Biometrics 73, 1179–1188 DOI: 10.1111/biom.

12675
December 2017

Parametric Modeling of Quantile Regression Coefficient Functions

with Censored and Truncated Data

Paolo Frumento* and Matteo Bottai**

Karolinska Institutet, Institute of Environmental Medicine, Unit of Biostatistics Nobels väg
13, 17177 Stockholm, Sweden
∗ email: [email protected]
∗∗ email: [email protected]

Summary. Quantile regression coefficient functions describe how the coefficients of a quantile regression model depend on
the order of the quantile. A method for parametric modeling of quantile regression coefficient functions was discussed in a
recent article. The aim of the present work is to extend the existing framework to censored and truncated data. We propose an
estimator and derive its asymptotic properties. We discuss goodness-of-fit measures, present simulation results, and analyze
the data that motivated this article. The described estimator has been implemented in the R package qrcm.
Key words: Censored and truncated quantile regression; Conditional quantile function; Quantile regression coefficients
modeling; R package qrcm.

1. Introduction For example, we could define θ = (θ00 , θ01 , θ02 , θ10 ) and
Quantile regression (qr) permits estimating the conditional assume
quantiles of a response variable, given covariates. For example,
a linear model can be used to describe the conditional quantile β0 (p | θ) = θ00 + θ01 p + θ02 p2 ,
function,
β1 (p | θ) = θ10 .
Q(p | x) = x β(p).
T
(1)
This approach, referred to as quantile regression
For any given p ∈ (0, 1), β(p) is a vector of quantile regres- coefficients modeling (qrcm) in the rest of the article,
sion coefficients associated with x. For varying p, β(p) can has two main advantages: (i) it permits describing the
be seen as a set of quantile regression coefficient functions conditional quantile function using a parsimonious set of
describing how each regression coefficient depends on the parameters, while keeping the quantile regression structure
order of the quantile. The large majority of the literature on defined by (1); and (ii) it enables imposing some degree
quantile regression (e.g., Koenker, 2005) has focused on the of smoothness on β(p), thus facilitating interpretation,
problem of estimating a single quantile of interest. When sev- prediction, and inference.
eral quantiles are estimated, results are typically summarized Frumento and Bottai (2016b) propose estimating θ by
graphically. minimizing the integral, with respect to p, of the loss func-
The motivation for this article comes from a study relating tion of ordinary quantile regression. The aim of this work
survival after the age of 65 to birth weight. An estimat- is to extend their estimator to censored and truncated
ing equation for censored, truncated quantile regression data.
(Frumento and Bottai, 2016a) was used to estimate centiles The article is structured as follows. In Section 2, we briefly
0.01, 0.02, . . . , 0.75. The estimated coefficients associated with describe how quantile regression coefficient functions can be
the intercept ( β0 ) and the low birth weight indicator (
β1 ) are modeled. In Section 3, we describe how θ can be estimated
displayed in Figure 1. From a graphical inspection of the fig- with censored and truncated data. We derive the asymptotic
ure, 
β0 (p) appears to be smooth. For example, it could be properties of the proposed estimator in Section 4. We discuss
described by a polynomial function. On the other hand,  β1 (p) goodness-of-fit measures in Section 5, and present simulation
is characterized by a large random variation. Visual inter- results in Section 6. In Section 7, we analyze the dataset that
polation suggests that the regression coefficient is oscillating motivated the article.
around the constant value −2.5.
In their recent article, Frumento and Bottai (2016b) 2. Modeling Quantile Regression Coefficient
suggested using a parametric model to describe quantile Functions
regression coefficient functions through a finite-dimensional We assume that model (1) holds and that β(p) is indexed by
parameter vector θ such that a parameter θ such that

β(p) = β(p | θ). (2) β(p) = β(p | θ) = θb(p) (3)

© 2017, The International Biometric Society 1179

1180 Biometrics, December 2017

Intercept Birth weight < 2.5 Kg

25

0
−1
20

−2
β0(p)

β1(p)
15
^

−3
10

−4
−5
5

0.00 0.25 0.50 0.75 0.00 0.25 0.50 0.75

p p

Figure 1. Quantile regression coefficients associated with intercept (left) and low birth weight indicator (right) at centiles
0.01, 0.02, . . . , 0.75, with pointwise 95% confidence intervals. The dotted line indicates the value zero.

where b(p) = [b1 (p), . . . , bk (p)]T is a set of k known functions θ10 + θ13 (when p = 1). Also note that β1 (p), being a lin-
of p, and θ is a q × k matrix, where q denotes the dimension ear function of p, is constrained to be monotone. Special
of x. The conditional quantile function can be rewritten as cases of interest are θ01 = θ02 = 0, corresponding to a Uni-
form distribution with support (θ00 + θ10 x, θ00 + (θ10 + θ13 )x),
Q(p | x, θ) = xT β(p | θ) = xT θb(p). (4) and θ13 = 0, that imposes homoscedasticity and generates
the asymmetric Logistic distribution, reducing to the conven-
The quantile regression coefficient functions are given by tional Logistic if θ01 = −θ02 , and to the (shifted) Exponential
if θ01 = 0.
βj (p | θ) = θj1 b1 (p) + · · · + θjk bk (p), This simple example demonstrates the flexibility of the pro-
posed modeling approach. Possible choices for b(p) include
j = 1, . . . , q. Neither assumption (1) nor (3) is crucial to the polynomials (p2 , p3 , . . .), known quantile functions (e.g., that
validity of the proposed approach, and non-linear models of a Normal distribution), trigonometric functions (e.g.,
could be assumed instead. cos(2πp)), splines. In most applied situations, the quantile
We now illustrate the potential of this modeling approach regression coefficient functions β(p) can be expected to be
with an example. Consider the following model with a single well approximated by rather simple functions of p that can
covariate x be modeled with relatively few parameters. Frequently, some
of the coefficients can be reasonably assumed to be linear
functions of p, or not to depend on p at all. Various examples
Q(p | x, θ) = β0 (p | θ) + β1 (p | θ)x
of parametric quantile functions are described in the book
by Gilchrist (2000), in Frumento and Bottai’s (2016b) arti-
and assume that cle, and in the documentation of the qrcm R package. More
examples of model building are offered in Section 7.
β0 (p | θ) = θ00 + θ01 log(p) + θ02 log(1 − p),
β1 (p | θ) = θ10 + θ13 p. 3. The Estimator
In this section, we first describe the existing approaches to
Using matrix notation, we write
estimate β(p) for a given value of p, showing how standard
⎛ 1
⎞ quantile regression estimators (Section 3.1) can be gener-

