Long-Term Efficiency of Public Service Provision in A Context of Budget PDF

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

Socio-Economic Planning Sciences xxx (xxxx) xxx

Contents lists available at ScienceDirect

Socio-Economic Planning Sciences


journal homepage: http://www.elsevier.com/locate/seps

Invited paper

Long-term efficiency of public service provision in a context of budget


restrictions. An application to the education sector
Laura López-Torres a, Diego Prior b, *
a
Department of Economics and Business, Faculty of Economics and Business Studies, University of Alcala, Plaza de La Victoria, 28802, Alcalá de Henares, Spain
b
Department of Business, Faculty of Economics and Business Studies, Universitat Autònoma de Barcelona, Building B, Campus Bellaterra, 08193, Barcelona, Spain

A R T I C L E I N F O A B S T R A C T

JEL classification: This paper proposes an extension of the non-parametric long-term evaluation of efficiency, the conditional panel
C14 data DEA model, which takes into account the panel structure of the data and, at the same time, incorporates the
C33 role of contextual factors in the estimations. Its application to the education sector for the period analyzed
C61
(2009–2014) shows the utility of this method, since it obtains more representative efficiency scores for the
C67
complete time-period, is more robust to external shocks, and allows improvements to the decision-making
I21
process in the allocation of the budget available for the public education sector. The results are clear and pre­
Keywords:
Public education
sent an evolution towards the convergence of the efficiency scores, precisely in a time period when hard budget
Efficiency constraints severely reduced the resources available for public schools.
DEA
Panel data assessment
Budget constraints
Convergence

1. Introduction time-variant efficiency scores, thus making it possible to break down the
overall time-invariant efficiency score into yearly time-variant effi­
This paper contributes to the literature on measuring educational ciency scores based on stable input and output weights.
efficiency by extending the model proposed by Surroca et al. [1] to Specifically, this paper extends the panel data DEA model in several
compute long-term efficiency scores: the data panel DEA method. The ways. On the one hand, we adopt a conditional approach capable of
main advantage of this model over intertemporal frontiers ([2,3]) is controlling the influence of contextual factors in shaping the efficiency
that, while preserving the intertemporal technology condition, it en­ frontier. To our knowledge, previous literature using a DEA model with
ables us to estimate a single efficiency score per DMU (decision making panel data has not addressed this important issue. On the other hand, we
unit) for the period of analysis, taking into account the panel structure of build on selective convexity in the panel data DEA model, as we are
the data. In addition, the primal version of this model ensures that input dealing with volume and ratio measures in our model. The ability to
and output weights are invariant and representative for the period of allow for heterogeneity in the definition of variables is another great
analysis, conditions that cannot be obtained with other models that are advantage of this proposed extension. In order to obtain acceptable es­
not panel-based. Therefore, we obtain a consistent efficiency estimate timations, we follow the proposal by Olesen et al. [5] to treat the vari­
over time. The stability of the weights is an attractive property when ables presented in ratio form while refining the reference set technology.
long-term evaluations are to be carried out. Furthermore, this model In order to illustrate the benefits of our methodological proposal, we
avoids the volatility of running cross-sectional DEA, as it provides a use a complete panel database composed of public primary schools from
robust estimate. Thus, significant variations (shocks) in the period of an area in the region of Catalonia (Spain). Specifically, and to test the
analysis are smoothed out in the estimate, since average values of the robustness of the model estimating efficiency when external shocks are
entire period are used. We therefore calculate the time-invariant effi­ presented, we examine the period between academic years 2009/2010
ciency score representative of the study period. In addition, building on and 2013/2014, a period when the public education system in Spain, as
the extension described by Pérez-López et al. [4], we estimate in the rest of the world, was affected by budget cuts as a consequence of

* Corresponding author.
E-mail addresses: Laura.Lopezt@uah.es (L. López-Torres), Diego.Prior@uab.es (D. Prior).

https://doi.org/10.1016/j.seps.2020.100946
Received 26 December 2019; Received in revised form 23 July 2020; Accepted 9 September 2020
Available online 15 September 2020
0038-0121/© 2020 Elsevier Ltd. All rights reserved.

Please cite this article as: Laura López-Torres, Diego Prior, Socio-Economic Planning Sciences, https://doi.org/10.1016/j.seps.2020.100946
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

the global financial crisis. Table 1


The economic crisis that hit most countries after 2008 forced public Evolution of the average expenditure per student according to the levels of the
finances to be efficient and sustainable. In this scenario, the role of International Standard Classification of Education. (Data in euros and in relation
public administrations, responsible for the provision of public services, to GDP per capita).
has gained attention in the literature, as several economic reforms have Panel A
been implemented with the aim of conditioning the funding and In Euros 2009 2010 2011 2012 2013 2014
fostering the performance of public services ([6]). In parallel, scientific
Total 6,546 6,427 6,273 5,819 5,593 5,529
literature has centered the debate on defining the elements making up Primary schools 5,298 5,257 5,141 4,846 4,685 4,620
the objective function of public services, as well as on disentangling the High schools 7,175 6,928 6,783 6,230 5,738 5,653
underlying technology of public services for evaluation purposes. These University 10,071 10,147 9,829 8,887 9,100 8,924
concerns are most evident in studies dealing with efficiency assessments Professional education 7,800 7,488 7,084 6,402 6,119 5,822
in a wide range of public services, including education, health care or Panel B
local governments, among others (extensive literature reviews can be In proportion to the 2009 2010 2011 2012 2013 2014
found in Hollingsworth [7] for health care, De Witte and López-Torres GDP pc
[8] for education and Narbón-Perpiñá and De Witte [9,10] for local
Total 28.5 28.2 27.7 26.1 24.8 24.7
services). Primary schools 23.1 23.0 22.7 21.7 20.8 20.6
Education was one of the sectors most affected by the great recession High schools 31.2 30.4 29.9 27.9 25.4 25.3
due to government control of public spending. In the specific case of University 43.8 44.5 43.3 39.9 40.3 39.9
Professional education 33.9 32.8 31.2 28.7 27.1 26.0
Spain, and during the period we analyze (2009–2014), the government
significantly reduced public spending on education by 16.7%, from Source: Spanish Ministry of Education (Estadística del Gasto Público en Edu­
53,375 million euros in 2009 to 44,461 million euros in 2014, to save cación, Ministerio de Educación y Formación Profesional).
almost 9000 million euros (see Fig. 1).
In terms of expenditure per student, Table 1 shows that in the case of the budget for school meals or textbook grants. Under these conditions,
primary education, the average expenditure in 2014 was €4620 per it is likely that this basic public service will not be provided at the same
student, which represented 20.6% of GDP per capita. If we compare level in all regions, which will presumably jeopardize the equity and the
these figures with those for 2009, we observe that the average expen­ principle of equal opportunities. In this context, any action aimed at
diture per student amounted to €5298 (23.1% of GDP per capita), rep­ assessing efficiency and performance in the public education sector is a
resenting a fall in percentage terms of 12.8%. In the period we analyze, priority for economic policy, as it can highlight strengths and weak­
there were continued reductions in educational spending that were nesses in current practices, reveal directions for improvement and, ul­
homogeneous throughout the entire public education system; cuts were timately, lead to a better use of the resources dedicated to the provision
therefore linear and equal (in relative terms) for all public educational of public services ([12]).
institutions. Consequently, our main research question is to test whether or not
The consequences of the budget cuts were especially severe in the the economic crisis has modified the long-term efficiency of public
region of Catalonia (northeast Spain), where investment in public edu­ schools. In other words, the empirical goal of this paper is to determine
cation per student for non-university levels of education remained below whether the severe budget constraints in public expenditure on educa­
the Spanish average (Fig. 2) between 2009 and 2014. tion have led to a reduction in the long-term efficiency score of schools.
Budget cuts had a direct impact on the amount of educational re­ To do so, we apply the extended version of the panel data DEA presented
sources. As an example, and according to the national statistics on public before using data from public primary schools in Catalonia, for the
education from the Ministry of Education in Spain [11] during this period 2009–2014. We will analyze this phenomenon with the most
period, some schools lost up to 2.8% of their teaching staff and one-third stable comparison conditions possible and by extending the innovative
of the hours worked by the educational psychology team. In addition, efficiency model that we explain in full below.
substantial cuts were made to schools’ operational expenditures and to Our second goal, following Guccio et al. [13,14], is to analyze the

Fig. 1. Public spending on education in Spain during the period of analysis. (data in thousands of €)
Source: Spanish Ministry of Education (Subdirección General de Estadística y Estudios del Ministerio de Educación y Formación Profesional).

2
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

Fig. 2. Public expenditure per student. Non-university education (in euros).


Source: Spanish Ministry of Education (Estadística del Gasto Público en Educación, Ministerio de Educación y Formación Profesional).

convergence patterns in the efficiency scores across the analyzed period. 6 provides concluding remarks.
To address the second objective, we use the β-convergence and
σ -convergence tests developed by Barro and Sala-i-Martin [15,16]. We 2. Review of the literature on efficiency in education
aim to expand the evidence in the literature by analyzing the link be­
tween efficiency and convergence during a period of crisis. This will be The analysis of efficiency in the education sector has a long tradition
of great interest to policymakers who need to assess the stability and in academic literature (see historical reviews in Worthington [19],
convergence of the education system and promptly decide corrective Johnes [20], Emrouznejad et al. [21] and, more recently, in De Witte
interventions. Only a few papers have attempted to investigate effi­ and López-Torres [8] and Emrouznejad and Yang [22]). Following
ciency in terms of convergence/divergence, albeit with respect to higher Hanushek [23–25], previous works used an approach based on esti­
education institutions ([13,14,17,18]). Thus, our goal is to provide mating the educational production function. However, this approach
additional findings on efficiency convergence patterns in the Catalan does not explore the presence of potential inefficiency, since it is based
public primary education system during the crisis period. on average measures. Frontier techniques (parametric and
The results indicate that inefficiency is high, as we found an output- non-parametric) overcome this limitation and entail constructing a
oriented average inefficiency score of 1.38 during the analysis period boundary, composed of the most efficient units, that serves as a
(conditional time-invariant estimation); that is, with the same level of benchmark for comparison purposes. Among these models, the
educational resources, schools should be able to improve student out­ non-parametric approach has been most frequently adopted in the
comes (measured by students’ grades in standardized skill tests at the educational context due to its ability to handle multiple intermediate
end of primary education) by 38%, on average. Regarding the time- outputs and its flexibility, which makes it particularly suitable in an area
variant efficiency, our findings show that in a period of economic where the production technology is difficult to define ([26]). In this
recession schools tend to improve their management efficiency as a sense, data envelopment analysis (DEA), free disposal hull (FDH), and
result of measures to control and tighten public spending. In addition, partial frontiers (mainly order-m and order-α) are the most frequently
we found convergence in efficiency scores due to schools’ efforts to use used models in empirical studies.
the available resources efficiently, especially during the toughest years Starting from the pioneering works by Bessent and Bessent [27],
of the economic crisis. Charnes et al. [28,29] and Bessent et al. [30], empirical studies on ef­
This paper contributes to enrich the existing literature on efficiency ficiency in education have grown in importance. Not only has this body
in education in the following ways: first, from the methodological point of literature contributed to a better understanding of the educational
of view, we develop a novel methodology that extends the panel data factors that influence students’ academic achievement, but it also pro­
DEA model proposed by Surroca et al. [1] by adding the role of vides useful information for public administration decision makers.
contextual factors into the efficiency estimation and allowing for the The application of this type of methodology has evolved towards
inclusion of heterogeneity in the definition of variables. The paper also more sophisticated and robust models that allow a more accurate esti­
advances in the study of convergence patterns of efficiency in the mate of efficiency. Regarding the role of contextual factors, the meth­
educational environment. This joint approach is not very common in odology has advanced from the application of second-stage models [31]
previous education literature. Second, in terms of practical implications, to conditional models [32,33] in order to incorporate environmental
the results of the analysis can shed light on the public education system factors into the estimation of efficiency scores and explore their poten­
at the institutional level, mainly with regard to the appropriate alloca­ tial influence without assuming the ungranted separability condition of
tion of the available budget, especially in a context of budgetary two-stage approaches. Applications of this method in the education
constraints. sector can be found in Cherchye et al. [34], Haelermans and De Witte
The rest of this paper is organized as follows: the following section [35], De Witte and Kortelainen [36], López-Torres and Prior [37],
presents a brief summary of the literature on efficiency in education. Cordero et al. [38] and López-Torres et al. [39].
Section 3 presents the methodological development we propose. Section Nevertheless, most of these studies adopt a cross-sectional approach,
4 describes the main characteristics of the dataset and the variables which makes it difficult to obtain a long-term view of the efficiency over
included in the analysis. Section 5 discusses the results. Finally, Section time and conclude with a stable efficiency score regardless of the period

