European Journal of Operational Research 282 (2020) 1177–1184

Contents lists available at ScienceDirect

European Journal of Operational Research

journal homepage: www.elsevier.com/locate/ejor

Interfaces with Other Disciplines

Quantile Stochastic Frontiers

Mike G. Tsionas
Lancaster University Management School, LA1 4YX, United Kingdom

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, based on Jradi and Ruggiero (2019). Stochastic Data Envelopment Analysis: A Quantile Re-
Received 11 June 2019 gression Approach to Estimate the Production Frontier. European Journal of Operational Research, 278 (2),
Accepted 7 October 2019
385–393] we propose a novel quantile Stochastic Frontier Model (SFM) and develop Markov Chain Monte
Available online 22 November 2019
Carlo techniques for numerical Bayesian inference. In an empirical application to US large banks we doc-
Keywords: ument important differences between the Quantile and the traditional SFM, in terms of several aspects
Productivity and competitiveness of the data. We also document considerable heterogeneity among different quantiles in terms of returns
Efficiency to scale, technical change, efficiency change, technical efficiency, as well as productivity growth.
Quantile Stochastic Frontier model
Bayesian Inference
© 2019 Elsevier B.V. All rights reserved.

1. Introduction Jradi and Ruggiero (2019) and Jradi, Parmeter, and Ruggiero
(2019a,b) propose a Stochastic Data Envelopment Analysis(SDEA)
Quantile regression models are based on conditional quantiles problem, and they introduce the concept of the most likely quan-
(including the median) as a function of certain covariates and, in tile. As they mention: “The usefulness of the SDEA estimator de-
this sense, they are more flexible than least squares which focuses pends on the ability to choose the appropriate τ [quantile]. The
exclusively on the conditional mean. Certainly, although less gen- model can be solved for any quantile but additional information
eral than nonparametric regression, it is an interesting alternative is needed to determine which quantile is most likely.” Their ap-
as economic restrictions can be imposed easily, the method han- proach is based on the Afriat inequalities so that the functional
dles outliers in a transparent manner, and heterogeneity is allowed form of the frontier is left unspecified. Of course, the approach re-
as we focus attention on modeling the entire distribution of the lies on a known distribution for the composed error term, say a
dependent variables given the covariates. normal/half-normal convolution. Jradi and Ruggiero (2019) incor-
Previous work includes Wang and Wang (2013), who ap- porates the Afriat inequalities but leave the functional form un-
plied shape-restricted support vector regression with pinball loss, specified. While it is true that this paper presents the stochas-
Aragon, Daouia, and Thomas-Agnan (2005), Daouia and Simar tic DEA version with the Afriat constraints, the paper also con-
(2007) and Martins-Filho and Yao (2008), who proposed the use siders parametric quantile regression techniques. The key point of
of conditional quantiles of random output. An alternative is pro- the paper was establishing the optimal quantile consistent with
vided by Wang, Wang, Dang, and Ge (2014) in which a two-step the underlying assumed distribution of the composed error. More-
approach is proposed: first, identifying fitted values that minimize over, Jradi et al. (2019a,b), extended the quantile approach for the
asymmetric absolute loss under shape restrictions; and second, normal-exponential stochastic frontier model. I believe these re-
constructing an estimator that links these fitted values. Of course, cent papers should be cited. Also, Jradi et al. (2019a,b) focus on es-
quantiles and quantile estimation have a long history of use in ap- timating the most likely quantile based on a new approach, mainly
plied studies. exploiting the characteristics of a half normal distribution of the
As Behr (2010) noted in the context of banking: “[T]he quan- absolute value of the normal-half normal error term.
tile regression approach allows efficient or almost efficient banks In this paper, we consider explicitly quantile regression in
to apply production relations which may differ strongly from av- the context of stochastic frontier modeling. Specifically, for each
erage or low efficiency banks. Indeed, the empirical regression re- quantile of the conditional distribution of the dependent variable
sults hint for strong differences in regression parameters at differ- (output, say) we allow for the possibility that there is quantile-
ent quantiles, that is for less or more efficient banks. Additionally, specific inefficiency. In particular, the quantile-specific model con-
conditional quantile regressions are very robust compared to con- sists of a specification that relates inputs to output but there is a
ditional mean regressions against outliers.” quantile-specific one-sided error term as well, which accounts for
technical inefficiency. Therefore, the quantile-specific model is a
convolution of asymmetric Laplace-half-normal errors rather than
E-mail address: [email protected]

