Carlo simulations assess the finite-sample performance of the proposed
JEL classification: methods. We apply the methods to the investment equation model using a firm-level data with repeated
C14 measures of investment demand, Tobin’s q. We document strong heterogeneity in the sensitivity of
C23 investment to Tobin’s q and cash flow across the conditional distribution of investment. The cash flow
G31 sensitivity is relatively larger at the lower part of the distribution, providing evidence that these firms are
more exposed to and dependent on fluctuations in internal finance.
Keywords:
© 2017 Elsevier B.V. All rights reserved.
Quantile regression
Measurement errors
Investment equation

1. Introduction particularly if information is collected through one-time retrospec-

tive surveys, which are notoriously susceptible to recall errors.
Quantile regression (QR) models have provided a valuable tool Recently, the topic of ME in variables has received considerable at-
in economics and statistics as a way of capturing heterogeneous tention in the QR literature. The standard QR estimator suffers from
effects of covariates on the outcome of interest, exposing a wide bias in the presence of ME (see, e.g., He and Liang (2000)). Chesher
variety of forms of conditional heterogeneity under weak distri- (2001) studies the impact of covariate ME on quantile functions
butional assumptions. Under some assumptions on the unobserv- using a small variance approximation argument. Schennach (2008)
able factors, QR can also be interpreted as providing a structural investigates identification of a nonparametric quantile function
relationship between the outcome of interest and its observable when there is an instrumental variable measured on all sampling
and unobservable determinants. Also importantly, QR provides a units. Wei and Carroll (2009) develop a QR method that corrects
framework for robust inference. the ME bias by constructing joint estimating equations that simul-
Measurement errors (ME) have important implications for the taneously hold for all quantile levels. The method makes use of
reliability of general standard estimation and testing. Variables an iterative algorithm and requires parametric pre-specification
used in empirical analysis are frequently measured with error, of a conditional density to obtain a consistent estimator for the
parameters of interest. Hu and Schennach (2008) establish that
✩ The authors would like to express their appreciation to Stephane Bonhomme, the availability of instruments enables the identification of non-
Tim Conley, Silvia Goncalves, Stefan Hoderlein, Roger Koenker, Arthur Lewbel, classical errors-in-variables models. The resulting identification
Salvador Navarro, Alex Poirier, Yuya Sasaki, Susanne Schennach, Liang Wang, Zhijie induces an estimating equation that encompasses QR methods.
Xiao, and seminar participants at Boston College, Syracuse University, University of
Schennach (2014) studies a model defined by moment conditions
Arizona, University of Western Ontario, 2015 Midwest Econometrics Group, 2015
CMStatistics, 2016 Latin American Workshop in Econometrics of the Economet- with unobservables, which covers QR with ME. The parameters
ric Society, 2016 Meeting of the Brazilian Econometric Society, and 2016 North of interest are obtained by averaging the moment functions over
America Summer Meeting of the Econometric Society for helpful comments and the unobservables using a least favorable distribution obtained
discussions. Vitor Possebom helped us with great research assistance. We also through an entropy maximization procedure.1
would like to thank Jianqing Fan, an associate editor, and two anonymous referees
for their constructive comments and suggestions. All the remaining errors are ours.
1 Recently, Torres-Saavedra (2013) and Hausman et al. (2014) study ME in the
*Corresponding author.
E-mail addresses: firpo@[Link] (S. Firpo), antonio-galvao@[Link] dependent variable of QR models. We refer to Ma and Yin (2011), Wang et al. (2012),
(A.F. Galvao), suyong-song@[Link] (S. Song). and Wu et al. (2015) for other recent developments in QR models with ME.

[Link]
0304-4076/© 2017 Elsevier B.V. All rights reserved.
 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164 147

Thus, in the analysis of QR with mismeasured covariates, it has Monte Carlo simulations assess the finite-sample properties
been common to employ estimation methods that either impose of the proposed methods. We evaluate the estimator in terms of
parametric restrictions on nuisance functionals or use exogenous empirical bias, standard deviation, and mean squared error, and
information as those provided by instrumental variables (see, compare its performance with methods that are not designed for
e.g., Wei and Carroll (2009), Schennach (2008), and Chernozhukov dealing with ME issues. The experiments suggest that the proposed
and Hansen (2006)). Nevertheless, methods relying on parametric approach performs relatively well in finite samples and effectively
assumptions are very sensitive to misspecification of such condi- removes the bias induced by ME.
tions in practical inference. In addition, finding exogenous instru- Our procedure should be useful for those empirical settings
mental variables is known to be a nontrivial task in most empirical based on QR models in which ME in the independent variables
applications. is a concern. To motivate and illustrate the applicability of the
This paper contributes to both the QR and ME branches of the methods, we apply the developed methods to Fazzari et al. (1988)
literature by developing estimation and inference methods for QR investment equation model, where a firm’s investment is regressed
models in the presence of ME in the covariates. This is achieved by on a proxy for investment demand (Tobin’s q) and cash flows.
exploring repeated measures of the true regressor.2 The main con- This is a well-known model in the corporate investment litera-
tributions are the following. First, we propose a simple and easily- ture. The QR approach is a useful tool in this example because it
implementable two-step semiparametric estimation procedure for allows us to capture the heterogeneity in the Tobin’s q and cash
QR models that preserves the semiparametric distribution-free flows along the conditional distribution of investment. Concerns
and heteroscedastic features of the model. The first step employs about ME have been emphasized in the context of the empirical
a general nonparametric estimation of the density function. The investment model. Theory suggests that firm’s investment demand
second step uses the estimated densities as weights in a weighted is captured by marginal q, but this quantity is unobservable and
QR estimation. For the first step, we propose a nonparametric esti- researchers use instead its measurable proxy, average q. Because
mator for the conditional density without imposing distributional average q measures marginal q imperfectly, a measurement prob-
assumptions on the ME. Specifically, we show that two mismea- lem naturally arises (see, e.g., Hayashi (1982), Poterba (1988),
sured covariates are sufficient to identify the conditional density Erickson and Whited (2000), and Almeida et al. (2010)). Within
of interest in the presence of ME. In turn, this result guarantees that framework, finding valid and strong instrumental variables to
consistent estimation of the structural parameters of interest. solve the endogeneity problem is not, in general, an easy task.
In the second main contribution we establish the asymp- In the empirical example we use a data set taken from Cummins
totic properties of the two-step estimator, assuming that the et al. (2006), where there are two measures of average q. The first
conditional densities satisfy smoothness conditions and can be measure conforms with prior research and is constructed using the
estimated at an appropriate nonparametric rate. The third con- standard equity prices. The second proxy for the firm’s intrinsic
tribution is to develop practical statistical inference and testing value is based on analysts’ earnings expectations. Thus, our method
procedures for general linear hypotheses based on the Wald statis- is a natural alternative solution to the ME problem where repeated
tic. To implement these tests in practice the critical values are measures on Tobin’s q are available. The results document strong
computed using a bootstrap method. We provide sufficient condi- evidence of substantial heterogeneity in the sensitivity of invest-
tions under which the bootstrap is theoretically valid, and discuss ment to Tobin’s q and cash flow across the conditional distribution
an algorithm for its practical implementation. Our method leads of investment. The empirical results show that larger cash flow
to a simple algorithm that can be conveniently implemented in sensitivity occurs at the lower part of the investment distribution,
empirical applications. showing evidence that these firms are more exposed to and de-
Compared to the existing procedures for QR models with ME, pendent on fluctuations in internal finance. Our empirical findings
our approach has several distinctive advantages. First, the method support the idea that the proposed methods are a useful alternative
employs a nonparametric estimator in the first step and does not to existing approaches in economic applications in which ME is an
assume global linearity at all quantile levels for the estimation important concern.
of the conditional density function. Such feature makes the pro- The rest of the paper is organized as follows. Section 2 presents
cedure applicable to any τ -quantile of interest, thus relaxing the the model and discusses identification of the parameters of inter-
requirement of a joint estimation and providing more flexibility. est in the presence of ME. Section 3 proposes the two-step QR
In contrast, Wei and Carroll (2009) require a parametric pre- estimator. Section 4 establishes the asymptotic properties of the
specification of a conditional density for implementation of their estimator. Inference is discussed in Section 5. Section 6 presents
estimator. Second, our algorithm is computationally simple and the Monte Carlo experiments. In Section 7, we illustrate empirical
easy to implement in practice because estimation of the weights usefulness of the new approach with an application to the invest-
does not require recursive algorithms allowing the weights for all ment equation model. Finally, Section 8 concludes the paper.
observations to be obtained from one single step. As a result, the
quantile estimation in the second step is attained by minimizing 2. Model and identification
only one single convex objective function at the quantile of in-
terest. On the other hand, Wei and Carroll (2009) make use of 2.1. Model
an iterative algorithm and require the estimating equations to be
solved jointly for all quantiles, which increases the dimensionality We first introduce the model studied in this paper. Given a
of the problem substantially. Finally, the estimated weights exhibit quantile τ ∈ (0, 1), we define the following quantile regression
a property of uniform consistency, implying that it is feasible to (QR) model,
establish both the consistency and asymptotic normality of the re-
sulting estimators of the parameters of interest. Hence, the method Yi = Xi β0 (τ ) + Zi⊤ δ0 (τ ) + εi (τ ), (1)
provides standard inference and testing procedures. where Yi is the scalar dependent variable of interest, Xi is a
potentially-mismeasured scalar continuous covariate,3 Zi is a
2 Identification and estimation of conditional average regression models and L-vector of correctly-observed covariates, and εi (τ ) is the innova-
conditional density with repeated measures of the true regressor have been studied tion term whose τ th quantile is zero conditional on (Xi , Zi ). The
in Li and Vuong (1998), Li (2002), Schennach (2004a), Delaigle et al. (2008), and Hu
and Sasaki (2015a) among others. We extend this literature to estimation and
inference for QR models using repeated measures of the true regressor. 3 Below we describe the extension to the multi-dimensional case.
148 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164

structural parameters of interest are θ0 (τ ) = (β0 (τ ), δ0 (τ )⊤ )⊤ . In where fYXZ (y, x, z) and fYZ (y, z) are the joint density of (Y , X , Z ) and
general, each β0 (τ ) and δ0 (τ ) will depend on τ , but we assume τ (Y , Z ), respectively, and where fX |YZ (x | y, z) ≡ f0 is the conditional
to be fixed throughout the paper and suppress such a dependence density of X given (Y , Z ). By replacing the outer expectation with
for notational simplicity. its empirical counterpart, we write the sample analog of the pop-
Suppose (Yi , Xi , Zi ) are i.i.d. random variables defined on a com- ulation objective function (4) as:
plete probability space (Ω , F, P). Denote the support of a random n ∫
variable by supp(·). Define the population objective function for the 1∑
Q̃n (β, δ, f ) := ψτ (Yi − xβ − Zi⊤ δ )[x Zi ] · fX |YZ (x | Yi , Zi )dx
τ th conditional quantile as n x
i=1

Q (β0 , δ0 ) := E ψτ (Yi − Xi β0 − Zi⊤ δ0 )[Xi Zi ] = 0,

[ ]
(2) = 0. (5)

where ψτ (u) := (τ − I {u < 0}) with the indicator function I {·}. The integration in (5) makes the function continuous in its
When the true covariates (X , Z ) are observed, β0 and δ0 in (1) can argument. The summand of (5) is Ex [ψτ (Yi − xβ − Zi⊤ δ )[x Zi ] |
be consistently estimated from the standard QR model with sample Yi , Zi ], the conditional mean of the original score function given the
analog of Q (β, δ ) in (2) as observed Y and Z . Moreover, (5) is an unbiased estimating function,
that is, has mean zero, and will be the basis for constructing esti-
n
1∑ mating equations to obtain consistent estimates of the parameters
Qn (β, δ ) := ψτ (Yi − Xi β − Zi⊤ δ )[Xi Zi ] = 0. (3)
n of interest.
i=1
Therefore, one would solve the new estimating Eq. (5) to esti-
The presence of the indicator function in Eq. (3) implies that the mate the parameters of interest. In empirical applications, how-
solution may not be an exact zero. It is usual to write the estimator ever, the true conditional density fX |YZ (x | y, z) is unknown and to
as a minimization problem, and use linear programming to solve implement the estimator (5) in practice one needs to replace it with
the optimization. Thus, the above moment condition is a slight fX |YZ (x | y, z), a consistent estimate of fX |YZ (x | y, z). Thus, a (feasi-
ˆ
abuse of notation, but since everything else involving observed ble) estimator would first estimate fX |YZ (x | y, z). The fitted density
data is an estimating equation that will have a zero, we will function from this step would be used to estimate the coefficients
use the estimating equation nomenclature. For more details on of interest in a second step. Finally, with a consistent estimate
Z -estimator with non-smooth objective functions, see e.g., He and of the conditional density, (β0 , δ0 ) can be consistently estimated.
Shao (1996, 2000). However, in general, the conditional density is not identified due
to the unobservability of the true X .
2.2. Measurement error bias and its solution In this paper, we make use of the repeated measures, X1 and X2 ,
and show that two mismeasured covariates are sufficient to iden-
Under the assumption of perfectly-measured regressors, the tify the conditional density in the presence of ME on the covariate.
solution of (3) can be shown to produce consistent estimates of We also propose a nonparametric estimator for the conditional
(β0 , δ0 ). Nevertheless, it is commonly observed that researchers density without imposing assumptions on known distributions of
have to use the regressor X measured with error. Using mismea- the ME. In turn, this result guarantees consistent estimation of the
sured X in the standard QR estimation in (3) induces bias in the structural parameters of interest.
estimates of the coefficients of interest (see, e.g., He and Liang, In a related model, Wei and Carroll (2009) make use of an
2000). Thus, estimation of the standard QR model under measure- iterative algorithm to obtain a consistent estimator of the con-
ment errors (ME) leads to inconsistent estimates. To overcome this ditional density fX |YX1 (x | y, x1 ) in the presence of ME on X .4
drawback we propose a methodology that makes use of repeated They focus on model with one measurement of true X (here X1 )
measures. Both variables are mismeasured observables of the true and with no other observed covariates Z for simplicity. Although
covariate. their approach can be useful in some applications, it has important
Suppose that true covariate X is unobservable due to ME. In- technical challenges. First, to implement the estimator, one needs
stead, a researcher observes two error-laden measurements which to estimate the conditional density fX |YX1 (x | y, x1 ) which requires
are noisy measures of X and defined as follows pre-specified parametric form of fX |X1 (x | x1 ). This suffers from po-
X1i = Xi + U1i tentially serious model misspecification. Second, and related to the
first problem, there is a problem to solve the estimating equations,
X2i = Xi + U2i ,
since estimating the conditional density fX |YX1 (x | y, x1 ) involves
where U1i and U2i are ME. Therefore, the observed random vari- estimation of the entire process β0 (τ ) over quantiles τ . In other
ables are (Yi , X1i , X2i , Zi ), and one seeks to estimate the parameters words, the estimating equations in Wei and Carroll (2009) need to
(β0 , δ0 ). be solved jointly for all the τ ’s, which increases the dimensionality
We show how to use information from the measures X1 and X2 of the problem substantially and makes implementation consid-
to obtain consistent estimates of the parameters of interest. For erably difficult. This is reflected in the tractability of inference for
that purpose, it is useful to rewrite Q (β, δ ) as a function of the their method.
density function as well as (β, δ ): In the next section we propose a procedure that yields a consis-
tent estimator of (β0 , δ0 ) in (5). We develop a method for QR with
Q̃ (β0 , δ0 , f0 ) := E[ψτ (Y − X β0 − Z ⊤ δ0 )[X Z ]]
∫ ME, which relies on estimating the conditional density function
= ψτ (y − xβ0 − z ⊤ δ0 )[x z ] · fYXZ (y, x, z)dydxdz nonparametrically. The method is a two-step estimator, where the
∫ first step estimates the density nonparametrically, and the second
step employs a standard weighted QR procedure. Before proceed-
= ψτ (y − xβ0 − z ⊤ δ0 )[x z ] ing to estimation, we show an identification result for the density
· fX |YZ (x | y, z)fYZ (y, z)dydxdz function which is essential in the estimation. For expositional ease,
[∫ ] we use fX |YZ (x | y, z) and f (x | y, z) synonymously.
=E ψτ (Y − xβ0 − Z ⊤ δ0 )[x Z ] · fX |YZ (x | Y , Z )dx
x 4 We note that their conditional density is slightly different from ours since there
= 0, (4) is a mismeasured covariate, X1 , in their conditioning set.
 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164 149