3
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

of analysis. Some papers in the literature try to surpass this limitation by [y0 = y01 , …, y0m , …, y0M ∈ RM
+ ] denote the variables corresponding to the
applying a dynamic index to reveal total factor productivity changes observed unit under analysis. With a panel data structure, we define a
over time and decompose several measures of productivity growth into new variable τ (τ = 1, …, T) representing the corresponding time period
components of technological and efficiency change through ratio-based for the inputs and the outputs: [xs,1 τ , …, xs,n τ , …, xs,Nτ ∈ RN+ ] and [ys,τ = ys1τ ,…
productivity indexes. One example is the Malmquist Index ([40]).1
,ys,mτ ,…,ys,Mτ ∈ RM
+ ]. Following the proposal of Surroca et al. [1], the primal
Despite its multiple applications, this index presents some drawbacks
(see Bjurek et al. [43] for more details). An alternative is the application variable returns-to-scale time-invariant panel data DEA evaluation is as
of the Hicks-Moorsteen index, proposed by Bjurek [44], which is defined follows:
as a ratio of a Malmquist output-oriented index over a Malmquist ∑
N

input-oriented index. Despite its good theoretical properties, this index Min
ti ti
hti,vrs = vti0 + vtin x̃on ;
um ,vn
is seldom applied in the literature on efficiency in education. Empirical n =1

applications can be found in Arjomandi et al. [45] and Aparicio et al.


s.t.
[46].
Although the Hicks-Moorsteen index overcomes the limitations of ∑
M
the Malmquist index, it is based on an efficiency estimate that uses utim ỹom = 1; (1)
different weights for inputs and outputs in each period/year of analysis. m=1

This fact makes it difficult to interpret the efficiency score and the total

M ∑
N
factor productivity change index as indicators of long-term evaluation, utim ys,mτ − vti0 − vtin xns,τ ≤ 0; s = 1, …, S; τ = 1, …, T;
since the weights are not stable throughout the analyzed period. The m=1 n =1
panel data DEA model presented below in Section 3 overcomes this
limitation and is a structurally more robust model, since the weights are utim , vtin ≥ 0; m = 1, …, M; n = 1, …, N.
the same throughout the period of analysis, thus favoring the interpre­
tation of the efficiency index obtained as an indicator of long-term where ỹom is the average value, corresponding to the output m, for the
( )
evaluation. ∑ /
complete time period T ỹom = Tτ=1 yo, τ
m T and x̃on is the average value,
3. Methodology (
corresponding to the input n, for the complete time period T x̃on =
An important decision to be made before starting to estimate in­ )
efficiency levels relates to the specification of the prevalent technology. ∑T o,τ
/
τ=1 xn T . Applying program [1], one time-invariant efficiency co­
This specification is not trivial, as it has direct implications for the ef­
ficiency levels to be found. With non-parametric technology, DEA efficient (hti,vrs ), M output weights (uti1 , …, utiM ), and N input weights, (vti1 ,
models operate with virtual points, thus establishing linear combina­
…, vtiN ), are estimated for each unit. Besides determining the
tions among real observations. As a result, each inefficient unit is related
time-invariant stable common set of weights, program [1] has three
to another one, or a convex combination of other units, that is more
additional properties: a) it is less dependent on the specific values of the
efficient.
variables in one particular year; b) it ensures that no changes in the
The literature highlights some important limitations of non-
valuation system (the input and output weights) take place across time
parametric frontier estimation methods: the curse of dimensionality,
periods; and c) the consideration of average values does not imply any
their lack of statistical properties – as they are deterministic in nature –
loss of information, because the reference technology has not been
and the potential impact of outliers in the estimations ([47,48]).
modified. However, despite these advantages, we must consider that this
Various proposals have been put forward to overcome these limita­
methodology is based on Tulkens’ intertemporal analysis ([2]), and
tions. In this paper, we are interested in a phenomenon––the conver­
therefore it is assumed that there is no technological change over time.2
gence patterns in the efficiency scores across the analyzed period––that
For computation reasons, it is convenient to define the dual program
in optimal conditions would be analyzed in the most homogeneous
of [1], which provides the output-oriented variable returns-to-scale time-­
possible way. For this reason, we extend the recent proposal to use panel
invariant panel data DEA model:
data DEA ([1]) to consider the estimation of efficiency scores with the
following characteristics: i) ability to operate with unconditional as well Max
s,τ
hti,vrs = αti ;
as conditional estimations; ii) offering time invariant and time variant
λ

efficiency scores; and iii) allowing for technologies that present vari­ s.t.
ables in volume and in ratio form; in other words, the ability to define a
technology with selective convexity, following Podinovski [49]. One of ∑
S ∑
T
the outstanding characteristics of this approach is that the results are λs,τ ys,mτ ≥ αti ỹom ; m = 1, …, M; (2)
robust to the presence of outliers and temporal random shocks, and s=1 τ = 1

provide a specific efficiency score for each unit, representative of the



S ∑
T
complete time period being analyzed. Hence, the interpretation of the λs,τ xs,n τ ≤ x̃on ; n = 1, …, N;
results obtained is not far from what can be obtained from a fixed-effects s=1 τ = 1

parametric regression.

S ∑
T
λs,τ = 1;
3.1. Panel data DEA model s=1 τ = 1

λs,τ ≥ 0.
We now introduce some notations. Assume that for S units (s = 1,…,
S) there are N inputs [xs = xs1 , …, xsn , …, xsN ∈ RN+ ]producing M outputs where λs,τ stands for the activity vector in an intertemporal technology.
[ys = ys1 , …, ysm , …, ysM ∈ RM 0 0 0 N
+ ]. Let [x1 , …, xn , …, xN ∈ R+ ]and

2
In an analysis of productive processes affected by important technological
1
See Färe et al. [41] and Grosskopf [42] for historical overviews. changes, it would be necessary to control for this source of heterogeneity.

4
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

3.2. Time-varying panel data efficiency scores b is the smoothing bandwidth. The conditional efficiency estimation
implies the restriction of the technology to consider only those units that
Following Pérez-López et al. [4] the efficiency coefficients estimated happen to have their contextual variables in between the two extremes
in programs [1,2] can be defined as time-invariant as they are unique and delimited by the respective bandwidth:
representative of the complete time period under analysis. However, as
(z − b) ≤ z0 ≤ (z + b) (6)
we are interested in analyzing to what extent there is convergence on the
levels of efficiency, it could be of interest to know the variations in the The conditional output-oriented variable returns-to-scale time-invariant
efficiency coefficients during the different time periods. In this panel data DEA model:
sub-section, we demonstrate that from the time-invariant efficiency coef­ ti,vrs
ficient it is possible to develop a time-variant series of efficiency coefficients.
ti
Max̂
h =α
̂ ;
λŝ,τ̂
To do so, we define the following Proposition:
Proposition. Assuming the existence of just one output (M = 1), and s.t.
that ỹ* is the maximum level of output at the frontier, the time-invariant
panel data efficiency coefficient hti,vrs can be defined as hti,vrs = ỹ*/ o . In ̂S ∣(z− b≤z0 ≤z+b) ̂
ỹ ∑ T ∣(z− b≤z0 ≤z+b)

(7)
ti
these circumstances, the time-invariant panel data efficiency coefficient λŝ,τ̂ yŝ,τ̂
m ≥̂ α ỹom ; m = 1, …, M;
is dependent on the specific levels of efficiency coefficients corre­ ̂s∣(z− b≤z0 ≤z+b) ̂τ ∣(z− b≤z0 ≤z+b)

sponding to each year (the time-variant panel data efficiency coefficients).


Proof. Assuming the existence of one output, the time-invariant panel ̂S ∣(z− b≤z0 ≤z+b) ̂
∑ T ∣(z− b≤z0 ≤z+b)

data efficiency coefficient is: λŝ,τ̂ xnŝ,τ̂ ≤ x̃on ; n = 1, …, N;
̂s∣(z− b≤z0 ≤z+b) ̂τ ∣(z− b≤z0 ≤z+b)
y*
hti,vrs = (3)
ỹo ̂S ∣(z− b≤z0 ≤z+b) ̂
T ∣(z− b≤z0 ≤z+b)
∑ ∑
Both the numerator and the denominator are averages of the values λŝ,τ̂ = 1;
corresponding to each of the years: ̂s∣(z− b≤z0 ≤z+b) ̂τ ∣(z− b≤z0 ≤z+b)

∑T *,τ ∑T o,τ
y y
ỹ*m = τ=1 m ; ỹom = τ=1 m (4) λŝ,τ̂ ≥ 0.
T T
It is clear that the unconditional efficiency score from expression [2]
After introducing these averages into the original expression of the
will always exhibit equal or more inefficiency than the conditional ef­
time-invariant panel data efficiency coefficients and some manipulations, ti,vrs
the mathematical expression [2] concludes as follows: ficiency score from expression [7] (meaning, hti,vrs ≥ ̂
h ) because,
depending on the value of the smoothing bandwidth b, the left-hand side

T
hti,vrs = htv,1 w1 + … + htv,τ wτ + … + htv,T wT = htv,τ wτ (5) of program [7] will have fewer components than the left-hand side of
τ=1 program [2].
The proof demonstrates that time-invariant panel data efficiency is the
weighted average of the time-variant panel data efficiency coefficient. 3.4. Extending the panel data DEA: selective convexity
Q.E.D.
Section 3.1 presents the variables selected as output and inputs. This
selection is well underpinned in previous literature, but it presents a
3.3. Extending the panel data DEA: the conditional model problem that must be considered in the definition of the models to be
estimated. Table 4 in Section 4.2 below shows that most of the variables
Primal and dual programs defined in sub-section 2.1 can be labeled can be characterized as volume variables (meaning they have a
as unconditional models, as they do not restrict any possible unit to be the magnitude that expresses the size of the unit). These variables are y1, x1
benchmark of the DMU under analysis. This implies that, when the level and x2. However, the third input, x3, is an ordinal variable that reflects
of efficiency achieved is dependent from specific environmental condi­ the family input —Parental Education— and goes from 0 to 8 but
tions, unconditional models can overestimate the level of inefficiency if without providing information about the level of the inputs consumed.
the benchmark of an inefficient unit is operating in a more favorable This situation can bias the results3 and according to Olesen et al. [5],
environment. To control for these circumstances, Cazals et al. [47] and there are two possible ways to control this effect. One of them is to
Daraio and Simar [32,33] developed a conditional efficiency approach to define a non-convex technology, which means comparing real units with
account for those variables in the same model of efficiency estimation. real units, but this solution requires a large sample of DMUs and un­
The basic rationale behind conditional models is to compare the unit fortunately this is not our case. Olesen et al. [5] propose a second
under analysis with other units that present similar levels in the possible way forward, known as ‘selective convexity’. In our specific
contextual or environmental factors (Z ∈ Rr ). case, the selective convexity criterion implies the requirement to define
The conditional approach has become very popular in the recent an additional constraint when defining the technology: being ̂ n an
literature on efficiency measurement in education (e.g. Ref. [34,36, ordinal input, the selective convexity criterion is similar to what is
50–53]). To estimate the conditional efficiency scores, smoothing considered in a non-convex technology, say xn̂ s,τ < x̃n̂ o , which means
techniques are needed such that, in the reference samples, observations that the benchmark chosen as a peer of the unit under analysis has to
with comparable z-values have to be chosen. Therefore, this approach operate with a level of this input not greater than the unit under
relies on the estimation of a non-parametric kernel function to select the assessment. Introducing this additional requirement to model [7] gives
appropriate reference partners and a bandwidth parameter h using a rise to our selective conditional output-oriented variable returns-to-scale
method with some bandwidth choice. As all our Z variables are time-invariant panel data DEA model, which is as follows:
continuous, we apply the method proposed by Badin et al. [54], based
on the least squares cross validation procedure (LSCV) developed in
Racine and Li [55], Hall et al. [56] and Li and Racine [57]. 3
When the output is a volume variable and the input a ratio, large DMUs
Following De Witte and Kortelainen [36], we employ the Epa­ appear to be more efficient. When the output is a ratio and the input a volume
nechnikov kernel density function (Kb̂ (z,zi ) = b− 1 {K[(Z − zi ) /b)]}where variable, small DMUs will present a better level of efficiency.

