WWW.JAQM.

RO

JOURNAL
OF
APPLIED
QUANTITATIVE
METHODS

Quantitative Methods Inquires Vol. 10

No. 1
Spring
2015
ISSN 1842–4562
 Editorial Board

JAQM Editorial Board

Editors
Ion Ivan, Bucharest University of Economic Studies, Romania
Claudiu Herteliu, Bucharest University of Economic Studies, Romania
Gheorghe Nosca, Association for Development through Science and Education, Romania

Editorial Team
Cristian Amancei, Bucharest University of Economic Studies, Romania
Catalin Boja, Bucharest University of Economic Studies, Romania
Radu Chirvasuta, Imperial College Healthcare NHS Trust, London, UK
Ştefan Cristian Ciucu, Bucharest University of Economic Studies, Romania
Irina Maria Dragan, Bucharest University of Economic Studies, Romania
Eugen Dumitrascu, Craiova University, Romania
Matthew Elbeck, Troy University, Dothan, USA
Nicu Enescu, Craiova University, Romania
Bogdan Vasile Ileanu, Bucharest University of Economic Studies, Romania
Miruna Mazurencu Marinescu, Bucharest University of Economic Studies, Romania
Daniel Traian Pele, Bucharest University of Economic Studies, Romania
Ciprian Costin Popescu, Bucharest University of Economic Studies, Romania
Aura Popa, YouGov, UK
Marius Popa, Bucharest University of Economic Studies, Romania
Mihai Sacala, Bucharest University of Economic Studies, Romania
Cristian Toma, Bucharest University of Economic Studies, Romania
Erika Tusa, Bucharest University of Economic Studies, Romania
Adrian Visoiu, Bucharest University of Economic Studies, Romania

Manuscript Editor
Lucian Naie, SDL Tridion

I
 Advisory Board

JAQM Advisory Board

Luigi D’Ambra, University of Naples “Federico II”, Italy

Kim Viborg Andersen, Copenhagen Business School, Denmark
Tudorel Andrei, Bucharest University of Economic Studies, Romania
Gabriel Badescu, Babes-Bolyai University, Romania
Catalin Balescu, National University of Arts, Romania
Avner Ben-Yair, SCE - Shamoon College of Engineering, Beer-Sheva, Israel
Ion Bolun, Academy of Economic Studies of Moldova
Recep Boztemur, Middle East Technical University Ankara, Turkey
Constantin Bratianu, Bucharest University of Economic Studies, Romania
Ilie Costas, Academy of Economic Studies of Moldova
Valentin Cristea, University Politehnica of Bucharest, Romania
Marian-Pompiliu Cristescu, Lucian Blaga University, Romania
Victor Croitoru, University Politehnica of Bucharest, Romania
Gurjeet Dhesi, London South Bank University, UK
Cristian Pop Eleches, Columbia University, USA
Michele Gallo, University of Naples L'Orientale, Italy
Angel Garrido, National University of Distance Learning (UNED), Spain
Anatol Godonoaga, Academy of Economic Studies of Moldova
Alexandru Isaic-Maniu, Bucharest University of Economic Studies, Romania
Ion Ivan, Bucharest University of Economic Studies, Romania
Adrian Mihalache, University Politehnica of Bucharest, Romania
Constantin Mitrut, Bucharest University of Economic Studies, Romania
Mihaela Muntean, Western University Timisoara, Romania
Peter Nijkamp, Free University De Boelelaan, The Nederlands
Bogdan Oancea, Titu Maiorescu University, Romania
Victor Valeriu Patriciu, Military Technical Academy, Romania
Dan Petrovici, Kent University, UK
Gabriel Popescu, Bucharest University of Economic Studies, Romania
Mihai Roman, Bucharest University of Economic Studies, Romania
Satish Chand Sharma, Janta Vedic College, Baraut, India
Ion Smeureanu, Bucharest University of Economic Studies, Romania
Nicolae Tapus, University Politehnica of Bucharest, Romania
Timothy Kheng Guan Teo, University of Auckland, New Zeeland
Daniel Teodorescu, Emory University, USA
Dumitru Todoroi, Academy of Economic Studies of Moldova
Nicolae Tomai, Babes-Bolyai University, Romania
Pasquale Sarnacchiaro, Unitelma Sapienza University, Italy
Vergil Voineagu, Bucharest University of Economic Studies, Romania

II
 Contents

Page
Quantitative Methods Inquires

Emanuela RAFFINETTI, Isabella ROMEO

Evaluating Social Tracking in the Primary School: Evidence from the Lombardy 1
Region (Italy)

Smaranda CIMPOERU
A Logistic Model on Panel Data for Systemic Risk Assessment – Evidence from 15
Advanced and Developing Economies

Silvia DEDU, Florentin SERBAN

Stochastic Optimization using Interval Analysis, with Applications 30
to Portfolio Selection

Adriana AnaMaria DAVIDESCU

Bounds Test approach for the Long Run Relationship between Shadow Economy 36
and Official Economy. An Empirical Analysis for Romania

Kalyan MONDAL, Surapati PRAMANIK

The Application of Grey System Theory in Predicting the Number of Deaths of 48
Women by Committing Suicide- A Case Study

Ion PARTACHI, Vitalie MOTELICA

Methods of Measuring Core Inflation in Inflation Targeting Countries 56

Angel-Alex HAISAN, Vasile Paul BRESFELEAN

Connections between Will to Emigrate and Attachment 67
Theory – A Data Mining Approach

Diana-Silvia ZILISTEANU, Ion Radu ZILISTEANU, Mihai VOICULESCU

A Study of Survival Modelling in Dialysis Patients 85
Applying Different Statistical Tools

Eva MILITARU
The Redistributive Effect of the Romanian Tax-Benefit 93
System: A Microsimulation Approach

Eduard Gabriel CEPTUREANU

Survey Regarding Resistance to Change in Romanian 105
Innovative SMEs from IT Sector

III
 Quantitative Methods Inquires

EVALUATING SOCIAL TRACKING IN THE

PRIMARY SCHOOL: EVIDENCE FROM THE
LOMBARDY REGION (ITALY)1

Emanuela RAFFINETTI2
PhD, Post-Doc Research Fellow, Department of Economics,
Management and Quantitative Methods
Università degli Studi di Milano, Italy

E-mail: [email protected]

Isabella ROMEO3
PhD, Post-Doc Research Fellow,
Department of Statistics and Quantitative Methods
University of Milano-Bicocca, Italy

E-mail: [email protected]

Abstract
Recently, the Italian schools were deeply affected by the “social tracking” phenomenon,
intended as the process of segregating students into socio-economic classes. Typically, this
phenomenon occurs within the lower secondary school. In such a perspective, the study
reported in the paper is innovative, since addressed to investigate the actual presence of the
social tracking phenomenon as an event starting from the primary school. For this purpose,
we considered data provided by Invalsi (Istituto Nazionale per la Valutazione del Sistema di
Istruzione e Formazione) with regard to students of the fifth grade of primary schools in the
Lombardy region (Italy). The study was carried out following two different approaches. First,
a preliminary descriptive analysis of the segregation phenomenon was carried out by
computing the Gini coefficient of the the socio-economic status average at class level.
Second, due to the usual hierarchical structure of educational data, multilevel models were
considered with the aim of partitioning the pupils’ socio-economic status variability within
the student, class and school level. In this way, school and class social segregation indicators
were obtained. Subsequently, a conditional multilevel model including school and class
social segregation indicators as explanatory variables was built. Results underline that even
though in general social tracking is not an actual threat for the Lombardy primary schools, a
remarkable socio-economic heterogeneity among classes appears especially in some
provinces of the Lombardy region.

Key words: social tracking phenomenon, class heterogeneity, Gini coefficient, segregation
indices, multi-level modeling, Invalsi data

1
 Quantitative Methods Inquires

1. Introduction

Interest in evaluating the Italian education systems is manifest in a large number of

recent publications and in the diffusion of standardized tests (e.g., Haladyna, 1991; Ballard
and Bates, 2008). Typically, the content of these contributions focuses on the main pupils
and schools’ determinants affecting the learning levels of students. If on one hand the
educational research field stresses the impact of such factors on the students’ attainments,
on the other hand only a few works addressed the issue of equal opportunity in education
(i.e. each state must provide the same opportunities for everyone who attends school
regardless of gender, race or nationality). Even though the Italian law imposes the “equity
principle” which should be preserved by composing the most possible heterogeneous
classes, recent studies highlight that the practice of segregating students with similar features
is particularly widespread, especially in the lower secondary schools (e.g. Ferrer-Esteban,
2011). Such a sociological issue falls under the name of “formal tracking” phenomenon. In
some cases, school staff may generate a great deal of selection by assigning children with
similar achievement to the same classroom, in order to minimize teaching difficulty, or by
placing all of the “problematic” students in a certain teacher’s class because he is good at
dealing with them. However, the segregation phenomenon can be generated in several
ways and at different levels. Specifically, the increasing participation of the pupils’ parents
to the dynamics of the school is leading to a kind of “informal tracking” phenomenon,
allowing families to influence the classroom composition in order to better respond to their
social features, such as for instance their socio-economic status (e.g. Dupriez et al., 2008).
Social tracking gives rise to homogeneity within classes (social segregation) that in
turn may come out in inequality of education opportunities (e.g., Checchi and Flabbi, 2007;
Hindriks et al., 2010). Children with different family background, race and ability will have
different access to knowledge. It was proven (for example, Loveless, 1999) that whether the
curriculum is adjusted to better match ability level of students, while high ability students
may receive a boosted achievement, low ability students may suffer from assignment to
lower tracks. Thus, homogeneity within classes negatively affects disadvantages students.
Classroom environment is then really important for student achievement, as stated by Hill
and Rowe (1996): “How much a student learns depends on the identity of the classroom on
which the student is assigned”. Indeed, a student’s innate ability can affect his peers, not only
through knowledge spillovers but also through his behavior. On the contrary, a student who
has not learned self-discipline at home may bother the classroom.
The study presented in this paper is innovative since it attempts to explore the
actual presence of the social segregation phenomenon in Italy as an event starting from the
primary schools. Indeed, to the best of our knowledge, no research contributions illustrating
the existence of an informal tracking phenomenon in the Italian primary schools are
currently provided in literature. More precisely, our research question is the following. Since
primary schools represent the first education compulsory stage after the kindergarten, the
segregation process of kids can probably be encouraged by parents on the basis of their
socio-economic features. Kindergarten has a relevant role in the process of contact among
the families of kids. Thus, the pupils' families may wish that their children were kept together
with their kindergarten friends, when accessing to the primary school.
The analysis was carried out on data provided by the National Evaluation
Committee (Istituto Nazionale per la Valutazione del Sistema di Istruzione e Formazione,

2
 Quantitative Methods Inquires

henceforth Invalsi). In Italy the National Evaluation Committee has been established with the
specific aim of evaluating the Italian schools through the analysis of the students’
achievement at different levels of education; second and fifth year of the primary school (age
7 and 10, respectively), first and third of the lower-secondary (age 11 and 13), second and
fifth of the upper-secondary (age 15 and 18). The collection of such data started from the
school year 2008-2009 and represents the first time that a law imposes a national
evaluation by using standardized tests in all students population. Here, we considered a
unique dataset that tracks the performance in Reading of students of the fifth grade of
primary schools in the Lombardy region for the school year 2009-2010.
On the statistical point of view, our proposal was pursued through two different
approaches. First, a preliminary investigation of the social tracking phenomenon was
provided by resorting to a descriptive inequality index, the Gini coefficient, which is widely
used for studying inequality in education attainments (e.g., Leckie et al., 2012). The Gini
coefficient was computed by taking into account the class average value of the variable
representing the socio-economic status (henceforth denoted by SES) over all the classes in
every province of the Lombardy region. Second, to shed light on how the heterogeneity of
the students’ performance and SES are portioned out between school and class level,
different multilevel models were considered both to properly take into account the
hierarchical structure of data with pupils nested in classes and schools (e.g., Snijders and
Bosker, 1999) and to define social segregation indices at school and class level. Finally, a
conditional multilevel model with even the social segregation indices is performed.
The remainder of the paper is organized as follows. In Section 2 the examined
Invalsi dataset is illustrated and some descriptive statistics provided. In Section 3 a
preliminary analysis of the social tracking phenomenon is introduced by resorting to a
descriptive approach based on the Gini coefficient. In Section 4 an overview of the proposed
multilevel methodology is presented. In Section 5 school and class level social segregation
indices are computed and commented. Section 6 is devoted to the discussion of the obtained
results. Finally, Section 7 concludes.