https://doi.org/10.1016/j.ejor.2019.10.012
0377-2217/© 2019 Elsevier B.V. All rights reserved.
1178 M.G. Tsionas / European Journal of Operational Research 282 (2020) 1177–1184

30
10%
25%
50%
25 75%
90%

20
density

15

10

0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
returns to scale
Fig. 1. Typical conditional posterior densities of q.

a model based exclusively on the asymmetric Laplace model. This then the qth regression quantile (0 < q < 1) is any solution to the
provides more flexibility and, at the same time, it allows for het- linear programming problem:
erogeneity as the technology is allowed to be different across dif-  
min q |yi − xi β| + (1 − q ) |yi − xi β|. (2)
ferent quantiles. The model unifies the problem of selecting the β
{i|yi −xi β ≥0} {i|yi −xi β <0}
most likely quantile of Jradi and Ruggiero (2019) and estimation of
frontier parameters, as well as technical inefficiency in a common Here, xi is a k × 1 vector of regressors, and β is a k × 1 pa-
statistical framework. Moreover, as Jradi and Ruggiero (2019) sug- rameter vector. We denote observations of the dependent variable
gest, we impose the no-quantile-crossing property of Wang et al. as y = [y1 , . . . , yn ] , X = [x1 , . . . , xn ] . In this paper, we extend the
(2014). As the problem of the functional form remains, we pro- model to allow for efficiency estimation:
pose the use of flexible and globally concave Symmetric General- yi = xi β + vi − ui , i = 1, . . . , n, (3)
ized McFadden form (Kumbhakar, 1994). Moreover, regarding the
Bayesian approach, “The use of Bayesian inference in generalized where ui is a non-negative error component representing technical
linear and additive models is quite standard these days. The rela- inefficiency. We denote u = [u1 , . . . , un ] . Roughly, what we want to
tive ease with which Markov chain Monte Carlo (MCMC) methods do is represent a stochastic frontier model in terms of quantiles of
may be used for obtaining the posterior distributions, even in com- the distribution of the dependent variable:
plex situations, has made Bayesian inference very useful and at- yi = xi β (q ) + vi (q ) − ui (q ), i = 1, . . . , n, (4)
tractive. Unlike conventional methods, Bayesian inference provides
with some abuse of notation to indicate that parameters β , noise
one with the entire posterior distribution of the parameter of in-
and inefficiency differ by quantile. This is a general form of quan-
terest. In addition, it allows for parameter uncertainty to be taken
tile regression and allows parameters and inefficiency to vary by
into account when making predictions.” (Yu & Moyeed, 2001, p.
quantile without necessarily using the concept of most likely quan-
438). Additionally, by treating the quantile as a parameter we have
tile. This is done in the interest of users who are interested in the
access to its posterior distribution and, therefore, the most likely
behavior of parameters, inefficiency, productivity growth, etc., for
quantile can be estimated using the posterior mean, median or
all quantiles of the distribution of the dependent variable.
mode. Of course, access to the complete set of quantile estimates
Quantile regression can be given a statistical interpretation if
is retained (for fixed quantile q ranging from, say, 0.1–0.9 with step
we consider the asymmetric Laplace density:
0.1 or lower).   
p(v ) ∝ τ −1 exp −τ −1 |v| qI[0,∞ ) (v ) + (1 − q )I(−∞,0] (v ) . (5)
The likelihood function is:

2. Model 
n 
Lq (β , τ ; y, X ) ∝ τ −n exp −τ −1 |yi − xi β| qI[0,∞) (yi − xi β )
i=1
Quantile regression was introduced by Koenker and Bassett 
(1978) and is based on the idea that given the linear model: 
+ (1 − q )I(−∞,0] (yi − xi β ) . (6)
yi = xi β + vi , i = 1, . . . , n, (1)
 M.G. Tsionas / European Journal of Operational Research 282 (2020) 1177–1184 1179

30 90
10% 10%
25% 25%
50%
80 50%
25 75% 75%
90% 90%
70

20 60

50
density

density
15

40

10 30

20

10

0 0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 -0.04 -0.02 0 0.02 0.04 0.06 0.08
returns to scale technical change
90 45
10% 10%
25% 25%
80 50%
40 50%
75% 75%
90% 90%
70 35

60 30

50 25
density

density

40 20

30 15

20 10

10 5

0 0
-0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 0.1 0.12
efficiency change productivity growth

Fig. 2. Aspects of the model, I.

Next we consider the normal mixture representation of the Based on the likelihood, we have the following posterior:
Laplace distribution. Specifically, we consider the following prob-
pq (β , τ , σu |y, X ) ∝ τ −n σu−n · p(β , τ , σu )
ability density for the disturbance, with the new scale parameter 
σv = (2τ )1/2 : ∞ 
n 
· exp −τ −1 |yi − xi β + ui | qI[0,∞) (yi − xi β + ui )
−1/2 0
p( v i | w i ) ∝ σv2 wi i=1

v2i    
n

× exp − qI[0,∞ ) (vi ) + (1 − q )I(−∞,0] (vi ) , (7) + (1 − q )I(−∞,0] (yi − xi β + ui ) − 2σ1 2 u2i dui , (9)
2σv2 wi u
i=1

where wi follows a standard exponential distribution, viz. p(wi ) = where p(β , τ , σ u ) is the prior. Alternatively, we consider the aug-
exp (−wi ). For Bayesian inference in quantile regression, see mented posterior:
Tsionas (2003), Yang and He (2010) and Yang, Wang, and He pq (β , τ , σu , {ui }ni=1 |y, X ) ∝ τ −n σu−n · p(β , τ , σu )
(2016). 
We assume that, independently of vi and xi , ui follows a half- 
n 
· exp −τ −1 |yi − xi β + ui | qI[0,∞) (yi − xi β + ui )
normal distribution, viz. ui ∼ N+ (0, σu2 ). The composed error, εi =
i=1
vi − ui has density: 
 
n
∞ + (1 − q )I(−∞,0] (yi − xi β + ui ) − 2σ1 2 u2i . (10)
p(εi ) ∝ τ −1 σu−1 u
i=1
 0
  
× exp −τ −1 |εi | qI[0,∞ ) (εi ) + (1 − q )I(−∞,0] (εi ) − 2σ1 2 u2i dui . Notice that since q is considered as a parameter, MCMC draws
u
become available and, therefore, we can have access to the
(8) marginal posterior density of q. As a result, the mean or mode of
1180 M.G. Tsionas / European Journal of Operational Research 282 (2020) 1177–1184

45 10
new model
10%
classical SFM
25% 9
40 50%
75%
90% 8
35

7
30

6
25
density

density
5

20
4

15
3

10
2

5 1

0 0
0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
efficiency efficiency
Classical SFM Quantile SFM
90 450 18

10% 10%
25% 25%
80 400 16 50%
50%
75% 75%
90% 90%
70 350 14

60 300 12

50 250 10
density

density
density

40 200 8

30 150 6

20 100 4

10 50 2

0 0 0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0 1 2 3 4 0 0.2 0.4 0.6 0.8 1
rank correlation of efficiency with classical SFM Heteroskedasticity test p-value 10 -3 Heteroskedasticity test p-value

Fig. 3. Aspects of the model, II.