5
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

̂ ti,vrs ̂ ti ̂ tv,τ− 1
Max ̂
h =α
̂ ; Ln ̂
hs is the natural logarithm of the time-variant selective condi­
λŝ,τ̂
tional output-oriented efficiency score of the DMU s in period τ − 1; α
s. t. and β are the parameters to be estimated, being β the convergence
parameter; and ετs is the error term. Following Kumar and Russell [64],
̂S ∣(z− b≤z0 ≤z+b) ̂T ∣(z− b≤z0 ≤z+b) we used generalized least squares (GLS) to run the estimations in case
( ) ( ) errors are correlated and/or there is inequality in variances.
̂S ∣ xn̂ s,τ <x̃n̂ o ̂T ∣ xn̂ s,τ <x̃n̂ o
∑ ∑ ti
The σ-convergence represents the estimation of the cross-sectional
λŝ,τ̂ yŝ,τ̂
m ≥ ̂
̂
α ỹom ; m = 1, …, M; (8) dispersion. It indicates the speed with which the efficiency of a DMU
̂s∣(z−( b≤z0 ≤z+b) ) ̂τ ∣(z−( b≤z ≤z+b)
0
) converges with the average efficiency level of the sample. In our case we
̂s∣ xn̂ s,τ <x̃n̂ o ̂τ ∣ xn̂ s,τ <x̃n̂ o can define the σ -convergence as:
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√ ( tv,τ
)2
√∑S
̂S ∣(z− b≤z0 ≤z+b) ̂T ∣(z− b≤z0 ≤z+b)
√ s=1 Ln ̂ ̂
h s − uτ

̂S ∣(xn̂ s,τ <x̃n̂ o ) ̂T ∣(xn̂ s,τ <x̃n̂ o ) στ = (11)
∑ ∑ S
λŝ,τ̂ xnŝ,τ̂ ≤ x̃on ; n = 1, …̂
n …, N;
̂ tv,τ
̂s∣(z− b≤z0 ≤z+b) ̂τ ∣(z− b≤z0 ≤z+b) where Ln ̂ h s is the natural logarithm of the time-variant selective
̂s∣(xn̂ s,τ <x̃n̂ o ) ̂τ ∣(xn̂ s,τ <x̃n̂ o ) conditional output-oriented efficiency score of the DMU s in period τ, S is
the total number of DMUs considered in the study, and uτ is the sample
̂ tv,τ
̂S ∣(z− b≤z0 ≤z+b) ̂T ∣(z− b≤z0 ≤z+b)
( ) ( ) mean of Ln ̂ h s . Therefore, there will be σ -convergence if στ decreases
̂S ∣ xn̂ s,τ <x̃n̂ o ̂T ∣ xn̂ s,τ <x̃n̂ o over time.
∑ ∑
λŝ,τ̂ = 1;
̂s∣((z− b≤z0 ≤z+b) ̂
4. Data and variables
) τ ∣((z− b≤z0 ≤z+b)
)
̂s∣ xn̂ s,τ <x̃n̂ o ̂τ ∣ xn̂ s,τ <x̃n̂ o
The methodological approach described above was applied to
calculate long-term efficiencies in the Catalan public education sector.
ŝ,τ̂
λ ≥ 0. We selected the education sector for the analysis in view of its major
As the left-hand side component of restrictions in model [8] will have importance in terms of cost, economic and social returns, because it is a
fewer potential DMUs to be combined than programs [2,7], we can basic public service ([65]), and because its performance needs to be
expect to estimate a lower level of inefficiency, which implies that: studied in greater detail.

ti,vrs ̂ ti,vrs
hti,vrs ≥ ̂
h ≥̂
h (9)
4.1. Sample description

3.5. Analysis of convergence In this study we use data from public schools in the Spanish region of
Catalonia. The Spanish education system is decentralized across 17 re­
In this subsection we describe the tests used to analyze the patterns of gions (called Comunidades Autónomas) that have the necessary compe­
convergence in the efficiency of schools. The idea is to extend the evi­ tences for funding and monitoring public and semi-private (publicly
dence found in the efficiency analysis by looking at the dynamics of funded) schools. The role of the national government in this decentral­
efficiency in terms of convergence of the Catalan public schools in the ized system is limited to legislating and establishing a minimum com­
considered period. In this line, Barro and Sala-i-Martin [15,16] proposed mon educational framework for the country. Public schools are funded
two tests for convergence: β-convergence and σ-convergence. These tests by taxpayers and managed by the Department of Education of each re­
have been previously applied in the efficiency literature (see examples gion. As we stated in the introduction, our analysis is restricted to public
for the banking sector in Refs. [58,59,60,61and62] and, for the educa­ education institutions, as they were more likely to suffer from the cut­
tional context, in Ref. [13,14,63and46]). backs introduced in response to the economic crisis. Moreover, we focus
The β-convergence is determined from the regression of the growth our analysis on primary schools. As in the rest of the country, public
rate on the initial level of any variable which, in our case, is the effi­ primary education in Catalonia is compulsory and free from six to twelve
ciency score. In our context, β-convergence establishes a relation be­ years old. Primary education encompasses six academic years divided
tween the efficiency score with respect to the initial period of analysis. into three cycles, each of two academic years.4
The purpose is to investigate whether the performance of schools that Within the Catalan system, there are several Regional Educational
had poorer levels of efficiency in the first period grows at a faster rate Areas (REAs), which are agencies––like educational dis­
than that of schools with a better initial score. Therefore, if the coeffi­ tricts––answerable to the Department of Education in Catalonia. These
cient is negative and statistically significant, there is a convergence; if REAs are managed by a group of education inspectors who provide
the relationship is positive, there is a divergence. The regression func­ schools with assistance and help them conduct the homogeneous stan­
tion for computing β-convergence in the case of the analysis of efficiency dardized skills tests at the end of primary education (when students
using the conditional panel data DEA time-variant estimates is as finish the sixth course).
follows:

̂ tv,τ ̂ tv,τ− 1
̂ tv,τ− 1
Ln ̂h s − Ln ̂
hs = α + βLn ̂
hs + ετs (10) 4
Primary education in Spain consists of six academic courses. These six
courses are divided into three cycles, each of two courses. The first cycle, called
̂ tv,τ the initial cycle, includes the first and second years of primary education; the
where Ln ̂h s is the natural logarithm of the time-variant selective
second, the intermediate cycle, covers the third and the fourth courses, and the
conditional output-oriented efficiency score of the DMU s in period τ;
third one, or upper cycle, the fifth and the sixth years of primary education.
According to the Spanish Educational Law, LOMCE 2013, students should ac­
quire specific competences in each cycle. The law also establishes that a student
can repeat only one course during the entire primary education stage.

6
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

To carry out our analysis, we use school level data from the Vallès In line with previous papers, grades obtained by students in stan­
Occidental REA.5 This REA consists of a total of 23 municipalities and is dardized tests are used as an output indicator ([71]).9 As Hoxby [72]
located in the surroundings of Barcelona, the capital of Catalonia. We states, the most important argument for opting for this proxy as an
consider that it is a representative sample to run our analysis, as this REA output is possibly that both policy makers and parents use this criterion
includes more than 100 public primary schools. In addition, it offers a to evaluate school performance. At this point it is worth noting that as
representative picture of the population of public primary schools in we are working at school level, we computed an aggregated indicator of
Catalonia and Spain, as it is a mixture of metropolitan cities and agri­ grades by school. Thus, for each school we first calculated the average
cultural towns. In order to contextualize and externally validate our grade obtained by students considering the four modules assessed in the
results, Table 2 presents some figures to show the similarity between the standardized skill tests at the end of primary education (mathematics,
selected REA, Vallès Occidental, Catalonia and Spain during the period of Catalan, Spanish and English language). The score ranges from 0 to 100
analysis.6 points and the pass mark for each competence assessed is 50 points.
Our data comes from the Evaluation Council of the Education System Then, we computed a single output labeled aggregated average grade by
in Catalonia (Consell Superior d’Avaluació del Sistema Educatiu). As multiplying the school’s average grade by the number of students
stated before, in the sixth year of primary education all students are enrolled in the sixth grade of primary education, the ones who took the
tested on their basic skills in mathematics, Catalan, Spanish and a tests.
foreign language (English). The Evaluation Council and the inspectors in The selection of variables for inputs and environmental factors can
each REA conduct a comprehensive uniform standardized diagnostic be complex and, in some cases confusing, given that the literature is
assessment of students.7 We use students’ grades in these standardized diverse and does not specify precisely how to discriminate between
tests as our main output variable, together with other variables (inputs them. In this study, we based our decision on the following criterion:
and environmental factors) that we present below, in order to analyze input variables must fulfill the requirement of monotonicity (i.e., ceteris
the effect of the economic crisis on long-term efficiency in public pri­ paribus, more input implies an equal or higher level of output). Thus, the
mary schools. selected input variables should present a significant positive correlation
Our unit of analysis is the school, and the dataset covers five aca­ with the output vector in addition to having theoretical support from
demic years, from 2009/2010 to 2013/2014, an extensive period to previous works.
analyze the consequences of the economic crisis. Using this information, With respect to input variables, our decision is based on the approach
we built a complete and balanced five-year panel data set composed of used in Mancebón and Muñiz [73], in which an efficiency notion is
124 public primary schools from Vallès Occidental. Table 3 summarizes estimated taking into account the relationship between students’ aca­
the number of students and schools per academic year. In total, our demic grades, their background and school resources. According to this
balanced panel is composed of 620 observations, 124 schools per year, criterion, we evaluate whether the student is performing to the best of
encompassing 231,971 students.8 their ability. In this sense, most empirical papers attempting to measure
schools’ efficiency usually include some measure of students’ back­
4.2. Description of variables ground, such as family input, together with indicators of human and
physical resources at school ([74] for an early review or [75] for a recent
The dataset provided by the Evaluation Council provides some in­ update).
formation on students’ educational achievement in the comprehensive In our study we consider three input variables. Firstly, the number of
uniform standardized tests, their families, the educational environment, teachers employed in the school (teachers), both permanent and non-
and school management. From this information and following the permanent, is selected following the criterion of previous studies (e.g.
literature on efficiency in education (see Ref. [8] for a systematic liter­ Ref. [52,76–79]). Second, we include information on school operational
ature review), we selected the following variables of analysis. Table 4 expenditures (expenses), excluding teachers’ salaries, as a measure for
contains the definitions of the variables included in the analysis. physical resources in schools ([78,80]). In our case, operating expenses
refer to schools’ structural expenses that guarantee their proper func­
tioning, specifically supplies, cleaning, dining room, etc. In other words,
5
We have data for the entire population of public primary schools in Cata­ we do not include a variable as an indicator of capital investment, which
lonia for the period of analysis. However, the computation of the conditional could be endogenous, but rather the operating expenses required to
version of the panel data DEA requires powerful resources to compute the cover the daily basic needs of schools. In this case, schools do not
bandwidths for each environmental variable in a reasonable time. For that negotiate this level of spending as it is set by the competent regional
reason, we only consider Vallès Occidental, which is representative of the government according to their size. Finally, the inclusion of students’
population of public primary schools in Catalonia and Spain, so it would be background as inputs (usually represented by an indicator to proxy their
feasible to extrapolate our results to the situation of public primary schools in socioeconomic status) has been widely discussed in the literature, since
the rest of the country.
6 some papers have considered it as an environmental factor ([81]), others
We thank one of the anonymous reviewers for raising this point.
7 as an input reflecting the students’ raw material ([77,78,82]), while
Note that these comprehensive standard tests are graded by external
teachers and education inspectors.
others test both definitions ([83]). In this paper, we use mother’s
8
Before setting up the final sample, we performed a test to detect outliers
following the approach by Prior and Surroca [66]. The procedure of identifying
extreme observations is continued as long as the null hypotheses of equality
9
between successive efficiency distributions for the complete sample cannot be Scores obtained in the different waves are comparable over time, since the
rejected. This procedure identified only one school in the sample as an outlier, structure, the content, the response system and the evaluation criteria remained
which was therefore removed. Other approaches for outlier correction can be stable during the analysis period. In academic year 2017/2018 the Evaluation
found in Andersen and Petersen [67], Wilson [68], Simar [69] and Chen and Council included a fifth competence to assess students’ knowledge of the nat­
Johnson [70]. ural environment. However, our sample is not affected by this change as this
academic year falls outside our analysis period. We thank one of the anonymous
reviewers for their comment on this point.