2. Data

Our proposal is based on data coming from the survey led by Invalsi at the end of
the school year 2009-2010 and referring to students of the fifth grade (students of about 11
years old). Coherently with our research scope, the variable under study is here detected by
the pupils school achievement in Reading, expressed as the proportion of correct answers
provided in the administered test by each student. Such data cover the whole population (it
is not a sample) made up of 77.200 students belonging to 4.488 classes that in turn belong
to 1.050 primary schools located in different provinces of the Lombardy region. The
administered test is built on 41 multiple-choice items and is composed by two parts: the
former is related to the comprehension of two texts and the latter is related to the grammar
issues. The testing time is of one hour. The test reserves even a set of questions concerning
the students’ personal information (e.g. gender, ethnicity, grade retention and so on).
Further information about the social, economic and cultural conditions of students are
collected through additional questionnaires filled by the School Principals and students'
parents. Variables considered for the analysis are enlisted below and include:
 demographic variables: i.e. gender, ethnicity, year of birth;

3
 Quantitative Methods Inquires

 sociocultural variables: in this case, a synthetic index, named SES is made directly
available by Invalsi. It is computed analogously to the OECD’s procedure, that is by
considering the parents’ occupation and education, possession of some kinds of
goods such as, for instance, the availability of an encyclopedia or an Internet
connection, the number of books at home and so on (Campodifiori et al., 2010);
 school variables: school size (number of students), type of school administration
(private or public), number of female students, number of students repeating one or
more grades and number of students belonging to ethnic minorities;
 geographical area of the school, specified in the provinces of the Lombardy region:
Bergamo (BG), Como (CO), Lecco (LC), Lodi (LO), Milano (MI), Pavia (PV), Varese
(VA), Brescia (BS), Cremona (CR), Mantova (MN) and Sondrio (SO).
A note about the type of school administration (private or public) is needed. For
private school we mean schools with private involvement in managing and funding. Here,
we only focused on private schools following the ministerial program and thus considered
equivalent to the public ones.
Before proceeding to the construction of the statistical model, an analysis of
missing data was done for all the variables that potentially may be included in it. The
reference dataset is characterized by variables which present missing values at random.
However, the main trouble appears with the pre-school (i.e. kindergarten attendance)
variable whose lack of information is consistent, since missing values amount to the 10.4%.
In such a context, the problem of missing data was easily solved by directly deleting the pre-
school variable from the model. This is because, the ejection of the pre-school variable from
the model found reason in its low contribution in explaining the Reading scores variability.
In order to provide more interpretable parameters, all the variables were
standardized and a reference level was defined (e.g., Snijder and Bosker, 1999).
Furthermore, to better clarify the role of the categorical variables included into the model
and concerning the demographic characteristics of pupils (i.e. gender, ethnicity, and grade
retention) and the school features (public or private status), a related description is presented
in Table 1, where the corresponding reference categories are reported.

Table 1. Description of the pupil and school categorical variables

Variables Description
Demographic
Gender Male (reference category); Female
Ethnicity Italian (reference category); Ethnic minorities of first or second generation
Student that has not repeated a year (reference category, pupils born in
Grade Repetition 1998);
Student that has reapeated at least a year (grade repetition)
Educational
School Administration Public (reference category); Private

With regard to class and school level, we considered variables representing the
proportion of students being female, repeating one or more grades and belonging to ethnic
minorities. These variables were already available in the dataset at school level and relate to
students belonging to all grades in the school. On the contrary, variables at class level were
derived as aggregation of individual covariates at class level. Thus, the latter are related only
to students participating to the survey of the fifth grade. Moreover, the school and class
average of the students’ SES index were computed as aggregation of individual SES index.

4
 Quantitative Methods Inquires

The main key statistics about variables at class and school level are displayed in
Table 2. It is worth noting that variables at school level were centered on the grand mean
and variables at student level were centered on the school average. As shown in Table 2, the
average score4 amounts to 73.20 with a standard deviation equal to 16.63, the average
percentage of female is the 49% at class level and the student SES average is 0.03 at class
level and 0.04 at school level. In addition, almost the 9% of schools are private, the average
percentage of ethnic minority students amounts to the 13% both at class and school level,
while the average percentage of students repeating the year is the 3% at class level and
smaller than the 1% at school level.

Table 2. Descriptive Statistics

Number of
units Mean St Dev Min Max
Score in Reading 77,200 73.20 16.63 0.00 100.00
% Ethnic Minorities 4,466 0.13 0.55 0.00 1.00
Class mean SES 4,487 0.03 2.17 -2.05 2.16
Class
% Females 4,485 0.49 0.51 0.00 1.00
% Student Repeating the year 4,487 0.03 0.23 0.00 1.00
Class size 4,488 20.00 10.00 6.00 28.00
% Ethnic Minorities 1,050 0.13 0.18 0.00 0.83
School mean SES 1,050 0.04 0.86 -1.28 2.04
% Females 1,050 0.48 0.07 0.00 1.00
% Student Repeating the year 1,050 0.003 0.01 0.00 0.03
School size 1,050 532.00 428.00 28.00 1,338
School Administration: Public 74,265 91.17
School Administration: Private 7,191 8.83
Province: BG 10,020 12.30
Province: BS 11,401 14.00
School
Province: CO 4,869 5.98
Province: CR 2,833 3.48
Province: LC 2,867 3.52
Province: LO 1,923 2.36
Province: MN 3,477 4.27
Province: PV 3,991 4.90
Province: SO 1,558 1.91
Province: VA 7,412 9.10
Province: MI 31,105 38.19

3. Preliminary Analysis: the Gini coefficient

In the literature, a wide range of indices are proposed for assessing the actual
presence of the social tracking phenomenon. As deeply discussed by Leckie et al. (2012),
Hutchens (2004) and Reardon and Firebaugh (2002), researchers typically resort to
descriptive indices such as, for instance dissimilarity and square root indices (e.g., Duncan
and Duncan, 1955; Jenkins et al., 2008), in order to detect possible scenarios of inequality
in education opportunity. Since our aim is not limited to detect the presence of inequality in
opportunity but to measure its extent, within the large set of available descriptive indices, the
Gini coefficient was considered (e.g., Gini, 1921). More in detail, the idea here is to provide
a measure of the heterogeneity between classes in term of the socio economic status of
students. For this purpose, we propose to consider as variable of interest the average SES at
class level. For all the classes within each school and each province of the Lombardy region,

5
 Quantitative Methods Inquires

we computed the average value of the students’ SES index. We remark that for every single
student, the SES index ranges between -3 and +3. Thus, it is reasonable to believe
that the average SES at class level may take even negative values. In such a context, the
reliability of the classical Gini coefficient may come less since requiring the considered
variable to be characterized by non-negative values. Indeed, in case of negative values, the
Gini coefficient may violate the normalization principle and thus take values greater than
one. A solution to this problem was recently provided by Raffinetti et al. (2014), who
introduced a new Gini coefficient adjusted for the presence of negative values. The new Gini
coefficient, expressed as the ratio between the absolute mean difference
1⁄ ∑ ∑ | | and 2/ ∑ | |, fulfills the normalization principle. This allows us
to provide a measure of inequality in opportunity which can occur as a consequence of the
class composition process conditioned to the pupils’ socio-economic status. Indeed, if the
Italian schools actually respected the legislative principle of “equal-heterogeneity” in the
composition of classes, the Gini coefficient should be close to zero. This does not happen, as
shown by results in Table 3, where the Gini coefficient of the average SES at class is reported
for every province.

Table 3. Gini coefficient of the average SES at class level per province
Province BG BS CO CR LC LO MI MN PV SO VA
Gini
coefficient 0.65 0.67 0.71 0.69 0.72 0.71 0.72 0.63 0.73 0.69 0.70

The Gini coefficient is greater than 0.60 in all the provinces. More precisely, over
the 50% of the provinces presents a Gini coefficient greater than 0.70. The province of PV
has the higher heterogeneity between classes with a Gini coefficient equal to 0.73. Such
findings are made more evident by the boxplots in Figure 1 which show a remarkable
variability of the average SES at class level in every province of the Lombardy region.
Even though these descriptive statistics seem to confirm the presence of the social
tracking phenomenon, they are obtained without taking into account the hierarchical
structure of classes nested in schools. Thus, the high heterogeneity between classes at
province level may reflect the high heterogeneity between schools within the province. For
this reason, one may assume this variability to be explained by the gaps across the territorial
areas where schools are located. Indeed, schools located in more disadvantaged areas catch
more disadvantaged students. Further investigations were carried out by distinctly computing
the Gini coefficient of the SES average at class level within each school across all the
provinces. Also in this case, the Gini coefficient reaches very high values, leading us to
believe that heterogeneity between classes is a real threat for the equality in opportunity in
the Italian primary schools. In order to validate such a conclusion, the multilevel modeling
approach (e.g., Goldstein, 2011) was considered to take into account the complexity of the
educational systems organized in school and class level. First, we assessed how the
variability of SES portions out among the different considered levels in order to define
segregation status indicators at class and school level, as suggested by Ferrer-Esteban
(2011). Subsequently, we analyzed the partition of the variability of the scores in the Reading
test among the different levels. Finally, a conditional multilevel model was built in order to
evaluate the effects of both the SES index and segregation status indicators, after controlling
for the aforementioned variables, with the purpose of detecting the actual presence of the
social tracking phenomenon across the Lombardy primary schools.