2.3. Conditional density estimating equations are simply adjusted to these corresponding
changes.
As described above, f (x | y, z) is an important element for
the identification of the parameters of interest in the QR with 3. Estimation
ME. This section describes the identification of the conditional
density function f (x | y, z) which is required to compute the two- Given the identification condition in Eq. (6) of Theorem 1, we
step estimator. The identification is based on the assumption that are able to estimate the structural parameters of interest, (β0 , δ0 ).
repeated measures of the true regressor are observed. We state the We propose a semiparametric estimator that involves two-step
following assumptions to obtain the main identification result. estimation. Implementation of the estimator is simple in practice.
Assumption A.I: (i) E[U1 | X , U2 ] = 0; (ii) U2 ⊥ (Y , X , Z ). In the first step, one estimates the nuisance parameter, the condi-
tional density, using a nonparametric method which requires no
Assumption [Link]: (i) E[|X |] < ∞; (ii) E[|U1 |] < ∞; optimization. In the second step, by plugging in these estimates, a
(iii) |E[exp(iζ X2 )]| > 0 for any finite ζ ∈ R. general weighted quantile regression (QR) is performed.
Assumption [Link]: (i) sup(x,y,z)∈supp(X ,Y ,Z ) f (x | y, z) < ∞; Specifically in Eq. (6), we can rewrite Q̃ (β, δ, f ) as:
(ii) f (x | y, z) is integrable on R for each (y, z) ∈ supp(Y , Z ).
[∫ ]
Q̃ (β, δ, f ) = E ψτ Y − xβ − Z ⊤ δ [x Z ] · fX |YZ (x | Y , Z )dx ,
( )
Assumption A.I imposes restrictions on the repeated measures
x
of X . Assumption A.I(i) requires conditional mean zero of ME on
X1 , but allows dependence of the ME and (X , U2 ). Assumption which does not depend on data on X . Thus, estimation of (β0 , δ0 )
A.I(ii) requires that ME on X2 is independent of true X as well follows from solving a feasible version of Q̃n (β, δ, f ):
as other variables. However, it does not necessarily require zero n ∫
1∑
mean of U2 . Thus, our setting on the repeated measures can be Q̃n (β, δ,ˆ fX |YZ (x | Yi , Zi )dx,
ψτ Yi − xβ − Zi⊤ δ [x Zi ] · ˆ
( )
f) =
useful for an example such that there is a drift or trend in the mis- n x
i=1
measured covariates. Assumption [Link] imposes mild restrictions on
where
the existence of the first moments of X and U1 , and nonvanishing ∫
characteristic function of X2 . These have been commonly assumed 1
fX |YZ (x | Yi , Zi ) =
ˆ φ̂ (ζ , Yi , Zi ) exp(−iζ x)dζ ,
in the deconvolution literature (see, e.g., Fan (1991) and Fan and 2π ζ
Truong (1993)). Assumption [Link] is trivially satisfied in commonly-
and the only feature of this sample objective function that has not
used conditional densities.
yet been presented is φ̂ , the estimate of φ , which is defined in
Let φ (ζ , y, z) ≡ E[eiζ X | Y = y, Z = z ] be the conditional
the next section. In practice, as we discuss next, we approximate
characteristic function of X given Y and Z . The following theorem
integrals by sums, and end up with a double sum (on observations
presents the identification of f (x | y, z).
and on grid values of X ). Importantly on that representation is the
Theorem 1. Suppose Assumptions A.I–[Link] hold. Then, for (x, y, z) ∈ fact that the estimates (β̂, δ̂ ) will be obtained by a weighted QR,
supp(X , Y , Z ), whose weights will be given by the estimate ˆ fX |YZ .
∫
1
f (x | y, z) = φ (ζ , y, z) exp(−iζ x)dζ , (6) 3.1. Estimation of nuisance parameter
2π
where for each real ζ , In this subsection we discuss the estimation of the nuisance
ζ parameter in the first step, i.e., the conditional density f (x | y, z).
E[eiζ X2 | Y = y, Z = z ] iE[X1 eiξ X2 ]
(∫ )
φ (ζ , y, z) = exp dξ . We propose a nonparametric method by adapting the class of flat-
E[ eiζ X2 ] 0 E[eiξ X2 ] top kernels of infinite order suggested by Politis and Romano
(1999). Consider the following assumption.

Proof. See Appendix A. ■ Assumption [Link]:∫The real-valued kernel x → k(x) is measurable

and symmetric with k(x)dx = 1, and its Fourier transform ξ →
The theorem implies that conditional density f (x | y, z) can κ (ξ ) is bounded, compactly supported, and equal to one for |ξ | < ξ̄
be written as a function of purely-observed variables. For this, we for some ξ̄ > 0.
use useful properties of Fourier transform. Namely, we write f (x | From Assumption [Link], we allow for a kernel of the form (e.g., Li
y, z) as the inverse Fourier transform of φ (ζ , y, z). This simplifies and Vuong (1998))
identification since φ (ζ , y, z) is easily identified from Assumptions
sin(x)
A.I–[Link] by removing the ME, U1 and U2 , in the frequency do- k(x) = ,
mains (ζ , ξ ). It is worth noting that the identification result is πx
similar to Kotlarski (1967) who identifies the density of X from with its Fourier transform such that
its repeated measurements by assuming mutual independence of ∫
1 (x)
X , U1 , and U2 . Our approach rests on weaker assumptions than κ (hx ζ ) = k exp(iζ x)dx,
hx hx
their mutual independence, which is highlighted in condition A.I.
In addition, the result in Theorem 1 is related to that in Schennach for a bandwidth hx . This flat-top kernel of infinite order has the
(2004b) who identifies the conditional mean of Y given X . Since we property that its Fourier transform is equal to one over the [−1, 1]
are interested in conditional quantile effects we need to identify interval and zero elsewhere, which guarantees that the bias goes
the conditional density of X , which requires a stronger indepen- to zero faster than any power of the bandwidth. We note that the
dence condition (A.I(ii)). ill-posed inverse problem occurs when one tries to invert a convo-
We note that the current identification setup can be extended lution operation. This is true to our proposed estimator because it
to a K -vector of X . In this case, the integral in Eq. (5) becomes a is divided by a quantity which converges to zero as the frequency
multiple integral over the K -vector of X . In addition, the integral parameter goes to infinity by Riemann–Lebesgue lemma. By esti-
in Eq. (6) should be an integral over a K -vector of ζ , while the inte- mating the numerator using the kernel whose Fourier transform
gral in the definition of φ (·) should be over a piecewise smooth path is compactly supported, one can guarantee that the ratio is under
connecting 0 and ζ , where ξ and ζ are K -dimensional vectors. The control. This is because that the numerator can decay to zero
150 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164

before the denominator converges to zero. This compact support Step 2. To compute Eq. (5) in practice, we perform a numerical
of the Fourier transform of the kernel can be easily implemented approximation to the integral over x. We do this via translating the
by preserving most of the properties of the original kernel. For problem into a weighted QR problem. Let x̃ = (x̃1 , x̃2 , . . . , x̃m ) be a
instance, one can transform any given kernel k̃ into a modified fine grid of possible xj values, akin to a set of abscissas in Gaussian
kernel k with compact Fourier support by using a window function quadrature. For each τ , θ̂ (τ ) = (β̂ (τ ), δ̂ (τ )⊤ )⊤ can be computed by
that is constant in the neighborhood of the origin and vanishes solving
beyond a given frequency. n
∑ m
∑
The following theorem summarizes the result. ψτ (Yi − x̃j β − Zi⊤ δ )[x̃j Zi ] · ˆ
f (x̃j | Yi , Zi ; h) = 0, (8)
i=1 j=1
Theorem 2. Suppose Assumptions A.I–[Link] hold, and let k(·) satisfy
Assumption [Link]. For (x, y, z) ∈ supp(X , Y , Z ) and hx > 0, let where ˆ f (x̃j | Yi , Zi ; h) is obtained from Step 1. The weighted QR
∫ ( ) of Yi on x̃j and Zi with corresponding weights ˆ f (x̃j | Yi , Zi ; h) can
1 x̃ − x
f (x | y, z ; hx ) ≡ k f (x̃ | y, z)dx̃. be readily computed using the function called ‘‘rq’’ in R package
hx hx quantreg.
Then we have
∫ 4. Asymptotic properties
1
f (x | y, z ; hx ) = κ (hx ζ )φ (ζ , y, z) exp(−iζ x)dζ .
2π This section investigates the large sample properties of the
proposed two-step estimator. While these methods seem similar
Proof. See Appendix A. ■
to the ones discussed by Wei and Carroll (2009), the nonparametric
(2) (2)
Let hn ≡ (hxn , hn ) with hn≡ (hyn , hzn ) be a set of smoothing estimation of the conditional density function raises some new
∑n issues for the asymptotic analysis of the estimator. First, we es-
parameters. Let Ê[·] denote a sample average, i.e., 1n i=1 [·]. Finally,
tablish the asymptotic results for the estimator of the conditional
we introduce a consistent nonparametric estimator of f (x | y, z)
density function given in (7). Second, we establish consistency and
motivated by Theorem 2.
asymptotic normality of the second step estimator in (8).
Definition 2.3. The estimator of f (x | y, z) is defined as
∫ 4.1. Asymptotic properties of the density estimator
1
f (x | y, z ; hn ) ≡
ˆ κ (hxn ζ )φ̂ (ζ , y, z , h(2)
n ) exp(−iζ x)dζ , (7)
2π In this subsection we establish the asymptotic properties of
the density[ function estimator in Eq. (7). Let µ(ζ ) ≡ E[eiζ X ],
for hn → 0 as n → ∞, where
ω1 (ζ ) ≡ E eiζ X2 , and χ (ζ , y, z) ≡ E [eiζ X2 | Y = y, Z = z ] f (y, z).
]
(∫ ζ
Ê[eiζ X2 | Y = y, Z = z ] iÊ[X1 eiξ X2 ] We denote the kernel density estimator of f (y, z) by ˆ f (y, z) ≡
)
φ̂ (ζ , y, z , h(2)
n ) ≡ exp dξ .
Ê[eiζ X2 ] 0 Ê[eiξ X2 ] Ê[khy (Y − y)khzn (Z − z)]. Let X ≡ R be the support of X , and
n
Y × Z be a compact set contained in the support of (Y , Z ), and
The above estimator is useful to compute the structural param-
Dζ denote a partial derivative with respect to ζ . We impose the
eters of interest. Since it has an explicit closed form, it requires following assumptions.
no optimization routine unlike other likelihood-based approaches. Assumption B.I: (i) There exist constants C1 > 0 and γµ ≥ 0 such
Estimation of conditional mean, Ê[eiζ X2 | Y = y, Z = z ], can that
be achieved via any nonparametric method. For instance, one
⏐ Dζ µ(ζ ) ⏐
⏐ ⏐
n k (·/hn )
1
might use popular kernel estimation with khn (·) ≡ h− |Dζ ln µ(ζ )| = ⏐
⏐ ⏐ ≤ C1 (1 + |ζ |)γµ .
(e.g., Epanechnikov kernel) defined as µ(ζ ) ⏐
Ê[eiζ X2 khy (Y − y)khzn (Z − z)] (ii) There exist constants Cφ > 0, αφ ≤ 0, νφ ≥ 0, and γφ ∈ R such
Ê[eiζ X2 | Y = y, Z = z ] ≡ n
, that νφ γφ ≥ 0 and
Ê[khy (Y − y)khzn (Z − z)]
n
sup |φ (ζ , y, z)| ≤ Cφ (1 + |ζ |)γφ exp(αφ |ζ |νφ ),
where khy (Y − y) is a univariate kernel function and khzn (Z − z) is a (y,z)∈Y ×Z
n
y
L-multivariate kernel function with corresponding bandwidths hn and if αφ = 0, then γφ < −1.
z
and hn , respectively. (iii) There exist constants Cω > 0, αω ≤ 0, νω ≥ νφ ≥ 0, and
γω ∈ R such that νω γω ≥ 0 and
3.2. Estimation of the structural parameters
min{ inf |χ (ζ , y, z)|, |ω1 (ζ )|} ≥ Cω (1 + |ζ |)γω exp(αω |ζ |νω ).
(y,z)∈Y ×Z
This section describes the general estimator for QR models with
ME. We propose a Z -estimator that involves two-step estimation. Assumption [Link]: (i) E[|X1 |2 ] < ∞; (ii) E[|X1 ||X2 |] < ∞;
(iii) E[|X2 |] < ∞.
Given the identification condition in Eq. (5) and the estimator of
(ln n)1/2
(
the density function described in the previous section, we are able f (y, z) − f (y, z)| = Op
Assumption [Link]: sup(y,z)∈Y ×Z |ˆ (nhy hz )1/2
to estimate the structural parameters of interest. We estimate the )
parameters, θ0 = (β0 , δ0⊤ )⊤ , for a selected τ of interest, from the
s 2
∑
+ s∈{y,z } (h ) .
following two steps: Assumptions B.I–[Link] are standard for nonparametric deconvo-
f (xj | Yi , Zi ; h) for each ith observation and jth
Step 1. Estimate ˆ lution estimators, since the rates of convergence depend on the
grid as in Eq. (5) where j ∈ J ≡ {1, 2, . . . , m} with m number tails of the Fourier transforms (see, e.g., Fan (1991) and Fan and
of grids for approximating the numerical integral. The choice of Truong (1993)).
kernels and bandwidths are provided in Definition 2.3. The inte- Assumption B.I concerns the smoothness of the (conditional)
grals in Eq. (7) are performed using the fast Fourier transforms characteristic functions of the true regressor X and the ob-
(FFT) algorithm. Well-behaving performance of the algorithm is served measure X2 . The literature commonly adopts two types
guaranteed by the smoothness of the characteristic function φ (·) of smoothness assumptions: ordinary and supersmoothness. Or-
and the finiteness of the moments. dinary smoothness admits a Fourier transform whose tail decays
 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164 151