7
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

Table 2
Comparison between Vallès Occidental, Barcelona, Catalonia and Spain. Key figures.
Percentage of immigrant population Family Income (thousands of €)

Vallès Occidental Barcelona Catalonia Spain Vallès Occidental Barcelona Catalonia Spain

2009 12.1% 16.5% 14.9% 12.2% 16.8 20.5 16.7 15.5


2010 12.1% 16.5% 15.0% 12.4% 16.3 19.7 16.2 15.6
2011 11.8% 16.2% 14.7% 12.3% 16.0 19.9 16.1 15.1
2012 11.4% 16.5% 14.7% 12.3% 16.5 20.3 16.5 15.0
2013 11.0% 16.4% 14.3% 11.9% 17.0 20.6 16.8 14.2
2014 10.3% 16.0% 14.5% 10.8% 17.4 20.7 17.0 15.5

GDPpc (thousands of €)

Vallès Occidental Barcelona Catalonia Spain

2009 27.2 26.9 27.9 23.1


2010 26.4 27.0 27.1 23.0
2011 26.6 26.6 27.2 22.8
2012 27.4 25.8 27.8 22.0
2013 28.0 25.8 29.0 21.9
2014 29.3 26.5 30.1 22.2

Source: Spanish National Statistics Institute, INE (Instituto Nacional de Estadística)

Table 3 Table 4
Sample decomposition. Definition of variables.
Academic 2009/ 2010/ 2011/ 2012/2013 2013/ Label Variable Description Type
year 2010 2011 2012 2014
Output Y Aggregated Aggregated average grade Volume
Students 44,055 45,285 46,831 47,986 47,814 average grade obtained by students in the
Schools 124 124 124 124 124 standardized test
conducted at the end of
Source: the authors. primary education.
Inputs X1 Teachers Number of teachers in the Volume
educational level as a proxy for family input (parental education) and we school, both permanent
and non-permanent
treat it as an input.10 Our decision is justified not only because it satisfies
teachers.
the monotonicity condition (mothers with a higher level of education X2 Expenses Operational expenditure Volume
may have favorable cultural habits that can help their children cultivate (excluding teachers’
study habits which foster their academic achievement, in contrast to salary).
mothers who do not have that cultural habit), but also because the X3 Parental Mother’s average Ratio
education educational level.
literature has revealed that mothers are crucial in determining students’ 0. No education. 1. Primary
academic achievement ([84]). In addition, this variable remains the education. 2. Secondary
same for the student, even if they change from one school to another. education. 3. Professional
Following Hanushek [74,75], we consider it as a kind of family input training (upper secondary)
4. Upper secondary
bearing in mind that not all the inputs in the education production
education (pre-university).
functions are inside the school. Overall, according to Santín and Sicilia 5. Professional training
[85], these three input variables represent the educational inputs (postsecondary). 6. First
required to carry out the learning process: students’ characteristics (raw degree. 7. Postgraduate. 8.
material), teachers (human capital) and public expenditure (physical Doctorate.
Environmental Z1 Educational Percentage of students with Ratio
resources). factors needs educational needs at the
Regarding the variables representing the environment in which the school (autism,
school is operating, we consider three different sources of heterogeneity personality-behavior
at the school level. Specifically, we decided to include information about disorder, psychological or
mental disability, severe
contextual conditions, in terms of schooling, that are imposed to schools
hearing or visual disability,
because they represent characteristics of their student body. Moreover, exceptionally gifted
it is important to note that schools have no control over the students students).
enrolled, as public schools in Spain do not select their student body; Z2 Immigrants Percentage of non-Spanish Ratio
rather, it is imposed by the regional government. We therefore ensure students at the school
Z3 Annual Percentage of student Ratio
absenteeism absences during the
academic year (students
10
absent more than 75% of
Another option would have been to use the professional category of one or all school days)
both parents. However, we rejected this indicator because the measurement
scale for professional categories changed during the period of analysis. In order Source: the authors.
to reduce noise and volatility, we therefore opted to use the mother’s educa­
tional level. Indeed, this indicator is more stable to professional changes due to that there are no self-selection problems in our sample. Where these
the economic crisis, as it is a family input indicator that does not decline if the contextual conditions exist, either additional educational resources may
mother (father) is unemployed as a consequence of the crisis. In other words, be required or school management may be more complex; we therefore
the mother (father) is not a worse mother (father) if she (he) is unemployed; on
the contrary, she (he) will be able to spend more time with their child/children
and dedicate more time to educational tasks. We thank an anonymous reviewer
for raising this point.

8
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

consider that these factors can be treated as exogenous. The variables we methodological approach presented in Section 3; that is, the output-
selected are the percentage of students with educational needs (educa­ oriented panel data DEA model with selective convexity.13 Firstly, we
tional needs), the percentage of immigrants11 (immigrants) and the degree show the results for selective time-invariant output-oriented efficiency
of absenteeism12 (annual absenteeism). scores obtained from the unconditional and conditional panel data DEA
Table 5 reports the descriptive statistics for all these variables. As can model. Subsequently, we decompose the efficiency scores to show the
be seen, students’ scores in the basic skills tests are, on average, higher time-variant estimates (see Refs. [4] for more details on this estimation).
than 50 points out of 100. At this educational level, most students Then, we complement the long-term efficiency analysis with the results
usually pass. Additionally, the number of teachers hired remains stable in terms of convergence, with the objective of determining whether
up to the school year 2011/2012, where we find a decline, suggesting schools tend to converge and quickly adapt to the budget changes caused
that the budgetary constraints during the crisis also affected the number by the financial crisis.
of teachers employed. The trend for the operational expenditures is
similar, decreasing over the period and recovering slightly at the end. 5.1. Long-term efficiency estimates
Parental education is, on average, constant over time, indicating that
more often than not, mothers have a bachelor’s degree or professional In this subsection we present the results of the selective output-
training. With respect to contextual factors, the percentage of students oriented panel data DEA model distinguishing between introducing or
with special educational needs is lower than 2% during the period of not the influence of contextual variables (unconditional vs conditional
analysis. Something similar occurs with the level of absenteeism, model). As can be seen in Table 6, on average, the time-invariant con­
although the percentage is slightly higher. The proportion of immigrants ditional efficiency for the academic years between 2009 and 2014 for
follows a clearly negative trend during the period, although it increases primary schools in Vallès Occidental is 1.38; that is, schools have, on
slightly in the 2013/2014 school year. average, an improvement margin of 38%. In other words, with the same
During the period of analysis, which coincides with the economic level of educational resources, and considering their contextual condi­
crisis, the number of immigrants arriving in Spain fell and, in turn, the tions, they should be able to improve student outcomes by that per­
crisis led some immigrant families to leave Spain and/or return to their centage (obviously, efficient schools have lower targets while inefficient
countries of origin, moving away from high unemployment and a schools will be expected to meet more demanding targets). If we
waning economy. According to the National Statistics Institute, Spain compare this result with the unconditional model, we observe that in­
had a negative migratory balance during the entire crisis period, and efficiency is triggered up to 1.50 on average. These results are in line
between 2009 and 2013 the census recorded a loss of more than 100,000 with those previously obtained in the literature, which confirms that the
immigrants. This fall was reflected in the number of immigrant students inclusion of contextual variables in the efficiency estimation smooths
enrolled in schools, since many of these migratory flows involve whole the results and improves efficiency by restricting the scope of compar­
families. In short, fewer migrants arrived and some left Spain, one ison to DMUs with similar conditions to the unit of analysis ([86]).
possible net effect being a reduction in the number of immigrant chil­ The following tables show the time-invariant conditional efficiency
dren enrolled in schools. However, in the academic year 2013/2014, a scores classified by schools’ complexity level, population size and
modest increase was observed, possibly indicating that either the chil­ market concentration.14
dren of immigrant families already living in Spain had reached school Firstly, we measure the level of schools’ complexity according to the
age, or that immigration to Spain had started to rise again and families Catalan Directorate General of Public School Teachers and Staff
were arriving with school-age children. (Dirección General de Profesorado y Personal de Centros Públicos). More
complex schools are defined as those located in environments with
5. Empirical results especially disadvantaged socio-economic and socio-cultural character­
istics. They account for 25% of schools in Catalonia. Factors that lead to
In this section we present the main results obtained by applying the the classification of a school as especially vulnerable are variables
related to parents (their income level, whether they are unemployed
and/or whether they are national minimum wage workers) and vari­
11
The proportion of immigrants could have been considered as an input rather ables regarding students (number of pupils with late school starting age,
than an environmental factor. However, we consider that it does not meet the etc.).
monotonicity criterion, since a higher proportion of immigrant students does Table 7 indicates that the average long-term inefficiency is similar in
not necessarily mean an improvement in educational performance. In fact, schools with lower and higher level of complexity, 1.37 and 1.39
students from immigrant families tend to perform worse than ‘native’ students, respectively, as confirmed by the Mann-Whitney test (p-value = 0.48).
at least during the first years of schooling. We thank one of the anonymous
Although more complex schools are slightly more inefficient, the dif­
reviewers for pointing this out.
12 ference is not statistically significant. However, as can be seen, both the
The degree of absenteeism could also be considered as a bad output, or the
complement (attendance) could even be taken and considered as an input in the
standard deviation and the upper quartiles are different in both cate­
model. However, there are several reasons that justify our decision to consider gories of schools. Interestingly, the dispersion in the efficiency scores is
it as an environmental factor. The first is due to a technicality of the method­
ology used. In order to reduce the complexity of the mathematical formulation,
13
our proposed conditional panel data DEA model only requires inputs and It is worth noting that the results are shown in terms of inefficiency as the
desirable (good) outputs. The second reason has to do with the information score goes from 1 to infinite.
14
collected by this variable. The degree of absenteeism is a proxy for heteroge­ We also carried out an analysis of the time-invariant efficiency according to
neity, i.e., ethnic minorities or low-income families usually generate more the age of the school. The most inefficient schools are the newest (3.21, on
absenteeism than medium/high income families, and this is exogenous to average) despite having new facilities (electronic whiteboards, better equip­
schools. Therefore, it is a variable of complexity in educational management. ment, etc.). We understand that these schools start from zero and face an
Thirdly, we do not consider the inverse of this variable as an input due to the important entry barrier: lack of experience in school management. Even if this
need to comply with the monotonicity condition not only in terms of “more is the case, new schools are nevertheless not created at random but are typically
input, more output” but also in terms of “more input, better output”. Our output built in areas where the demand has increased, and their students have specific
variable is a measure of the quality of educational results, i.e., average grade in characteristics (we thank one of the anonymous reviewers for raising this
the homogeneous tests, not about the quantity, i.e., the number of students point). On the other hand, schools with 12 years’ experience are the most
passing the tests. Including school attendance as an input does not guarantee efficient, since they have an improvement margin of just 8.25%. This may show
better grades. Obviously, more students will be evaluated, but not for this that schools with considerable experience know how to manage resources more
reason will the grades be better. efficiently in times of crisis.