6
 Quantitative Methods Inquires

Figure 1. Boxplots of the average SES at class in every province of the Lombardy region

4. An overview about three-level models

As mentioned above, models suitable in treating hierarchical data are the

multilevel models since they allow relationships to be simultaneously assessed at several
levels (e.g., Snijders and Bosker, 1999), represented by pupils, classes and schools. Some
details about the multilevel algebraic specification are briefly provided below.
Let us consider a three-level multilevel model in educational context, where level 1
is represented by students, level 2 by classes and level 3 by schools. The relationship
between the -th student’s achievement, belonging to the -th class, which in turn belongs to
the -th school, is expressed by:

, (1)

, (2)

where: is the random effect at school level, an allowed-to-vary departure from the grand
mean, is the random effect at student level, a departure from the school effect and
is the random effect at student level, a departure from the class effect within a school. The
variance components at each level are defined as follows: variance between schools,
; variance between students within classes within schools ;
variance between students within classes within schools ; and variance
between classes . Different forms of variance shares are derived: the share of
variance due to gaps between schools, corresponding to the intra-school correlation (level 3)

7
 Quantitative Methods Inquires

, and the share of variance due to differences between classes,

corresponding to the intra-class correlation (level 2) .

Since multilevel models allow to decompose the variability of a specific
phenomenon among the different involved levels, they provide information about the
heterogeneity associated to each considered level. To have an idea of such heterogeneity,
first the intra-class correlation coefficient (ICC) was computed. Indeed, through such a
coefficient, the extent of the outcome variation related to gaps between units of each
considered level was obtained. Secondly, a model with variables described in Section 2 was
built to show both covariates really affecting the students’ achievement and their impact.
Furthermore, also segregation status indicators at class and school level were considered.
The latter is identified as the between classes and schools variance, when an unconditional
multilevel model for the SES variable is fitted. Finally, a comparison between the
unconditional model (empty) and the conditional (full) model was introduced to show the
contribution of the same model in explaining the performance variability at each level of the
analysis. The residual variance located at different levels was interpreted as the result of
unobserved factors, as discussed in more detail in Section 6.

5. Segregation indices

As suggested by Ferrer-Esteban (2011), social segregation at class and school level

are typically measured trough the between-class variance and the the between-school
variance, respectively.
A fully unconditional three-level model for the SES index allows to portion the SES
variability among the considered level: within classes, between classes within schools and
between schools. A high variability of SES between classes underlines more heterogeneity
among different classes within the same school, meaning that classes are more
homogeneous in respect of their social background. Conversely, a high SES variability
between schools underlines more heterogeneity among schools, implying aggregation of
students with similar social background within the school. These indicators give an idea of
the extent both schools and classes within schools are socially dissimilar. Ferrer-Esteban
(2011) analyzed the Italian secondary schools and found out that the SES variability at school
level reaches a value of 32% in some Italian provinces, while the SES variability at class level
reaches a value of 12%. Furthermore, they stressed that while the SES variability at school
level is connected with the presence of metropolitan areas, the SES variability at class level
has a clear pattern of territorial distribution that responds to a north-south gradient, with
higher values of class segregation in the South of Italy.
For what concerns the primary schools we expected a remarkable SES variability
between schools, given that this kind of school is particularly widespread across the Italian
territory. For this reason, primary schools usually catch students of the area in which they are
located. So, schools located in areas with more disadvantaged families will catch
disadvantaged students. Furthermore, the high diffusion of the primary schools involves
schools to be composed by one or few classes for each grade, leading to expect a low SES
variability between classes. In Table 4 the segregation indicators at province and regional
level for the Lombardy primary schools are reported.

8
 Quantitative Methods Inquires

Table 4. Social segregation indicators

Province Between class variance (in %) Between school variance (in %)
BG 2.24% 15.28%
BS 2.71% 14.85%
CO 5.33% 11.03%
CR 2.88% 11.76%
LC 2.20% 17.53%
LO 3.80% 12.36%
MI 3.46% 25.20%
MN 3.12% 7.06%
PV 4.47% 17.49%
SO 2.49% 8.53%
VA 3.67% 16.53%
Lombardy 3.23% 19.58%

In the Lombardy region, the variability of SES between schools is equal to 19.6%,
while between classes it is equal to 3.2%. In particular, the Lombardy provinces highlight a
SES variability at school level ranging between the 7.1% of Mantova and the 25.2% of
Milano, and a SES variability at class level ranging between the 2.2% of Bergamo and Lecco
and the 5.3% of Como and Pavia. These values, compared to the findings illustrated by
Ferrer-Esteban (2011) across the whole Italy and for the lower secondary schools, are to be
considered non-low. Indeed, it is well-known that the social segregation is a phenomenon
appearing more marked in the lower secondary school and in the South of Italy. As
expected, the metropolitan area of Milan presents a high variability of SES between schools.
The SES variability between classes is low on average, but with non-low values for some
provinces. To evaluate if such heterogeneity between schools and classes provides an actual
impact on the students’ achievement, a multilevel model built on the Reading score was
considered. The related results are discussed in the following section.

6. Multilevel model results

The content of this section is focused on both the analysis of the partition of the test
performance variability at individual and group level and identifying the presence of social
segregation. Typically, in the education literature the study of variability in achievements is
based on two-level models characterized by student and school level. Our aim is a little bit
wider since addressed to identify the share of variability attached both to school and class
level. The need of including the class level is supported by our research question, that is
investigating the actual presence of social tracking within the Italian primary school. For this
reason, a three-level model was applied to account for the class level.
In order to define the performance variability partition among the three involved
levels, an empty model without explanatory variables was applied. The related results in
terms of variance decomposition and intra-class correlation coefficient (ICC) are reported in
Table 5.

9
 Quantitative Methods Inquires

Table 5. Variance decomposition of the implemented multilevel model – fifth grade of

primary school
Three level - School-Class-Student
Empty model ICC (%)
Var. Between Schools 13.0 4.7
Var. Between Classes 21.3 7.7
Var. Within Classes 243.0 87.6
Total Var. 277.3 100.0

According to such findings, in primary school the well known prevailing variability
in performance depends on the students’ characteristics ( 243.0 . In addition, variability
between classes is greater than variability between schools, in line with the evidence
gathered from several studies (e.g., Hill and Rowe, 1996). Indeed, the former ( 21.3 is
almost double with respect to the latter ( 13.0 . Despite we expected a low variability
between classes when only a grade is considered, we found a high percentage of this
variability equal to more than the 7% of the total one. Variability between classes appears as
the consequence of several factors which may be strictly related to grade, teacher effect and
unobservable variables as well as peer effect.
To shed light on individual, class and school variables really impacting on the
students’ achievement, a conditional multilevel model was considered. Results about
variables significance are displayed in Table 6. Variables marked by three asterisks in Table
6 are significant at a confidence level α 0.01, variables marked by two asterisks at a
confidence level α 0.05 and finally, variables marked by only an asterisk at a confidence
level α 0.1. Variables representing the percentage of female students, the percentage of
students repeating a year, the size of class and school are non-significant in the analysis at
both school and class level. The intercept value represents the Reading score for an
“average” student who is defined as an Italian male student, with no grade retention, a SES
index equal to the average value of all the students, and whose class presents a percentage
of students belonging to ethnic minorities and an average SES index corresponding to the
mean value of all the classes. Furthermore, his attended school is public, located in Milan
and characterized by a percentage of students belonging to ethnic minorities and an
average SES index equal to the mean value of all the schools. Such a student achieves a
performance in Reading equivalent to 77.21 (i.e. an “average” student correctly answers to
the 77% of the test). To be a male implies a reduction in Reading score equal to 1.64. As it is
trivial to believe, when focusing on individual student variables, the consistent decrease of
the Reading score is related to students belonging to ethnic minorities of first generation (-11
points). Student belonging to ethnic minorities of second generation provides a smaller
decrease in Reading achievements, corresponding to 7.80 points. Definition of ethnic
minorities of first and second generation is needed. Ethnic minorities of first generation are
students born in their origin country from parents belonging to ethnic minorities, while ethnic
minorities of second generation are students born in Italy from parents belonging to ethnic
minorities. The same negative results on the Reading performance are associated to class
and school variables concerning the percentage of students belonging to ethnic minorities.
Obviously, also the grade retention involves a worsening of -7.43 points. Conversely, a
positive trend in Reading performance is associated to the SES index. Indeed, an unitary
increment in the SES index value provides an increase of 4.38 points in the Reading score.
This happens also for the school and class SES index variable. Students attending a school in

10
 Quantitative Methods Inquires

the province of Lodi, Mantova and Brescia reach worse results with respect to those
attending a school in the province of Milan.

Table 6. Three-level Multilevel model Effects

Levels Variables Estimate
Intercept 77.21***
Gender (Female) -1.64***
Ethnic minority - First Generation (Ref. Italian) -10.99***
Student
Ethnic minority - Second Generation (Ref. Italian) -7.80***
Grade Repetition -7.43***
Student SES 4.38***
% Ethnic minority in class -0.05**
Class mean SES 0.98***
Class % Female students in class 0.00
% Repeating the year in class 0.03
Class size -0.04
% Ethnic minority at school -0.08***
School mean SES 3.18***
% Female students in school -0.01
% Repeating the year in school 0.16
School size 0.05
Private school (Ref. Public) -0.98
School Segregation -8.83*
Class Segregation -53.03
Province: BG (Ref. MI) 0.26
School
Province: BS (Ref. MI) -0.93*
Province: CO (Ref. MI) 1.37
Province: CR (Ref. MI) -1.31
Province: LC (Ref. MI) 0.09
Province: LO (Ref. MI) -2.44**
Province: MN (Ref. MI) -2.75***
Province: PV (Ref. MI) 0.24
Province: SO (Ref. MI) 0
Province: VA (Ref. MI) 0

Both school and class segregation indicators, computed in Section 5, were

considered for every province and included as explanatory variables into the model. While
the school segregation indicator is significant in the model, the class segregation indicator
does not. This finding involves the presence of school segregation and the absence of class
segregation in the Lombardy primary schools. Such a conclusion arises as the consequence
of the widespread of the primary schools across the Italian territory. Indeed, since the
primary schools usually receive students living in the area where the schools are located, the
socio-economic status of the area strongly affects the socio-economic status of the school.
The consistent SES variability between schools is in general an expected result which further
stands out for the metropolitan area of Milan. In particular, it is worth noting that the effect
of segregation on achievement is negative causing a reduction of over 8 points on the
Reading performance.

11
 Quantitative Methods Inquires

To clarify the contribution of such variables in explaining the students' achievement

variability in the Reading test, an analysis on the variance reduction at school and class level
was carried out. As shown in Table 7, the full multilevel model provides a contribution of
about only the 33.7% of the students’ performance variability at school level, about the
17.7% at class level and about the 15% within class level. These outcomes have not to be
considered as a failure of our proposed approach, since typically in the primary school the
main share of variability in achievement is the consequence of non-observable students’
characteristics such as, for instance, hours working on homework and/or students’ interest in
school matters.

Table 7. Decomposition of Variance

Empty Model Conditional Model
Variance Factors
ICC Reduction ICC
1.3% Observed school factors
Between Schools 4.7% 33.7%
2.5% Other unobserved factors
1.3% Observed class factors (class composition)
Between Classes 7.7% 17.7%
6.2% Teacher effect and other unobserved factors
13.3% Observed individual factors
Within Classes 87.6% 15.0%
75.5% Other unobserved individual factors
Total 100.0% 16.1% 100.0%

The explained variance has to be ascribable to the presence of observed factors at

different levels, while the residual variance can be ascribable to the presence of unobserved
factors. For instance, at class level the residual variance may include the impact of teacher
and/or other unobserved factor. Definitely, observed school factors explain only the 1.3% of
the performance variability. Unobserved school level factors account for the largest
differences in variability in school performances, but in this case they capture just the 2.5%
of the overall variability in achievements. Compositional factors at class level account for the
1.3% of the overall performance variability. Thus, it is reasonable to believe that, the impact
of unobservable variables between classes (within schools) on gaps in achievement, is much
more marked amounting to the 6.2%. Finally, the unobserved individual factors account for
the 75.5% of the overall variability highlighting that the unobservable student’s
characteristics represent the largest differences in the non-explained variability.