alized to the presence of random censoring and truncation
⎜ log(p) ⎟ θ00 θ01 θ02 0
⎜
b(p) = ⎝ ⎟ and θ = .
(Section 3.2). We then illustrate how to directly estimate θ
log(1 − p)⎠ θ10 0 0 θ13 using Frumento and Bottai’s (2016b) method (Section 3.3),
p and describe how to extend their estimator to censored and
truncated data (Section 3.4).
The resulting conditional quantile function does not have
a closed-form probability density function and does not cor- 3.1. Standard Quantile Regression
respond to any known distribution. The model’s intercept is Throughout the article, T indicates the response variable
unbounded, as log(p) and log(1 − p) have asymptotes, while of interest with quantile function defined by (1). Based on
the coefficient of x is bounded between θ10 (when p = 0) and a N-dimensional sample (ti , xi ), the p-th quantile regression
 Quantile Regression Coefficients Modeling 1181

coefficients are estimated by minimizing function defined by


N 1
−1
LN (β(p)) = N (p − ωti ,p )(ti − xi β(p)),
T
(5) L̄N (θ) = LN (β(p | θ))dp, (8)
i=1 0

where ωti ,p = I(ti ≤ xT

i β(p)). The corresponding gradient
or, equivalently, by finding the zeroes of the corresponding
function is integrated gradient


N 1
−1
SN (β(p)) = N xi (ωti ,p − p) (6) S̄N (θ) = SN (β(p | θ))b(p)T dp, (9)
i=1 0

and generates an unbiased estimating equation for β(p). Note where LN and SN are defined by (5) and (6), respectively.
that SN (β(p)) is not a continuous function of β(p). An effi- Their methods permits estimating the entire quantile process
cient algorithm for this estimation problem was developed by at once, describing it by a finite-dimensional vector θ of model
Barrodale and Roberts (1973) and is implemented in widely parameters. Unlike (6), equation (9) is a smooth function of
available statistical software. its arguments.
Standard quantile regression can be seen as the nonpara-
3.2. Quantile Regression with Censored and Truncated metric version of this estimator and is obtained as a special
Data case of it in two different ways: (i) by carrying out the integrals
Suppose a pair (C, Z) of random variables is such that one in (8) and (9) in a sufficiently small interval (p − , p + ),
can only observe Y = min(T, C) and δ = I(T ≤ C), subject to letting β(p | θ) to be constant in the interval; and (ii) by allow-
Y > Z. The pair (C, Z) is assumed to be independent of T , ing β(p | θ) to be arbitrarily flexible, for example, by using a
given x. As a result of the truncation mechanism, a sample piecewise linear function of p with a large number of knots.
of n observations is drawn from an original sample of size N,
3.4. Quantile Regression Coefficients Modeling with
n ≤ N. The observed quantities are denoted by (zi , yi , δi , xi ),
Censored and Truncated Data
yi > zi , i = 1, . . . , n.
There is no obvious generalization of (5), the loss function Following the same logic of equation (9), we propose estimat-
of standard quantile regression, that could handle censoring ing θ with censored and truncated data by finding the zeroes
and truncation. However, a generalization of (6) exists and is of the integrated estimating equation
given by
1

n  S̄n (θ) = Sn (β(p | θ))b(p)T dp, (10)
−1 p−1
Sn (β(p)) = n xi ωyi ,p + (1 − δi )ωyi ,p
1 − F (yi | xi ) 0
i=1
where Sn corresponds to the estimating equation for censored,
p−1
−ωzi ,p − ωzi ,p (7) truncated quantile regression defined by (7).
1 − F (zi | xi ) To derive an expression for S̄n (θ), we define

where F (· | x) is the conditional cumulative distribution func- p p

tion of T , ωyi ,p = I(yi ≤ xT
i β(p)) and ωzi ,p = I(zi ≤ xi β(p)).
T
Bh (p) = bh (u)du, B̄h (p) = Bh (u)du, h = 1, . . . , k,
Equation (7) contains two set of parameters, β(p), the
0 0
parameter of interest, and F (· | x), that incorporates infor-
mation about all quantiles in (0, p). Estimation is carried and denote B(p) and B̄(p) the corresponding k × 1 vectors.
out in two steps: first, a nonparametric estimate F  (· | x) is Simple algebra shows that (10) can be written as
computed; then, (7) is solved for β(p) letting F = F .
The unbiasedness of (7) is proved in the Appendix. A  
1 − δi  
n