9
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

Table 5
Descriptive statistics of variables included in the analysis.
Variable Label 2009/2010 2010/2011 2011/2012 2012/2013 2013/2014

Output
Average gradea y Mean 65.82 72.18 72.89 68.36 72.10
SD 9.45 9.95 9.14 8.78 9.03
Inputs
Teachers x1 Mean 27.23 27.81 26.64 23.64 25.21
SD 7.89 7.48 6.78 6.32 6.18
Expenses x2 Mean 10,130.00 10,292.25 8105.84 9827.60 9059.37
SD 2633.43 2691.91 2096.73 2706.16 2212.73
Parental education x3 Mean 4.98 4.49 4.39 4.60 4.90
SD 0.89 0.91 0.87 0.87 0.77
Contextual factors
Educational needs z1 Mean 1.67 1.30 1.19 1.26 1.03
SD 2.08 1.93 1.28 1.93 1.22
Immigrants z2 Mean 12.60 9.99 9.28 6.98 7.73
SD 11.21 9.81 8.60 7.75 8.44
Annual absenteeism z3 Mean 1.40 2.89 2.23 1.64 1.24
SD 2.93 3.03 2.96 2.93 2.87
N 124 124 124 124 124
a
Note that the average grade at school level is shown here for interpretation purposes, instead of the aggregated average grade. In the model we introduce the
aggregated average score as an output.
Source: the authors.

Table 6
Selective time-invariant data panel efficiency scores.
N Min Max 1◦ Quartile Median 3◦ Quartile Mean SD

Conditional model 124 1.00 3.91 1.13 1.24 1.51 1.38 0.42
Unconditional model 124 1.00 4.39 1.16 1.30 1.63 1.50 0.55

Source: the authors

Table 7
Selective time-invariant data panel efficiency scores by complexity.
N Min Max 1◦ Quartile Median 3◦ Quartile Mean SD

Low complexity 124 1.00 3.91 1.13 1.20 1.52 1.37 0.44
High complexity 124 1.00 2.46 1.13 1.28 1.50 1.39 0.37

Source: the authors

greater in schools with less complex school environments, while schools population density, we found that these are large schools (measured by
with more complex environments present a less dispersed distribution of the number of students enrolled in the school) compared to the average
efficiency (0.44 and 0.37, respectively). In addition to the above, the size of schools in our full sample from Vallès Occidental (the average size
beta convergence analysis conducted below in section 5.3 shows that of these schools is 400 students compared to an average of 372 students
more inefficient schools experienced greater growth (improvement in in Vallès Occidental).
efficiency) than schools with a higher initial level of efficiency While the above is true, we realize that the number of schools in this
(Table 11). Based on the above empirical evidence, we therefore category is very small (only three schools) compared to the largest group
consider that complexity matters and our findings suggest that schools that is represented by schools in municipalities with more than 100,000
with higher levels of complexity make a great effort to manage their inhabitants (55 schools). As can be seen in Table 8, these schools have an
more diverse student bodies, which brings them closer to the average improvement margin of 34% (in line with the average inefficiency found
level of efficiency of schools in less complex environments. for the whole sample in Table 6, 38%).
With respect to the analysis by population size, Table 8 shows that The differences in time-invariant efficiency score by population size
schools located in rural populations (Pop<5000 inhabitants) are the are not significant, as confirmed by the results of the Kruskal-Wallis test
most inefficient (2.09, on average) especially due to the difficulties they (p-values>0.12 in all cases).
face (few students, lower resources, etc.).15 Lastly, in terms of competition, measured through the market con­
At the other extreme, schools located in populations between 5001 centration index, the Herfindahl index,16 it is observed that, in general,
and 10,000 inhabitants are the least inefficient in our sample (1.17, on schools with a higher level of competition are less inefficient than those
average). Despite operating in relatively small municipalities in terms of operating in environments with no competition (Table 9).

15 16
Another possible source of inefficiency in smaller schools could be the high The Herfindahl index is the sum of the squared market shares of schools
structural costs needed to guarantee their proper functioning; these are fixed located in the same municipality (market share is defined as the ratio between
costs that all schools incur regardless of their size. In the case of bigger schools the students in a specific school divided by the total number of students in the
this fixed cost is shared among a larger group of students, while in smaller municipality). This index has been used previously in the literature to measure
schools the relative fixed cost per student is higher, making the school more competition in public education (see Harrison and Rouse [87] and references
inefficient. We thank one of the anonymous reviewers for raising this point. within).

10
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

Table 8
Selective time-invariant data panel efficiency scores by population size.
N Min Max 1◦ Quartile Median 3◦ Quartile Mean SD

Pop≥100,001 55 1.00 2.54 1.11 1.20 1.49 1.34 0.37


50,001 ≤ Pop≤100,000 20 1.00 1.65 1.15 1.23 1.31 1.24 0.20
20,001 ≤ Pop≤50,000 31 1.00 2.51 1.16 1.27 1.61 1.39 0.34
10,001 ≤ Pop≤20,000 13 1.08 3.91 1.13 1.29 1.70 1.65 0.81
5001 ≤ Pop≤10,000 3 1.00 1.31 1.10 1.19 1.25 1.17 0.16
Pop≤5000 2 2.03 2.15 2.06 2.09 2.12 2.09 0.08

Source: the authors

Table 9
Selective time-invariant data panel efficiency scores by market concentration.
N Min Max 1◦ Quartile Median 3◦ Quartile Mean SD

Very high concentration (HI ≥ 0.76) 3 1.00 2.15 1.52 2.03 2.09 1.73 0.63
High concentration (0.51 ≤ HI ≤ 0.75) 2 1.19 1.31 1.22 1.25 1.28 1.25 0.08
Moderate concentration (0.26 ≤ HI ≤ 0.5) 8 1.08 3.91 1.18 1.36 1.89 1.78 0.97
Low concentration (0.11 ≤ HI ≤ 0.25) 43 1.00 2.51 1.15 1.27 1.58 1.38 0.34
Very low concentration (HI ≤ 0.1) 68 1.00 2.54 1.11 1.20 1.45 1.31 0.34

Source: the authors

This result is in line with the literature confirming that the public competitive environments, where they hold barely 10% of the market.
education system is not without competition. Hoxby [88] and Barrow
and Rouse [89] note that even without taking into account private or
5.2. Time-variant efficiency estimates
semi-private schools, the public education system would not be devoid
of competition. As Millimet and Collier [90, p.135] conclude “compe­
With the estimation of the time-variant conditional efficiency we can
tition from other public schools is the most remarkable type of compe­
split the time-invariant conditional efficiency reported in Table 6 for
tition in the status quo.” In addition, Hanushek and Rivkin [91] argue
each year in the study period. In this case, we compare the time-variant
that “the most important element of competition comes from other
conditional efficiency computed with stable weights (as stated previ­
public schools.” A substantial body of literature has emerged attempting
ously in Section 3) with an intertemporal (cross-sectional) conditional
to uncover the factors that influence the relationship between compe­
model, which suffers from disregarding the panel data structure.
tition and efficiency in public education (see López-Torres et al. [39] and
Table 10 and Fig. 3 summarize the results of both estimations for the full
references within). As Millimet and Collier state [90], increased
sample.
competition positively impacts on the efficiency of public schooling.
As can be seen, the conditional model and the intertemporal version
Public schools compete mainly to attract students. Given that public
generate similar efficiency scores. However, the intertemporal efficiency
schools cannot engage in price competition, competition between public
schools focuses on attracting and retaining a larger and better student
Table 10
body [92]. Public schools that do not attract students will have their
Time-variant estimates. Comparison selective panel data DEA and intertemporal
resources reduced and teachers may be relocated. In this scenario,
DEA.
Rincke [93] states that public schools in metropolitan areas have
TV conditional panel data DEA model
stronger competition because they are divided into more school dis­
tricts. With more schools available, it is easier for parents to choose a 2009/ 2010/ 2011/ 2012/ 2013/
school that matches their preferences.17 Hoxby [95] also states that “this 2010 2011 2012 2013 2014

traditional form of school choice is by far the most pervasive and N 124 124 124 124 124
important form of choice in elementary and secondary schooling today.” Min 0.97 0.91 0.66 0.52 0.77
Max 8.29 4.34 3.17 3.09 4.48
Based on the above evidence, we contend that our result for the case
1◦ 1.38 1.27 1.16 1.21 1.18
of Vallès Occidental supports the hypothesis of a positive relationship Quartile
between efficiency and competition [96] as, on average, schools in more Median 1.58 1.37 1.27 1.37 1.28
competitive environments ( HI ≤ 0.1) are less inefficient than those 3◦ 1.87 1.68 1.44 1.65 1.55
operating in environments with no competition (HI ≥ 0.76) (Table 9 Quartile
Mean 1.83 1.59 1.38 1.45 1.41
reports average inefficiency scores of 1.31 vs. 1.73, respectively). These SD 0.96 0.57 0.38 0.36 0.44
results reinforce the idea that competition stimulates teaching practices
that improve educational results and, therefore, fosters efficiency. Intertemporal conditional DEA Model
Despite the above, we find the lowest level of inefficiency in those
2009/ 2010/ 2011/ 2012/ 2013/
schools whose market share covers between 51% and 75% of the market 2010 2011 2012 2013 2014
(1.25, on average), although these findings correspond to just two
N 124 124 124 124 124
schools in our sample. Most of the schools in our setting operate in Min 1.00 1.00 1.00 1.00 1.00
Max 3.37 3.02 2.27 4.49 1.99
1◦ 1.26 1.18 1.00 1.09 1.00
Quartile
17
The public education system in Spain does allow parents to freely decide Median 1.45 1.30 1.11 1.23 1.10
from up to 10 schools for their children. After submitting their choices in order 3◦
1.71 1.57 1.21 1.42 1.20
of preference, a set of established criteria determines the final allocation (for Quartile
Mean 1.57 1.45 1.17 1.31 1.14
instance, distance to the school, siblings attending the school, etc.). Interested
SD 0.46 0.45 0.25 0.39 0.17
readers can refer to the paper by Calsamiglia and Güell [94] for more details
about the school choice process in Catalonia. Source: the authors

11
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

Fig. 3. Evolution of time-variant panel data estimates and intertemporal estimatesSource


the authors.