7. Conclusions

In this paper we investigated the presence of the social segregation phenomenon

by analyzing education data provided by Invalsi and concerning the achievement in the
Reading test, obtained in the school year 2009-2010, by students attending the fifth grade
of the primary schools in the Lombardy region (Italy). From this point of view, our study is
innovative since it attempted to detect the social segregation phenomenon as an event
starting from the primary schools. For this purpose, two different approaches were
considered. First, a preliminary investigation of the social tracking phenomenon was
provided by resorting to the Gini coefficient computed by taking into account the class
average value of the SES variable over all the classes in every province of the Lombardy

12
 Quantitative Methods Inquires

region. Results show a high heterogeneity between classes and would seem to validate the
hypothesis of social tracking inside the primary schools. However, to account for the
hierarchical data structure, a multilevel model was carried out. First of all, segregation
indices at class and school level through a fully unconditional three-level model for the SES
index were found out. Such indices are defined in terms of the SES index (representing the
pupils’ socio-economic background) variability among the considered levels (i.e. within
classes, between classes within schools and between schools). Findings highlight that even
though the SES index presents a low variability on average, such variability is consistent in
value across some provinces. Then, a conditional multilevel model including both indicators
of between class and school segregation as explanatory variables for every province was
built. While the school segregation index is significant in the model, the class segregation
index does not. These results suggest that the segregation phenomenon mainly occurs at
school level, neglecting the actual threat of the social tracking phenomenon in the primary
schools of the Lombardy region. However, from a descriptive point of view the presence of a
consistent class heterogeneity is an evidence especially in some provinces of the Lombardy
region. This issue encourages us to believe that such a phenomenon may represent an
actual event in early education in the areas of Italy (South and Islands), where inequality in
households’ socio-economic status is known from the literature to be more marked.

Bibliography

1. Ballard, K. and Bates, A. Making a connection between student achievement, teacher

accountability and quality classroom instruction, The Qualitative Report, Vol. 13,
no. 4, 2008, pp. 560-580
2. Campodifiori, E., Figura, E., Papini, M. and Ricci, R. Campodifiori, E. Un indicatore di status
socio-economico-culturale degli allievi della quinta primaria in Italia (in Italian),
Invalsi, 2010
3. Checchi, D. and Flabbi, L. Intergenerational mobility and schooling decisions in Germany
and Italy: the impact of secondary school tracks, IZA Discussion Paper No.
20(2876), 2007, pp. 210-217
4. Duncan, O. and Duncan, B. A methodological analysis of segregation indexes, American
Sociological Review, Vol. 20, no. 2, 1955, pp. 210-217
5. Dupriez, V., Dumay, X. and Vause, A. How do school systems manage pupils’
heterogeneity?, Comparative Education Review, Vol. 52, no. 2, 2008, pp. 245-273
6. Ferrer-Esteban, G. Rapporto sulla scuola in Italia (in Italian), Laterza, Bari, 2011
7. Gini, C. Measurement of inequality and income, The Economic Journal, Vol. 31, 1921, pp.
124-126
8. Goldstein, H. Multilevel Statistical Models, Wiley, 2011
9. Haladyna, T. Raising standardized achievement test scores and the origins of test score
pollution, Educational Researcher, Vol. 20, no. 5, 1991, pp. 2-7
10. Hill, P. and Rowe, K. Multilevel modelling in school effectiveness research, School
Effectiveness and School Improvement, Vol. 7, 1996, pp. 1-34
11. Hindriks, J., Verschelde, M., Rayp, G., and Schoors, K. Ability tracking, social segregation and
educational opportunity: evidence from Belgium, Discussion Paper, Center for
Operations Research and Econometrics, Bélgica, 2010
12. Hutchens, R. One measure of segregation, International Economic Review, Vol. 45, no. 2,
2004, pp. 555-578

13
 Quantitative Methods Inquires

13. Jenkins, S., Micklewright, J. and Schnepf, S. Social segregation in secondary school: how
does England compare with other countries?, Oxford Review of Education, Vol. 34,
no. 1, 2008, pp. 21-37
14. Leckie, G., Pillinger, R., Jones, K. and Goldstein, H. Multilevel modeling of social segregation,
Journal of Educational and Behavioral Statistics, Vol. 37, no. 1, 2012, pp. 3-30
15. Loveless, T. Will tracking reform promote social equity?, Understanding Race, Class and
Culture, Vol. 56, no. 7, 1999, pp. 28-32
16. Raffinetti, E., Siletti, E. and Vernizzi, A. On the Gini coefficient normalization when attributes
with negative values are considered, Statistical Methods & Applications, DOI
10.1007/s10260-014-0293-4, 2014
17. Reardon, S. and Firebaugh, G. Measures of multigroup segregation, Sociological
Methodology, Vol. 32, no. 1, 2002, pp. 33-67
18. Snijders, T. and Bosker, R. Multilevel Analysis. An introduction to basic and advanced
multilevel modeling, London: Sage Publications, 1999

1
Acknowledgements
Thanks are due to the Italian National Evaluation Committee (Invalsi - Istituto Nazionale per la Valutazione del
Sistema di Istruzione e Formazione) for providing us with the data for this study.

2
Emanuela Raffinetti has a Master Degree in Economics (Finance) at University of Pavia (Italy); International
Postgraduate Master in Complex Systems at IUSS (Istituto Universitario Studi Superiori) of Pavia (Italy) and Ph.D. in
Statistics at Bocconi University of Milan (Italy). Currently, she is Post-Doc Research Fellow at Università degli Studi di
Milano (Italy). Her research activity concerns: dependence analysis, concordance and discordance measures,
categorical variables treatment, models for ordinal variables, inequality measures in income distribution, statistical
techniques for high-dimensional dataset, customer satisfaction assessment and evaluation of educational systems.

3
Isabella Romeo has Bachelor Degree in Statistics at University of Milano-Bicocca (Italy); a Master Degree in
Biostatistic and Applied Statistics at University of Milano-Bicocca (Italy), Ph.D. in Statistics at University of Milano-
Bicocca (Italy). Currently, she is Post-Doc Research Fellow at University of Milano-Bicocca (Italy). Her research
activity concerns: models for ordinal and categorical variables, causal inference in a counterfactual framework,
policy evaluations, longitudinal analysis, education and labour market applications, statistical techniques to manage
administrative data and high-dimension dataset.

4
The score on Reading test corresponds to the total percentage of correct answers provided by the students. Thus, it
lies between 0 and 100.

14
 Quantitative Methods Inquires

A LOGISTIC MODEL ON PANEL DATA FOR SYSTEMIC

RISK ASSESSMENT – EVIDENCE FROM ADVANCED
AND DEVELOPING ECONOMIES1

Smaranda CIMPOERU
PhD, University Assistant,
Department of Statistics and Econometrics,
Bucharest University of Econmic Studies, Romania

E-mail: [email protected]; [email protected]

Abstract
The present paper proposes a framework for developing a new early warning system (EWS) for
identifying systemic banking risk and finding the macroeconomic indicators which turn to be
the best indicators in predicting stressful situation in the economic environment. The research
problem is very much debated in the specialty literature, as the exposure of the financial
system is generally derived from deteriorating systemic conditions. We propose a logistic model
applied on two panel data sets – advanced and emerging economies. Results are satisfactory,
as apart from the GDP Growth or Debt level, as main triggers for financial stress situation, we
also find the Output Gap as a significant early warning signal for predicting financial and
economic crisis.

Keywords: systemic risk, early warning systems, financial crisis, binary variables panel data

1. Introduction

A well-functioning financial system is mandatory for an efficient economy.

However, the fragility of financial systems can cause financial crisis and have significant
impact in the real economy. The topic of financial crisis is highly relevant in terms of policy,
as outlined by Kauko (2014). Crises trigger output losses and social costs, with an average
production loss of 20% of annual Gross Domestic Product (GDP). It is very important to have
a good understanding of the past crisis events, of the mistakes made and to learn the
lessons from the crisis that happened over time because, as time showed us, history could
repeat itself.
In the last twenty years, the world economy has been faced with a significant
number of financial crises, from Latin America, to Asia, from Nordic Countries to East and
Central European countries, it all culminated with the financial tsunami which burst in 2007.
A new and critical need for the Early Warning Systems has appeared since 2008: an updated
EWS that would correctly include in the model the way financial markets are affected by
changes in risk factors and risk transmission. Since the Great Financial Crisis, it has come as

15
 Quantitative Methods Inquires

an evidence the exposure to systemic risk is affected by propagation effects and links among
financial institutions which are strongly determined by the structure of the financial system.
Considering the increased complexity of the financial systems and risk associated
within, attention is drawn by the specialists that a new EWS tool should be used an
orientation rather than a signaling technique. The main role and value of the EWS is
providing a systemic overview and functioning as a monitor for the systemic risk. As
mentioned in Gramlich et al (2010), the results of an EWS should not be overestimated.
However, once critical signals are emitted, the supervisory authorities would need support
”on the basis of an expected, but not yet realized, deterioration”.
In the present paper we propose a logistic macroeconomic model for panel data
with the aim of finding the macroeconomic leading indicators of distress. We carry out two
models – for advanced and emerging economies and find which are the macroeconomic
variables having the highest weight in the probability of a crisis. The rest of the paper is
organized as follows. In section 2 we review literature in what concerns the construction of
early warning systems for banking crisis – the role of the EWS, the main concepts and
techniques used to model the systems, with reference to the latest findings in the literature;
section 3 – gives an overview on the particularities of modeling binary outcomes for panel
data – which is the methodology employed in the case study. In section four, we propose two
models – one for advanced, one for emerging economies and find which are the
macroeconomic indicators of systemic risk. The study is innovative as it includes data for both
types of economies and also the whole period of the last 5 years since the burst of the
Global Financial Crisis. Including the output gap variable in the list of signals is a new
concept in the literature and proves to be a significant early trigger for systemic risk. Section
five presents the conclusions.