complete description of this estimating equation and its S̄n (θ) = n−1 xi −δi B(pyi ) + B̄(pyi ) − B̄(1)
asymptotic properties can be found in Frumento and Bottai 1 − pyi
i=1
(2016a), to which we refer the reader for details. This method T
1  
for censored and truncated quantile regression is implemented − B̄(pzi ) − B̄(1) (11)
in the ctqr R package, in combination with the pch package 1 − pzi
that is employed to estimate F (· | x) using a form of piecewise
constant hazards model. where pyi = F (yi | xi , θ) and pzi = F (zi | xi , θ). In particular,
pyi corresponds to the order of the quantile such that Q(pyi |
3.3. Standard Quantile Regression Coefficients Modeling xi , θ) = xTi β(pyi | θ) = yi . The value of pzi can be interpreted
Frumento and Bottai (2016b) proposed modeling β(p) as analogously, but can be zero if zi is smaller than the minimum
in (3) and estimating θ by minimizing the integrated loss of the support of T . Note that θ only enters (11) through pyi
1182 Biometrics, December 2017

and pzi . Since pyi and pzi are not generally available in closed uses b(p) = tan[π(p − 0.5)] and describes the quantile func-
form, they can be evaluated using a bisection algorithm as tion of a Cauchy distribution with scale θ. Although this is
described in Frumento and Bottai (2016b). p
a well-defined distribution for any θ > 0, Theorem 1 does not
Unlike (7), equation (11) is a smooth function of its argu- apply, since B(p) = 0 b(u)du is not finite. A proof of Theo-
ments and does not involve nuisance parameters, since θ rem 1 can be found in Newey and McFadden (1994), Theorem
determines the full parametric form of both Q(· | x, θ) and 2.6 (p. 2132).
F (· | x, θ). Estimation is carried on in a single step, by directly To establish asymptotic normality and write the (q ×
solving (11) for θ. This also leads to a much simpler inference k) × (q × k) asymptotic covariance matrix, we stack θ 0 and
compared with the two-steps, semiparametric theory that is θ n , that are q × k matrices, into vectors of length q × k
needed to investigate the asymptotic behavior of equation (7). of the form (θ11 , . . . , θjh , . . . , θqk ), and we redefine S̄n (θ)
Due to censoring, large quantiles may not be identified. The accordingly.
integral in (10) could be computed in an interval (0, pmax ).
However, the parameter θ would not be identified if a lower Theorem 2. (Asymptotic Normality). Suppose that the
bound pmin > 0 was used, because expression (7) carries infor- assumptions of Theorem 1 are satisfied, and that (i) θ 0 is
mation on lower quantiles. an interior point of ; (ii) S̄n (θ) is continuously differen-
Apart from the differences in how estimation is performed, tiable in a neighborhood N of θ 0 , with probability approaching
the censored, truncated quantile regression estimator defined one; (iii) E[||S̄n (θ 0 )||2 ] is finite; (iv) E[supθ∈N ||∇θ S̄n (θ)||]
by (7) can be thought of as the nonparametric version of < ∞; (v) H = E[∇θ S̄n (θ 0 )] is a full rank matrix. Then, for
that defined by (11), obtained by treating θ as an infinite- = E[S̄n (θ 0 )S̄n (θ 0 )T ],
dimensional parameter.

4. Asymptotic Properties √  
n(
d
θ n − θ 0 ) → N 0, (H T −1 H)−1 .
To investigate the asymptotic behavior of the proposed esti-
mator we apply standard method-of-moments theory. Since
Conditions (i) and (ii) are standard; conditions (iii) and (iv)
the model is exactly identified (i.e., there are as many param-
ensure that a central limit theorem can be applied to S̄n (θ).
eters as estimating equations), we take the weighting matrix
For condition (v) to hold, we assume that x and b(p) are non
to be the identity.
singular. The proof of Theorem 2 can be found in Newey and
We denote  as θ 0 ∈  the true population parameter sat-
isfying E S̄n (θ 0 ) = 0, and we write as 
McFadden (1994), Theorem 3.4 (p. 2148).
θ n the estimator
The analytical expressions of H and are rather com-
of θ 0 based on a censored, truncated sample (zi , yi , δi , xi ),
plicated and include unknown quantities, mainly the joint
i = 1, . . . , n. We derive consistency and asymptotic normal-
distribution of (Z, C), fZC (z, c | x), and the conditional prob-
ity by letting n → ∞. Since n ≤ N, the results also hold for
ability of inclusion, P(min(T, C) > Z | x). Instead, the sample
N → ∞.
counterparts of H and can be used to estimate the
asymptotic covariance matrix. The Jacobian matrix, which
 
Theorem 1. (Consistency). Assume that (i)  is a com-
estimates H and is needed to implement the Newton–Raphson
pact set; (ii) E S̄n (θ) = 0 only at θ = θ 0 ; (iii) S̄n (θ)
is  continuous at each θ ∈  with probability one; (iv) algorithm, can be written as
E supθ∈ ||S̄n (θ)|| < ∞. Then,
∂S̄n (θjh )  n  
= n−1 xij xij a(pyi | xi , θ)−1 δi bh (pyi )bh (pyi )
∂θj h
θ n →
p
θ0 . i=1

(1 − δi )bh (pyi )