to zero at a geometric rate |ζ |γ , γ < 0 whereas supersmooth- We note that the uniform convergence rate for the first stage
ness admits a Fourier transform whose tail decays to zero at an is obtained over the compact set Y × Z contained in the support
exponential rate exp (α|ζ |γ ), α < 0, γ > 0.5 Assumption of (Y , Z ). However, we highlight that this does not necessarily
B.I simultaneously imposes ordinary and supersmoothness condi- imply that the random variables must be compactly supported
tions for notational ease. Assumption B.I(i) imposes a restriction (see, e.g., Schennach (2008)). Alternatively, it is possible to obtain a
on the tail behavior of µ(ζ ), the characteristic function of X . A uniform convergence rate and establish the asymptotic properties
term exp (α1 |ζ |ν1 ) is omitted with merely a small loss of generality over an expanding set (see, e.g., Newey (1994b), Andrews (1995),
since ln µ (ζ ) is indeed a power of ζ for all common ordinary and and Hansen (2008)). We consider uniformity over expanding sets
supersmoothness functions. Assumption B.I(ii) concerns φ (ζ , y, z), that diverge slowly to infinity. Detailed derivation of the uniform
the conditional characteristic function of X given Y = y and Z = z. convergence rate over expanding sets is provided in the Online
It states that the rate of decay of φ (ζ , y, z) is governed by the Supplementary Appendix (see Appendix B), and the results in
conditional density of X given Y = y and Z = z (i.e., f (X | Y = Theorem 3 remain the same.
y, Z = z)), the parameter of interest in the first step. Assumption
B.I(iii) jointly imposes a restriction on χ (ζ , y, z), the conditional 4.2. Asymptotic properties of the two-step estimator
characteristic function of X2 given Y = y and Z = z weighted by
the joint density of (Y , Z ) (i.e., E[eiζ X2 | Y = y, Z = z ]f (y, z)), and In this subsection, we derive the asymptotic properties of the
ω1 (ζ ), the characteristic function of X2 . Since ω1 (ζ ) = E[eiζ X2 ] = second step estimator of parameters of interest in (8). We establish
E[eiζ X ]E[eiζ U2 ], the smoothness of ω1 (ζ ) is determined by the com- consistency and asymptotic normality for a given quantile τ ∈
bination of X and U2 . As commonly imposed in deconvolution-type (0, 1).
estimators, χ (ζ , y, z) and ω1 (ζ ) need to be bounded below because
4.2.1. Consistency
they appear in the linearization of the estimator. The restriction
Consistency is a desirable property for most estimators. We
on χ (ζ , y, z) typically requires the joint density of (Y , X2 , Z ) to be
wish to establish consistency of the estimator θ̂ = (β̂, δ̂ ⊤ )⊤
bounded away from zero on its support. This is generally satisfied
defined in Eq. (8), where ˆf , given in (7), is an estimator of f0 = f (x |
by common continuous distributions supported on the real line.
y, z).
Assumption [Link] imposes mild moment restrictions required
First, notice that from the estimating equation in (5) we have
for consistency. Assumption [Link] imposes a standard condition on
nonparametric estimator of the joint density of f (y, z).
n ∫
1∑
The next result establishes the asymptotic properties of the Q̃n (β, δ, f ) = ψ (Yi − xβ − Zi⊤ δ )(x Zi ) · f (x | Yi , Zi ) dx,
n
density function estimator. i=1

and its expectation is

Theorem 3. Let Assumptions A.I–[Link] and B.I–[Link] hold. Then for ∫
Q̃ (β, δ, f ) = E ψ (Yi − xβ − Zi⊤ δ )(x Zi ) · f (x | Yi , Zi ) dx.
y
(x, y, z) ∈ X × Y × Z and h > 0 satisfying max{(hn )−1 , (hzn )−1 } =
η
O (n ) and
The estimator θ̂ = (β̂, δ̂ ⊤ )⊤ is obtained by equating Q̃n (β, δ,ˆ f ) to
(hxn )−1 = O (ln n)1/νω −η νω ̸= 0,
( )
if
zero. Note that Q̃ (β, δ, f0 ) = 0 if and only if (β, δ ⊤ )⊤ = (β0 , δ0⊤ )⊤
(hxn )−1 = O n(1−20η)/2(γµ −γω ) νω = 0, ∈ Θ.
( )
if
Now we state the following sufficient conditions for the estima-
for some η > 0, we have
tor to be consistent.
f (x | y, z ; h) − f (x | y, z)|
[∫ C.I: For any δ > 0, there exists] ϵδ > 0 such that
sup |ˆ Assumption
(x,y,z)∈X ×Y ×Z
inf∥θ−θ0 ∥>δ ∥E x ψτ (Y − xβ − Z ⊤ δ )[x Z ] · f0 dx ∥ > ϵδ .
−1 γB −1 νB
= O (hx ) exp αB (hx )
(( ) ( ( ) ))
( −1/2 x −1 γ )νL )) Assumption [Link]: (i) E[|X |2+ε ] < ∞; (ii) E[∥Z ∥2+ε ] < ∞.
((h ) ) L exp αL (hx )−1 ,
( (
+ Op n
Assumption [Link]: E[|X | | Y , Z ] < ∞.
with αB ≡ αφ ξ̄ νφ , νB ≡ νφ , γB ≡ γφ + 1, αL ≡ αφ 1{νφ =νω } − αω , Assumption C.I is a standard identification condition in the QR
νL ≡ νω , γL ≡ 1 + γφ − γω , and δL ≡ γµ + γφ − γω + 2. literature (see, e.g., Chen et al. (2003) and Kato et al. (2012)). [Link]
is also standard in QR, and requires the second moment of latent
Proof. See Appendix A. ■ variable and well-measured regressor to be finite, see e.g., Angrist
et al. (2006) and Koenker (2005). [Link] requires the first conditional
The theorem above establishes consistency and the uniform moment of the latent variable to be finite.
convergence rate of the proposed estimator. The conditions on the The following theorem derives consistency of the proposed
bandwidths are imposed to guarantee that asymptotic behavior of estimator, θ̂ = (β̂, δ̂ ⊤ )⊤ .
the linear approximation of the expression ˆ f (x | y, z ; h) − f (x |
y, z) is essentially determined by a variance term since a nonlinear Theorem 4. Suppose that ]θ0 is the unique solution of E x ψτ (Y
[∫
remainder term is asymptotically negligible. The result also shows − xβ0 − Z ⊤ δ0 )[x Z ] · f0 dx = 0, and assumptions C.I–[Link], and
that convergence rate depends on the tail behaviors of the asso- conditions of Theorem 3 are satisfied. Then, as n → ∞
ciated quantities. For instance, consider the case that φ (ζ , y, z) is p
ordinarily smooth. When χ (ζ , y, z) and ω1 (ζ ) in Assumption B.I θ̂ → θ0 .
is also ordinarily smooth (i.e., νω = 0), one can choose smaller
bandwidth so that resulting convergence rate of the estimator is Proof. See Appendix A. ■
faster than when one of them is supersmooth. A uniform law of large numbers for the first step estimator,ˆf (x |
y, z), is a standard requirement in two-step estimation literature to
5 The typical examples of ordinarily smooth functions are uniform, gamma, establish consistency of the second step estimator; see, e.g., Newey
symmetric gamma, Laplace (or double exponential), and their mixtures. Normal, and McFadden (1994). We note that, for the proposed estimator,
Cauchy, and their mixtures are supersmooth functions. this requirement is satisfied by Theorem 3.
152 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164

] θ0 is the unique solution of E x ψτ (Y

[∫
4.2.2. Asymptotic normality Theorem 5. Suppose that
Now we derive the limiting distribution of the two-step estima- − xβ0 − Z ⊤ δ0 )[x Z ] · f0 dx = 0, and assumptions C.I–[Link], G.I–[Link],
tor in (8). Let ∥ · ∥∞ be the supremum norm over the argument. and conditions of Theorem 3 are satisfied. Then, as n → ∞
We impose the following assumptions to establish asymptotic √ d
normality. n(θ̂ − θ0 ) → N(0, Λ)

Assumption G.I: The distribution function GY (y | X = x, Z = z) for some positive definite matrix Λ = Γ1−1 V Γ1−1 .
is absolutely continuous, with continuous densities gY (y | X = x, Z =
z) uniformly bounded away from 0 and ∞. Proof. See Appendix A. ■

Assumption [Link]: Let Γ1 := E[gY (X β0 + Z ⊤ δ0 | X , Z )(X , Z ) It is worth noting that the density estimator in Eq. (7) requires
(X , Z ⊤ )⊤ ] be positive definite and Vn := v ar [Qn (θ0 )]. There exists a multivariate kernel functions when X and Z are multi-dimensional.
nonnegative definite matrix V such that Vn → V as n → ∞. This generates slower rate of convergence of the density estimator.
As in the conventional kernel density estimation, it would be diffi-
f − f0 ∥∞ = op (n−1/4 ).
Assumption [Link]: ∥ˆ cult to resolve the problem associated with large dimensionality of
Assumption [Link]: The function f (x | y, z) ∈ F is a uni- X . On the other hand, one could use semiparametric single index
formly smooth function with the uniform smoothness order δ > models to mitigate the problem from large dimensionality of Z . The
convergence rate in the second stage would not be affected by the
dim(x, y, z)/2, and Lipschitz.
multi-dimensionality of X and Z as long as Assumption [Link] still
Conditions G.I and [Link] are standard in the QR literature; see,
holds.
e.g., Koenker (2005).
Condition [Link] imposes that the estimator of the nuisance pa-
5. Inference
rameter converges at a rate faster than n−1/4 . A similar condition
appears in condition (2.4) in Theorem 2 of Chen et al. (2003). We In this section, we turn our attention to inference in the quantile
note that Assumption [Link] is verifiable for particular examples regression (QR) with measurement errors (ME) model. Important
through Theorem 3. As shown in Theorem 3, the convergence rate questions posed in the econometric and statistical literatures con-
is controlled by the smoothness of quantities such as φ (ζ , y, z), cern the nature of the impact of a policy intervention or treatment
χ (ζ , y, z), and ω1 (ζ ). Thus, the rate of convergence depends on on the outcome distributions of interest; for example, whether a
the possible combinations of smoothness of these quantities. For policy exerts a significant effect, a constant versus heterogeneous
instance, if φ (ζ , y, z) is ordinarily smooth and if χ (ζ , y, z) and effect, or a non-decreasing effect. It is possible to formulate a wide
ω1 (ζ ) are supersmooth, a convergence rate of the form (ln n)−υ variety of tests using variants of the proposed method, from simple
for some υ > 0 is achieved. This case illustrates a very slow rate tests on a single QR coefficient to joint tests involving many co-
of convergence. On the other hand, a faster convergence rate, n−υ variates simultaneously. We suggest a bootstrap-based inference
for some υ > 0, which satisfies Assumption [Link], can be achieved procedure to test general linear hypotheses.
when φ (ζ , y, z) is also supersmooth. In addition, if all three quan-
tities, φ (ζ , y, z), χ (ζ , y, z), and ω1 (ζ ), are ordinarily smooth, the 5.1. Test statistic
slow convergence problem is easily avoided. To illustrate this in
a specific example, consider the case that φ (ζ , y, z) and ω1 (ζ ) are General hypotheses on the vector θ (τ ) = (β (τ ), δ (τ )⊤ )⊤ can
supersmooth and both follow a normal distribution with variance be accommodated by standard tests. The proposed statistic and
σ 2 . Then, νφ = νω ̸= 0 in Assumption B.I. From Theorem 3, one can the associated limiting theory provide a natural foundation for the
αB
+η hypothesis Rθ (τ ) = r. Consider the following null hypothesis for a
show that the convergence rate is n 2(αL −αB ) for some small η > 0. given τ of interest
Because the normal distribution has the )characteristic function
whose tail is of the form exp −(σ 2 /2)|ζ |2 , we have αφ = αω = H0 : Rθ (τ ) − r = 0,
(
−σ 2 /2 and νφ = νω = 2. As a result, the convergence rate becomes where R is a full-rank matrix imposing q number of restrictions on
1 1
n− 2 +η which is fast enough to achieve n− 4 in Assumption [Link].6 the parameters, and r is a column vector of q elements.
Therefore, the required rate of convergence in [Link] is attainable The following are examples of hypotheses that may be consid-
under proper combination of smoothness conditions (see, e.g., Fan ered in this framework.
(1991), Fan and Truong (1993), and Schennach (2004b)).
Assumption [Link] is a smoothness condition on the conditional Example 1 (No Effect of the Mismeasured Variable). For a given τ , if
density function. A similar assumption appears in Newey (1994a) there is no mismeasured variable effect in the model, then under
and Chernozhukov and Hansen (2006). Condition [Link] allows for a H0 : β (τ ) = 0, θ (τ ) = (β (τ ), δ (τ )⊤ )⊤ , R = [1 0] and r = 0.
wide variety of nonparametric estimators, including the estimator
described in Section 4.1. The role of [Link] is to allow for an estimated Example 2 (Location Shifts). The hypotheses of location shifts for
density, which is the weight in the estimating equation in (8). β (τ ) and δ (τ ) can be accommodated in the model. For the first case,
This condition together with Theorem 3 ensures that the weight H0 : β (τ ) = β , so θ (τ ) = (β (τ ), δ (τ )⊤ )⊤ , R = [1 0] and r = β . For
is asymptotically well behaved to obtain the limiting distribution the latter case, H0 : δ (τ ) = δ , so that R = [0 1] and r = δ .
of the estimator of the structural parameters. Practical implementation of testing procedures for the null
Asymptotic normality of the estimator, θ̂ = (β̂, δ̂ ⊤ )⊤ , is estab- hypothesis can be carried out based on the following test statistic
lished in the following result.
Wn (τ ) = Rθ̂ (τ ) − r . (9)
6 Similarly, the convergence rate satisfying Assumption [Link] can be proved for From Theorem 5, at given τ , and under the null hypothesis, it
√ d
the case of an ordinary smooth distribution. For instance, consider the Laplace follows n(Rθ̂ (τ ) − r) → N(0, RΛR⊤ ). If we are interested in test-
distribution with mean µ and variance σ 2 . The tail of the characteristic function
of a Laplace distribution is of the form |ζ |−2 . To establish the results, first, the
ing H0 , a Chi-square test could be conducted based on the statistic
convergence rate can be pinned down, with νω = 0. Second, the condition can be in Eq. (9). However, to carry out practical inference procedures,
derived by plugging in γφ = γω = −2. even for a fixed quantile of interest, and construct a Wald statistic,
 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164 153

one would need to first estimate Λ consistently, and consequently where ε ∼ N(0, 0.25), and β1 and β2 are the parameters of
nuisance parameters which depend on both the unknown θ0 and interest.7 We set them as (β1 , β2 ) = (0.5, −0.5). The true variable
f0 in a complicated way. The estimation of Λ is potentially difficult X is not observed by the researcher, and we use additive forms
because it contains additional terms from the effect of θ on the of measurement errors (ME) to generate the mismeasured X as
objective function indirectly through f0 . An alternative method follows:
is to use the statistic Wn directly and the bootstrap to compute
X1i = Xi + U1i ,
critical values and also form confidence regions. Therefore, to make
practical inference we suggest the use of bootstrap techniques to X2i = Xi + U2i ,
approximate the limiting distribution. where we generate X ∼ N(0, 1), and we use a Laplace distribu-
tion density as L(0, 0.25) to generate both measurement errors,
5.2. Implementation of testing procedures U1 and U2 . We compute and report results for the proposed QR
estimator. For comparison, we compute the density ˆ fX |Y using
Practical implementation of the proposed tests is simple. To test different procedures. First, we construct our proposed estimator
H0 , one needs to compute the test statistics Wn (τ ) for a given τ of to control for ME, using the variables (Y , X1 , X2 ), where the density
interest. The steps for implementing the tests are as follows: is estimated by the Fourier Estimator. Second, we use the variables
First, the estimates of θ (τ ) are computed by solving the problem (Y , X ) to construct an ‘‘infeasible’’ kernel estimator of fX |Y in the
in Eq. (8). Second, Wn (τ ) is calculated by centralizing θ̂ (τ ) at r. first step. Finally, the variables (Y , X1 ) are used for ‘‘naive’’ kernel
Third, after obtaining the test statistic, it is necessary to compute estimator of fX |Y which still suffers from ME. For all estimators, we
the critical values. We propose the following scheme. Take B as a consider fourth-order Gaussian kernel. We approximate the inner
large integer. For each b = 1, . . . , B: summation in Eq. (8) using Gauss–Hermite quadrature which is
useful for the indefinite integral. We perform 1000 simulations
(i) Obtain the resampled data {(Yib , X1ib , X2ib , Zib ), i = 1, . . . , n}. with n = 500 and n = 1000. We scan a set of bandwidths for X
(ii) Estimate θ̂ b (τ ) and set Wnb (τ ) := R(θ̂ b (τ ) − θ̂ (τ )). and Y in order to find empirical optimal bandwidths in terms of
(iii) Go back to step (i) and repeat the procedure B times. minimizing mean squared error.