the marginal posterior distribution of q provides direct access to prior. Finally, assuming that we have a flat prior on q, viz.
the most likely quantile.
p( q ) ∝ 1 , 0 ≤ q ≤ 1 , (13)
Based on (7) we can consider an alternative augmented poste-
rior where the weights {wi }ni=1 are treated as parameters: it is easy to see from (11) that we have the following conditional
posterior:
pq (β , σv , σu , {ui }ni=1 , {wi }ni=1 |y, X ) ∝ σv−n σu−n · p(β , σv , σu )
 2 p(q|β , τ , σu , {ui , wi }ni=1 , y, X ) ∝ exp (Dq ), 0 ≤ q ≤ 1, (14)
n 
1  yi − xi β + ui  n  
· exp − qI[0,∞ ) (yi − xi β + ui ) where D = −τ  
i=1 I[0,∞ ) (yi − xi β + ui ) − I(−∞,0] (yi − xi β + ui ) .
−1
2 σv
2 wi
i=1 Typical conditional densities are shown in Fig. 1 to motivate our

 
n 
n choice for sampling in Section 3.2.
+ (1 − q )I(−∞,0] (yi − xi β + ui ) − 2σ1 2 u2i − wi . (11) It should be noted that treating q as a parameter in (13) and
u
i=1 i=1 (14), allows the most likely quantile (say q∗ ) to be found using the
median, mode or mean of its posterior distribution via standard
Our prior is as follows:
numerical methods based on MCMC.
p(β|σv , σu ) ∝ 1,
  3. Numerical Bayesian inference
(nv +1 ) q
p( σ v ) ∝ σ v exp − 2σv2 , (12)
v
  3.1. General
(nu +1 ) q
p( σ u ) ∝ σ u exp − 2σu2 ,
u
We use MCMC methods of inference, and especially the Gibbs
where nv , nu , qv , qu ≥ 0 are prior parameters. The priors are condi- sampler with data augmentation. The Gibbs sampler provides ac-
tionally conjugate and the prior for β is flat. In our empirical work, cess to the posterior by drawing from the posterior conditional dis-
we set nv , nu = 1, qv , qu = 10−4 which corresponds to an almost flat tributions.
 M.G. Tsionas / European Journal of Operational Research 282 (2020) 1177–1184 1181

5
SGM
Cobb-Douglas
4.5
translog

3.5

3
density

2.5

1.5

0.5

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
*
q
Fig. 4. Marginal posterior densities of the optimal or most likely quantile, q∗ .
 2
Based on (11) we derive the following posterior conditional dis-  
1  yi − xi β + ui
n
tributions: ∝ exp − qI[0,∞ ) (yi − xi β + ui )
2σv2 i=1
wi
pq (σv |β , σu , {ui }ni=1 , {wi }ni=1 , y, X ) ∝ σv
−(n+nv +1 )
exp −
Qv
, (15) 
2σ 2
v + (1 − q )I(−∞,0] (yi − xi β + ui ) . (19)

 (y −x β +u )2 Therefore,
where Qv = qv + ni=1 i iw i [qI[0,∞ ) (yi − xi β + ui ) + (1 − q )
i
I(−∞,0] (yi − xi β + ui )], pq (β|σv , σu , {ui }ni=1 , {wi }ni=1 , y, X )
 

Qu ∝ exp − 2σ1 2 (y + u − X β ) W −1 (y + u − X β ) qno (β ,u )
pq (σu |β , σv , {ui }ni=1 , {wi }ni=1 , y, X ) ∝ σu−(n+nu +1) exp − , (16) v
2σu2 
+ (1 − q )n−no (β ,u ) , (20)
n
where Qu = qu + i=1 u2i , n
  where no (β , u ) = i=1 I[0,∞ ) (yi − xi β + ui ), W = diag[w1 , . . . , wn ]
pq (wi |β , σv , σu , {ui }ni=1 , y, X ) ∝ exp −wi − Ai w−1
i
, wi > 0, (17) The distribution in (20) resembles a multivariate normal, since
when q = 1 (or zero) we have:
where Ai = 1
2σv2
(yi − xi β + ui )2 [qI[0,∞) (yi − xi β + ui ) + (1 − q )  
I(−∞,0] (yi − xi β + ui )], β|σv , σu , {ui }ni=1 , {wi }ni=1 , y, X ∼ Nk βˆ , V , (21)