2. Literature review – early warning systems for

systemic banking crisis

As the cost of the most recent financial crisis was estimated at app. USD 12 trillion
(reaching 20% of the GDP in most affected countries), the forward-looking instruments of
supervisory banks gain more and more importance as the amplitude of financial crisis
increases. With the crisis becoming more prominent, the literature on EWS models has
grown significantly. However, the existing EWS models failed to predict the recent global
crisis and this is mainly due to the fact that they do not fully reflect the way that financial
markets are affected by changes in risk factors and risk transmission.
Basically, an early warning system (EWS) has the role of anticipating whether an
economy will be affected by a financial crisis by developing a framework which would allow
for predicting financial stressful situations. In the literature there are three approaches for
constructing an Early Warning System for predicting banking crisis: the bottom-up approach,
the aggregate approach and the macroeconomic approach. In the first approach mentioned,
the probability of insolvency is estimated for each bank and the signal for systemic instability
is triggered when the probability of insolvency becomes significant for a high proportion of
the banking assets in the respective economy. For the second approach, the same model is
applied to aggregate bank data instead of individual bank data. In what concerns the third
approach, the attention is focused on establishing a relationship between economy wide
variables, based on the fact that a number of macroeconomic variables are expected to

16
 Quantitative Methods Inquires

affect the financial system and reflect its condition. The third approach will also be used in
the case study that we are proposing in the paper.
Gramlich et al. (2010) make a critical review of earlier EWS literature and highlight
the main components of a EWS risk model:
 Risk measures – stress assessment; in the literature this can take the form of a binary
index (Kaminsky, Reinhart – 1999; Edison – 2003), three-state index (Bussiere;
Fratzscher – 2002) or continous index (Illing Liu – 2003, Hanschel, Monnin – 2005);
 Risk factors – risk indicators – usually chosen between micro risks, macro risks (most
cited being the work of Reinhart, Rogoff – 2009) and structural risks;
 The risk model – a theory on how to combine the risk measures and the risk factors.
Basically there are two approaches for this: the leading indicator (or signal theory)
and data-focused regression models.
The approaches of the EWS models are mainly statistical driven. First models are
proposed by Diebold and Rudebusch (1989) for constructing economic indexes. The
technique was adapted by Kaminsky and Reinhart (1999) who propose the signal approach:
a potential crisis is signaled when a risk factor exceeds a predefined threshold. The threshold
is adjusted to balance type I errors (model failed to predict crises when they actually take
place) and type II errors (models wrongly predicts crises that do not occur). This technique
has also been approached by Borio (2002, 2009). Demirguc-Kunt and Detragiache (1998)
are the first to use regression analysis for evaluating the predictive power of risk factors. In
their later study (2005), the compare both techniques and conclude that the logit model is
the most suitable in assessing financial risk. We also note, the neuro-fuzzy approach of Lin et
al (2006) for identifying the drivers of currency crisis and find that this artificial intelligence
tool improves the prediction of crisis. Still, the black-box pattern of these methods remains a
disadvantage for understanding the big picture of the crisis mechanisms.
Other approaches from the specialty literature include : a non-parametric method
based upon K-means clustering to predicting crisis events (Fuertes, Kalotychou, 2004) - in
their study they find that the optimal model can be constructed based on the decision-
makers preferences regarding the desired trade-off between missed defaults and false
alarms; Kalman filter estimation of state space models (Mody, Taylor, 2003) – with the aim
of extracting a measure of regional vulnerability for emerging economies; factor model with
Markov regime switching dynamics (Chauvet, Dong, 2004) for the prediction of nominal
exchange rates in the East Asian countries.

3. Binary outcome models – particularities for panel data

Considering that in our case, the dependent variable takes the form of a binary
variable (presence or absence of the crisis event), we will turn our attention to the binary
choice models. In this case, the model will have the following form :
1| ,
0| 1 ,
where x is the vector of explanatory factors and β is the vector of parameters that reflect the
changes in x on the probability. The problem that arises is to find a suitable model for the
function F. If we would use the familiar liner regression model, we would encounter a series
of problems. First of all, the disturbances in the model would be heteroscedastic due to the
restriction imposed to have the dependent variable 0 or 1. Assuming that this problem can

17
 Quantitative Methods Inquires

be solved by a GLS estimation, a more serious problem is that we cannot be assured that the
predictions in the model will look like probabilities. That is the main reason for which we
have to use another type of function, that would have the following properties:
lim 1| 1
→

lim 1| 0
→

As stated in Greene, in principle, any “proper, continuous probability distribution

defined over the real line will suffice”. If the normal distribution is being used, the probit
model is obtained:

1|

Due to its mathematical advantages, the logistic distribution is also often used,
determining the logit model:

1|
1
where ∙ indicates the logistic cumulative distribution function. The question arises on
which one of the two models to use. The two distributions have similar bell shaped
distributions, with the difference that the tails are heavier in the logistic one. The logistic
distribution tends to give larger probabilities to 1 for extremely small values of than
the normal distribution would. Or otherwise said, the conditional probability approaches 0 or
1 at a slower rate in logit than in probit. One would expect to obtain different predictions
from the two models if the sample contains very few favorable cases ( Y’s equal to 1) or very
few un-favorable cases (Y’s equal to 0). “There are practical reasons for favoring one the
other in some cases for mathematical convenience, but it is difficult to justify the choice of
one distribution or another on theoretical grounds” (Greene). Most applications would state
the models generally give similar results, with the limitations expressed before.
An important thing to note for logit and probit models is that the parameters in the
model are not necessarily the marginal effects like in the classical regression models. This
happens because the marginal effect of a regressor in the logit model depends not only on
the coefficient of that regressor, but also on the value of all regressors in the model. For
computing marginal effects, we can evaluate the expression for the samples means of the
data or evaluate the marginal effects at every observation and use the sample average of
the individual marginal effects.
The literature dedicated to the binary choice models for panel data is rapidly
growing. An overview is given in Greene (2011). We distinguish between random and fixed
effects models by the relationship existing between the unobserved, individual specific
hetereogeneity and the vector of regressors. The effect model has the following form:
∗
, 1, … , ; 1, … ,
∗
1 0, 0 .
As per Greene (2011), the assumption that is unrelated to produces the
random effects model. However, this places a restriction on the distribution of the
heterogeneity. If the model permits correlation between and , then we have a fixed
effects model. The disadvantage of the fixed effect model is that the maximum likelihood
estimator becomes inconsistent, while in the random effects model strong assumptions
regarding heterogeneity should be made.

18
 Quantitative Methods Inquires

4. Case study

In this part of the paper, we propose a framework that could be used a starting
point for developing an early warning signals system comprising macroeconomic indicators
for monitoring and maintaining financial stability in an economy.
In the first part we describe the data used. As data sources we relied on
macroeconomic data publicly available at World Bank and International Monetary Fund. Due
to significant discrepancy regarding data availability across countries, but also based on
particularities of emergent versus advanced economies, we decided to split the initial sample
into two data sets. That is, one data set contains the information for the advanced
economies: Austria, Germany, Denmark, Spain, Finland, France, United Kingdom, Greece,
Ireland, Italy, Netherlands, Norway, Portugal, Sweden and Belgium. The variables included
for this sample are: Cash deficit, GDP growth, Exports, Stocks, Inflation, Output Gap and
Debt. The observation period is 1990 – 2012, that is the entire panel for advanced
economies contains 315 observations in 15 groups. The second sample will include data on
emerging economies: Bulgaria, Czech Republic, Croatia, Hungary, Iceland, Israel, Lithuania,
Poland, Romania, Slovak Republic, Slovenia, Latvia. The variables included for the emergent
countries sample are: M2 growth, GDP growth, Exports, Stocks, Inflation. Fewer variables are
included due to issues regarding data availability. That is also one of the reasons the
observation period is reduced to 1995 – 2012. Another reason for reducing the observation
period is the particularities of the emergent economies included in the sample, economies
which are mainly from the ex-communist bloc and in the first years of the 1990s developed
abnormal values of the macroeconomic indicators. Total panel for the emergent economies
contains 216 observations in 12 groups. In the next table we present a detailed description
of the indicators included, as they are given on the official sites cited.
The dependent variable used in the model is a binary variable and takes the value
1 if the country has been reported as experiencing a banking crisis in the respective year.
Data for the banking crises has been taken from official sources in IMF (Leaven, 2008 and
further extended).

Table 1. Indicators description

Indicator Indicator Description Observations
GDP growth Annual percentage growth rate of GDP at market Observation period for
(annual %) prices based on constant local currency. Aggregates advanced economies
are based on constant 2005 U.S. dollars. GDP is the 1990 – 2012;
sum of gross value added by all resident producers in for emergent economies
the economy plus any product taxes and minus any 1995 – 2012.
subsidies not included in the value of the products. It
is calculated without making deductions for
depreciation of fabricated assets or for depletion and
degradation of natural resources.
Cash surplus/deficit Cash surplus or deficit is revenue (including grants) Observation period for
(% of GDP) minus expense, minus net acquisition of nonfinancial advanced economies
assets. This cash surplus or deficit is closest to the 1990 – 2012.
earlier overall budget balance (still missing is lending
minus repayments, which are now a financing item
under net acquisition of financial assets).
Inflation, consumer Inflation as measured by the consumer price index Observation period for
prices (annual %) reflects the annual percentage change in the cost to advanced economies
the average consumer of acquiring a basket of goods 1990 – 2012;
and services that may be fixed or changed at for emergent economies

19
 Quantitative Methods Inquires

Indicator Indicator Description Observations

specified intervals, such as yearly. The Laspeyres 1995 – 2012.
formula is generally used.
Money and quasi Average annual growth rate in money and quasi Observation period for
money growth money. Money and quasi money comprise the sum of emergent economies
(annual %) currency outside banks, demand deposits other than 1995 – 2012.
those of the central government, and the time,
savings, and foreign currency deposits of resident
sectors other than the central government. This
definition is frequently called M2. The change in the
money supply is measured as the difference in end-
of-year totals relative to the level of M2 in the
preceding year.
Exports of goods and Exports of goods and services represent the value of Observation period for
services (% of GDP) all goods and other market services provided to the advanced economies
rest of the world. They include the value of 1990 – 2012;
merchandise, freight, insurance, transport, travel, for emergent economies
royalties, license fees, and other services, such as 1995 – 2012.
communication, construction, financial, information,
business, personal, and government services. They
exclude compensation of employees and investment
income (formerly called factor services) and transfer
payments.
Stocks traded, total Stocks traded refers to the total value of shares Observation period for
value (% of GDP) traded during the period. This indicator complements advanced economies
the market capitalization ratio by showing whether 1990 – 2012;
market size is matched by trading. for emergent economies
1995 – 2012.
Output gap (% of Output gaps for advanced economies are calculated Observation period for
potential GDP) as actual GDP less potential GDP as a percent of advanced economies
potential GDP. 1990 – 2012.
General government Net debt is calculated as gross debt minus financial Observation period for
net debt (% of GDP) assets corresponding to debt instruments. These advanced economies
financial assets are: monetary gold and SDRs, 1990 – 2012.
currency and deposits, debt securities, loans,
insurance, pension, and standardized guarantee
schemes, and other accounts receivable.
Source: World Bank Data, International Monetary Fund

4.1. Estimation results for the advanced economies

Before estimating the model, we analyze a graphic representation of the variables
included. Although all variables experienced a drop in the 2007 – 2008 period, the most
representative evolution is the one of the GDP growth. The graphs for the first panels are
reproduced in Figure 1. We notice the evolution of the GDP growth for Greece which
remains on a descendent path, although the rest of the economies experience a drop in the
GDP growth in 2008 followed by a modest recovery in the next years.
Next step is to test the stationarity of the time series included. For this, we apply
specific unit – root tests for panel data. For consistency of results we use four tests: Levin –
Lin – Chen , Breitung, Im – Pesaran – Shin, Hedri LM Test. In the first two tests the null
hypothesis is that the panels contain unit roots with the alternative hypothesis that panels
are stationary, while in the last two tests the null hypothesis is that all panels contain unit
roots with the alternative hypothesis that some panels are stationary.
Results are presented in Figure 2 (example for a unit root estimation output –
results for the test Levin – Lin – Chen applied to GDP growth) and in tables 2, 3 and 4 which
summarize the statistics and p-values for the four tests, for all variables included in the
analyze.