Let ĉ1B−α denote the empirical (1 − α )-quantile of the simulated 6.2. Monte Carlo results
sample {Wn1 , . . . , WnB }, where α ∈ (0, 1) is the nominal size. We
reject the null hypothesis if Wn is larger than ĉ1B−α . Confidence We report results for the following statistics of the coefficient
intervals for the parameters of interest can be easily constructed β2 : bias (B), standard deviation (SD), and mean squared error
by inverting the tests described above. (MSE). First of all, in order to illustrate the problem of ME in prac-
We provide a formal justification of the simulation method. tice, we consider a model estimation where the researcher ignores
Let ˆf ∗ be the bootstrap version of the estimator ˆ f . Consider the the ME problem and performs a parametric median regression of
following condition. Y on X1 without correcting for the ME in X . This simple regression
Assumption [Link]: With P ∗ -probability tending to one, ∥ˆ
f∗ − provides the bias of 0.1686, the standard error 0.02655 and the MSE
f ∥∞ = op∗ (n−1/4 ).
ˆ of 0.02586. These results highlight the importance of correcting for
Condition [Link] is a condition on the density estimation with the ME problem.
the bootstrapped sample and could be verified under the same Now we discuss and present the results for the nonparametric
assumptions implying condition [Link]. estimators with(out) correction of ME. Tables 1–3 report finite-
sample performance of three different two-step estimators at the
Lemma 1. Under Assumptions C.I–[Link], G.I–[Link], [Link] with ‘‘in median: (i) our proposed estimator (Fourier estimator); (ii) infea-
probability’’ replaced by ‘‘almost surely’’, and conditions of Theorem 3, sible kernel estimator; (iii) naive kernel estimator. These results
√ √ d are for n = 500, but the results for n = 1000 are similar. At the
the bootstrap estimator of the θ0 is n-consistent and n(θ̂ ∗ − θ̂ ) →
bottom of each table, B, SD, and MSE from optimal bandwidth are
N(0, Λ) in P ∗ -probability.
reported. In Table 4 we vary the quantiles and present results for
the different estimators across different deciles with n = 1000.
Proof. See Appendix A. ■ Table 1 shows that the proposed estimator is effective in re-
Lemma 1 establishes the consistency of the bootstrap proce- ducing the bias when true X is measured with errors and repeated
dure. It is important to highlight the connection between this measures of the mismeasured covariate are available. These results
result and the previous section. Lemma 1 shows that the limiting are comparable to the infeasible kernel estimator in Table 2. On
distribution of the bootstrap estimator is the same as that of Theo- the other hand, the results in Table 3 from the naive kernel esti-
rem 5, and hence the above resample scheme is able to mimic the mator ignoring ME in X show much larger bias over all selected
asymptotic distribution of interest. Thus, computation of critical bandwidths. Therefore, our estimator outperforms the naive kernel
values and practical inference are feasible. estimator in terms of both bias and MSE. The minimum MSE for
our proposed method is 0.00674 while the minimum MSE from
the naive kernel estimator is 0.01008. This result confirms that
6. Monte Carlo simulations
the methods proposed in this paper are beneficial in finite samples
when repeated measures of the mismeasured regressor are avail-
6.1. Monte Carlo design able to the researcher.
Table 4 reports finite-sample performance of three estimators
In this section, we describe the design of a small simulation over various quantiles with n = 1000. For simplicity, we use the
experiment that has been conducted to assess the finite-sample optimal bandwidths obtained from the simulation results above.
performance of the proposed two-step estimator discussed in the The results confirm that our proposed estimator performs well
previous sections. We consider the following model as a data over different level of quantiles.
generating process:

Yi = β1 + β2 Xi + εi , 7 For simplicity, the perfectly-observed covariate Z is absent here.

154 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164

Table 1 Table 4
Fourier estimator. Simulation results over various quantiles.
hx \ hy 0.2 0.3 0.4 0.5 0.6 τ \ estimator Fourier Infeasible Naive
B 0.09592 0.07751 0.09131 0.06610 0.09373 B 0.07068 0.00510 0.09976
1.5 SD 0.09765 0.08403 0.11718 0.06085 0.08298 τ = 0.2 SD 0.06196 0.02002 0.02746
MSE 0.01874 0.01307 0.02207 0.00807 0.01567 MSE 0.00883 0.00043 0.01070
B 0.09770 0.07553 0.08439 0.08457 0.09696 B 0.06372 0.00515 0.09785
1.6 SD 0.09206 0.07425 0.08380 0.07819 0.07446 τ = 0.3 SD 0.05685 0.01775 0.02568
MSE 0.01802 0.01122 0.01414 0.01327 0.01495 MSE 0.00729 0.00034 0.01023
B 0.10225 0.06963 0.07425 0.07394 0.09968 B 0.06014 0.00480 0.09737
1.7 SD 0.10788 0.04347 0.09206 0.08228 0.08720 τ = 0.4 SD 0.05530 0.01734 0.02507
MSE 0.02209 0.00674 0.01399 0.01224 0.01754 MSE 0.00667 0.00032 0.01011
B 0.09728 0.07464 0.07571 0.08745 0.10426 B 0.05943 0.00457 0.09778
1.8 SD 0.08762 0.05453 0.06794 0.07288 0.07359 τ = 0.5 SD 0.05487 0.01770 0.02383
MSE 0.01714 0.00855 0.01035 0.01296 0.01629 MSE 0.00654 0.00033 0.01013
B 0.08473 0.09663 0.08869 0.10269 0.10458 B 0.06029 0.00479 0.09818
1.9 SD 0.02795 0.09039 0.08893 0.11265 0.08798 τ = 0.6 SD 0.05546 0.01792 0.02391
MSE 0.00796 0.01751 0.01577 0.02324 0.01868 MSE 0.00671 0.00034 0.01021
Optimal hx hy B SD MSE B 0.06326 0.00542 0.09915
1.7 0.3 0.06963 0.04347 0.00674 τ = 0.7 SD 0.05681 0.02003 0.02422
MSE 0.00723 0.00043 0.01042
B 0.07213 0.00458 0.10130
Table 2
τ = 0.8 SD 0.05947 0.02265 0.02564
Infeasible kernel estimator.
MSE 0.00874 0.00053 0.01092
hx \ hy 0.1 0.2 0.3 0.4 0.5
B 0.00706 0.01670 0.01124 0.01097 0.03331
0.1 SD 0.04271 0.04013 0.04270 0.04468 0.04288
MSE 0.00187 0.00189 0.00195 0.00212 0.00295
relationship by introducing other factors that influence the firm’s
B 0.01065 0.00200 0.00537 0.01181 0.03037 optimal investment level. More specifically, financial constraints
0.2 SD 0.03394 0.03158 0.03454 0.02594 0.02880 create a wedge between internal and external funding that invali-
MSE 0.00127 0.00100 0.00122 0.00081 0.00175 dates theoretical arguments in the spirit of Modigliani and Miller
B 0.00513 0.00737 0.00777 0.01121 0.01890 (1958) capital structure irrelevance proposition. In this scenario
0.3 SD 0.02867 0.02620 0.02737 0.02755 0.02652
MSE 0.00085 0.00074 0.00081 0.00088 0.00106
firm’s cash flows reflect the presence of financial constraints and
B 0.01320 0.00935 0.01242 0.01729 0.02802 may contain information relevant for explaining the differences in
0.4 SD 0.02411 0.02800 0.02527 0.02289 0.02464 investment demand across firms. Following Fazzari et al. (1988),
MSE 0.00076 0.00087 0.00079 0.00082 0.00139 investment–cash flow sensitivities became a standard metric in the
B 0.02175 0.01778 0.01631 0.02742 0.03347
literature that examines the impact of financing imperfections on
0.5 SD 0.02270 0.02195 0.02541 0.02582 0.02875
MSE 0.00099 0.00080 0.00091 0.00142 0.00195 corporate investment (Stein, 2003).
Optimal hx hy B SD MSE Empirical models proposed to assess the sensitivity of invest-
0.3 0.2 0.00737 0.02620 0.00074 ment demand to firm characteristics are usually fraught with the
presence of measurement error (see e.g., Erickson and Whited
Table 3 (2000), and Almeida et al. (2010)). A typical example is the use of
Naive kernel estimator. the average Tobin’s q for describing the investment–capital ratio.
hx \ hy 0.1 0.2 0.3 0.4 0.5 Theory suggests that firm’s investment demand is captured by
B 0.11254 0.11132 0.10409 0.10296 0.13249 marginal q, but this quantity is unobservable and researchers use
0.1 SD 0.04776 0.04581 0.04199 0.04449 0.04084 instead its measurable proxy, average q. Since the two variables are
MSE 0.01495 0.01449 0.01260 0.01258 0.01922 not the same, a measurement problem naturally arises (Hayashi,
B 0.09695 0.09553 0.10231 0.10989 0.12630
1982; Poterba, 1988). The introduction of error when measuring
0.2 SD 0.03399 0.03093 0.03381 0.02886 0.03153
MSE 0.01055 0.01008 0.01161 0.01291 0.01695 these variables causes bias in least squares estimators. This bias
B 0.10133 0.09800 0.10012 0.10388 0.11663 can lead to erroneous interpretations of the effect of firm charac-
0.3 SD 0.02957 0.02854 0.02820 0.02557 0.02600 teristics on investment demand. Thus, Poterba (1988) introduces
MSE 0.01114 0.01042 0.01082 0.01145 0.01428 the idea that errors in measuring Tobin’s q may be responsible for
B 0.10243 0.09939 0.10476 0.10609 0.11884
the observed investment–cash flow sensitivities. If cash flow were
0.4 SD 0.02715 0.03008 0.02441 0.02245 0.02630
MSE 0.01123 0.01078 0.01157 0.01176 0.01481 correlated with investment opportunities not well measured by
B 0.10757 0.10313 0.10299 0.11340 0.11782 the proxy for the marginal Tobin’s q, investment–cash flow sensi-
0.5 SD 0.02196 0.02276 0.02703 0.02525 0.02700 tivities could arise. This argument minimizes the role of financing
MSE 0.01205 0.01115 0.01134 0.01350 0.01461 constraints in determining the relationship between firms’ cash
Optimal hx hy B SD MSE
flows and investment.
0.2 0.2 0.09553 0.03093 0.01008
A number of studies intend to control for the measurement
error (ME) in Tobin’s q, while analyzing the relationship between
investment and cash flow. A common approach has been to use
7. Empirical application
the standard instrumental variables together with ordinary least
squares (OLS) and generalized method of moments (GMM) esti-
This section illustrates the usefulness of the proposed methods mators to correct the ME problem (see, e.g., Almeida et al. (2010)
with an empirical example. We study Fazzari et al. (1988) invest- and Lewellen and Lewellen (2016)). It has been common to use lags
ment equation model, where a firm’s investment is regressed on of the observed Tobin’s q as instruments, by assuming that they
a proxy for investment demand (Tobin’s q) and cash flows. Theory are uncorrelated with the error term in the regression equation.
suggests that a correct measure for firm’s investment demand is Almeida et al. (2010) use lagged Tobin’s q as instruments to resolve
marginal Tobin’s q. This measure stems from the relationship that the ME problem. They estimate standard conditional average mod-
equates firm’s marginal benefit with marginal cost in equilibrium. els and the results show the importance of both Tobin’s q and cash
Nevertheless, the presence of financial constraints may distort this flow in investment equation models.
 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164 155

Table 5
Summary statistics (standardized data).
Min. First quartile Median Mean Third quartile Max.
IK −1.7590 −0.7140 −0.1595 0.0000 0.4972 6.0110
qe −1.1130 −0.6819 −0.3478 0.0000 0.3019 7.9500
q̂ −1.2590 −0.6898 −0.3018 0.0000 0.3591 9.2710
CFK −7.6380 −0.7284 −0.1864 0.0000 0.4988 5.4140
Number of observations 11,431