scores are lower (meaning they find less inefficiency) than the panel and it offers consistent efficiency indexes over time, using stable weights
ones, which could lead to erroneous conclusions, since a school may for inputs and outputs. In other words, the intertemporal DEA estima­
seem less inefficient with a cross-sectional estimation than with one that tion hides inefficiency by generating temporal inconsistency on the
takes into account the panel structure of the data. Here we observe the input and output weights; that is, the negative effect of the crisis on
importance of maintaining the consistency of the weights in the long inefficiency is underestimated when the analyst uses the standard
run. The intertemporal approach artificially reduces the inefficiency by intertemporal DEA models.
annually modifying the importance (weight) assigned to each variable, If we disaggregate the time-variant estimates according to schools’
so the volatility of the shadow weights of the variables is considerably complexity, number of inhabitants in the municipality and degree of
higher. For their part, the panel data estimations define a stable set of market concentration, we find that the general conclusion does not
weights for the complete time period under assessment. This reduces the change; that is, the intertemporal model hides inefficiency in all the
volatility of the valuation of virtual outputs and inputs by considering cases.
not just a specific year, but the aggregated magnitudes in the long run. As an example, Fig. 4 presents the evolution of efficiency scores ac­
This temporal consistency on the weights implies increases in the in­ cording to the level of schools’ complexity (difficulty of management).
efficiency score estimates for specific years as long as their input and As can be seen, the previous trend holds; that is, the conditional model
output mix does not coincide with that corresponding to the aggregated with panel data shows more inefficiency than the intertemporal DEA
levels. model, regardless of the level of schools’ complexity. We can also
To better explain this fact, we graphically observe that the panel data confirm that, in both estimates, schools with a higher level of complexity
DEA model estimates efficiency scores that are more stable, smoothed are more inefficient than those with a lower level of complexity
and robust to random shocks such as the economic crisis. In contrast, we throughout the period.
observe more volatility over time with intertemporal estimates. When However, the graph shows a very interesting phenomenon that will
we want to compare important phenomena over time (such as the eco­ later be reinforced with the study of convergence and, at the same time,
nomic crisis and the consequent cuts in public resources), the panel it leads us to confirm that the estimates of the conditional panel data
estimates are more robust than the intertemporal estimates, which DEA are more robust than those obtained by intertemporal analysis. As
introduce more volatility and noise into the results. This could lead to can be seen, there is convergence in the efficiency scores obtained with
misinterpretations, since intertemporal DEA estimates do not take into the selective conditional panel data DEA between 2009 and 2011, and
account the panel structure of the data and could hide the lack of con­ between schools with higher and lower level of complexity. From 2012
sistency of the yearly movements, as they consider weights for which the onwards, schools began to diverge in levels of efficiency, confirming that
DMU appears more efficient, and these weights are different every year. not all of them are able to manage the available resources in time of
However, the panel data DEA model takes into account a stable structure crisis since schools with less complexity became less inefficient (this

Fig. 4. Evolution of efficiency scores according to the level of complexitySource


the authors.

12
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

same conclusion is obtained with the β-convergence test for the full the reality.
sample). In contrast, the intertemporal DEA estimates do not reflect this In summary, we can conclude that we find relatively high in­
trend; rather the opposite is found as the efficiency indexes show greater efficiency over the period. However, in times of crisis, schools that
convergence at the end, the opposite conclusion to that obtained in the operate in a more stable environment (low complexity), and with the
convergence analysis (p-value = 0.106, if we compare 2013 with 2012). pressure of competition around them, tend to improve their manage­
Finally, as we concluded before in Table 9, the least inefficient ment efficiency due to the use of tight measures to control public
schools are those operating in more competitive environments. This spending. Schools may also improve their efficiency when faced with
trend is supported by the time-variant estimates, regardless of the model possible penalization, by the government, in the form of closure for
used (panel data or intertemporal) (Fig. 5). inefficient management of spending. This possibility is highlighted in
This finding confirms the existence of competition in the public ed­ López-Torres and Prior [97], who stated that during academic year
ucation sector and shows it is beneficial, as schools strive to improve the 2012/2013 seven schools were closed in Catalonia and plans were made
way they manage educational resources to foster student outcomes ([38, to cut 73 school groups due to lack of resources.
96]). For the cases of low and very low concentration, the trend to
improve the level of efficiency in the first years of the crisis holds, a sign 5.3. Dynamics of efficiency: patterns of convergence
that schools are quickly adjusting to a lower available budget. The
opposite occurs for those schools operating in more concentrated envi­ Once the results for efficiency have been obtained by applying the
ronments (moderate and very high concentration) whose inefficiency selective conditional output-oriented panel data DEA model, the next
score rose considerably from 2011 onwards with the panel data esti­ step is to explore the patterns of convergence in efficiency scores.
mation, while the intertemporal model shows an improvement, hiding β-convergence ([15,16]) establishes a relationship between the variable

Fig. 5. Evolution of the efficiency score according to the level of market concentrationSource
the authors.

13
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

of interest (efficiency) with respect to the initial period considered Table 12


(2009 in this study). The goal is to study whether the most inefficient σ-convergence coefficients (time-variant efficiency model).
schools in the first period experience greater growth (improvement) σ 2009/ σ 2010/ σ 2011/ σ 2012/ σ 2013/
than schools with the highest initial level of efficiency. Thus, if the β 2010 2011 2012 2013 2014
coefficient is negative and statistically significant, we confirm the exis­ σ-coefficient 0.336 0.280 0.231 0.229 0.227
tence of convergence, while if the relationship is positive there is Change in σ − 0.168 − 0.177 − 0.006 − 0.009
divergence. The regression function to calculate the β-convergence is
Source: the authors.
shown above in equation [10]. Let us note at this point that there are two
options regarding the reference period. First, and following the main
period in which all schools significantly converged and improved their
line of analysis of the paper, we set 2009 as the baseline period, as 2009
efficiency (values closer to 1 in the scores) in order to comply with the
represents one year before the beginning of the economic recession.
new public budget restrictions and continue to provide education to the
Therefore, we proceeded to calculate the β-convergence in every period
best of their ability, despite having fewer public resources. By contrast,
with respect to the base year. Second, we ran the test for each year with
in 2012/2013 the levels of convergence slowed down, which may be due
respect to the previous one in order to find additional evidence of
to the fact that some schools are inefficient at this stage, a sign that not
convergence from one year to the next one. The results of the
all of them adapted to the change in the level of resources available to
β-convergence are shown below in Table 11.
efficiently support their educational management over time. However,
In this case, we used the results obtained from applying the selective
in 2013/2014 we again see a modest improvement in efficiency levels
conditional output-oriented time-variant panel data DEA model. We first
along with a trend to convergence, although it is very low. This could be
ran the test considering 2009/2010 as the baseline year and then with
interpreted as a first sign of recovery, which has a positive effect on the
respect to the previous academic year.
efficiency of educational management.
In both cases convergence is shown; that is to say, the most ineffi­
cient schools in the first period of analysis (2009/2010) experienced
greater growth than schools with a higher initial level of efficiency. This 5.4. Robustness check
occurs regardless of whether we take into account the base year (2009/
2010) as a reference or if we take the year before. Therefore, another The specification of the variables in an empirical model is a crucial
possible explanation for the improvement in efficiency is the existence of aspect, not only because it conditions the functioning of the model, but
convergence in efficiency scores due to the schools’ efforts to make also because it is essential to correctly define the production function
better use of available resources during the period of crisis. In this sense, behind it. In the case of efficiency in the education literature, it is true
our results strongly support the hypothesis of convergence among that there is no consensus regarding the classification of some types of
Catalan schools during the crisis. Additionally, we confirm that variables, such as the case of the socio-economic status of families.
convergence is greater over time, reinforcing the idea of improving ef­ Indeed, some papers consider it as an input variable, and others as an
ficiency at the end of crisis. environmental factor (see Ref. [8] for a review). Our perspective, in line
With respect to σ -convergence, which represents the estimate of with our earlier explanation in Section 4.2, is to consider our variable
cross-sectional dispersion and indicates the speed with which the level parental education as an input variable in the conditional DEA model
of efficiency of a school converges with the average of all schools in the with panel data, following Hanushek [75].
same period, Table 12 reports the results for the model considered in Nevertheless, we consider that it is worth testing what would happen
equation [11]. with our results if we changed the definition of the model and consid­
The results for the σ -convergence test substantiate convergence; that ered this variable as a contextual factor.18 Therefore, we re-estimated
is, low-efficiency schools in the first period converge with the average the model by changing the consideration of the input parental educa­
efficiency score faster than the rest of the sample. As Table 12 shows, the tion (x3) to consider it as an environmental factor (z4) that is beyond the
dispersion of the efficiency index dropped by 16.8% from 0.336 in school’s control, as we did with annual absenteeism, immigrants and stu­
2009/2010 to 0.280 in 2010/2011, and the trend was similar in the dents with special educational needs. In this model (labeled new model),
following period, with a sharper descent in 2011/2012 (− 17.7%) fol­ which serves as a robustness check, the criterion for considering these
lowed by a slight decrease in the rest of the period. Therefore, the fastest variables as contextual factors is consistent and everything that is
period of convergence was the academic year 2011/2012, coinciding outside the school’s control is considered an environmental factor.
with the year of the harshest budget cuts due to the economic crisis. This Despite the above, we must recognize that the definition of the educa­
reinforces the idea that in times of crisis when control of public spending tional production function under this scenario is limited to inputs that
is greater, schools tend to improve their management efficiency by are exclusively inside, not outside, the school, an aspect that assumes a
optimizing the use of the resources they have. clear limitation of reality. In line with Hanushek [75] not all the inputs
Linking these results with the levels of efficiency obtained in sections in the education production function are inside the school. The findings
5.1 and 5.2, it is confirmed that the academic year 2011/2012 was the are summarized below.
First of all, Table 13 reports the results obtained for the time-
Table 11 invariant conditional efficiency. Compared with the original results in
β-convergence coefficients (time-variant efficiency model). Table 6, it is evident that the change in the model leads to a change in the
efficiency estimates. We observe that the time-invariant efficiency
Considering 09/10 as the baseline
10/11 with 11/12 with 12/13 with 13/14 with scores change in both estimations, unconditional and conditional and,
respect to 09/ respect to 09/ respect to 09/ respect to 09/ overall, we find more inefficiency in the new model. However, the
10 10 10 10 efficient units remain the same, except for one school in the sample in
β-coefficient − 0.271*** − 0.492*** − 0.658*** − 0.776*** the case of the conditional estimation. In short, the efficient frontier is
Considering the previous academic year
10/11 with 11/12 with 12/13 with 13/14 with
essentially composed of the same DMUs. This finding reinforces the
respect to respect to respect to respect to results obtained with the original model.
09/10 10/11 11/12 12/13 Additionally, we observe the same trend in the results when
β-coefficient − 0.271*** − 0.286*** − 0.203*** − 0.137 comparing the conditional and the unconditional estimations. As in the
Note: ***, **, and *: below 1%, 5% and 10% statistical significance thresholds,
respectively.
18
Source: the authors. We thank one of the anonymous reviewers for the suggestion on this point.