20
 Quantitative Methods Inquires

-1 0 -5 0 5 1 0 Austria Germany Denmark Spain

Finland France UK Greece

-1 0 -5 0 5 1 0
G D P _ g ro w th

Ireland Italy Netherlands Norway

-1 0 -5 0 5 1 0

1990 1995 2000 2005 2010

Portugal Sweden Belgium

-1 0 -5 0 5 1 0

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

Year
Graphs by ID

Figure 1. GDP growth evolution in the period 1990 – 2012 for advanced economies

Figure 2. Results of test Levin – Lin – Chu for the GDP growth

21
 Quantitative Methods Inquires

Table 2. Results of Unit Root Tests for Cash – Deficit and GDP Growth
Unit Root Test Statistic P-Value Unit Root Test Statistic P-Value
Levin-Lin-Chu* -4,6678 0,0000 Levin-Lin-Chu* -4,6545 0,0000
Breitung* -7,5674 0,0000 Breitung* -5,8305 0,0000
Im-Pesaran-Shin ** -3,2867 0,0005 Im-Pesaran-Shin ** -4,6086 0,0000
Hadri LM test ** 7,8664 0,0000 Hadri LM test ** 9,6757 0,0000
*null hypothesis panels contain unit roots / alternative hypothesis that panels are stationary
**null hypothesis that all panels contain unit roots / alternative hypothesis that some panels are stationary

Table 3. Results of Unit Root Tests for Exports and Stocks

Unit Root Test Statistic P-Value Unit Root Test Statistic P-Value
Levin-Lin-Chu* -1,6862 0,0459 Levin-Lin-Chu* -3,9922 0,0000
Breitung* -0,4516*** 0,3258*** Breitung* -3,2177 0,0006
Im-Pesaran-Shin -2,8915*** 0,0019*** Im-Pesaran-Shin -1,6853 0,0460
** **
Hadri LM test ** 16,5181 *** 0,0000 *** Hadri LM test ** 23,9276 0,0000
*null hypothesis panels contain unit roots / alternative hypothesis that panels are stationary
**null hypothesis that all panels contain unit roots / alternative hypothesis that some panels are stationary
*** including time trend; if the time trend component were not included, the series would contain unit roots

Table 4. Results of Unit Root Tests for Inflation and Output Gap
Unit Root Test Statistic P-Value Unit Root Test Statistic P-Value
Levin-Lin-Chu* -8,7241 0,0000 Levin-Lin-Chu* -4,3547 0,0000
Breitung* -1,3878*** 0,0826*** Breitung* -2,9368 0,0017
Im-Pesaran-Shin ** -6,8070 0,0000 Im-Pesaran-Shin ** -2,4560 0,0070
Hadri LM test ** 24,7587 0,0000 Hadri LM test ** 6,1074 0,0000
*null hypothesis panels contain unit roots / alternative hypothesis that panels are stationary
**null hypothesis that all panels contain unit roots / alternative hypothesis that some panels are stationary
*** including time trend; if the time trend component were not included, the series would contain unit roots

Considering the results above, we can conclude, based on all four tests applied that
the variables: Cash Deficit, GDP growth, Stocks, Inflation and Output Gap are stationary
(for a significance level of maximum 5%). However, for variable exports, the results of the
test show the presence of the unit root if trend component is not included (null hypothesis
cannot be rejected – Table 4) and for variable Debt all tests have associated p-values larger
than 0.1 concluding that the series is not stationary. For these two variables we take the first
difference of the variables and obtain that the resulted series are stationary. Results are
summarized in table 5.

Table 5. Results of Unit Root Tests for D(Exports) and D(Debt)

Unit Root Test Statistic P-Value Unit Root Test Statistic P-Value
Levin-Lin-Chu* -8,5419 0,0000 Levin-Lin-Chu* -2,1655 0,0152
Breitung* -9,3639 0,0000 Breitung* -5,3703 0,0000
Im-Pesaran-Shin ** -8,5182 0,0000 Im-Pesaran-Shin ** -4,8709 0,0000
Hadri LM test ** -1,2660*** 0,1027*** Hadri LM test ** 6,3528 0,0000
*null hypothesis panels contain unit roots / alternative hypothesis that panels are stationary
**null hypothesis that all panels contain unit roots / alternative hypothesis that some panels are stationary
*** including time trend

Next, we begin estimating the models. As stated before, we have the option of
estimating logit or probit models with the random or fixed effects (random effects possible
only for logistic models). Considering the nature of the “early warning signals” model we are
proposing, we include in our list of variables all the variables with lagged for two periods.

22
 Quantitative Methods Inquires

However, we obtain that only three of them are significant, that is : the GDP growth with lag
one, the Output gap with lag two and the first difference of variable Debt with lag one.

Table 6. Results for the estimation of the logistic model for advanced economies
(random effects)
Random-effects logistic regression Number of obs = 315
Group variable: id Number of groups = 15

Random effects u_i ~ Gaussian Obs per group: min = 21

avg = 21.0
max = 21

Integration method: mvaghermite Integration points = 12

Wald chi2(10) = 49.11

Log likelihood = -82.941931 Prob > chi2 = 0.0000

bc Coef. Std. Err. z P>|z| [95% Conf. Interval]

cash_deficit -.0684804 .0846879 -0.81 0.419 -.2344657 .0975049

gdp_growth -1.630855 .2672036 -6.10 0.000 -2.154564 -1.107146
stocks .0132952 .0060939 2.18 0.029 .0013514 .025239
inflation .0941213 .1648881 0.57 0.568 -.2290534 .417296
outputgap 1.320462 .3052972 4.33 0.000 .7220906 1.918834
d_exports .0791391 .114939 0.69 0.491 -.1461373 .3044155
d_debt .1446802 .0426859 3.39 0.001 .0610174 .228343

gdp_growth
L1. -1.321017 .2696935 -4.90 0.000 -1.849607 -.7924275

outputgap
L2. -1.227266 .2600205 -4.72 0.000 -1.736897 -.7176349

d_debt
L1. .0907671 .0441878 2.05 0.040 .0041606 .1773736

_cons 1.843755 .9490806 1.94 0.052 -.016409 3.703918

/lnsig2u 1.019219 .6333285 -.2220818 2.26052

sigma_u 1.664641 .5271323 .8949022 3.096462

rho .457198 .1571719 .1957724 .7445347

Likelihood-ratio test of rho=0: chibar2(01) = 18.28 Prob >= chibar2 = 0.000

Results for the logit model with random effects are presented in table 6. The model
is valid, considering the likelihood-ratio test for rho (p-value = 0.0000). Considering the p-
values of the variables included in the sample, at a 0.05 significance level, the following
variables are significant: GDP growth and GDP growth lagged one period (both coefficients
with negative signs, as expected): Stocks ( positive sign), Output Gap and Output Gap
lagged two periods (first with a positive sign and second with a negative sign); first difference
of the governmental debt and first difference of the governmental debt lagged with one
period (both coefficients positive).

23
 Quantitative Methods Inquires

The results show that cash deficit, inflation level and variation in exports are not
significant early warning signs for predicting crisis. The signs of the significant variables are
related to economic theory. A decrease in the GDP growth and the increase in the output
gap are the most significant early warning signs for the advanced economies. Also, an
increase in the variation of governmental debt (one year prior to crisis) and the increase in
volumes of stocks traded can be viewed as early warning indicators, but with smaller
contributions to the probability of a crisis appearance.
Apart for this model, we also estimate (using same variables) a logit model with
fixed effects. The results of the estimation are presented in a comparative manner in Table 7
below.

Table 7. Comparative results of Logit models (fixed / random effects)

Variable Coefficient Std. Error Model
GDP Growth -1.6308 0.2672 Logit Random Effects
-1.6779 0.2690 Logit Fixed Effects
Stocks 0.0132 0.0060 Logit Random Effects
0.0078 0.0071 Logit Fixed Effects
Output Gap 1.3204 0.3052 Logit Random Effects
1.6748 0.3572 Logit Fixed Effects
D(Debt) 0.1446 0.0426 Logit Random Effects
0.1459 0.0396 Logit Fixed Effects
GDP Growth (L1) -1.3210 0.2696 Logit Random Effects
-1.4626 0.2862 Logit Fixed Effects
Output Gap (L2) -1.2272 0.2600 Logit Random Effects
-1.3160 0.2732 Logit Fixed Effects
D (Debt) (L1) 0.0907 0.0441 Logit Random Effects
0.1310 0.0463 Logit Fixed Effects

The results are similar for the two types of models. However, considering the
estimated probabilities of the model (probability that the outcome is positive), we conclude
that the random effects model is much more suitable for the underlying data. The post
estimation results are in Table 8. As per IMF statistics used, the only countries that did not
experience crisis in 2009 from the advanced economies selected are Finland and Norway.
That is, the probability estimated for Norway is very good, but the one estimated for Finland
is associated to a crisis situation, although the country has not been reported as so. We also
note, the low probability reported for Sweden, although the country has been reported as
affected by the crisis. Greece, Italy and Ireland, as well as Portugal have estimated
probabilities very close to one – these being the countries the most affected by the crisis, thus
with the level of the macroeconomic variables most eroded.

Table 8. Post estimation results for the logistic model – advanced economies
(random effects)
Country Year Exp Prob
Austria 2009 0.5582764
Germany 2009 0.8579196
Denmark 2009 0.9491678
Spain 2009 0.7292508
Finland 2009 0.8606029
France 2009 0.9507059
UK 2009 0.9999521
Greece 2009 0.9997895

24
 Quantitative Methods Inquires

Country Year Exp Prob

Ireland 2009 0.9878655
Italy 2009 0.9950404
Netherlands 2009 0.6415532
Norway 2009 0.0431651
Portugal 2009 0.9893235
Sweden 2009 0.2477488
Belgium 2009 0.5460162

4.2. Estimation results for the emergent economies

The graphic representation of the GDP growth’s evolution for the emergent
economies in the panel is found in Figure 3 below. The graph analyze is similar with the one
that we had for the advanced economies. However, we note some particularities – the
countries from the former communist block experienced a drop GDP also in the period 1995
– 1996, due to transition period. Also, Baltic Countries (Latvia and Lithuania) experienced
the most severe drops in GDP in the crisis years, as can be easily observed from Figure 3.
In Table 9 and 10 we have the results for the unit root tests applied to the variables
M2 growth, GDP growth, Exports, Stocks and Inflation. We find that M2 growth, GDP growth
and Stocks are all stationary. For Inflation, all tests (except the Breitung test) confirm the that
the variable is stationary. However, considering that for Exports, the null hypothesis that
panels contain unit roots cannot be rejected for three of the four tests, we decide to use the
first difference of exports in the model – where we accept the stationarity of the variable in
three out of four tests (results of unit root tests before and after differentiation are presented
in Table 11).