− Bh (pyi )(1 − pyi ) + B̄h (pyi ) − B̄h (1)
Conditions (i) and (ii) ensure model identification and rely (1 − pyi )2
on the assumption that Q(p | x, θ) describes a proper, well-
bh (pzi )

defined quantile function. This requires the existence of a non- + a(pzi | xi , θ)−1 Bh (pzi )(1 − pzi )
(1 − pzi )2
empty subset  ⊂ Rq×k , in which the quantity 
+B̄h (pzi ) − B̄h (1) (13)

a(p | x, θ) = x θb (p),
T
(12)

representing the derivative of Q(p | x, θ) with respect to p, where xij is the j-th element of xi , (j, j  ) = 1, . . . , q,
is strictly positive. For example, Q(p | θ) = θ cos(2πp) is not (h, h ) = 1, . . . , k, and a(p | x, θ) is defined by (12).
well-defined, being non monotone for all θ. For condition (iii) The sample counterpart of is given by the outer product
to hold, F (· | x, θ) must be a continuous function of θ. This of the vector containing the summands of S̄n ( θ n ).
is generally true provided that b(p) is continuous and the For any given p, the estimated quantile regression
distribution is non-degenerate. Condition (iv) is not trivial: coefficients are  βn (p) = β(p | 
θn ) = 
θ n b(p). This is just a
although b(p) can be unbounded, as is the case when the 
linear transformation of θ n and its covariance matrix is
outcome variable is unbounded, B(p) and B̄(p) must be finite. estimated accordingly. Inference on a two-step semipara-
For example, the following model metric estimator based on (7) would be much more
complicated and computationally demanding (see, e.g.,
Q(p | θ) = θ tan[π(p − 0.5)] Ackerberg et al. , 2012).
 Quantile Regression Coefficients Modeling 1183

Negative values of a(p | xi ,  θ n ) = dQ(p | xi , θ n )/dp indicate enrollment. In real data, both mechanisms can be present,
that the fitted conditional quantile function is not a monotoni- and a different model could be fitted for each cause of
cally increasing function of p. In particular, it is recommended censoring.
to evaluate a(pyi | xi , 
θ n ), where  pyi = F (yi | xi , 
θ n ), that cor- Note that −Z is a left-truncated random variable with
responds to the fitted probability density function of T , truncation time −Y . Under model (a), U is a right-censored
evaluated at yi . variable with indicator of event 1 − δ. Under model (b), C
is a left-truncated, right-censored outcome with truncation
5. Goodness-of-Fit Measures time Z, and indicator of event 1 − δ. To estimate the con-
ditional distributions of Z and C (or U), we used a form
Frumento and Bottai (2016b) propose assessing the model fit
of piecewise constant hazards model implemented by the R
by comparing the distribution of F (ti | xi , 
θ n ) with a U(0, 1)
package pch, version 1.2, to which we refer for technical
distribution. They suggest using a measure of distance, such
details. The described procedure is implemented in the qrcm
as the Kolmogorov–Smirnov or Cramér-von Mises statistic,
package.
and describe a Monte Carlo approach to compute correct
p-values.
6. Simulation Study
The analogous testing procedure can be applied to censored
and truncated data and is described by the following steps. A variety of simulation scenarios were implemented to assess
the finite-sample properties of the described estimator. We
pzi = F (zi | xi ,  pyi = F (yi | xi , 
present results from three different scenarios. To describe the
(1) Define  θn ) and  θ n ),
simulation, we use the notation N(μ, σ 2 ) for the Normal dis-
i = 1, . . . , n. Use the censored, truncated dataset
tribution; U(a, b) for the Uniform distribution; B(a, b) for
(
pzi , 
pyi , δi ) to estimate the distribution of
the Beta distribution; Ber(π) for the Bernoulli distribution;
pti = F (ti | xi , 
 θ n ) nonparametrically (e.g., by calcu-
G(a, b) for the Gamma distribution; and Exp(λ) for the Expo-
lating the baseline survival function of a null Cox
nential distribution; we denote by z(p) the quantile function
model), and compare it with a U(0, 1) distribution. A
of a N(0, 1) distribution.
graphical visualization is recommended and permits
All scenarios included two covariates x1 ∼ U(0, 3) and x2 ∼
discarding severely misspecified models. A formal testing
Ber(0.5), such that the conditional quantile function of T was
procedure involves computing a test statistic V , and
carrying out steps 2–5 to compute the p-value.
(2) Estimate the joint distribution of (Z, C) given x, Q(p | x, θ) = β0 (p | θ) + β1 (p | θ)x1 + β2 (p | θ)x2 .
and use it to compute an estimate ψ xi of ψxi =
P (min(T, C) > Z | xi ), the conditional probabilities of The scenarios differed for the specification of β(p | θ) =
inclusion. Estimate the sample size before truncation as
 −1 [β0 (p | θ), β1 (p | θ), β2 (p | θ)]T , and for the joint distribution
n
 = n n−1
N i=1
 xi
ψ . of (Z, C), given x. The simulation scenarios are described
below. For each scenario, we report the expected truncation
(3) Sample N observations from x with probability 1/ψ x , gen- rate, P(min(T, C) < Z), and the expected proportion of cen-
erating a vector x∗ of covariate values. soring after truncation, P(T ≤ C | min(T, C) > Z).
(4) Simulate N observations t ∗ , z∗ , and c∗ from their con-
ditional distributions, evaluated at x∗ . Apply censoring Scenario (i). We defined β0 (p) = p, β1 (p) = (p − 0.1)2 ,
and truncation, obtaining a final sample (z∗ , y∗ , δ∗ , x∗ ) of and β2 (p) = 1 + (p − 1)3 . The truncation variable Z was gen-
random size n∗ . erated from a U(−1, 1), while the censoring variable was
(5) Estimate the model on the simulated data and compute obtained as C = Z + 2 + W with W ∼ G(2.5, 5). The expected
the corresponding value V ∗ of the test statistic. truncation rate was 0.13, and the expected censoring was 0.16.