pq (ui |β , σv , σu , {wi }ni=1 , ; y, X )

where βˆ = (X W −1 X )−1 X W −1 (y + u ), V = σv2 (X W −1 X )−1 . The
(ri + ui )2   “scale” factor of (20) depends on β through no (β , u). In prac-
∝ exp − qI[0,∞ ) (ri + ui ) + (1 − q )I(−∞,0] (ri + ui )
2σv2 wi tice, we can use (21) as a proposal distribution. Suppose the
proposed candidate is β c and MCMC is currently at state β (s) .
1
− u2i , ui ≥ 0, (18) Then we accept the candidate with the Metropolis–Hastings
2σu2 probability:1
where ri = yi − xi β .
⎧    ⎫

⎨ exp − 2σ1 2 (y + u − X β c ) W −1 (y + u − X β c ) qno (β
c
,u )
+ (1 − q )n−no (β
c
,u )
−1 ⎬
  ⎭.
v
min 1, 
⎩ 
exp − 2σ1 2 y + u − X β (s ) W −1 (y + u − X β (s ) ) qno (β (s) ,u ) + (1 − q )n−no (β (s) ,u ) − 1
v

The conditional posterior distribution of β is given by:

pq (β|σv , σu , {ui }ni=1 , {wi }ni=1 , y, X ) 1

For a different algorithm, see Tsionas (2003).
1182 M.G. Tsionas / European Journal of Operational Research 282 (2020) 1177–1184

8 18
SGM SGM
* *
q q
16
7

14
6

12
5

10

3
6

2
4

1
2

0 0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 1.05
returns to scale efficiency
50 60 35
SGM SGM SGM
* * *
q q q
45
30
50
40

25
35
40

30
20

25 30

15
20

20
15
10

10
10
5
5

0 0 0
-0.06 -0.04 -0.02 0 0.02 0.04 0.06 -0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05 -0.04 -0.02 0 0.02 0.04 0.06 0.08
technical change efficiency change productivity growth

Fig. 5. Comparison of classical stochastic frontier and optimal quantile frontier.