Bulgaria Czech Republic Croatia Hungary

10
0
-10
-20

Iceland Israel Lithuania Poland

GDP_growth
10
0
-10
-20

Romania Slovak Republic Slovenia Latvia

10
0
-10
-20

1995 2000 2005 2010 1995 2000 2005 2010 1995 2000 2005 2010 1995 2000 2005 2010

Year
Graphs by ID

Figure 3. GDP growth evolution in the period 1995 – 2012 for advanced economies

25
 Quantitative Methods Inquires

Table 9. Results of Unit Root Tests for M2 Growth and GDP Growth
Unit Root Test Statistic P-Value Unit Root Test Statistic P-Value
Levin-Lin-Chu -3.4923 0.0002 Levin-Lin-Chu -4.8273 0.0000
Breitung -2.3557 0.0092 Breitung -5.5205 0.0000
Im-Pesaran-Shin -4.1504 0.0000 Im-Pesaran-Shin -3.7274 0.0001
Hadri LM test 7.7165 0.0000 Hadri LM test 2.0937 0.0181

Table 10. Results of Unit Root Tests for Stocks and Inflation
Unit Root Test Statistic P-Value Unit Root Test Statistic P-Value
Levin-Lin-Chu -4.9641 0.0000 Levin-Lin-Chu -5.5655 0.0000
Breitung -3.4122 0.0003 Breitung 0.1903 0.5754
Im-Pesaran-Shin -1.7552 0.0396 Im-Pesaran-Shin -4.6811 0.0000
Hadri LM test 6.3827 0.0000 Hadri LM test 3.3155 0.0005

Table 11. Results of Unit Root Tests for Exports and Variation in exports (first difference)
Unit Root Test Statistic P-Value Unit Root Test Statistic P-Value
Levin-Lin-Chu -1.6202 0.0526 Levin-Lin-Chu -5.0471 0.0000
Breitung 0.5703 0.7158 Breitung -6.9226 0.0000
Im-Pesaran-Shin 0.3539 0.6383 Im-Pesaran-Shin -5.5524 0.0000
Hadri LM test 20.7622 0.0000 Hadri LM test -0.7522 0.7740

In what follows, we proceed to the same steps as for the sample of advanced
economies. We estimate the logistic model – with random and fixed effects. As we did
previously, we include in the list of variables all the variables lagged for two periods. This
time, we obtain that only the GDP growth with lag one, the variation of exports with lag two
are significant in the model. The results of the estimation with random effects are presented
in Table 12.
The model is valid, considering the likelihood-ratio test for rho (p-value = 0.001).
Considering the p-values of the variables included in the sample, at a 0.05 significance
level, the following variables remain significant: M2, GDP growth, Variation in Exports and
GDP growth lagged one period (both coefficients with negative signs, as expected). Inflation
level and stocks, as well as the variation in exports lagged with two periods are not
significant early warning signs for predicting crisis in the case of emerging economies. The
signs of the significant variables are related to economic theory. A decrease in the GDP
growth or a decrease in the money supply can be considered the most significant early
warning signals for the emergent economies. In Table 13 we present the post-estimation
results for the random effects logistic model. We notice that the model give weaker results
than the one for the advanced economies. This could be mainly due to the lower number of
variables included in the model. The expected probabilities for the Baltic Countries
(Lithuania, Latvia) are, as expected, the most close to one, as these are countries which
experienced the most dramatic fall in the economy (as also shown from the graph).

26
 Quantitative Methods Inquires

Table 12. Results for the estimation of the logistic model for emerging economies
(random effects)
Random-effects logistic regression Number of obs = 180
Group variable: id Number of groups = 12

Random effects u_i ~ Gaussian Obs per group: min = 15

avg = 15.0
max = 15

Integration method: mvaghermite Integration points = 12

Wald chi2(7) = 27.20

Log likelihood = -56.892479 Prob > chi2 = 0.0003

bc Coef. Std. Err. z P>|z| [95% Conf. Interval]

m2_growth -.1219335 .0436757 -2.79 0.005 -.2075362 -.0363308

gdp_growth -.3857805 .104889 -3.68 0.000 -.5913591 -.1802019
d_exports .1089936 .0505533 2.16 0.031 .0099109 .2080763
stocks -.0069754 .0180328 -0.39 0.699 -.0423191 .0283683
inflation -.0020855 .0561766 -0.04 0.970 -.1121897 .1080186

gdp_growth
L1. -.1313661 .0790617 -1.66 0.097 -.2863241 .0235919

d_exports
L2. .0880516 .0593502 1.48 0.138 -.0282727 .2043759

_cons .4819724 .7813385 0.62 0.537 -1.049423 2.013368

/lnsig2u 1.478609 .832565 -.1531886 3.110406

sigma_u 2.094478 .8718945 .9262655 4.736048

rho .571448 .2038912 .2068471 .8720892

Likelihood-ratio test of rho=0: chibar2(01) = 10.26 Prob >= chibar2 = 0.001

Table 13. Post estimation results for the logistic model – emerging economies
(random effects)
Country Year Exp Prob
Bulgaria 2009 0.5983654
Czech 2009 0.7480773
Croatia 2009 0.9052944
Hungary 2009 0.8786198
Iceland 2009 0.9756301
Israel 2009 0.1613406
Lithuania 2009 0.9918979
Poland 2009 0.1706047
Romania 2009 0.7642968
Slovak Republic 2009 0.3201566
Slovenia 2009 0.5925208
Latvia 2009 0.9994201

27
 Quantitative Methods Inquires

5. Conclusions

In the present paper, we propose a framework to be used for developing an Early

Warning System for assessing systemic risk. We find important insight regarding the
macroeconomic variables that could be considered early triggers of banking distress. On one
hand, for advanced economies, the cash deficit, the variation in exports and inflation are not
significant signals for situation of crisis, while for emerging economies, inflation and value of
stocks traded turn out to have no prediction power for predicting crisis (a note should be
made here that the indicator value of stocks traded is significant for the advanced economies
– this could be explained by the still immature stock market in emerging economies). On the
other hand, the evolution on GDP growth is the most important signal for a crisis situation,
that is one year prior to crisis eruption. Moreover, the paper ads important contribution to
the specialty literature by considering the Output Gap in the model – which is find to be a
significant trigger for the inefficiency of the economy and a good predictor of crises. The
model has very good estimates of the probability of default, confirming the set of most
affected economies by the Financial Crisis (Greece, Italy, Ireland, Portugal, Baltic Countries)
and stable economies – the Nordic Countries.
Paper is subject to further development – quarterly data could be used instead on
annually for a more dynamic picture of the crisis development; also, instead of the binary
variable, a continuous index for banking or financial stability would offer much more
information for the economy’s evolution.

References

1. Borio, C. and Lowe, P. Assessing the Risk of Banking Crises, Bank for International
Settlements Quarterly Review, 2002, pp. 43-54
2. Bussiere, M. and Fratzscher, M. Towards a new early warning system of financial crises,
European Central Bank Working Paper, No. 145, Frankfurt, 2002
3. Chauvet, M. and Dong, F. Leading indicators of country risk and currency crises: the Asian
experience, Economic Review, Federal Reserve Bank of Atlanta, Issue Q1, 2004, pp.
25-37
4. Davidescu, A. A. M. Evaluating the relationship between official economy and shadow
economy in Romania. A Structural Vector Autoregressive approach, Journal of
Social and Economic Statistics, vol.3, no.2, 2014, pp. 57-65
5. Demirgüç-Kunt, A. and Detragiache, E. The Determinants of Banking Crises in Developing
and Developed Countries, IMF Staff Papers, Vol. 45, No. 1, 1998, pp. 81-109
6. Diebold, F. X. and Rudebusch, G. D. Scoring the Leading Indicators, Journal of Business, Vol.
62, No. 3, 1989, pp. 369-391
7. Edison, H. Do Indicators of Financial Crises Work? An Evaluation of an Early Warning
System, International Journal of Finance and Economics, Vol. 8, Iss. 1, 2003, pp. 11-
53
8. Fuertes A. and Kalotychou, E. Elements in the Design of an Early Warning System for
Sovereign Default, Computing in Economics and Finance, Society for Computational
Economics, Vol. 231, 2004
9. Gramlich, D., Miller, G.L., Oet, M.V. and Ong, S.J. Early warning systems for systemic banking
risk: critical review and modelling implications, Banks and Bank Systems , Volume
5, Issue 2, 2010
10. Greene, W.H. Econometric Analysis – 7th Edition, Prentice Hall, 2011

28
 Quantitative Methods Inquires

11. Hanschel, E. and Monnin, P. Measuring and Forecasting Stress in the Banking Sector:
Evidence from Switzerland, Bank for International Settlements Working Paper, Basel,
No. 22, 2005
12. Illing, M. and Liu, Y. An Index of Financial Stress for Canada, Bank of Canada Working Paper,
Ottawa, No. 14, June 2003
13. Kaminsky, G. and Reinhart, C. The Twin Crises: The Causes of Banking and Balance-of-
Payments Problems, American Economic Review, Vol. 89, No. 3, 1999, pp. 473-500
14. Kauko, K. How to forsee banking crises? A survey of the empirical literature, Economic
Systems, 2014 in press
15. Laeven, L. and Valencia, F. Systemic Banking Crises: a New Database, IMF Working Paper
No. 08, 2008, pp. 1-78
16. Miricescu, E. C. Investigating the determinants of long-run sovereign rating, Financial
Studies, Volume 18, issue 3, 2014, pp. 25-32
17. Mody A. and Taylor M.P. Common Vulnerabilities, Centre for Economic Policy Research (CEPR)
Discussion Papers 3759, 2003
18. Moscalu, M. The impact of interest rate spreads for Euro denominated loans on the
leverage ratio of Romanian listed companies, Proceedings of the 16th
International Scientific Conference Finance and Risk, Vol. I, Bratislava, Publishing
House Ekonom, 2014, pp. 138–145
19. Zaman, G., Goschin, Z., Partachi, I. and Herteliu, C. The contribution of labour and capital to
Romania's and Moldova's economic growth, Journal of applied quantitative
methods, 2(1), 2007, pp. 179-185
20. Zaman, G., Goschin, Z., and Herteliu, C. Analysis Of The Correlation Between The Gdp
Evolutions And The Capital And Labor Factors In Romaniam, Romanian Journal
for Economic Forecasting, 2(3), 2005, pp. 5-21

1
Acknowledgement
This work was cofinanced from the European Social Fund through Sectoral Operational Programme Human
Resources Development 2007-2013, project number POSDRU/159/1.5/S/134197 „Performance and excellence in
doctoral and postdoctoral research in Romanian economics science domain”.