A related method in the literature is to use different proxies to (IV-OLS). The QR estimation strategy follows Koenker and Bassett
the marginal Tobin’s q as alternative instruments. Cummins et al. (1978), and does not correct for ME. The IV-QR estimates use the
(2006) estimate the investment models using two sources of firm- method developed by Chernozhukov and Hansen (2006). For the
level data. They construct an analyst-based measure of average q IV-QR and IV-OLS, we use the variable q̂ as an instrument for qe . The
as well as a market-based measure of average q, and compare esti- IV strategy for QR is based on the assumption that the q̂ is strongly
mation results based on each measure to assess the robustness of related to the qe (i.e., IV) but the IV is independent of unobservable
the findings in the literature. Estimating conditional mean models, factors of investment as well as ME. We conjecture that the IV
they find no evidence that the cash flow is a statistically significant approach delivers different estimates than our proposed ME esti-
determinant of investment in US companies. mator since both procedures rely on different set of conditions. Our
More recently, Agca and Mozumdar (2017) propose using the method is particularly useful for the data set where it is unlikely
two measures of q contemplated in Cummins et al. (2006) as that the IV is independent of the regression error which contains
instruments for controlling for ME in marginal Tobin’s q. Their ME on q̂, since the IV is also potentially mismeasured.9
approach consists of using the lags of one of the variables as in- To identify the effects of Tobins’ q and cash flow we work with
struments for the other in the linear error-in-variables model. This the conditions given in Section 2. Assumption A.I(i) requires the ME
leads them to conclude that cash flow is a statistically significant on the q̂ (U1 ) to have zero conditional mean given the true Tobin’s
cause of investment and investment–cash flow sensitivity is higher q (X ) and the ME of qe (U2 ). This excludes correlation between U1
for financially constrained firms.8 and X or between U1 and U2 . Hence, a nonclassical reporting error
We use quantile regression (QR) methods in the investment assumption (e.g., Hu and Sasaki (2015b)) is not allowed in our
equation model. The proposed QR estimator is designed to correct setting.10 Nevertheless, Assumption A.I(i) allows nonlinear depen-
for the ME problem while exploring heterogeneous covariate ef- dence of the ME (U1 ) and the true Tobin’s q (X ) and the ME of qe (U2 ).
fects across the conditional quantile functions. The QR model has This assumption is particularly useful when the miscalculation is
two advantages. First, our QR method proposes a solution to the ME changing over the level of Tobin’s q. Assumption A.I(ii) requires
problem in Tobin’s q by using repeated measures of this variable. that the ME of qe (U2 ) to be independent of true Tobin’s q (X ) as
Second, we accommodate possible heterogeneity on the Tobin’s q well as other variables (Y , Z ). This does not necessarily require
and cash flow in the conditional distribution of investment. Indeed, zero mean of U2 . Hence, this condition allows a possibility that, on
this heterogeneity is not revealed by conventional least squares average, analysts’ earnings expectations on the Tobins’ q could be
procedures. either larger or smaller than the standard equity prices.
The objective is to estimate the following conditional quantile The results for both mean and quantile estimates of the sen-
function: sitivity of investment to Tobin’s q and cash flow are presented in
the left panels and right panels, respectively, of Figs. 1–3. Fig. 1
QIKi (τ |qi , CFKi ) = α (τ ) + β (τ )qi + δ (τ )CFKi , (10) presents the results for QR and OLS. Fig. 2 displays the results for
where the quantity IK denotes the ratio of investment, I, to capital IV-QR and IV-OLS. Finally, Fig. 3 collects the results for ME–QR.
stock, K ; CFK is the ratio of cash flow, CF , to capital stock; and q All figures contain point estimates as well as the corresponding
is the observed measure of average q, which is a measure of the 95% pointwise confidence bands. In the nonparametric estimation
(latent) true marginal Tobin’s q. step in our proposed estimator, for the choice of the bandwidth
We use a data set taken from Cummins et al. (2006). In this (hxn in Eq. (7)), we use an informal rule where the estimates are not
data there are two measures of average q. The first measure is sensitive to marginal changes in the neighborhood of the optimal
constructed using the standard equity prices (qe ). The second proxy bandwidth. Other bandwidths are chosen based on Silverman’s
for the firm’s intrinsic value is based on analysts’ earnings expec- rule of thumb. The number of bootstrap replications to construct
tations (q̂). Thus, the data on investment, the capital stock, the the confidence intervals is 250.
market-based measure of average q, and cash flow are standard The left panels of Figs. 1–3 show significant positive estimates
from Compustat, while the data on expected earnings are from for Tobin’s q estimates. All the quantile estimates in Figs. 1–3
I/B/E/S International Inc. The construction of qe and q̂ are detailed display evidence of positive and increasing effect of the investment
in Appendix B.2 of Cummins et al. (2006). The sample consists of demand, Tobin’s q, on the investment spending, across quantiles
11,431 observations over the 1982–1999 period. For each firm, in a of the conditional distribution of investment. This result docu-
given year, we have information on investment, cash flow, and two ments empirical evidence of strong heterogeneity in the effects
measures of Tobin’s q. For practical estimation we standardize all of Tobin’s q across the distribution of investment. Relative to the
variables by subtracting the mean and dividing by the correspond- QR in Fig. 1, the IV-QR in Fig. 2 shows virtually no difference in
ing standard deviation of each variable. The summary statistics (for the importance of Tobin’s q. This result might be interpreted as
the standardized data) are described in Table 5.
We present results for the estimates, and corresponding 95% 9 We note that the independence condition required by the IV method implies
confidence intervals, using our proposed methods (ME–QR). For independence between ME on q̂ and qe . Because the analysts’ earnings expectations
completeness, we also provide results from the standard QR and (q̂) are dependent on the standard equity prices (qe ), their ME are likely to be related.
OLS, as well as the instrumental variables QR (IV-QR) and OLS However, our approach requires a weaker assumption of conditional mean zero as
in Assumption A.I(i).
10 We acknowledge that although our conditions partially relax the classical
8 Erickson and Whited (2000) suggest another alternative solution which relies ME assumptions, it would be interesting to extend the current results to a fully
on the high-order moments. nonclassical ME model in QR.
156 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164

Fig. 1. Quantile regression and ordinary least squares. Left box plots the coefficients on Tobins’s q. Right box plots the coefficients on Cash Flow.

Fig. 2. Quantile regression instrumental variables and two stage least squares. Left box plots the coefficients on Tobins’s q. Right box plots the coefficients on Cash Flow.

the IV approach being ineffective. On the other hand, the ME– results show positive and monotonically-increasing cash flow sen-
QR in Fig. 3 shows different results. We note that, according to sitivity at low quantiles of the distribution; estimated coefficient
the theory of investment equation, see, e.g., Fazzari et al. (1988), is increasing from 0.30 to 0.36 up to approximately 45th quantile.
and previous empirical studies, e.g., Kaplan and Zingales (1997), The cash flow sensitivity starts to decease for larger quantiles,
on average, the investment demand has positive effect on the and becomes relatively smaller at the very top of the conditional
investment spending. Thus, this result has been verified in our investment distribution. These findings uncover several important
ME–QR estimations, although after correcting for ME the estimates features. First, they document important heterogeneity on the
decrease relative to the standard QR. Moreover, we find evidence response of investment spending to cash flow along the conditional
that firms have heterogeneous responses to changes of investment quantile function. Firms in different quantiles of the conditional
demand across the conditional distribution of investment. distribution of investment respond differently to marginal changes
The results regarding the sensitivity of investment to cash flow in the cash flow. Second, Fig. 3 shows evidence strong heterogene-
are presented in the right panels of Figs. 1–3. The right panel of ity with an inverted U-shape. This result could be interpreted at the
Fig. 1 presents the results for standard QR and shows decreasing light of the effects of financial constraints on corporate policies as
point estimates of cash flow effects on investment over the quan- in Almeida et al. (2004). The investment spending is more sensitive
tiles. The mean regression estimate is represented by the horizon- to cash flow (large magnitude of the coefficients) for firms at lower
tal straight line, which shows a positive effect close to 0.1, and is quantiles. The large coefficients for lower quantiles is an intuitive
statistically different from zero at usual levels of significance. The result. The cash flow coefficient captures the potential sensitivity
results for IV-QR in Fig. 2 are virtually the same as those in Fig. 1. of investment to fluctuations in available internal finance—after
Thus, again these results might be interpreted as the instruments investment opportunities. Thus, the results show evidence that,
being invalid to resolve the measurement error problem. on the one hand, firms at low levels of investment spending, and
The right panel results in Fig. 3 for the ME–QR estimates exhibit thus likely financially constrained, are in fact more exposed to
a distinct inverted U-shape, implying larger cash flow sensitivity and dependent on fluctuations in internal finance. On the other
for those firms in lower quantiles, that is, the lower part of the hand, firms at upper quantiles, which have higher levels of invest-
conditional distribution of investment. In particular, the ME–QR ment spending and are financially unconstrained, are less exposed
 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164 157

Fig. 3. Quantile regression measurement error. Left box plots the coefficients on Tobins’s q. Right box plots the coefficients on Cash Flow.

to and dependent on fluctuations in internal finance. Third, this private signals in auction, earnings data in earnings dynamics
heterogeneous effect across quantiles also indicates that, for a would appear to be a natural laboratory for further development
fixed level of q, the variability of the investment spending across of quantile regression models with repeated measures.
the conditional distribution increases as the level of cash flow
increases. Intuitively, firms with larger cash flow are entitled to Appendix A. Mathematical appendix
invest in a larger range in contrast to the firms with smaller cash
flow. Proof of Theorem 1. Given Assumption [Link], we have
Overall, the application illustrates that QR method is an impor-
φ (ζ , y, z) ≡ E[eiζ X | Y = y, Z = z ]
tant tool to study investment equation models. It allows us to study ∫
the impacts of Tobins’ q and cash flow at different quantiles of = E[eiζ X | Y = y, Z = z , X = x]f (x | y, z)dx
the conditional distribution of investment. The empirical results
∫
document findings that larger cash flow sensitivity occurs at the
lower part of the investment distribution, showing evidence that = f (x | y, z)eiζ x dx
these firms are more exposed to and dependent on fluctuations in
internal finance. where the last expression is the Fourier transform of f (x | y, z).
Note that for (x, y, z) ∈ supp(X , Y , Z ),
∫
8. Conclusion 1
φ (ζ , y, z) exp(−iζ x)dζ
2π
This paper develops estimation and inference for quantile re-
is the inverse Fourier transform of φ (ζ , y, z). Thus we have
gression models with measurement errors. We propose a semi- ∫
parametric two-step estimator assuming availability of repeated 1
f (x | y, z) = φ (ζ , y, z) exp(−iζ x)dζ .
measures of the true covariate. The asymptotic properties of the 2π
estimator are established. We also develop statistical inference
We now need to show that
procedures and establish the validity of a bootstrap approach
ζ
E[eiζ X2 | Y , Z ] iE[X1 eiξ X2 ]
(∫ )
to implement the methods in practice. Monte Carlo simulations
φ (ζ , Y , Z ) = exp dξ .
assess the finite-sample performance of the proposed methods E[eiζ X2 ] 0 E[eiξ X2 ]
and show that the proposed methods have good finite-sample
From Assumptions A.I–[Link]
performance. We apply the methods to an empirical application
to the investment equation model. The results document strong iE[Xeiξ X ]
Dξ ln(E[eiξ X ]) =
heterogeneity in the sensitivity of investment to Tobin’s q and E[eiξ X ]
cash flow across the conditional distribution of investment, and
iE[Xeiξ X ]E[eiξ U2 ]
illustrate that our methods are useful in empirical models where =
measurement error is an important issue. E[eiξ X ]E[eiξ U2 ]
Many issues remain to be investigated. In this paper the quan- iE[Xeiξ (X +U2 ) ]
=
tile of interest is fixed, τ ∈ (0, 1). The extension of the results to the E[eiξ (X +U2 ) ]
uniform case is desirable and important for uniform inference over iE[Xeiξ (X +U2 ) ] + iE[E(U1 | X , U2 )eiξ (X +U2 ) ]
the entire conditional quantile function over τ . Such extension =
E[eiξ X2 ]
would require generalizing the current results. One of the key
steps for the derivations would be to establish stochastic equicon- iE[Xeiξ (X +U2 ) ] + iE[E(U1 eiξ (X +U2 ) | X , U2 )]
=
tinuity of the appropriate centralized scores uniformly over the E[eiξ X2 ]
iξ (X +U2 )
quantiles. In addition, the analysis of the quantile regression with iE[Xe ] + iE[U1 eiξ (X +U2 ) ]
nonclassical measurement error is also a critical direction for fu- =
E[eiξ X2 ]
ture research. There are many potential applications for the pro- iξ X2
iE[X1 e ]
posed methods. Examples as employer–employee matched sample = .
for wages, matched federal agency and firm-level data, bidder’s E[eiξ X2 ]
158 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164

Therefore, for each real ζ , where

iζ X n
φ (ζ , Y , Z ) ≡ E[e | Y , Z ] 1∑
χ̂ (ζ , y, z) = eiζ X2j khy (Yj − y)khz (Zj − z)
E[eiζ X | Y , Z ]E[eiζ U2 ] n
= E[eiζ X ] j=1
E[eiζ X ]E[eiζ U2 ] = Ê eiζ X2 khy (Y − y)khz (Z − z)
[ ]
E[eiζ X2 | Y , Z ]
E[eiζ X ]
n
= 1∑
E[eiζ X2 ] f (y, z) =
ˆ khy (Yj − y)khz (Zj − z)
n
E[eiζ X2 | Y , Z ] j=1
exp ln(E[eiζ X ]) − ln 1
( )
=
E[eiζ X2 ] = Ê [khy (Y − y)khz (Z − z)] ,
(∫ ζ
E[eiζ X2 | Y , Z ]
)
= exp D ξ ln(E [e iξ X
])d ξ and δ χ̂ (ζ , y, z) ≡ χ̂ (ζ , y, z) − χ (ζ , y, z) and δˆ
f (y, z) ≡ ˆ
f (y, z) −
E[eiζ X2 ] 0 f (y, z). We use the following representation
iζ X2
(∫ ζ
| Y , Z] iE[X1 eiξ X2 ]
)
E[ e
ξ , ω̂X1 (ζ ) ωX (ζ ) + δ ω̂X1 (ζ )
=
E[eiζ X2 ]
exp
E[eiξ X2 ]
d = 1 = qX1 (ζ ) + δ q̂X1 (ζ ) (11)
0 ω̂1 (ζ ) ω1 (ζ ) + δ ω̂1 (ζ )
where the third equality is obtained by U2 ⊥ (Y , X , Z ). ■ where qX1 (ζ ) = ωX1 (ζ )/ω1 (ζ ) and where δ q̂X1 (ζ ) can be written as
either
Proof of Theorem 2. Note that the inverse Fourier Transform of )−1
δ ω̂X1 (ζ ) ωX1 (ζ )δ ω̂1 (ζ ) δ ω̂1 (ζ )
( )(
κ (hx ζ ) is k(x/hx )/hx , and the inverse Fourier Transform of E[eiζ X | δ q̂X1 (ζ ) = − 1+
Y = y, Z = z ] is f (x | y, z) by Eq. (13). Also note that from the ω1 (ζ ) (ω1 (ζ ))2 ω1 (ζ )
convolution theorem, the inverse Fourier Transform of the product or δ q̂X1 (ζ ) = δ1 q̂X1 (ζ ) + δ2 q̂X1 (ζ ) with
of κ (hx ζ ) and E[eiζ X | Y = y, Z = z ] is the convolution between
the inverse Fourier Transform of κ (hx ζ ) and the inverse Fourier
δ ω̂X1 (ζ ) ωX1 (ζ )δ ω̂1 (ζ )
δ1 q̂X1 (ζ ) ≡ −
Transform of E[eiζ X | Y = y, Z = z ]. Because Assumptions [Link](iii)– ω1 (ζ ) (ω1 (ζ ))2
[Link] guarantee the existence of f (x | y, z ; hx ), we conclude that ωX (ζ ) δ ω̂1 (ζ ) 2
)−1
δ ω̂1 (ζ )
( ) (
δ2 q̂X1 (ζ ) ≡ 1 1+
ω1 (ζ ) ω1 (ζ ) ω1 (ζ )
∫ ( )
1 x̃ − x
f (x | y, z ; h ) ≡
x
k f (x̃ | y, z)dx̃ )−1
hx hx δ ω̂X1 (ζ ) δ ω̂1 (ζ ) δ ω̂1 (ζ )
(
∫ − 1+ .
=
1
κ (hx ζ )E[eiζ X | Y = y, Z = z ] exp(−iζ x)dζ ω1 (ζ ) ω1 (ζ ) ω1 (ζ )
2π Similarly,
∫
1
= κ (hx ζ )φ (ζ , y, z) exp(−iζ x)dζ . ■ 1 1
2π = = q1 (ζ ) + δ q̂1 (ζ ) (12)
ω̂1 (ζ ) ω1 (ζ ) + δ ω̂1 (ζ )
The following lemma is helpful to derive the result given in
Theorem 3. Recall that X ≡ R is the support of X , and Y × Z is where q1 (ζ ) ≡ 1/ω1 (ζ ), and where
a compact set contained in the support of (Y , Z ).
)−1
δ ω̂1 (ζ ) δ ω̂1 (ζ )
( )(
δ q̂1 (ζ ) = − 1+
(ω1 (ζ ))2 ω1 (ζ )
Lemma A.1. For (x, y, z) ∈ X × Y × Z and hn > 0,
or δ q̂1 (ζ ) = δ1 q̂1 (ζ ) + δ2 q̂1 (ζ ) with
f (x | y, z ; h) − f (x | y, z)
δ ω̂1 (ζ )
ˆ
= B(x, y, z ; hx ) + L(x, y, z ; h) + R(x, y, z ; h), δ1 q̂1 (ζ ) ≡ −
(ω1 (ζ ))2
where B(x, y, z ; hx ) is a nonrandom ‘‘bias term’’ defined as 1 δ ω̂1 (ζ )
( )2 (
δ ω̂1 (ζ )
)−1
δ2 q̂1 (ζ ) ≡ 1+ .
B(x, y, z ; hx ) ≡ f (x | y, z ; hx ) − f (x | y, z); ω1 (ζ ) ω1 (ζ ) ω1 (ζ )
L(x, y, z ; h) is a ‘‘variance term’’ admitting the linear representation And also
χ̂ (ζ , y, z) χ (ζ , y, z) + δ χ̂ (ζ , y, z)
L(x, y, z ; h) ≡ f¯ (x | y, z ; h) − f (x | y, z , hx ) = = q2 (ζ , y, z) + δ q̂2 (ζ , y, z)
f (y, z)
ˆ f (y, z) + δˆf (y, z)
= Ê [ℓ(x, y, z , h; Y , X1 , X2 , Z )] (13)
where ℓ(x, y, z , h; Y , X1 , X2 , Z ) is defined in the proof of the lemma, where q2 (ζ , y, z) ≡ χ (ζ , y, z)/f (y, z), and where
and R(x, y, z ; h) is a ‘‘remainder term’’,
δ χ̂ (ζ , y, z) χ (ζ , y, z)δˆ f (y, z)
( )
R(x, y, z ; h) ≡ ˆ
f (x | y, z ; h) − f¯ (x | y, z ; h). δ q̂2 (ζ , y, z) = −
f (y, z) (f (y, z))2
)−1
Proof of Lemma A.1. Let ωA (ζ ) ≡ E Aeiζ X2 where A = 1, X1 and δˆf (y, z)
[ ] (
× 1+
f (y, z)
ω(ζ , y, z) ≡ E eiζ X2 | Y = y, Z = z
[ ]
∫ or δ q̂2 (ζ , y, z) = δ1 q̂2 (ζ , y, z) + δ2 q̂2 (ζ , y, z) with
= eiζ x2 f (x2 | y, z)dx2
δ χ̂ (ζ , y, z) χ (ζ , y, z)δˆ f (y, z)
χ (ζ , y, z) δ1 q̂2 (ζ , y, z) ≡ −
f (y, z) (f (y, z))2
= ,
f (y, z) χ (ζ , y, z) δˆ
(
f (y, z)
)2 (
δˆf (y, z)
)−1
δ2 q̂2 (ζ , y, z) ≡ 1+
where χ (ζ , y, z) ≡ eiζ x2 f (x2 , y, z)dx2 . Also let ω̂A (ζ ) ≡ Ê Aeiζ X2
∫ [ ]
f (y, z) f (y, z) f (y, z)
and δ ω̂A (ζ ) ≡ ω̂A (ζ ) − ωA (ζ ), and let )−1
δ χ̂ (ζ , y, z) δˆf (y, z) δˆf (y, z)
(
ω̂(ζ , y, z) ≡ Ê eiζ X2 | Y = y, Z = z ≡ χ̂ (ζ , y, z)/ˆ
[ ]
f (y, z)
− 1+ .
f (y, z) f (y, z) f (y, z)
 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164 159