14
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

Table 13 Table 15
Time-invariant data panel efficiency scores with the new model. Time-variant efficiency estimates (new model).
Desc. stat. Unconditional efficiency Conditional efficiency 2009/ 2010/ 2011/ 2012/ 2013/
2010 2011 2012 2013 2014
N 124 124
Min 1.00 1.00 N 124 124 124 124 124
Max 4.39 3.82 Min 0.58 0.91 0.59 0.81 0.76
1◦ Quartile 1.19 1.16 Max 9.23 5.24 4.02 4.44 6.34
Median 1.35 1.29 1◦ 1.31 1.20 1.11 1.14 1.13
3◦ Quartile 1.66 1.56 Quartile
Mean 1.57 1.45 Median 1.52 1.33 1.23 1.29 1.24
SD 0.63 0.50 3◦ 1.98 1.69 1.40 1.61 1.49
Quartile
Source: the authors. Mean 1.88 1.62 1.39 1.45 1.43
SD 1.27 0.72 0.51 0.54 0.67
original model, the inefficiency is higher in the unconditional estima­ Source: the authors.
tion, 1.57 on average, compared to 1.45.
In order to confirm whether the results are significantly different or
not, we ran the nonparametric Wilcoxon and Li tests to allow the Table 16
equality of two distributions to be tested. In the two cases, the null hy­ Non-parametric tests results for time-variant data panel efficiency scores. P-
pothesis assumes equality of distribution in the two samples. The p- values.
values of the tests for the case of time invariant efficiency scores are Test Conditional efficiency estimation (original vs new)
presented in Table 14.
Wilcoxon 0.11
As can be seen from Table 14, the p-values are not significant in any Li 0.10
case. Therefore, the null hypothesis of equality of distributions cannot
Source: the authors.
be rejected. This shows that our results from the original efficiency
model are robust and do not significantly change, even if we change the
definition of the model. numbers, support services and operating expenses. Besides these budget
Moving on to the results of the new model for the case of the time- constraints, the obvious question is to determine their social impacts on
variant efficiency scores, we reach similar conclusions. Table 15 sum­ the welfare of population. On the one hand, we can expect regressions in
marizes the results. the levels of equity and the principle of equal opportunities (as expen­
As before with the time-invariant efficiency score, we find that the diture on public education in Catalonia is far below the Spanish average,
change in the definition of the model leads to a change in the time- occupying the 14th position out of 17 Spanish regions). On the other
variant efficiency estimates. Again, we find more inefficiency in the hand, harsh budget constraints can have an undesirable internal impact
new model, on average, compared with the original results showed in when distributions of inefficiency tend to diverge as a result of the
Table 10 (upper panel). In spite of this, the p-values of the two tests are asymmetric impact of the budget restrictions. In this study we focused
not statistically significant (Table 16), ensuring that the distribution of on the second issue, determining whether convergence or divergence in
the two estimations is not different. the distributions of efficiency is related to cuts in the public budgets.
Therefore, considering all of the above, we can conclude that the From the methodological perspective, this objective required us to
results are robust despite modifying the specification of one of the operate with robust estimations of efficiency scores, to ensure we control
variables. those factors that can introduce noise into the assessment, without
In view of the results of the re-estimated model, we prefer to continue providing any additional information (such as, for instance, the yearly
working with the original model, not only due to the absence of sig­ change in the inputs and output weights). This can be guaranteed by the
nificant differences in the efficiency results, but also because the original use of the panel data DEA model in its two variants: the time invariant
model considers a more complete educational production function that and the time variant models. Indeed, by introducing stability in the
is closer to reality, in line with Hanushek [75]. Additionally, and weights of inputs and outputs, panel data DEA estimations assure ab­
methodologically speaking, the original model considers the possibility solute comparability of the efficiency scores, and this helps to clarify the
of including variables in absolute and relative terms following the se­ dynamic analysis of efficiency in order to observe the existent patterns of
lective convexity approach of Olesen et al. [5], which represents a clear convergence.
advantage for future implementations of the models in other samples One important issue to control when we want to obtain comparable
and contexts. results in terms of schools’ efficiency is the impact of external or
contextual factors. Here, previous research clearly indicates that effi­
6. Conclusions ciency scores are dependent on the school’s specific environment. To
control for this, in the methods section we have defined a new condi­
This paper analyzes schools’ efficiency in a context of financial crisis. tional panel data DEA model. This new proposal avoids unacceptable
The immediate result of the financial crisis was the severe impact in and unbalanced comparisons of schools that operate in different envi­
terms of budget restrictions for all public sector bodies ([98]). In our ronmental conditions. Another theoretical contribution is related to the
case study (budget restrictions in Catalan public schools between 2009 objective of making an assessment operative when some variables are
and 2014) these restrictions implied sharp reductions in teacher expressed in volume, but other are expressed in ratio terms. To provide
an acceptable solution for this problem, a selective convexity panel data
DEA model was also developed.
Table 14 The empirical section makes an assessment of the long-run efficiency
Non-parametric tests for time-invariant efficiency scores (p-values). of 124 Catalan primary schools for the period 2009–2014 (generating a
database with 620 school/year units). The results provide evidence of
Test Unconditional efficiency Conditional efficiency
estimation (original vs new) estimation (original vs new) convergence in the efficiency distributions, indicating that differences
among efficient and inefficient schools are reduced over time. These
Wilcoxon 0.2695 0.1450
Li 0.2187 0.1205 results reflect the schools’ efforts to make better use of available re­
sources during the crisis period. Therefore, as highlighted by previous
Source: the authors.

15
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

literature (see for instance Refs. [99,100]), budget constraints are [18] Agasisti T, Pohl C. Comparing German and Italian public universities:
convergence or divergence in the higher education landscape? Manag Decis Econ
effective tools to reduce differences in terms of efficiency. This is what,
2012;33(2):71–85.
after applying all the methodological controls, provided our estimations. [19] Worthington AC. An empirical survey of frontier efficiency measurement
Two basic questions remain unanswered. Firstly, what impact did techniques in education. Educ Econ 2001;9(3):245–68.
these budget constraints have in terms of equity and the principle of [20] Johnes J. Efficiency measurement. In: Johnes G, Johnes J, editors. The
international handbook on the economics of education. Cheltenham: Edward
equal opportunities? To explore this question, the sample would have to Elgar; 2004. p. 613–742.
be expanded to include a control group of schools that was not exposed [21] Emrouznejad A, Parker BR, Tavares G. Evaluation of research in efficiency and
to these severe operational conditions, which would not be possible in productivity: a survey and analysis of the first 30 years of scholarly literature in
DEA. Soc Econ Plann Sci 2010;42(3):151–7.
the Spanish context. In this case, and complementarily to the efficiency [22] Emrouznejad A, Yang GL. A survey and analysis of the first 40 years of scholarly
analysis, additional research methods, such as impact evaluation tech­ literature in DEA: 1978–2016. Soc Econ Plann Sci 2018;61:4–8.
niques, would be needed. The second logical extension of this study is to [23] Hanushek EA. Conceptual and empirical issues in the estimation of educational
production functions. J Hum Resour 1979;14(3):351–88.
determine the impact of the budget constraints in terms of innovations [24] Hanushek EA. The economics of schooling: production and efficiency in public
and technical change, which also requires additional methodological schools. J Econ Lit 1986;24(3):1141–77.
tools that are beyond the scope of this paper but that, of course, can be [25] Hanushek EA. The failure of input-based schooling policies. Econom J 2003;113
(485):F64–98.
explored in the future. Our expectations on these two issues are rather [26] Mancebón MJ, Bandrés E. Efficiency evaluation in secondary schools: the key role
pessimistic but, as social researchers, we first need to find empirical of model specification and of ex post analysis of results. Educ Econ 1999;7(2):
evidence of the phenomenon under scrutiny before providing any 131–52.
[27] Bessent A, Bessent W. Determining the comparative efficiency of schools through
acceptable explanation of the facts.
data envelopment analysis. Educ Adm Q 1980;16(2):57–75.
[28] Charnes A, Cooper WW, Rodhes E. Measuring the efficiency of decision making
CRediT authorship contribution statement units. Eur J Oper Res 1978;2(6):429–44.
[29] Charnes A, Cooper WW, Rodhes E. Evaluating program and managerial
efficiency: an application of Data Envelopment Analysis to program Follow
Laura López-Torres: Conceptualization, Methodology, Validation, through. Manag Sci 1981;27(6):668–97.
Formal analysis, Resources, Writing - original draft, Visualization. [30] Bessent A, Bessent W, Kennington J, Reagan B. An application of mathematical
programming to assess managerial efficiency in the Houston independent school
Diego Prior: Conceptualization, Methodology, Software, Formal anal­ district. Manag Sci 1982;28(12):1355–67.
ysis, Writing - original draft, Visualization, Funding acquisition. [31] Simar L, Wilson PW. Estimation and inference in two-stage, semi-parametric
models of production processes. J Econom 2007;136(1):31–64.
[32] Daraio C, Simar L. Introducing environmental variables in nonparametric frontier
Acknowledgements and funding models: a probabilistic approach. J Prod Anal 2005;24(1):93–121.
[33] Daraio C, Simar L. Advanced robust and nonparametric methods in efficiency
The authors are grateful to the Consell Superior d’Avaluació del Sis­ analysis. Methodology and applications. New York: Springer; 2007.
[34] Cherchye L, De Witte K, Ooghe E, Nicaise I. Efficiency and equity in private and
tema Educatiu for providing the data. This work was financially sup­ public education: a nonparametric comparison. Eur J Oper Res 2010;202(2):
ported by the Spanish Ministerio de Ciencia e Innovación [ECO2017- 563–73.
88241-R]. [35] Haelermans C, De Witte K. The role of innovations in secondary school
performance: evidence from a conditional efficiency model. Eur J Oper Res 2012;
223(2):541–9.
References [36] De Witte K, Kortelainen M. What explains performance of students in a
heterogeneous environment? Conditional efficiency estimation with continuous
and discrete environmental variables. Appl Econ 2013;45(17):2401–12.
[1] Surroca J, Prior D, Tribo Gine JA. Using panel data DEA to measure CEOs’ focus
[37] López-Torres L, Prior D. Should we charge environmental factors for poor
of attention: an application to the study of cognitive group membership and
students’ achievement? Appl Econ Lett 2015;22(17):1378–81.
performance. Strat Manag J 2016;37(2):370–88.
[38] Cordero JM, Santín D, Simancas R. Assessing European primary school
[2] Tulkens H, Vanden Eeckaut P. Non-parametric efficiency, progress and regress
performance through a conditional nonparametric model. J Oper Res Soc 2017;68
measures for panel data: methodological aspects. Eur J Oper Res 1995;80(3):
(4):364–76.
474–99.
[39] López-Torres L, Nicolini R, Prior D. Does strategic interaction affect demand for
[3] Cullinane K, Wang T. The efficiency analysis of container port production using
school places? A conditional efficiency approach. Reg Sci Urban Econ 2017;65(7):
DEA panel data approaches. Spectrum 2010;32(3):717–38.
89–103.
[4] Pérez-López G, Prior D, Zafra-Gómez JL. Temporal scale efficiency in DEA panel
[40] Caves DW, Christensen LR, Diewert WE. The economic theory of index numbers
data estimations. An application to the solid waste disposal service in Spain.
and the measurement of input, output, and productivity. Econometrica 1982;50
Omega 2018;76(4):18–27.
(6):1393–414.
[5] Olesen OB, Petersen NC, Podinovski VV. Efficiency analysis with ratio measures.
[41] Färe R, Grosskopf S, Roos P. Malmquist productivity indexes: a survey of theory
Eur J Oper Res 2015;245(2):446–62.
and practice. In: Index numbers: essays in honour of sten malmquist. Springer;
[6] OECD. Government at a glance. https://doi.org/10.1787/gov_glance-2017-en;
1998. p. 127–90.
2017.
[42] Grosskopf S. Some remarks on productivity and its decompositions. J Prod Anal
[7] Hollingsworth B. The measurement of efficiency and productivity of health care
2003;20(3):459–74.
delivery. Health Econ 2008;17(10):1107–28.
[43] Bjurek H, Førsund F, Hjalmarsson L. Malmquist productivity indices: an empirical
[8] De Witte K, López-Torres L. Efficiency in education: a review of literature and a
investigation. In: Fäare R, Grosskopf S, Russell R, editors. Index numbers: essays
way forward. J Oper Res Soc 2017;68(4):339–63.
in honour of sten malmquist. Boston: Kluwer; 1998. p. 217–39.
[9] Narbón-Perpiñá I, De Witte K. Local governments’ efficiency: a systematic
[44] Bjurek H. The Malmquist total factor productivity index. Scand J Econ 1996;98
literature review—part I. Int Trans Oper Res 2018;25(2):431–68.
(2):303–13.
[10] Narbón-Perpiñá I, De Witte K. Local governments’ efficiency: a systematic
[45] Arjomandi A, Salleh MI, Mohammadzadeh A. Measuring productivity change in
literature review—part II. Int Trans Oper Res 2018;25(4):1107–36.
higher education: an application of Hicks–Moorsteen total factor productivity
[11] Spanish Ministry of Education. Figures of education in Spain. School year 2013-
index to Malaysian public universities. J Asia Pac Econ 2015;20(4):630–43.
2014. Centro de Publicaciones. Ministerio de Educación, Cultura y Deporte; 2016.
[46] Aparicio J, López-Torres L, Santín D. Economic crisis and public education. A
http://www.educacionyfp.gob.es/dam/jcr:c8143066-8e16-4e16-b0b2-be9366e
productivity analysis using a Hicks-Moorsteen index. Econ Modell 2018;71(4):
d0744/resultados-ing.pdf.
34–44.
[12] Asmild M, Hollingsworth B, Birch S. The scale efficiency of hospitals: one size
[47] Cazals C, Florens JP, Simar L. Non-parametric frontier estimation: a robust
does not fit all. J Prod Anal 2013;40(2):197–206.
approach. J Econom 2002;106(1):1–25.
[13] Guccio C, Martorana MF, Mazza I. Efficiency assessment and convergence in
[48] Bădin L, Daraio C, Simar L. Explaining inefficiency in nonparametric production
teaching and research in Italian public universities. Scientometrics 2016;107(3):
models: the state of the art. Ann Oper Res 2014;214(1):5–30.
1063–94.
[49] Podinovski VV. Selective convexity in DEA models. Eur J Oper Res 2005;161(2):
[14] Guccio C, Martorana MF, Monaco L. Evaluating the impact of the Bologna Process
552–63.
on the efficiency convergence of Italian universities: a non-parametric frontier
[50] Bonaccorsi A, Daraio C, Simar L. Advanced indicators of productivity of
approach. J Prod Anal 2016;45(3):275–98.
universitiesAn application of robust nonparametric methods to Italian data.
[15] Barro RJ, Sala-i-Martin X, Blanchard OJ, Hall RE. Convergence across states and
Scientometrics 2006;66(2):389–410.
regions. Brookings Pap Econ Activ 1991;91(1):107–82.
[51] De Witte K, Rogge N. Accounting for exogenous influences in performance
[16] Barro RJ, Sala-i-Martin X. Convergence. J Polit Econ 1992;100(2):223–51.
evaluations of teachers. Econ Educ Rev 2011;30(4):641–53.
[17] Agasisti T, Dal Bianco A. Reforming the university sector: effects on teaching
efficiency—evidence from Italy. High Educ 2009;57(4):477–98.