Scenario (ii). We defined β0 (p) = − log(1 − p), β1 (p) =

If the data are not truncated, step 2 only requires to estimate √ √
3 p, and β2 (p) = p. We generated Z ∼ B(2, 1), and C =
the conditional distribution of C, and step 3 can be either
 = n (the actual 1 + x1 /3 + x2 /3 + W with W ∼ Exp(1/3). The expected trun-
skipped (letting x∗ = x) or carried out with N
 cation rate was 0.09, and the expected censoring was 0.25.
sample size) and ψx = 1. Steps 3–5 should be repeated a suffi-
cient number of times to estimate the distribution of the test Scenario (iii). We defined β0 (p) = 5 + 2z(p), β1 (p) = 2p +
statistic under the null hypothesis. The p-value of the test is 0.5 cos(2πp), and β2 (p) = 5(p − 0.5)I(p > 0.5). Z was sim-
then computed as the proportion of cases in which V ∗ > V . ulated from a N(x1 /3 + x2 , 32 ) distribution, while C was
In our simulation and data analysis, we chose V to be the U(0, 25). The expected truncation rate was 0.14, and the
Kolmogorov–Smirnov statistic. To implement step 2, we con- expected censoring was 0.24.
sidered two possible characterizations of the joint distribution
of (Z, C) given x: (a) C = Z + U, where U ⊥ Z | x, U > 0; and In scenario (i), C and Z are not independent and C > Z.
(b) Z ⊥ C | x. Situation (a) implies that Z and C are cor- The distribution of C depends on x in scenario (ii), while that
related with P(C > Z) = 1. It represents the most common of Z does in scenario (iii).
scenario, occurring when individuals are enrolled at time Z To fit the correctly specified model, we used b(p) =
√ √
and followed up during a time U, which may also be constant. [1, p, p2 , p3 ]T in scenario (i); b(p) = [log(1 − p), 3 p, p]T in
Model (b) describes censoring due to drop-out (e.g., death scenario (ii); and b(p) = [1, z(p), p, cos(2πp), (p − 0.5)I(p >
by other causes, migration), which may also occur before 0.5)]T in scenario (iii). Based on 2500 Monte Carlo
1184 Biometrics, December 2017