3.2. Drawing from the conditional posterior of q When q is a parameter, the enforcement of NQCP is more com-
plicated but can be implemented
 easily as follows. Suppose we
Random draws from the posterior conditional of q, can be re- have the MCMC draws β (s ) , q(s ) , s = 1, . . . , s where s is the cur-
alized as follows. When D < 0, the posterior conditional is expo-  
rent draw. If q(s ) > q(s −1 ) , we need to have:
nential restricted in the interval (0,1) so random draws can be 
obtained easily. When D > 0, we use acceptance sampling from xi β (s ) < xi β (s ) ∀i = 1, . . . , n, (23)
the density f (x ) = 2x, 0 ≤ x ≤ 1, from which random draws are 
otherwise we need: xi β (s ) > xi β (s ) (the posterior probability of ex-
realized as: x = U 1/2 , where U is a standard uniform draw. The act equality is zero). This restriction can be imposed easily as it
draw is accepted with probability: exp2(qDq ) . From these draws, we involves only two sets of β s. If we allow the burn-in phase of the
can reconstruct the marginal posterior density p(q|y, X). Then Gibbs sampler to be long enough, so that we have convergence
we have two options: Either we use the draws for β , {ui }ni=1 , q to the posterior, then imposition of (22) at each observed point
(which means that uncertainty with respect to q is explicitly taken is much easier and holds at all data points.
.
into account), or we find the posterior mean q̄ = E[ p(q|y, X )] = Regarding the specification of the functional form, we notice
−1
S ( s )
S s=1 q , we condition on q = q̄, and we run again the Gibbs that (3) can be a translog or any other functional form which is
sampler. In the second case, we condition on the “most likely” linear in the parameters. Often, this assumption is restrictive and,
value of q. In the first case, which is the one that we recommend, as Jradi and Ruggiero (2019) remarked, SDEA performs much bet-
q is explicitly treated as a parameter and, as we remarked, uncer- ter under misspecification relative to a linear model, as it does
tainty is taken into account. not make assumptions about the functional form. As SDEA, like
the CNLS of Kuosmanen and Johnson (2010) and Kuosmanen and
Kortelainen (2012) allows for monotonicity and curvature, we pro-
3.3. Imposition of no-quantile-crossing property and non-parametric pose to use the Symmetric Generalized McFadden (SGM) func-
functional form tional form, proposed by Kumbhakar (1994) based on Mc Fadden
(1978) and introduced by Diewert and Wales (1987). Since in the
Regarding the no-quantile-crossing property (NQCP) of Wang next section we will use a cost function, we have:
et al. (2014), this can be imposed by using the constraints:
C (w, Y ) = a w + g(w )(bY )2 + w Y˜ + 12 (ϑ  w )Y˜  Y˜ , (24)
xi β (q ) ≥ xi β (q + 1 ) ∀i = 1, . . . , n ∀q ∈ Q ≡ {q1 , . . . , qH−1 }, (22) 1 w Dw
where a, ϑ ∈ RK , b ∈ RM , g(w ) = D ≡ [dkk , k, k = 2, . . . , K] 2 w e ,
where 0 < q1 < · · · < qH−1 < 1, where H is the number of quan- is a symmetric, negative semidefinite matrix, e ∈ RK++ ,  and 
tiles we consider. For any given value of q in Bayesian analysis, the ( =   ) are K × M and M × M matrices, respectively, Y˜ is the
NCQP can be enforced at all observed points via rejection sampling vector of outputs, possibly augmented with other variables (like
(i.e. simply discard the draws for which NCQP in (22) is not satis- a time trend), and M ≥ M is the dimensionality of Y.This function
fied). Alternatively, one can impose that (22) holds at a small num- is linearly homogeneous in prices, and globally concave with re-
ber of observations, and then check whether it holds at all other spect to prices. In practice, we set e = [1, . . . , 1] as the only pur-
points. In practice, this results in considerable savings of time rel- pose of w e is to make sure that linear homogeneity holds with-
ative to naive rejection. out choosing arbitrarily one of the prices as numeraire. The aim of
 M.G. Tsionas / European Journal of Operational Research 282 (2020) 1177–1184 1183