iδ ω̂X1 (ξ ) iωX1 (ξ )δ ω̂1 (ξ )

∫ζ ∫ζ ( )
Let QX1 (ζ ) ≡ 0 (iωX1 (ξ )/ω1 (ξ ))dξ and δ Q̂X1 (ζ ) ≡ 0 (iω̂X1 (ξ )/ω̂1 × − dξ
(ξ ))dξ − QX1 (ζ ). Note that for some random function δ Q̄X1 (ζ ) such ω1 (ξ ) (ω1 (ξ ))2
that |δ Q̄X1 (ζ )| ≤ |δ Q̂X1 (ζ )| for all ζ ,
∫
1
= κ (hx ζ ) exp(−iζ x)φ (ζ , y, z)
2π
exp QX1 (ζ ) + δ Q̂X1 (ζ ) = exp(QX1 (ζ ))
( )
δ ω̂1 (ζ ) δ χ̂ (ζ , y, z) δˆ f (y, z)
( )
dζ
( )
1[ )2 × − + −
1 + δ Q̂X1 (ζ ) + exp(δ Q̄X1 (ζ )) δ Q̂X1 (ζ ) . ω1 (ζ ) χ (ζ , y, z) f (y, z)
](
× (14)
2 ∫ ∫ ±∞
1
From Eqs. (11)–(14), we have + κ (hx ξ ) exp(−iξ x)φ (ξ , y, z)dξ
2π ζ

f (x | y, z ; h) − f (x | y, z ; hx ) iδ ω̂X1 (ζ ) iωX1 (ζ )δ ω̂1 (ζ )

( )
ˆ
× − dζ
ω1 (ζ ) (ω1 (ζ ))2
∫
1
= κ (hx ζ )φ̂ (ζ , y, z , h(2) ) exp(−iζ x)dζ
2π ∫ [{
1 1
∫
1
− κ (hx ζ )φ (ζ , y, z) exp(−iζ x)dζ = − κ (hx ζ ) exp(−iζ x)φ (ζ , y, z)
2π 2π ω1 (ζ )
(∫ ζ
ω̂(ζ , y, z) iω̂X1 (ξ ) 1 iωX1 (ζ )
∫ [ ) ∫ ±∞
1
}
= κ (hx ζ ) exp(−iζ x) exp dξ − κ (h ξ ) exp(−iξ x)φ (ξ , y, z)dξ δ ω̂1 (ζ )
x
2π ω̂1 (ζ ) 0 ω̂1 (ξ ) 2π (ω1 (ζ ))2 ζ
(∫ ζ
ω(ζ , y, z) iωX1 (ξ )
)] { ∫ ±∞ }
− exp dξ dζ 1 i
ω1 (ζ ) + κ (hx ξ ) exp(−iξ x)φ (ξ , y, z)dξ δ ω̂X1 (ζ )
0 ω1 (ξ ) 2π ω1 (ζ ) ζ
(∫ ζ
ω(ζ , y, z) iωX1 (ξ )
∫ [ )
1
κ (h ζ ) exp(−iζ x) −
x
dξ
{ }
= exp 1 1
2π ω1 (ζ ) 0 ω1 (ξ )
+ κ (hx ζ ) exp(−iζ x)φ (ζ , y, z) δ χ̂ (ζ , y, z)
2π χ (ζ , y, z)
χ (ζ , y, z) δ χ̂ (ζ , y, z)
{
{ } ]
+ + 1 1
f (y, z) f (y, z) + − κ (hx ζ ) exp(−iζ x)φ (ζ , y, z) δˆ f (y, z) dζ
2π f (y, z)
χ (ζ , y, z)δˆ f (y, z)
}
− + δ 2 q̂2 ( ζ , y , z)
∫ [
(f (y, z))2 Ψ1 (ζ , x, y, z , hx ) Ê[eiζ X2 ] − E[eiζ X2 ]
( )
=
δ ω̂1 (ζ )
{ }
1
× − + δ2 q̂1 (ζ ) × exp(QX1 (ζ ))
+ Ψ2 (ζ , x, y, z , hx ) Ê[X1 eiζ X2 ] − E[X1 eiζ X2 ]
( )
ω1 (ζ ) (ω1 (ζ ))2
∫ ζ ∫ ζ
+ Ψ3 (ζ , x, y, z , hx ) Ê[eiζ X2 khy (Y − y)khz (Z − z)]
{ (
× 1+ iδ1 q̂X1 (ξ )dξ + iδ2 q̂X1 (ξ )dξ
− E[eiζ X2 khy (Y − y)khz (Z − z)]
0 0
)
(∫ ζ )2 }]
1
exp(δ Q̄X1 (ζ )) iδ q̂X1 (ξ )dξ dζ . + Ψ4 (ζ , x, y, z , hx ) Ê[khy (Y − y)khz (Z − z)]
(
+
2 0 ]
f (x | y, z ; hx ) by f¯ (x | y, z ; hx ). Then
We denote the linearization of ˆ − E[khy (Y − y)khz (Z − z)]) dζ
[∫
L(x, y, z ; h) = Ê Ψ1 (ζ , x, y, z , hx ) eiζ X2 − E[eiζ X2 ]
( )

≡ f¯ (x | y, z ; h) − f (x | y, z ; hx )
+ Ψ2 (ζ , x, y, z , hx ) X1 eiζ X2 − E[X1 eiζ X2 ]
( )
χ (ζ , y, z) δ ω̂1 (ζ )
∫ [
1
= κ (hx ζ ) exp(−iζ x) exp(QX1 (ζ )) −
+ Ψ3 (ζ , x, y, z , hx ) eiζ X2 khy (Y − y)khz (Z − z)
(
2π f (y, z) (ω1 (ζ ))2
− E[eiζ X2 khy (Y − y)khz (Z − z)]
∫ ζ
χ (ζ , y, z) 1
)
+ iδ1 q̂X1 (ξ )dξ
f (y, z) ω1 (ζ ) 0 + Ψ4 (ζ , x, y, z , hx ) (khy (Y − y)khz (Z − z)
1 δ χ̂ (ζ , y, z) 1 χ (ζ , y, z)δˆ f (y, z)
] ]
+ − dζ − E[khy (Y − y)khz (Z − z)]) dζ
ω1 (ζ ) f (y, z) ω1 (ζ ) (f (y, z))2
∫
1 ≡ Ê [ℓ(x, y, z , h; Y , X1 , X2 , Z )] ,
= κ (hx ζ ) exp(−iζ x)φ (ζ , y, z)
2π
where the following identity was used in the fourth equality: for
δ ω̂1 (ζ ) δ χ̂ (ζ , y, z) δˆ f (y, z)
[
any absolutely integrable function g
× − + −
ω1 (ζ ) χ (ζ , y, z) f (y, z) ∫ ∞ ∫ ζ ∫ ∞ ∫ ∞
∫ ζ(
iδ ω̂X1 (ξ ) iωX1 (ξ )δ ω̂1 (ξ )
) ] g(ζ , ξ )dξ dζ = g(ζ , ξ )dζ dξ
+ − dξ dζ −∞ 0 0 ξ
0 ω1 (ξ ) (ω1 (ξ ))2 ∫ 0 ∫ −∞ ∫ ∫ ±∞

1
∫ + g(ζ , ξ )dζ dξ ≡ g(ζ , ξ )dζ dξ ,
= κ (hx ζ ) exp(−iζ x)φ (ζ , y, z) −∞ ξ ξ
2π
and where
δ ω̂1 (ζ ) δ χ̂ (ζ , y, z) δˆ f (y, z)
( )
× − + − dζ 1 1
ω1 (ζ ) χ (ζ , y, z) f (y, z) Ψ1 (ζ , x, y, z , hx ) ≡ − κ (hx ζ ) exp(−iζ x)φ (ζ , y, z)
2π ω1 (ζ )
1 iωX1 (ζ )
∫ ∫ ±∞ ∫ ±∞
1
+ κ (hx ζ ) exp(−iζ x)φ (ζ , y, z)dζ − κ (hx ξ ) exp(−iξ x)φ (ξ , y, z)dξ
2π ξ 2π (ω1 (ζ ))2 ζ
160 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164
∫ ±∞
1 i and
Ψ2 (ζ , x, y, z , hx ) ≡ κ (hx ξ ) exp(−iξ x)φ (ξ , y, z)dξ
2π ω1 (ζ ) ζ
∫
Ψ1+ (ζ , hx )dζ ⪯ (1 + (hx )−1 )γµ +γφ −γω +2 exp −αω ((hx )−1 )νω
( )
1 1
Ψ3 ( ζ , x , y , z , h ) ≡
x
κ (h ζ ) exp(−iζ x)φ (ζ , y, z)
x
2π χ (ζ , y, z) × exp αφ ((hx )−1 )νφ ,
( )
1 1
Ψ4 (ζ , x, y, z , hx ) ≡ − κ (hx ζ ) exp(−iζ x)φ (ζ , y, z).
∫
■
Ψ2+ (ζ , hx )dζ ⪯ (1 + (hx )−1 )γφ −γω +2 exp −αω ((hx )−1 )νω
( )
2π f (y, z)
We use the following convenient notation for expositional sim-
× exp αφ ((hx )−1 )νφ ,
( )
plicity. ∫
Ψ3+ (ζ , hx )dζ ⪯ an (1 + (hx )−1 )γφ −γω +1 exp −αω ((hx )−1 )νω
( )
an
Definition A.1. We write f (ζ ) ⪯ g(ζ ) for f , g : R ↦ → R when
there exists a constant C > 0, independent of ζ , such that f (ζ ) ≤ × exp αφ ((hx )−1 )νφ ,
( )
Cg(ζ ) for all ζ ∈ R (and similarly for ⪰). Analogously, we write ∫
an ⪯ bn for two sequences an , bn when there exists a constant C Ψ4+ (ζ , hx )dζ ⪯ an (1 + (hx )−1 )γφ +1 exp αφ ((hx )−1 )νφ .
( )
an
independent of n such that an ≤ Cbn for all n ∈ N.
Then we have
Proof of Theorem 3. In order to obtain the uniform convergence
f (x | y, z ; h), we derive asymptotic convergence rate of Ψ + (h) = O ((hx )−1 )γµ +γφ −γω +2
(
rate of ˆ
the bias term, divergence rate of the variance term, and rely on
× exp (αφ 1{νφ =νω } − αω )((hx )−1 )νω .
( ))
negligibility of the remainder term. First, from Parseval’s identity
and Assumption [Link], we have Note that by Minkowski inequality,
|B(x, y, z , hx )| = |f (x | y, z ; hx ) − f (x | y, z)| [ ]
= |f (x | y, z ; hx ) − f (x | y, z ; 0)| E sup |L(x, y, z , h)|
⏐ ∫ (x,y,z)∈X ×Y ×Z
⏐ 1
κ (hx ζ )φ (ζ , y, z) exp(−iζ x)dζ
[ ]
= ⏐⏐
2π =E sup |f¯ (x | y, z ; h) − f (x | y, z ; hx )|
∫ ⏐ (x,y,z)∈X ×Y ×Z
1
φ (ζ , y, z) exp(−iζ x)dζ ⏐⏐
⏐
− [ ⏐∫
2π ⏐ [Ψ1 (ζ , x, y, z , hx )δ ω̂1 (ζ )
⏐
=E sup
(x,y,z)∈X ×Y ×Z
⏐ ∫ ⏐ ⏐
⏐ 1
(κ (h ζ ) − 1)φ (ζ , y, z) exp(−iζ x)dζ ⏐⏐
x
⏐
=⏐⏐
2π + Ψ2 (ζ , x, y, z , hx )δ ω̂X1 (ζ )
∫
1 ⏐]
⏐(κ (hx ζ ) − 1)⏐ |φ (ζ , y, z)| dζ
⏐ ⏐
≤ + Ψ3 (ζ , x, y, z , h )δ χ̂ (ζ , y, z) + Ψ4 (ζ , x, y, z , h )δ f (y, z)]dζ ⏐⏐
x x
⏐
2π
ˆ
∫ ∞
1
⏐(κ (hx ζ ) − 1)⏐ |φ (ζ , y, z)| dζ
⏐ ⏐ ∫ [( )
= ⏐Ψ1 (ζ , x, y, z , hx )⏐ |δ ω̂1 (ζ )|
π ξ̄ /hx
⏐ ⏐
≤E sup
∫ ∞ (x,y,z)∈X ×Y ×Z