Table 1 Table 2
Simulation results Performance of goodness-of-fit test

N = 250 N = 500 N = 1000 N = 250 N = 500 N = 1000

Scenario coef. p QRCM QR QRCM QR QRCM QR Scenario α α 1 −

 βI 1 −  α 1 −
βII  βI 1 −  α 1 −
βII  βI 1 − 
βII

(i) β0 0.25 .04 .13 .03 .09 .02 .06 (i) .01 .01 .04 .45 .01 .36 .93 .01 .94 .99
0.50 .07 .14 .05 .10 .04 .07 .05 .03 .28 .73 .04 .82 .99 .04 .99 .99
0.75 .11 .12 .08 .09 .05 .06 .10 .07 .50 .82 .09 .93 .99 .09 .99 .99
β1 0.25 .05 .08 .03 .06 .02 .04 (ii) .01 .01 .05 .16 .01 .24 .51 .01 .74 .89
0.50 .08 .10 .05 .07 .04 .05 .05 .04 .18 .33 .04 .50 .69 .04 .90 .97
0.75 .09 .10 .06 .07 .04 .05 .10 .08 .30 .44 .09 .63 .78 .09 .95 .98
β2 0.25 .14 .15 .10 .11 .07 .08
(iii) .01 .01 .65 .11 .01 .95 .26 .01 .99 .61
0.50 .15 .17 .11 .12 .07 .08
.05 .05 .84 .27 .05 .99 .51 .05 .99 .83
0.75 .13 .15 .09 .11 .07 .08
.10 .10 .91 .38 .09 .99 .65 .10 .99 .90
(ii) β0 0.25 .04 .30 .03 .22 .02 .16
0.50 .09 .28 .07 .20 .05 .15 α) and power (1 − 
Empirical probability of type I error ( β) of the
0.75 .19 .35 .13 .25 .10 .19 Kolmogorov–Smirnov (ks) goodness-of-fit test discussed in Section
5, for three different nominal levels α = (0.01, 0.05, 0.10). To mea-
β1 0.25 .07 .15 .05 .10 .03 .08
sure the power, we fitted two different wrong models. In model I,
0.50 .09 .15 .06 .10 .04 .08
β1 (p) was assumed to be a linear function of p. In model II, β2 (p)
0.75 .10 .19 .07 .13 .05 .10
was assumed to be constant.
β2 0.25 .14 .22 .10 .16 .07 .12
0.50 .19 .24 .14 .17 .09 .12
0.75 .24 .32 .17 .23 .12 .16
evaluated the distribution of the Kolmogorov–Smirnov test
(iii) β0 0.25 .32 .42 .22 .29 .16 .20 statistic based on B = 250 Monte Carlo replications. For dif-
0.50 .43 .59 .30 .41 .21 .29 ferent nominal levels, we estimated the level of the test as
0.75 .58 .75 .41 .55 .29 .39 the empirical proportion of cases in which the true model was
β1 0.25 .14 .20 .10 .14 .07 .09
falsely rejected; and the power of the test as the chance of
0.50 .29 .38 .20 .27 .14 .19 rejecting two different misspecified models, namely: a model
0.75 .42 .56 .29 .40 .21 .29 I in which β1 (p) was wrongly assumed to be a linear function
of p; and a model II in which β2 (p) was wrongly assumed to
β2 0.25 .40 .34 .27 .22 .19 .15
not depend on p.
0.50 .40 .68 .27 .47 .19 .33
0.75 .76 .98 .52 .70 .37 .50
Results are summarized in Table 2.The level of the test
was reasonably close to its nominal value, and its power was
Standard errors of quantile regression coefficients estimators sufficiently large even with small sample size.
under simulation scenarios (i)–(iii). N denotes the sample size
before truncation. We compare the qrcm estimator defined by (10) 7. Association between Birth Weight and
and (11), and the qr estimator that solves (7) in two steps. The Survival after Age 65
bias was virtually zero and is not reported. The coverage of 95%
Low birth weight is a well studied risk factor for some diseases
asymptotic confidence intervals was between 0.94 and 0.96 and is
not reported. in adulthood (e.g., Curhan et al., 1996; Zeng et al., 2010)
and is generally associated with low BMI and constitutional
weakness (e.g., Sørensen et al., 1997).
replications, we evaluated the bias and the standard The goal of this study was to investigate the association of
errors of βn (p) = β(p | 
θ n ), the qrcm estimator of β(p), at birth weight with survival after age 65. We used data from a
p = (0.25, 0.50, 0.75), and for three different sample sizes population-based cohort of Swedish men (Laven et al., 2008;
before truncation, N = 250, 500, 1000. For comparison, we Larsson et al., 2015). We selected individuals aged 65–80 at
also evaluated the bias and the standard errors of the qr the time of enrollment, and computed the truncation times zi
estimator β̃n (p), obtained by solving (7) in two steps, using as the differences between the time at enrollment and 65. The
the implementation provided by the ctqr R package. censoring variable was equal to ci = zi + 15 years follow-up.
Results are illustrated in Table 1.The bias was virtually We excluded preterm born, twins, and individuals with
zero for both estimators and is not reported. Due to its under- extreme body mass index (BMI) (<13 or >35 kg/m2 ).
lying parametric structure,  βn (p) was generally more efficient The final sample size was 6854. During follow-up, 3610
than β̃n (p). This reflects the fact that, in practice, β̃n (p) is the deaths were recorded, corresponding to a censoring rate of
nonparametric version of  βn (p). The coverage of asymptotic about 0.47.
95% confidence intervals was very close to its nominal value The following variables were included in the model: birth
and is not reported. weight as a categorical variable (≤2.5, 2.5–4, 4–5, >5 kg);
We assessed the performance of the goodness-of-fit test and BMI measured at the age of 20, which was likely to be
described in Section 5. For each simulated dataset, we an important intermediate variable. We used the following
 Quantile Regression Coefficients Modeling 1185

Table 3 Table 4
Alternative model specifications Summary of model 14

Model b(p) n. of par. p-value ks 1 − − log(p)
1 p 12 .000
Intercept 34.96 (0.30)∗ 17.05 (0.35)∗
2 p, p2 18 .000
Birth weight <2.5 −2.23 (0.58)∗ −
3 p, p2 , p3 24 .000
Birth weight 4–5 −0.30 (0.32) −
4 p, p2 , p3 , p4 30 .004
Birth weight >5 −0.76 (0.70) −
5 p, p2 , p3 , p4 , p5 36 .022
6 p1/2 , (1 − p)1/2 18 .000 Estimates of θ and asymptotic standard errors (in brack-
7 p1/3 , (1 − p)1/3 18 .136
8 p1/4 , (1 − p)1/4 18 .324
ets) based on model 14. The estimated intercept is  β0 (p) =
9 p, cos(πp), sin(πp) 24 .000 34.96 − 17.05 − log(p), while all other coefficients are assumed
10 log(p), log(1 − p) 18 .205 to be constant. The asterisk (∗ ) denotes significance less than 0.05.
11 z(p)I(p ≤ 0.5), z(p)I(p > 0.5) 18 .432 The model is represented graphically in Figure 2. The effect of
the BMI at the age of 20 is modeled using natural cubic splines
12 −
log(1 − p) 12 .000 (coefficients not shown) and is illustrated in Figure 3.
13  − − log(p) 12 .528
14 − − log(p) (only β0 ) 7 .612 larly useful, as it provided a good fit with a small number of
parameters.
Different parametrizations based on the set of functions that define In Model 14, β0 (p) was parametrized as in model 13, while
b(p). An intercept b(p) = 1 was always included. For each model, all other coefficients were assumed not to depend on p:
we report the p-value of the Kolmogorov–Smirnov (ks) statistic,
computed using the procedure described in Section 5. In model 14, 
only the intercept was modeled, while all other coefficients were β0 (p | θ) = θ00 − θ01 − log(p)
assumed to not depend on p. Model 14 is summarized in Table 4
βj (p | θ) = θj0 , j = 1, . . . , 5.
and represented graphically in Figures 2 and 3.