the cost function is mainly to estimate returns to scale, technical ilar but, practically, flat showing that identification of q∗ is quite
change, efficiency change, and productivity growth. Any cost func- difficult, in this instance. We can attribute this fact to misspecifi-
tion is dual to a production function or more accurately a produc- cation of the functional form. The most likely quantile can be es-
tion transformation function and the characteristics of the second timated using the posterior mean or mode and, in this instance, it
can be derived using duality via the cost function. is slightly over 0.30 for the SGM.
Another interesting issue is the comparison of the various mea-
4. Empirical application sures between classical stochastic frontier (represented here by the
SGM functional form) and the corresponding measures for the op-
We apply the new techniques to a US banking data set, pre- timal quantile (q∗ ). This is done in Fig. 5.
viously analyzed by Malikov, Kumbhakar, and Tsionas (2016). The Evidently, returns to scale are much lower in the classical fron-
data is an unbalanced panel with 2397 bank-year observations for tier model (averaging 0.8) whereas they are close to unity, on the
285 large commercial banks operating in 2001–2010, whose total average, for the optimal quantile frontier. In terms of efficiency, the
assets were in excess of one billion dollars (in 2005 U.S. dollars) optimal quantile frontier has a much large concentration of poste-
in the first three years of observation. The data come from Call rior probability around 0.90 whereas SGM has a long left tail from
Reports available from the Federal Reserve Bank of Chicago. There about 0.55 to 0.70. For technical change, efficiency change and pro-
are five inputs (whose prices, W, are available) and five outputs (Y). ductivity growth the optimal quantile frontier yields systematically
We estimate a cost function of the form ln C = ln C (ln W, ln Y, Z ) + larger values compared to SGM, and productivity growth averages
v + u. The functional form is translog and logs of non-performing near 3% whereas the corresponding value for SGM is near zero.
loans as well as equity are included as quasi-fixed netputs (Z). We
also include a time trend to measure technical change.2 We im- Concluding remarks
plement MCMC using 150,0 0 0 passes the first 50,0 0 0 of which are
discarded to mitigate possible start-up effects. In this study, we proposed a Quantile SFM and developed
Reported in Fig. 2 are sample distributions of posterior means MCMC techniques for numerical Bayesian inference, extending
for returns to scale (RTS), technical change (TC), efficiency change (Jradi & Ruggiero, 2019). In an empirical application to US large
(EC) and productivity growth (P G = T C + EC). RTS is defined banks we document important differences between the Quan-
as the inverse cost elasticity, e−1 cy (see Hanoch, 1975), where
tile and the traditional SFM, in terms of several aspects of the
 ∂ C (W,Y ) Ym data. We also document considerable heterogeneity among differ-
ecy = M m=1 ∂ Ym C (W,Y ) . Technical change is defined as T C =
∂ C (W,Y ) 1 M ∂ log C (W,Y ) ent quantiles in terms of returns to scale, technical change, effi-
∂t C (W,Y )
. Notice that e cy = m=1 ∂ log Ym , so it is an elas- ciency change, technical efficiency, as well as productivity growth.
ticity and measures the percentage effect on cost when all outputs Our functional form is based on the Symmetric Generalized Mc-
increase by 1%. Similarly, T C = ∂ log C∂(tW,Y ) measures the percentage Fadden which is flexible and globally concave in input prices, by
change in cost arising from “exogenous factors” or time (viz. neu- construction. The Quantile SFM can be extended to allow for de-
tral technical change). terminants of inefficiency, also known as environmental variables.
Clearly, all measures differ substantially by quantile. Specifically, This is a simple, yet important generalization that could be taken
PG seems to be positive in the 10%, 25% and 50% quantiles, while up in future work.
for the 90% quantile the posterior is multimodal and there is evi-
dence that PG can be negative for some banks. This effect is more Acknowledgments
pronounced for banks in the 75% quantile. The distributions of EC
are also multimodal, and there is evidence of positive EC only for The author is grateful to three anonymous reviewers for several
banks in the 10% and 25% quantile. useful comments on an earlier version.
Overall cost efficiency distributions by different quantile, are re-
ported in the upper-left panel of Fig. 3. In the upper-right panel
Supplementary material
reported are the efficiency distributions for the entire sample from
quantile and traditional SFM. In the lower-left panel, we present
Supplementary material associated with this article can be
the distribution of rank correlation coefficients between efficiency
found, in the online version, at doi:10.1016/j.ejor.2019.10.012
estimates across different quantiles. In the lower-right panel, we
report the posterior distributions of p-values for a heteroskedas-
References
ticity test for the classical and quantile SFM. These p-values in-
dicate that the traditional SFM clearly suffers from heteroskedas- Aragon, Y., Daouia, A., & Thomas-Agnan, C. (2005). Nonparametric frontier estima-
ticity, while this is successfully resolved via the quantile SFM. To tion: A conditional quantile-based approach. Econometric Theory, 21, 358–389.
implement these tests in a Bayesian framework, the cost func- Behr, A. (2010). Quantile regression for robust bank efficiency score estimation. Eu-
ropean Journal of Operational Research, 200(2), 568–581.
tion residuals squared are regressed on all explanatory variables Daouia, A., & Simar, L. (2007). Nonparametric efficiency analysis: A multivariate con-
for each MCMC draw. In turn, we save the p-value for the F-test ditional quantile approach. Journal of Econometrics, 140, 375–400.
that all regressors, except the intercept, are zero. The distribution Diewert, W. E., & Wales, T. J. (1987). Flexible functional forms and global curvature
conditions. Econometrica, 55(1), 43–68.
of these p-values is reported in the lower-right panel of Fig. 3. Hanoch, G. (1975). The elasticity of scale and the shape of average costs. The Amer-
Clearly, heteroskedasticity should be interpreted as specification ican Economic Review, 65(3), 492–497.
error in this context. Finally, we report marginal posterior densi- Jradi, Parmeter, & Ruggiero (2019a). Quantile estimation of stochastic frontiers with
the normal-exponential specification. EJOR.
ties of q, in Fig. 4. We report marginal posterior densities under
Jradi, M., & Ruggiero, J. (2019b). Stochastic data envelopment analysis: A quantile
the assumption that the functional form is SGM, Cobb-Douglas or regression approach to estimate the production frontier. European Journal of Op-
translog. The SGM favors values close to 0.4 on the average. The erational Research, 278(2), 385–393.
marginal posterior densities for the Cobb-Douglas are quite sim- Jradi, S., Parmeter, C., & Ruggiero, J. (2019b). Quantile estimation of the stochastic
frontier model. Economics Letters, 182, 15–18.
Koenker, R. W., & Bassett, G. W. (1978). Regression quantiles. Econometrica, 46,
33–50.
2
We impose monotonicity and concavity at 100 different points randomly cho- Kumbhakar, S. C. (1994). A multiproduct symmetric generalized mcfadden cost func-
sen from the support of explanatory variables ln W, ln Y. In turn, we verified that tion. Journal of Productivity Analysis, 5, 349–357.
the constraints hold at all points in the sample. It was not possible to enforce the Kuosmanen, T., & Johnson, A. (2010). Data envelopment analysis as nonparametric
restrictions using less than about 50 points, in this instance. least square regression. Operations Research, 58, 149–160.
1184 M.G. Tsionas / European Journal of Operational Research 282 (2020) 1177–1184