⪯ |φ (ζ , y, z)| dζ . ( )
⏐Ψ2 (ζ , x, y, z , hx )⏐ ⏐δ ω̂X (ζ )⏐
⏐ ⏐ ⏐ ⏐
ξ̄ /hx + sup 1
(x,y,z)∈X ×Y ×Z
Then, by Assumption B.I(ii), we have ( )( )
∞
⏐Ψ3 (ζ , x, y, z , hx )⏐ |δ χ̂ (ζ , y, z)|
∫ ⏐ ⏐
γφ νφ + sup sup
sup |B(x, y, z , h )| ⪯ x
Cφ (1 + |ζ |) exp(αφ |ζ | )dζ (x,y,z)∈X ×Y ×Z (y,z)∈Y ×Z
(x,y,z)∈X ×Y ×Z ξ̄ /hx
∫ ∞ ( )( )]
⏐Ψ4 (ζ , x, y, z , hx )⏐ ⏐δˆ
f (y, z)⏐ dζ
⏐ ⏐ ⏐ ⏐
γφ νφ
⪯ (1 + |ζ |) exp(αφ |ζ | )dζ + sup sup
(x,y,z)∈X ×Y ×Z (y,z)∈Y ×Z
ξ̄ /hx
(( )γ +1 )ν )) ∫ [
= O ξ̄ /hx φ exp αφ ξ̄ /hx φ
( (
)}1/2
Ψ1+ (ζ , hx ) E |δ ω̂1 (ζ )|2
{ (
(( ) ( ( ) )) ≤
−γ −ν
= O hx B exp αB hx B . (15)
{ (⏐ ⏐2 )}1/2
For the asymptotic divergence rate of the variance term, let + Ψ2+ (ζ , hx ) E ⏐δ ω̂X1 (ζ )⏐
(ln n)1/2
an = (nhy hz )1/2 + s∈{y,z } (hs )2 and define
∑
{ (⏐ ( )⏐2 )}1/2
+ an Ψ3+ (ζ , hx ) E ⏐⏐a− sup δ χ̂ (ζ , y, z) ⏐⏐
⏐ 1
⏐
n ·
∫ ∫
Ψ (h) ≡
+
Ψ1 (ζ , h )dζ + Ψ2+ (ζ , hx )dζ
+ x
(y,z)∈Y ×Z

∫ ∫ { (⏐ ( )⏐2 )}]
+ an Ψ3 (ζ , h )dζ + an Ψ4+ (ζ , hx )dζ ,
+ x
+ an Ψ4 (ζ , h ) E ⏐an · sup δˆ f (y, z) ⏐⏐ dζ
+
⏐ −1 x
⏐
⏐
(y,z)∈Y ×Z

where ΨA+ (ζ , hx ) ≡ sup(x,y,z)∈X ×Y ×Z |ΨA (ζ , x, y, z , hx )| for A ∈ [∫

{1, 2, 3, 4}. From Assumptions [Link] and [Link], and from similar ar- ≤n −1/2
Ψ1+ (ζ , hx ) E n|δ ω̂1 (ζ )|2
{ ( )}1/2
dζ
guments above, one can show that
( ⏐ ⏐2 )
⏐2 )}1/2
∫ { ( ⏐
E n|δ ω̂1 (ζ )| ⪯ 1, E n⏐⏐an · sup (δ χ̂ (ζ , y, z))⏐⏐ ⪯ 1, Ψ2+ (ζ , hx ) E n⏐δ ω̂X1 (ζ )⏐ dζ
( 2
) ⏐ −1 ⏐
+
(y,z)∈Y ×Z
( ⏐ ⏐ ) { ( ⏐ )⏐2 )}1/2
)⏐2
( ⏐ ∫ (
⏐2 )
E n δ ω̂X1 (ζ ) ⪯ 1, E n⏐an · sup δˆ
f (y, z) ⏐⏐ ⪯ 1, Ψ3 (ζ , h ) E n⏐an · sup δ χ̂ (ζ , y, z) ⏐⏐ dζ
⏐ −1 ( + x
⏐ −1 ⏐
⏐ ⏐ ⏐ + an ⏐
(y,z)∈Y ×Z (y,z)∈Y ×Z
 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164 161

op (n−1/2 ) by Assumption [Link], and f (x|Yi , Zi ) being integrable by

∫ { ( ⏐ ( )⏐2 )} ]
Ψ4 (ζ , hx ) E n⏐⏐a− sup δˆ f (y, z) ⏐⏐ dζ
+
⏐ 1
⏐
+ an n · assumption [Link](ii). The last assertion follows from [Link] and ap-
(y,z)∈Y ×Z
plication of Markov inequality as 1n supx,i ∥x Zi ∥ = op (n−1/2 ), since
⪯ n−1/2 Ψ + (h). P(supx,i ∥x Z ∥ > n1/2 ) ≤ nP(∥x Z ∥ > n1/2 ) ≤ nE∥x Z ∥2+ε /n 2 =
2+ε

Thus, we have that by Markov’s inequality o(1).

(b) Condition (1.2) holds directly by the identification condi-
sup |L(x, y, z , h)| tion C.I.
(x,y,z)∈X ×Y ×Z
(c) Now we show that Condition (1.3) is satisfied by verifying
= Op n−1/2 ((hx )−1 )γµ +γφ −γω +2 that Q̃ (β, δ, f ) is continuous in f uniformly for all (β, δ ⊤ )⊤ ∈ Θ .
(
For any ∥f − f0 ∥∞ ≤ ϵ ,
× exp (αφ 1{νφ =νω } − αω )((hx )−1 )νω .
( ))
(16)
∥Q̃ (β, δ, f ) − Q̃ (β, δ, f0 )∥
From Assumptions [Link]–[Link], selection of the bandwidths in the  ∫
statement of the theorem, and minor adjustment of the argument
E ψ (Yi − xβ − Zi δ )[x Zi ] · f (x | Yi , Zi ) dx
 ⊤
=
for the variance term above, one can show that the remainder
term is asymptotically negligible. So detailed proof is omitted here ∫ 
− E ψ (Yi − xβ − Zi δ )[x Zi ] · f0 (x | Yi , Zi ) dx
⊤

for brevity. Then putting Eqs. (15) and (16) together yields the 
result. ■  ∫ 
ψ β δ , ,
 ⊤

= E (Yi − x − Z )[ x Zi ] · [ f (x | Y i Zi ) − f 0 (x | Y i Zi )] dx
Proof of Theorem 4. To show consistency of the estimator, we  i 
apply Theorem 1 of Chen et al. (2003). Thus, we need to verify ∫
Conditions (1.1)–(1.5’) in Chen et al. (2003). Recall that ≤ E ∥[x Zi ]∥ · ∥f (x | Yi , Zi ) − f0 (x | Yi , Zi )∥∞ dx.
n ∫
1∑ The inequality holds by the property of exchanging norms and inte-
Q̃n (β, δ, f ) = ψ (Yi − xβ − Zi⊤ δ )[x Zi ] · f (x | Yi , Zi ) dx,
n
i=1
gral, Cauchy inequality, and the fact that ψ (·) ≤ 1. By Assumptions
[Link] and [Link], E ∥[x Zi ]∥ · ∥f (x | Yi , Zi ) − f0 (x | Yi , Zi )∥∞ dx < C , and
∫
and
∫ given that ∥f − f0 ∥∞ ≤ ϵ , as ϵ → 0 the result follows. Therefore,
Q̃ (β, δ, f ) = E ψ (Yi − xβ − Zi⊤ δ )[x Zi ] · f (x | Yi , Zi ) dx. Condition (1.3) holds.
(d) Condition (1.4) is satisfied by our Theorem 3.
(e) It only remains to verify Condition (1.5’). For any ϵn = o(1),
(a) First, we verify Condition (1.1) by verifying that Q̃n (β̂, δ̂,ˆ
f)
= op (n−1/2 ), which implies the desired property. This result is usual
in the quantile regression literature, and generally follows for the sup ∥Q̃n (β, δ, f ) − Q̃ (β, δ, f )∥ = op (1).
boundness of the score function and the computational property (β,δ )∈Θ ,∥f −f0 ∥∞ ≤ϵn
of quantile regression. Consider the following By noting that
 n ∫ 
1 ∑
∥Q̃n (β, δ, f ) − Q̃ (β, δ, f )∥

∥Q̃n (β̂, δ̂, f )∥ =  ψ (Yi − xβ̂ − Zi δ̂ )[x Zi ] · f (x | Yi , Zi ) dx
⊤
ˆ  ˆ 
n  ∫ ( ∑ n
i=1 1
ψ (Yi − xβ − Zi⊤ δ )[x Zi ] · f (x | Yi , Zi )

 ∑ n ∫ =
1  n
= ψ (Yi − xβ̂ − Zi⊤ δ̂ )[x Zi ] i=1
 n ) 
i=1
− Eψ (Yi − xβ − Zi⊤ δ )[x Zi ] · f (x | Yi , Zi ) dx

f (x | Yi , Zi ) − f (x | Yi , Zi )) dx
· (ˆ 
n ∫
∫ ( ∑n
1∑ 1
, ψ (Yi − xβ − Zi⊤ δ )[x Zi ]

+ ψ (Yi − xβ̂ − Zi⊤ δ̂ )[x Zi ] ≤ sup f (x | Y i Zi )
n
 i n
i=1 i=1
 ) 
− Eψ (Yi − xβ − Zi⊤ δ )[x Zi ] dx

· f (x | Yi , Zi ) dx

 
n ∫
∫   ( ∑ n
1
 ∑
1
≤  sup f (x | Yi , Zi ) dx ψ (Yi − xβ − Zi⊤ δ )[x Zi ]
  
≤ ψ (Yi − xβ̂ − Zi⊤ δ̂ )[x Zi ] · (ˆ
f (x | Yi , Zi )  ·sup
 n i
  x n
i=1 i=1
 )
− Eψ (Yi − xβ − Zi⊤ δ )[x Zi ]  ,

− f (x | Yi , Zi )) dx


1 n we have
 ∑ ∫ 
ψ β̂ δ̂ ,
⊤

+ (Yi − x − Z i ) [x Z i ] · f (x | Y i Zi ) dx 
n sup ∥Q̃n (β, δ, f ) − Q̃ (β, δ, f )∥
 
∫i=1 (β,δ )∈Θ ,∥f −f0 ∥∞ ≤ϵn
1
f (x | Yi , Zi ) − f (x | Yi , Zi )dx∥∞ · sup ∥x Zi ∥
≤C· ˆ  n
1 ∑
n x, i ≤ C · sup  ψ (Yi − xβ − Zi⊤ δ )[x Zi ]
(β,δ )∈Θ ,x n
∫ 
1 i=1
+ C · f (x | Yi , Zi )dx∥∞ · sup ∥x Zi ∥ 
n x, i
− Eψ (Yi − xβ − Zi⊤ δ )[x Zi ]
,

= op (n−1/2 ),
since ∥ supi f (x | Yi , Zi ) dx∥ < C by Assumption [Link]. Denote
∫
where C is generic constant, the first inequality is given by the
triangle inequality, and the last equality is given by the fact that φβ,δ (Yi , x, Zi ) = ψ (Yi − xβ − Zi⊤ δ )[x Zi ]. We need to show that
ψ (·) ≤ 1 and the computational property of quantile regres- {φβ,δ : (β, δ ) ∈ Θ } is G–C. Because {ψ (Yi − xβ − Zi⊤ δ ) : (β, δ ) ∈ Θ }
sion (Koenker and Bassett, 1978), Theorem 3, 1n supx,i ∥x Zi ∥ = is bounded and VC, it is G–C. Also, E[|X |] < ∞ and E[∥Z ∥] < ∞
162 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164

by Assumption [Link]. Finally, Fϵn = {f : ∥f − f0 ∥∞ ≤ ϵ} is G–C − Eψ (Yi − xβ0 − Zi δ0 )[x Zi ] · f0 (x | Yi , Zi ) dx
⊤

by Theorem 3. Those conditions and Corollary 9.27(ii) of Kosorok 
(2008) lead to our conclusion. ■ ∫ n
1∑
, ψ (Yi − xβ − Zi⊤ δ )[x Zi ]

≤ sup f (x | Yi Zi ) ·
Proof of Theorem 5. We now apply Theorem 2 of Chen et al. (2003)
 i n
i=1
to establish weak convergence. We need to check their Conditions − Eψ (Yi − xβ − Zi⊤ δ )[x Zi ] − sup f0 (x | Yi , Zi )
(2.1)–(2.6). i
(a) Condition (2.1) was verified in the first part of Theorem 4. n
1∑
(b) To verify Condition (2.2), note that · ψ (Yi − xβ0 − Zi⊤ δ0 )[x Zi ]
n
i=1
∫
Q̃ (β, δ, f0 ) = E ψ (Yi − xβ − Zi⊤ δ )[x Zi ] · f0 (x | Yi , Zi ) dx

− Eψ (Yi − xβ0 − Zi⊤ δ0 )[x Zi ] dx


= EE[ψ (Yi − xβ − Zi⊤ δ )[x Zi ] · |Yi , Zi ]


1 n
 ∑
= Eψ (Yi − xβ − Zi⊤ δ )[x Zi ] ≤ C · sup ψ (Yi − xβ − Zi⊤ δ )[x Zi ]
n
= EE[ψ (Y − xβ − Z ⊤ δ )[x Z ]|X , Z ]
 x
i=1
n
= EE[ψ (Y − xβ − Z δ )|X , Z ][x Z ]
⊤
1∑
− Eψ (Yi − xβ − Zi⊤ δ )[x Zi ] − ψ (Yi − xβ0 − Zi⊤ δ0 )[x Zi ]
= E(τ − GY (xβ + Z ⊤ δ ))[x Z ]. n
i=1

The derivative with respect to (β, δ ), denoted by Γ1 (β, δ, f ), is
− Eψ (Yi − xβ0 − Zi δ0 )[x Zi ]
,
⊤

−EgY (xβ + Z ⊤ δ )[x Z ][x Z ⊤ ]⊤ . It is continuous in (β, δ ) at (β0 , δ0 )
and positive definite by Assumptions G.I and [Link].
for a generic C by Assumption [Link]. So we need to show
(c) Now we verify Condition (2.3). We first calculate the path-
wise derivative of Q̃ (β, δ, f ) at f0 : 1 n
 ∑
sup  ψ (Yi − xβ − Zi⊤ δ )[x Zi ]
Γ2 (β, δ, f0 )[f − f0 ] = [Q̃ (β, δ, f0 + ζ (f − f0 )) − Q̃ (β, δ, f0 )]/ζ ∥β−β0 ∥≤ϵn ,∥δ−δ0 ∥≤ϵn ,x n

i=1

− Eψ (Yi − xβ − Zi⊤ δ )[x Zi ]

∫
=E ψ (Yi − xβ − Zi⊤ δ )[x Zi ] · [f (x | Yi , Zi ) − f0 (x | Yi , Zi )] dx.
n 
1∑
ψ (Yi − xβ0 − Zi δ0 )[x Zi ] + Eψ (Yi − xβ0 − Zi δ0 )[x Zi ]
⊤ ⊤