This model has only seven parameters, and always gen-

quantile regression model for the survival time:
erates a well-defined conditional quantile function without
quantile crossing, provided that θ01 > 0. Model 14 showed
Q(p | x) = β0 (p) + β1 (p)w<2.5 + β2 (p)w4−5 a good fit and was selected as the final model. Parameters’
+β3 (p)w>5 + β4 (p)s1 + β5 (p)s2 (14) estimates are reported in Table 4and represented graphically
in Figures 2and 3.
where (w<2.5 , w4−5 , w>5 ) are indicators of birth weight, with The results indicate that low birthweight is associated with
normal weight (2.5–4 kg) serving as reference group. The vari- shorter survival. The loss is quantified as more than 2 years.
ables s1 and s2 form the basis of a natural cubic spline Large birthweight does not appear to be a prognostic factor. A
that models the effect of BMI at age 20 years, such that longer survival is observed when the BMI during young adult-
s1 (22) = s2 (22) = 0. hood is in the normal range, approximately 18–22 kg/m2 .
We first used qr to estimate quantiles of order
0.01, 0.02, . . . , 0.75, solving (7) in two steps using the ctqr 8. Discussion
package. Estimating larger quantiles was not possible, due Applying quantile regression to survival data is becom-
to the large proportion of censoring. Only 155 observations ing increasingly popular. Results are frequently summarized
(about 2% of the total sample size) had an event after the graphically by plotting the estimated coefficients versus the
estimated 75th centile. order of the quantile. The estimated coefficient functions,
We then formulated the parametric models summarized however, are generally non-smooth, which hinders interpre-
in Table 3.In model 1, quantile regression coefficients were tation and may discourage a wider use of quantile regression.
assumed to be linear function of p, as in a Uniform distri- We described a general framework for parametric model-
bution. Although this model may not approximate the true ing of quantile regression coefficient functions, and introduced
distribution in general, it permits testing if β(p) is, on aver- an estimator that can be applied to censored and truncated
age, an increasing or decreasing function. As expected, the data. This approach is more efficient than estimating condi-
intercept was found to be significantly increasing over p. All tional quantiles one at the time, leads to simpler inferential
other coefficients did not appear to depend on p. procedures, and facilitates interpretation of the results. It
Models 2–5 used polynomials of increasing order. Models also permits using distributions that do not have closed-form
6–8 used different roots of p. Model 9 used trigonometric probability density function, offering great flexibility while
functions, and model 10 defined an asymmetric Logistic dis- keeping the structure of quantile regression models.
tribution. Model 11 used the quantile function of a standard The qrcm package in R provides various functions that
Normal distribution, in which the scale parameter differed for can be used for model building, plotting, and goodness-of-
p ≤ 0.5 and p > 0.5. Model 12 used the quantile function of fit procedures. Starting from version 2.0 the package has been
a Rayleigh distribution, and model 13 was a variation of it. expanded to implement the described quantile regression coef-
Models 1–6, 9, and 12 were rejected by the goodness-of-fit ficients modeling approach for censored and truncated data.
test described in Section 5. Model 13 was considered particu- The package is available from the authors upon request.
1186 Biometrics, December 2017

Intercept Birth weight < 2.5 Kg

25

0
−1
20

−2
β0(p)

β1(p)
5

−3
^

−4
1
1

−5
5
0

0.00 0.25 0.50 0.75 0.00 0.25 0.50 0.75

p p

Birth weight 4−5 Kg Birth weight > 5 Kg

3
1

2
1
0

0
β2(p)

β3(p)
−4 −3 −2 −1
^

^
−1
−2

0.00 0.25 0.50 0.75 0.00 0.25 0.50 0.75

p p

Figure 2. Estimated quantile regression coefficient functions under model 14 (see Tables 3 and 4). Continuous lines represent

β(p) = β(p | 
θ ). Dashed lines represent confidence bands. The broken line connects the qr estimates computed at a grid of
quantiles, with confidence intervals indicated by the shaded area. The intercept corresponds to a BMI of 22. The effect of
BMI is illustrated in Figure 3.

QR QRCM
25

25
20

20
Fitted quantile

Fitted quantile
15

15
10

10
5

15 20 25 30 15 20 25 30

BMI at 20 BMI at 20

Figure 3. Association between the BMI at age 20 and the quantiles of survival after age 65, modeled through a natural
cubic spline. Lines correspond to deciles from 0.1 to 0.7, evaluated at the reference category of birth weight (2.5–4 kg). Left:
qr estimator. Right: qrcm estimator, specified as in model 14 (Tables 3 and 4).
 Quantile Regression Coefficients Modeling 1187