Kuosmanen, T., & Kortelainen, M. (2012). Stochastic non-smooth envelopment of Wang, Y., & Wang, S. (2013). Estimating α -frontier technical efficiency with shape
data: Semi-parametric frontier estimation subject to shape constraints. Journal restricted kernel quantile regression. Neurocomputing, 101, 243–251.
of Productivity Analysis, 38, 11–28. Wang, Y., Wang, S., Dang, C., & Ge, W. (2014). Nonparametric quantile frontier es-
Malikov, E., Kumbhakar, S. C., & Tsionas, M. G. (2016). A cost system approach to timation under shape restriction. European Journal of Operational Research, 232,
the stochastic directional technology distance function with undesirable out- 671–678.
puts: The case of u.s. banks in 2001–2010. Journal of Applied Econometrics, 31(7), Yang, Y., & He, X. (2010). Bayesian empirical likelihood for quantile regression. An-
1407–1429. nals of Statistics, 40(2), 1102–1131.
Martins-Filho, C., & Yao, F. (2008). A smooth nonparametric conditional quantile Yang, Y., Wang, H. X., & He, X. (2016). Posterior inference in Bayesian quantile
frontier estimator. Journal of Econometrics, 143, 317–333. regression with asymmetric laplace likelihood. International Statistical Review,
Mc Fadden, D. L. (1978). Cost, revenue, and profit functions. In D. L. McFadden, & 84(3), 327–344.
M. Fuss (Eds.), Production economics. Amsterdam: North-Holland Yu, K., & Moyeed, R. A. (2001). Bayesian quantile regression. Statistics & Probability
Tsionas, E. G. (2003). Bayesian quantile inference. Journal of Statistical Computation Letters, 54(4), 437–447.
and Simulation, 73(9), 659–674.