For any ϵn ↓ 0, such that ∥(β, δ ) − (β0 , δ0 )∥ ≤ ϵn and ∥f − f0 ∥∞ ≤ −
n 
ϵn : i=1
√
= op (1/ n).
∥Q̃ (β, δ, f ) − Q̃ (β, δ, f0 ) − Γ2 (β, δ, f0 )[f − f0 ]∥ = 0
We need to show that φβ,δ is Donsker. Because {ψ (Yi − xβ −
and Zi⊤ δ ) : (β, δ ) ∈ Θ } is bounded and VC, it is Donsker with constant
∥Γ2 (β, δ, f0 )[f − f0 ] − Γ2 (β0 , δ0 , f0 )[f − f0 ]∥ envelope. Also E[|X |] < ∞ and E[∥Z ∥] < ∞ by assumption
 ∫ [Link]. The product of φβ,δ with [x Zi ] also forms a Donsker class
E [ψ (Yi − xβ − Zi δ ) − ψ (Yi − xβ0 − Zi δ0 )][x Zi ]
 ⊤ ⊤
= with a square integrable envelope. Finally, the class F , as defined
 in assumption [Link], is Donsker with a constant envelope. Given
assumption [Link], the bracketing number of F by Corollary 2.7.4
· [f (x | Yi , Zi ) − f0 (x | Yi , Zi )] dx

 in van der Vaart and Wellner (1996) satisfies log N[·] (ε, F , L2 (P)) =
dim(x,y,z) ′
O(ε − δ ) = O(ε −2−δ ), for some δ ′ < 0. Those conditions and
∫
≤ E |ψ (Yi − xβ − Zi⊤ δ ) − ψ (Yi − xβ0 − Zi⊤ δ0 )| Corollary 9.32(iii) of Kosorok (2008) lead to our conclusion. √
(f) Finally, we verify Condition (2.6). Noting that nQ̃n
· ∥[x Zi ]∥ ∥f (x | Yi , Zi ) − f0 (x | Yi , Zi )∥∞ dx
(β0 , δ0 , f0 ) converges weakly and Assumption [Link], we only verify
= o(1). that
√ √
∫
The first inequality holds by the property of exchanging norm and
nΓ2 (β0 , δ0 , f0 )[ˆ
f − f0 ] = nE ψ (Y − xβ0 − Z ⊤ δ0 )[x Z ]
integrals. The last equality holds because the domain for integra-
tion is O(1) by the indicator function, Assumptions [Link] and [Link],
f (x | Y , Z ) − f0 (x | Y , Z )) dx,
· (ˆ
and ∥f − f0 ∥∞ ≤ ϵn . The result follows as ϵn → 0. √
(d) Condition 2.4 holds by Assumption [Link]. converges weakly. √ Also, since the bias of ˆ f is op (1/ n), we only
need to verify nE ψ (Y − xβ0 − Z δ0 )[x Z ] · (ˆ f (x | Y , Z ) − Eˆ
∫
(e) Now we verify Condition (2.5’): f (x |
Y , Z )) dx converges weakly:
sup ∥Q̃n (β, δ, f ) − Q̃ (β, δ, f )
∥β−β0 ∥≤ϵn ,∥δ−δ0 ∥≤ϵn ,∥f −f0 ∥∞ ≤ϵn √
∫
√ nE ψ (Y − xβ0 − Z ⊤ δ0 )[x Z ]
− Q̃n (β0 , δ0 , f0 )∥ = op (1/ n).
[ ∫ ]
1
Note that · κ (hxn ζ )(φ̂ − Eφ̂ )e−iζ x dζ dx. (17)
2π
∥Q̃n (β, δ, f ) − Q̃ (β, δ, f ) − Q̃n (β0 , δ0 , f0 )∥
∫ ( ∑ n
First,
1
ψ (Yi − xβ − Zi⊤ δ )[x Zi ] · f (x | Yi , Zi )
 ⏐ ⏐
= ⏐ 1 n iζ X
⏐ ∑ ⏐
n ⏐ p
e 2j − Eeiζ X2j ⏐⏐ → 0.

i=1 sup ⏐⏐
ζ ⏐n
− Eψ (Yi − xβ − Zi⊤ δ )[x Zi ] · f (x | Yi , Zi ) j=1 ⏐
n
This is because eiζ X2j = cos(ζ X2j ) + i sin(ζ X2j ) and those two terms
)
1∑
− ψ (Yi − xβ0 − Zi⊤ δ0 )[x Zi ] · f0 (x | Yi , Zi ) iζ X2j p
→ EX1 eiζ X2j . Therefore,
∑n
n
i=1
are Lipschitz in ζ . Similarly, 1
n j=1 X1 e
 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164 163

1 ∑n iζ X2j iζ X
√
n j=1 X1 e
→
p EX1 e 2j
. By the continuous mapping theorem, distribution of n{Q̃n∗ (β̂, δ̂,ˆ
f ) − Q̃ (β̂, δ̂,√
f )}, which is approxi-
ˆ
iζ X2j iζ X
1 ∑n
n j=1 e Ee 2j mately the same as the distribution of nQ̃n (β0 , δ0 , f0 ) by the
(∫ ∑n iζ X2j
) verification of condition (2.5’) of Theorem 5. For this we need that
ζ ζ
i 1n j=1 X1 e iEX1 eiζ X2j
(∫ )
p f (·) possesses the same smoothness as f0 (·), which is guaranteed
.
ˆ
exp → exp
iζ X2j
Eeiζ X2j
1 n
by Theorem 3 and condition [Link]. ■
∑
0 n j=1 e 0

Also we have
∑ iζ X Appendix B. Supplementary material
1
y
hn hzn n
[e 2 khyn (Y − y)khzn (Z − z)]
iζ X2
Ê[e | Y = y, Z = z ] ≡ 1
∑ . Supplementary material related to this article can be found
y
h hz n
[khyn (Y − y)khzn (Z − z)]
n n online at [Link]
So (17) equals
References
√
∫ ∫
n ψ (y − xβ0 − z ⊤ δ0 )[x z ]
Agca, S., Mozumdar, A., 2017. Investment cash flow sensitivity: fact or fiction?
[ ∫ ] J. Financ. Quant. Anal. (forthcoming).
1 −iζ x
· κ (hn ζ )(φ̂ − φ )e
x
dζ dx dydz Almeida, H., Campello, M., Galvao, A., 2010. Measurement errors in investment
2π equations. Rev. Financ. Stud. 23, 3279–3328.
[ Almeida, H., Campello, M., Weisbach, M.S., 2004. The cash flow sensitivity of cash.
√
∫ ∫ ∫
1 J. Finance 59, 1777–1804.
= n ψ (y − xβ0 − z δ0 )[x z ] ·
⊤
κ (hxn ζ ) Andrews, D.W.K., 1995. Nonparametric kernel estimation for semiparametric mod-
2π
els. Econometric Theory 11, 560–596.
Angrist, J., Chernozhukov, V., Fernandez-Val, I., 2006. Quantile regression under
(
1 ∑ iζ X2 misspecification, with an application to the U.S. wage structure. Econometrica
× y z [e khyn (Y − y)khzn (Z − z)] 74, 539–573.
hn hn n
Chen, X., Linton, O., Van Keilegom, I., 2003. Estimation of semiparametric models
iζ X2j
(∫
ζ
∑n )/ n
i 1n j=1 X1 e 1∑ when the criterion function is not smooth. Econometrica 71, 1591–1608.
× exp 1
∑ n iζ X2j
eiζ X2j Chernozhukov, V., Hansen, C., 2006. Instrumental quantile regression inference for
0 n j=1 e
n structural and treatment effects models. J. Econometrics 132, 491–525.
j=1
) Chesher, A. (2001) Parameter approximations for quantile regressions with mea-
1 ∑ surement error, Working Paper CWP02/01, Department of Economics, Univer-
× y [khyn (Y − y)khzn (Z − z)] − φ sity College London.
hn hzn n Cummins, J.G., Hasset, K.A., Oliner, S.D., 2006. Investment behavior, observable
] expectations, and internal funds. Amer. Econ. Rev. 96, 796–810.
Delaigle, A., Hall, P., Meister, A., 2008. On deconvolution with repeated measure-
−iζ x
×e dζ dx dydz ments. Ann. Statist. 36, 665–685.
Erickson, T., Whited, T., 2000. Measurement error and the relationship between
√
∫ ∫ investment and q. J. Political Econ. 108, 1027–1057.
= n ψ (y − xβ0 − z ⊤ δ0 )[x z ] Fan, J., 1991. On the optimal rates of convergence for nonparametric deconvolution
problems. Ann. Statist. 19, 1257–1272.
Fan, J., Truong, Y.K., 1993. Nonparametric regression with errors in variables. Ann.
[ ∫ (
1 1 ∑ iζ X2
· κ (hn ζ ) y z
x
[e khyn (Y − y)khzn (Z − z)] Statist. 21, 1900–1925.
2π hn hn n Fazzari, S., Hubbard, R.G., Petersen, B., 1988. Financing constraints and corporate
(∫ ζ
iEX1 eiζ X2j
[ ) ] investment. Brooking Pap. Econ. Act. 1, 141–195.
× exp + op (1) /[Eeiζ X2j f (y, z) Gine, E., Zinn, J., 1990. Bootstrapping general empirical measures. Ann. Probab. 18,
0 Eeiζ X2j 851–869.
) ] Hansen, B.E., 2008. Uniform convergence rates for kernel estimation with depen-
+ op (1)] − φ e−iζ x dζ dx dydz dent data. Econometric Theory 24, 726–748.
Hausman, J., Luo, Y. and Palmer, C. (2014) Errors in the dependent variable of
{ [ quantile regression models, Working Paper, MIT.
√ 1∑
∫ ∫
1
= n ψ (Yj − xβ0 − Zj⊤ δ0 )[x Zj ] · κ (hxn ζ ) Hayashi, F., 1982. Tobin’s marginal q and average q: a neoclassical interpretation.
n 2π Econometrica 50, 213–224.
(∫ ζ He, X., Liang, H., 2000. Quantile regression estimates for a class of linear and partially
iEX1 eiζ X2
[ ) ]
× [eiζ X2j ] exp + o p (1) /[Eeiζ X2 f (y, z) linear errors-in-variables models. Statist. Sinica 10, 129–140.
0 Eeiζ X2 He, X., Shao, Q.-M., 1996. A general bahadur representation of m-estimators and
] } its application to linear regression with nonstochastic designs. Ann. Statist. 6,
−iζ x
dζ dx − φ + op (n−1/2 ),
2608–2630.
+ op (1)]e He, X., Shao, Q.-M., 2000. On parameters of increasing dimensions. J. Multivariate
Anal. 73, 120–135.
which converges weakly, and the result follows. ■ Hu, Y., Sasaki, Y., 2015a. Closed-form estimation of nonparametric models with non-
classical measurement errors. J. Econometrics 185, 392–408.
Hu, Y., Sasaki, Y., 2015b. Identification of paired nonseparable measurement error
Proof of Lemma 1. The proof is a direct application of Theorem models, Working Paper, Johns Hopkins University.
B in Chen et al. (2003) and parallel to that of convergence in Hu, Y., Schennach, S.M., 2008. Instrumental variable treatment of nonclassical
distribution. measurement error models. Econometrica 76, 195–216.
Kaplan, S., Zingales, L., 1997. Do financing constraints explain why investment is
Condition (2.4B) in Chen et al. (2003) is directly implied by
correlated with cash flow? Quart. J. Econ. 112, 169–215.
condition [Link]. Let Kato, K., Galvao, A.F., Montes-Rojas, G., 2012. Asymptotics for panel quantile regres-
sion models with individual effects. J. Econometrics 170, 76–91.
sup ∥Q̃n∗ (β, δ, f ) − Q̃n (β, δ, f ) Koenker, R., 2005. Quantile Regression. Cambridge University Press, New York, New
∥β−β0 ∥≤ϵn ,∥δ−δ0 ∥≤ϵn ,∥f −f0 ∥∞ ≤ϵn York.
√
− [Q̃n∗ (β0 , δ0 , f0 ) − Q̃n (β0 , δ0 , f0 )]∥ = op∗ (1/ n). Koenker, R., Bassett, G.W., 1978. Regression quantiles. Econometrica 46, 33–49.
Kosorok, M., 2008. Introduction to Empirical Processes and Semiparametric Infer-
The verification of Condition (2.5’B) in Chen et al. (2003) is ence. Springer.
Kotlarski, I., 1967. On characterizing the gamma and the normal distribution. Pacific
parallel to that of (2.5’) of Theorem 5. Finally, to verify Condition J. Math. 20, 69–76.
(2.6’B) in Chen et √ al. (2003), from Gine and Zinn (1990), the Lewellen, J., Lewellen, K., 2016. Investment and cashflow: new evidence. J. Fin.
P ∗ -distribution of n{Q̃n∗ (β̂, δ̂,ˆ
f ) − Q̃n (β̂, δ̂,ˆ
f )} approximates the Quantitative Anal. 51, 1135–1164.
164 S. Firpo et al. / Journal of Econometrics 198 (2017) 146–164

Li, T., 2002. Robust and consistent estimation of nonlinear errors-in-variables mod- Schennach, S.M., 2004b. Nonparametric regression in the presence of measurement
els. J. Econometrics 110, 1–26. error. Econometric Theory 20, 1046–1093.
Li, T., Vuong, Q., 1998. Nonparametric estimation of the measurement error model Schennach, S.M., 2008. Quantile regression with mismeasured covariates.
using multiple indicators. J. Multivariate Anal. 65, 139–165. Econometric Theory 24, 1010–1043.
Ma, Y., Yin, G., 2011. Censored quantile regression with covariate measurement Schennach, S.M., 2014. Entropic latent variable integration via simulation.
errors. Statist. Sinica 21, 949–971. Econometrica 82, 345–385.
Modigliani, F., Miller, M., 1958. The cost of capital, corporation finance and the Stein, J., 2003. Agency information and corporate investment. In: Harris, M., Stulz,
theory of investment. Amer. Econ. Rev. 48, 261–297. R. (Eds.), Handbook of the Economics of Finance. Elsevier/North-Holland,
Newey, W.K., 1994a. The asymptotic variance of semiparametric estimators. Amsterdam.
Econometrica 62, 1349–1382.
Torres-Saavedra, P.A., 2013. Quantile regression for repeated responses measured
Newey, W.K., 1994b. Kernel estimation of partial means and a generalized variance
with error, North Carolina State University, mimeo.
estimator. Econometric Theory 10, 233–253.
van der Vaart, A., Wellner, J.A., 1996. Weak Convergence and Empirical Processes.
Newey, W.K., McFadden, D.L., 1994. Large sample estimation and hypothesis testing.
Springer-Verlag Press, New York, New York.
In: Engle, R.F., McFadden, D.L. (Eds.), Handbook of Econometrics, vol. 4. North
Wang, H.J., Stefanski, L.A., Zhu, Z., 2012. Corrected-loss estimation for quantile
Holland, Elsevier, Amsterdam.
Politis, D.N., Romano, J.P., 1999. Multivariate density estimation with general flat- regression with covariate measurement errors. Biometrika 99, 405–421.
top kernels of infinite order. J. Multivariate Anal. 68, 1–25. Wei, Y., Carroll, R.J., 2009. Quantile regression with measurement error. J. Amer.
Poterba, J., 1988. Financing constraints and corporate investment: comment. Statist. Assoc. 104, 1129–1143.
Brookings Papers on Economic Activity 1, 200–204. Wu, Y., Ma, Y., Yin, G., 2015. Smoothed and corrected score approach to censored
Schennach, S.M., 2004a. Estimation of nonlinear models with measurement error. quantile regression with measurement errors. J. Amer. Statist. Assoc. 110,
Econometrica 72, 33–75. 1670–1683.