16
L. López-Torres and D. Prior Socio-Economic Planning Sciences xxx (xxxx) xxx

[52] Haelermans C, De Witte K, Blank JLT. On the allocation of resources for [79] Essid H, Ouellette P, Vigeant S. Productivity, efficiency, and technical change of
secondary schools. Econ Educ Rev 2012;31(5):575–86. Tunisian schools: a bootstrapped Malmquist approach with quasi-fixed inputs.
[53] De Witte K, Rogge N, Cherchye L, Van Puyenbroeck T. Economies of scope in Omega 2014;42(1):88–97.
research and teaching: a non-parametric investigation. Omega 2013;41(2): [80] Agasisti T. The efficiency of public spending on education: an empirical
305–14. comparison of EU countries. Eur J Educ 2014;49(4):543–57.
[54] Bădin L, Daraio C, Simar L. Optimal bandwidth selection for conditional [81] Ruggiero J, Vitaliano DF. Assessing the efficiency of public schools using data
efficiency measures: a data-driven approach. Eur J Oper Res 2010;201(2): envelopment analysis and frontier regression. Contemp Econ Pol 1999;17(3):
633–40. 321–31.
[55] Racine JS, Li Q. Nonparametric estimation of regression functions with both [82] Crespo-Cebada E, Pedraja F, Santín D. Does school ownership matter? An
categorical and continuous data. J Econom 2004;119(1):99–130. unbiased efficiency comparison for regions of Spain. J Prod Anal 2014;41(1):
[56] Hall P, Racine JS, Li Q. Cross-validation and the estimation of conditional 153–72.
probability densities. J Am Stat Assoc 2004;99(468):1015–26. [83] Queiroz MVAB, Sampaio RMB, Sampaio LMB. Dynamic efficiency of primary
[57] Li Q, Racine J. Nonparametric econometrics: theory and practice. Princeton: education in Brazil: socioeconomic and infrastructure influence on school
Princeton University Press; 2007. performance. Soc Econ Plann Sci 2019;70:100738. https://doi.org/10.1016/j.
[58] Weill L. Convergence in banking efficiency across European countries. J Int seps.2019.100738.
Financ Mark Inst Money 2009;19(5):818–33. [84] Kong W-H, Fu T-T. Assessing the performance of business colleges in Taiwan
[59] Casu B, Girardone C. Integration and efficiency convergence in EU banking using Data Envelopment Analysis and student based value-added performance
markets. Omega 2010;38(5):260–7. indicators. Omega 2012;40(5):541–9.
[60] Zhang T, Matthews K. Efficiency convergence properties of Indonesian banks [85] Santín D, Sicilia G. Measuring the efficiency of public schools in Uruguay: main
1992–2007. Appl Financ Econ 2012;22(17):1465–78. drivers and policy implications. Latin American Economic Review 2015;24(1):
[61] Ayadi R, Boussemart J, Leleu H, Saidane D. Mergers and acquisitions in European 1–28.
banking higher productivity or better synergy among business lines? J Prod Anal [86] Cordero JM, Polo C, Santín D, Simancas R. Efficiency measurement and cross-
2013;39(2):165–75. country differences among schools: a robust conditional nonparametric analysis.
[62] Degl’Innocenti M, Kourtzidis S, Sevic Z, Tzeremes NG. Bank productivity growth Econ Modell 2018;74(8):45–60.
and convergence in the European Union during the financial crisis. J Bank [87] Harrison J, Rouse P. Competition and public high school performance. Soc Econ
Finance 2017;75(2):184–99. Plann Sci 2014;48(1):10–9.
[63] Guccio C, Martorana MF, Monaco L. Evaluating the impact of the Bologna Process [88] Hoxby CM. Does competition among public schools benefit students and
on the efficiency convergence of Italian universities: a non-parametric frontier taxpayers? Am Econ Rev 2000;90(5):1209–38.
approach. J Prod Anal 2015;45(3):275–98. [89] Barrow L, Rouse CE. Using market valuation to assess public school spending.
[64] Kumar S, Russell R. Technological change, technological catch-up, and capital J Publ Econ 2004;88(9–10):1747–69.
deepening: relative contributions to growth and convergence. Am Econ Rev 2002; [90] Millimet DL, Collier T. Efficiency in public schools: does competition matter?
92(3):527–48. J Econom 2008;145(1–2):134–57.
[65] Krueger AB, Lindahl M. Education for growth: why and for whom. J Econ Lit [91] Hanushek EA, Rivkin SG. Does public school competition affect teacher quality?
2001;39(4):1101–36. In: Hoxby CM, editor. The economics of school choice. Chicago: University of
[66] Prior D, Surroca J. Performance measurement and achievable targets for public Chicago Press; 2003. p. 23–47.
hospitals. J Account Audit Finance 2010;25(4):749–65. [92] Zanzig BR. Measuring the impact of competition in local government education
[67] Andersen P, Petersen NC. A procedure for ranking efficient units in data markets on the cognitive achievement of students. Econ Educ Rev 1997;16(4):
envelopment analysis. Manag Sci 1993;39(10):1261–4. 431–44.
[68] Wilson PW. Detecting outliers in deterministic nonparametric frontier models [93] Rincke J. Competition in the public school sector: evidence on strategic
with multiple outputs. J Bus Econ Stat 1993;11(3):319–23. interaction among US schools districts. J Urban Econ 2006;59(3):352–69.
[69] Simar L. Detecting outliers in frontier models: a simple approach. J Prod Anal [94] Calsamiglia C, Güell M. Priorities in school choice: the case of the Boston
2003;20(3):391–424. mechanism in Barcelona. J Publ Econ 2018;163:20–36.
[70] Chen WC, Johnson AL. A unified model for detecting efficient and inefficient [95] Hoxby CM. School choice and school competition: evidence from the United
outliers in data envelopment analysis. Comput Oper Res 2010;37(2):417–25. States. Swed Econ Pol Rev 2003;10(2):9–66.
[71] Aparicio J, Cordero JM, Ortiz L. Measuring efficiency in education: the influence [96] Agasisti T. How competition affects schools’ performances: does specification
of imprecision and variability in data on DEA estimates. Soc Econ Plann Sci 2019; matter? Econ Lett 2011;110(3):259–61.
68:100698. https://doi.org/10.1016/j.seps.2019.03.004. [97] López-Torres L, Prior D. Centralized allocation of human resources. An
[72] Hoxby CM. The effects of class size on student achievement: new evidence from application to public schools. Comput Oper Res 2016;73(9):104–14.
population variation. Q J Econ 2000;115(4):1239–85. [98] Martínez-Campillo A, Fernández-Santos Y. The impact of the economic crisis on
[73] Mancebón MJ, Muñiz MA. Private versus public high schools in Spain: the (in) efficiency of public Higher Education institutions in Southern Europe: the
disentangling managerial and programme efficiencies. J Oper Res Soc 2008;59 case of Spanish universities. Soc Econ Plann Sci 2020;71:100771. https://doi.org/
(7):892–901. 10.1016/j.seps.2019.100771.
[74] Hanushek EA. Educational production functions. In: Economics of education. [99] Grosskopf S, Hayes KJ, Taylor LL, Weber WL. Budget-constrained frontier
Research and studies. Pergamon; 1987. p. 33–42. measures of fiscal equality and efficiency in schooling. Rev Econ Stat 1997;79(1):
[75] Hanushek EA. Education production functions. In: The economics of education. 116–24.
Academic Press; 2020. p. 161–70. https://doi.org/10.1016/B978-0-12-815391- [100] Biørn E, Hagen TP, Iversen T, Magnussen J. The effect of activity-based financing
8.00013-6. on hospital efficiency: a panel data analysis of DEA efficiency scores 1992–2000.
[76] Cordero JM, Pedraja F, Santín D. Enhancing the inclusion of non-discretionary Health Care Manag Sci 2003;6(4):271–83.
inputs in DEA. J Oper Res Soc 2010;61(4):574–84.
[77] Thieme C, Prior D, Tortosa-Ausina E. A multilevel decomposition of school
Laura López-Torres is assistant professor at the Department of Economics and Business
performance using robust nonparametric frontier techniques. Econ Educ Rev
(University of Alcala, Spain). Her main research interests fall at the intersection between
2013;32(1):104–21.
economics of education and efficiency and productivity analysis.
[78] Brennan S, Haelermans C, Ruggiero J. Nonparametric estimation of education
productivity incorporating nondiscretionary inputs with an application to Dutch
schools. Eur J Oper Res 2014;234(3):809–18. Diego Prior is full professor at the Department of Business (Universitat Autònoma de
Barcelona, Spain). His research interests are efficiency and productivity analysis, research
in financial accounting and management in the public sector.

17

You might also like