9. Supplementary Material Larsson, S. C., Drca, N. Jensen-Urstad, M., and Wolk, A. (2015).
This submission includes the following supplementary Incidence of atrial fibrillation in relation to birth weight
and preterm birth, International Journal of Cardiology 178,
material.
149–152.
r the qrcm R package, version 2.0, for quantile regression
Laven, B. A., Orsini, N.Andersson, S. O., Johansson, J. E., Gerber,
G. S., and Wolk, A. (2008). Birth weight, abdominal obesity,
coefficients modeling,
r the ctqr R package, version 1.0, for censored, truncated
and the risk of lower urinary tract symptoms in a population-
based study of Swedish men, Journal of Urology, 179,
quantile regression, 1891–1895.
r the pch package, version 2.1, which is required by qrcm and Leng, C., and Tong, X. (2013). A quantile regression estimator for
ctqr, censored data, Bernoulli 19, 344–361.
r the R scripts simulations.R and power.R to run the sim- Newey, W. K. and McFadden, D. L. (1994). Large Sample Esti-
mation and Hypothesis Testing, in R. F. Engle and D. L.
ulations.
McFadden. (eds.), Handbook of Econometrics, Vol. 4, North-
Holland: Elsevier Science, pp. 2113–2148.
The R packages are available with this article at the Shen, P. S. (2014). A weighted quantile regression for left-truncated
Biometrics website on Wiley Online Library. The data used and right-censored data, Journal of Statistical Computation
in the article are not public and the authors do not have the and Simulation, 84, 596–604.
right to distribute them. Sørensen, H. T., Sabroe, S., Rothman, K. J., Gillman, M.,
Fischer, P., and Sørensen, T. I. (1997). Relation between
weight and length at birth and body mass index in young
Acknowledgment sdulthood: cohort study, British Medical Journal 315, 1137.
The Authors wish to thank Celia Garcı́a Pareja, whose bril- PMCID: PMC2127730.
liant insight was at the origin of this project. Wang, H. J. and Wang, L. (2009). Locally weighted censored
quantile regression, Journal of the American Statistical
Association 104, 1117–1128.
References Zeng, Y., Zhang, Z., Xu, T., Fan, Z., Xiao, X., Chen, X., et al.
(2010). Association of birth weight with health and long-
Ackerberg, D., Chen, X., and Hahn, J. (2012). A practical term survival up to middle and old ages in china, Journal of
asymptotic variance estimator for two-step semiparametric Population Ageing 3, 143–159. Epub 2011 Mar 5.
estimators, The Review of Economics and Statistics 94,
481–498.
Barrodale, I. and Roberts, F. D. K. (1973). An improved algorithm Received November 2016. Revised January 2017.
for discrete L1 linear approximation, S.I.A.M. Journal on Accepted January 2017.
Numerical Analysis 10, 839–848.
Cheng, J. Y., Huang, S. C., and Tzeng, S. J. (2016).
Quantile regression methods for left-truncated and right-
Appendix
censored data, Journal of Statistical Computation and
Simulation 86, 443–459. DOI: 10.1080/00949655.2015. This appendix shows how to derive expectations under
1016433 http://dx.doi.org/10.1080/00949655.2015.1016433 censoring and truncation, and demonstrate unbiasedness of
Curhan, G. C., Willett, W. C., Rimm, E. B., Spiegelman, D., estimating equations (7) and (10).
Ascherio, A. L., and Stampfer, M. J. (1996). Birth weight The data-generating process involves three random vari-
and adult hypertension, diabetes mellitus, and obesity in US
ables, T , C, and Z, such that Y = min(T, C) and δ = I(T ≤ C)
men, Circulation 94, 3246–3250.
Frumento, P. (2016). pch: Piecewise Constant Hazards Models for
are observed subject to Y > Z. We denote by F (t | x) the con-
Censored and Truncated Data. R package version 1.2, url: ditional cumulative distribution function (cdf) of T , and by
http://CRAN.R-project.org/package=pch F̄ (t | x) = 1 − F (t | x) the corresponding survival function. We
Frumento, P. (2016). qrcm: Quantile Regression Coefficients write the joint probability density function(pdf) of Z and
∞
Modeling. R package version 2.0, url: http://CRAN. C as fZC (z, c | x), and define fZC̄ (z, c | x) = c fZC (z, t | x)dt
z ∞
R-project.org/package=qrcm and fZ C̄ (z, c | x) = −∞ c fZC (t, s | x)dsdt.
Frumento, P. (2016). ctqr: Censored and Truncated Quantile Simple algebra shows that the joint truncated cdf of Z and
Regression. R package version 1.0, url: http://CRAN.
Y is given by
R-project.org/package=ctqr
Frumento, P., and Bottai, M. (2016a). An estimating equa-
tion for censored and truncated quantile regression, FZY (z, y | x) = P(Z ≤ z, Y ≤ y | Y > Z, x) (14)
Computational Statistics and Data Analysis, doi:
10.1016/j.csda.2016.08.015. [In press].
Frumento, P., and Bottai, M. (2016b). Parametric modeling of  
quantile regression coefficient functions, Biometrics, 72, = ψx−1 ψx (z) − F̄ (y | x)fZ C̄ (z, y | x)
74–84, Doi: 10.1111/biom.12410, [Epub ahead of print].
Gilchrist, W. (2000). Statistical Modeling with Quantile Func- subject to y > z. In the formula,
tions, Boca Raton, FL: Chapman & Hall/CRC Press, ISBN
1-58488-174-7.
u
Koenker, R. (2005). Quantile Regression, Econometric Soci-
ety Monograph Series, Cambridge: Cambridge University ψx (u) = F̄ (t | x)fZC̄ (t, t | x)dt
Press. −∞
1188 Biometrics, December 2017

and ψx ≡ ψx (∞) = P(Y > Z | x) is the conditional probability E [Sn (β(p))] = Ex ψx−1
of observing Y . For any function g(z, y | x),  !
⎡ ⎤ × fZ C̄ (xT β(p), xT β(p) | x) F (xT β(p) | x) − p
∞
E [g(z, y | x)] = Ex ⎣ g(z, y | x)dFZY (z, y | x)⎦ , (15) which is zero if and only if F (xT β(p) | x) = p. A similar proof
−∞ can be used to prove that equation (10) yields an unbi- 
ased estimating equation for θ. However, since E S̄n (θ) =
where Ex denotes expectation over the distribution of x. 1
To prove that (7) generates an unbiased estimating equa- 0
E [Sn (β(p | θ))] b(p)T dp, unbiasedness of (7) also implies
tion for β(p), we use (15) and show that that of (10).