29
 Quantitative Methods Inquires

STOCHASTIC OPTIMIZATION USING INTERVAL ANALYSIS,

WITH APPLICATIONS TO PORTFOLIO SELECTION1

Silvia DEDU
PhD, Bucharest University of Economic Studies, Department of Applied Mathematics;
Romanian Academy, Institute of National Economy

E-mail: [email protected]

Florentin SERBAN
PhD candidate, University of Bucharest, Faculty of Mathematics and Computer Science
PhD, Bucharest University of Economic Studies, Department of Applied Mathematics

Abstract
In this paper we study a class of optimization problems under uncertainty, with parameters
modeled by stochastic random variables. Interval analysis and multiobjective stochastic
programming concepts are introduced. Then these two concepts are combined to build a
stochastic programming model, with the coefficients of the constraints and the coefficients of
the objective function modeled by interval numbers and discrete interval random variables.
This model can be used to solve a portfolio optimization problem.
Keywords: interval analysis, multiobjective stochastic programming, uncertainty,
optimization
1. Introduction

The input parameters of the mathematical programming model are not exactly
known because relevant data are inexistent or scarce, difficult to obtain or estimate, the
system is subject to changes, and so forth, that is, input parameters are uncertain in nature.
This type of situations are mainly occurs in real-life decision-making problems. These
uncertainties in the input parameters of the model can characterize by interval numbers or
random variables with known probability distribution.
The occurrence of randomness in the model parameters can be formulated as
stochastic programming (SP) model. SP is widely used in many real-world decision-making
problems of management science, engineering, and technology. Also, it has been applied to
a wide variety of areas such as, manufacturing product and capacity planning, electrical
generation capacity planning, financial planning and control, supply chain management,
dairy farm expansion planning, macroeconomic modeling and planning, portfolio selection,
traffic management, transportation, telecommunications, and banking
An efficient method known as two-stage stochastic programming (TSP) in which
policy scenarios are desired for studying problems with uncertainty. In TSP paradigm, the
decision variables are partitioned into two sets. The decision variables which are decided
before the actual realization of the uncertain parameters are known as first stage variables.
Afterward, once the random events have exhibited themselves, further decision can be made

30
 Quantitative Methods Inquires

by selecting the values of the second-stage. The formulation of two-stage stochastic

programming problems was first introduced by Dantzig [5]. Further it was developed by Barik
[2] and Wallcup [15].This article proposes such of stochastic programming

2. Interval Analysis

Interval analysis was introduced by Moore [10] . The growing efficiency of interval
analysis for solving various real life problems determined the extension of its concepts to the
probabilistic case. Thus, the classical concept of random variable was extended to interval
random variables, which has the ability to represent not only the randomness character,
using the concepts of probability theory, but also imprecision and non-specifficity, using the
concepts of interval analysis. The interval analysis based approach provides mathematical
models and computational tools for modeling data and for solving optimization problems
under uncertainty.
The results presented in this chapter are discussed in more detail in [1, 11],
Let x L , xU be real numbers, x L  xU .
Definition 2.1. An interval number is a set defined by:

X  [ x]  x  R | x L  x  xU ; x L , xU  R . 
Remark 2.1. We will denote by x the interval number x L , xU , with   x L ,xU R .
We will denote by IR the set of all interval numbers.

Definition 2.2. Let (Ω,K,P) be a probability space and IR be the set of the real intervals. An
interval random variable [X] is an application [X]: ΩR, defined by:
[X]() =[XL(),XU()], where XL, XU: ΩR are random variables, such that XL  XU almost
surely. We say that the interval random variable [X] is a discrete interval random variable if it
takes values in a finite subset of the set of the real numbers Otherwise we say that [X] is a
continuous interval random variable.
Definition 2.3. The product between the real number a and the interval number x is
defined by:
 
 a  x L , a  xU , if a  0

 
a  x  a  x| x x   a  xU , a  x L , if a  0 .


0, if a  0

Let 
[x]  xL , xU  and [ y]  y , y 
L U
be interval numbers, with x , x , y , y R .
L U L U

Definition 2.4. The equality between interval numbers is defined by:

x   y if and only if x L  y L and x U  y U .
Definition 2.5.
     L L U U .
The summation of two interval numbers is defined by: x  y  x  y , x  y

The subtraction of two interval numbers is defined by: x   y  x  y , x  y  .

L U U L

The product between two interval numbers is defined by:

x y  minx L y L , x L yU , xU y L , x L yU , maxx L y L , x L yU , xU y L , x L yU 

31
 Quantitative Methods Inquires

1  1 1 
0 y , then
1
If is defined by: 
 y   y U y L
, ,
y 
If 0 y , then the division between two interval numbers is defined by:
x  min x L , xU , max x L , xU  .
 y    y U y L   U
 y y 
L 

Definition 2.6. x   y if f x  y and x  y ;

L L U U

3. Stochastic Programming

3.1. Multiobjective Stochastic Programming

In stochastic or probabilistic programming some or all of the parameters of the
optimization problem are described by stochastic or random variables rather than by
deterministic quantities. In recent years, multiobjective stochastic programming problems
have become increasingly important in scientifically based decision making involved in
practical problem arising in economic, industry, healthcare, transportation, agriculture,
military purposes, and technology. Mathematically, a multiobjective stochastic programming
problem can be stated as follows:
n
max z t  c
j 1
t
j ] x j , t  1, 2 ,..., T

n
subject to a
j 1
ij x j  b i , i  1, 2 ,..., m 1 (3.1)
n

d
j 1
ij x j  b m1  i , i  1, 2 ,..., m 2

x j  0 , j  1, 2 ,..., n
where the parameters aij , i  1,2,..., m1 , j  1,2,..., n and bi , i  1, 2,..., m1 are discrete
random variables with known probability distributions. The rest of the parameters
c tj , j  1,2,..., n , t  1,2,..., T , d ij , i  1,2,..., m2 , j  1,2,..., n and bm1 i , i  1,2,..., m2
are considered as known intervals.

3.2. Multiobjective Two-Stage Stochastic Programming

In two-stage stochastic programming (TSP), decision variables are divided into two
subsets:
(1) a group of variables determined before the realizations of random events are
known as first stage decision variables, and
(2) another group of variables known as recourse variables which are determined
after knowing the realized values of the random events.
A general model of TSP with simple recourse can be formulated as follows [3,9,15]:

32
 Quantitative Methods Inquires

n
 m1 
max z   c j x j  E   q i y i 
j 1  i 1 
n
subject to y i  bi   a ij x j , i  1,2,..., m1
j 1
n

d
j 1
ij x j  bm1  i , i  1,2,..., m 2 (3.2)

x j  0, j  1,2,..., n
y i  0, i  1,2,..., m1
where x j , j  1,2,...,n and yi , i  1,2,...,m1 are the first stage decision variables and second
stage decision variables respectively.
Further, qi , i  1,2,...,m1 are defined as the penalty costs associated with the
n
discrepancies between a x
j 1
ij j and bi and E is used to represent the expected value of a

random variable.
Multiobjective optimization problems appear in most of the real life decision
making problems. Thus, a general model of multiobjective stochastic programming model
(3.1) can be stated as follows:
n
 m1 
max z t   c tj x j  E   q it y i , t  1,2,..., T
j 1  i 1 
n
subject to y i  bi   a ij x j , i  1,2,..., m1
j 1
n

d
j 1
ij x j  bm1  i , i  1,2,..., m 2 (3.3)

x j  0, j  1,2,..., n
y i  0, i  1,2,..., m1

4. Random Interval Multiobjective Two-Stage Stochastic Programming

Optimization model incorporating some of the input parameters as interval random

variables is modeled as random interval multiobjective two-stage stochastic programming
(RIMTSP) to handle the uncertainties within TSP optimization platform with simple recourse.
Mathematically, it can be presented as follows:

33
 Quantitative Methods Inquires

 m1 
 
n

max z t   c tj x j  E   q it y i , t  1,2,..., T
j 1  i 1 
n
subject to y i  [ Bi ]   [ Aij ] x j , i  1,2,..., m1
j 1

 d x  
n

ij j  bm1  i , i  1,2,..., m 2 (4.1)

j 1

x j  0, j  1,2,..., n
y i  0, i  1,2,..., m1
where x j , j  1,2,...,n and yi , i  1,2,...,m1 are the first stage decision variables and second
stage decision variables respectively. Further, c , j  1,2,..., n , t  1,2 ,...,T
t
j are the costs

associated with the first stage decision variables and qit , i  1,2 ,...,m1 , t  1,2 ,...,T are the
n
penalty costs associated with the discrepancie between [ A ]x
j 1
ij j and [ Bi ] of the kth

objective function. The left hand side parameter [ Aij ] and the right hand side parameter

[ Bi ] are discrete interval random variables, with known probability distributions and
between E is used to represent the expected value associated with interval random
variables.

5. Conclusions

The new approach based on interval analysis provides mathematical models and
computational tools for modeling the imprecision of financial data and for solving decision
making problems under uncertainty. This article have proposed a random interval
multiobjective two-stage stochastic programming

References

1. Alefeld, G., Herzberger, J., Introduction to Interval Computation, Academic Press, 1983
2. Barik, S.K, Biswal, M.P. and Chakravartay, D. Multiobjective Two-Stage Stochastic
Programming Problems with Interval Discrete Random Variables, Advances in
Operations Research, Article ID 279181, 2012
3. Birge J.R. and Louveaux F. Introduction to Stochastic Programming, Springer, New York, 1997
4. Chinneck J.W. and Ramadan K. Linear programming with interval coefficients, Journal of the
Operational Research Society, Vol. 51, No. 2, 2000, pp. 209-220
5. Dantzig, G.B Linear programming under uncertainty, Management Science 1, pp. 3-4,1955
6. Hansen, E. and Walster, G.W. Global optimization using interval analysis, Marcel Decker Inc.,
2004
7. Ileanu, B. V., Isaic-Maniu, A. and Herteliu, C. Intellectual capital components as causes of
regional disparities. A case study in Romania, Romanian Journal of Regional
Science, Vol. 3, No. 2, 2009, pp. 39-53

34
 Quantitative Methods Inquires

8. Isaic-Maniu, A. an Herteliu, C. Ethnic and religious groups in Romania-Educational (co)

incidences, Journal for the Study of Religions and Ideologies, No. 12, 2005, pp. 68-75
9. Kambo, N.S. Mathematical Programming Techniques, Affiliated East-West Press, New York,
1997
10. Moore, R.E. Methods and Application of Interval Analysis, SIAM Philadelphia, 1979
11. Nobibon, F.T. and Guo, R. Foundation and formulation of stochastic interval programming,
PhD Thesis, African Institute for Mathematical Sciences, Cape Town, South Africa, 2006
12. Raducanu, A. M., Feraru, V., Herteliu, C. and Anghelescu, R. Assessment of The Prevalence of
Dental Fear and its Causes Among Children and Adolescents Attending a
Department of Paediatric Dentistry in Bucharest, OHDMBSC, Vol. 8, No. 1, 2009,
pp. 42-49
13. Shapiro, A., Dentcheva D. and Ruszczynski A. Lectures on Stochastic Programming: Modeling
and Theory, Book Series: MOS-SIAM Series on Optimization 9, 2009, pp. 1-436
14. Toma, A. and Dedu, S. Quantitative techniques for financial risk assesment: a comparative
approach using different risk measures and estimation methods, Procedia
Economics and Finance, Vol. 8, 2014, pp. 712-719
15. Walkup, D.W. and Wets R.J.B. Stochastic programs with recourse, SIAM Journal on Applied
Mathematics 15, 1967

1
Acknowledgement
This paper has been financially supported within the project entitled “Horizon 2020 - Doctoral and Postdoctoral
Studies: Promoting the National Interest through Excellence, Competitiveness and Responsibility in the Field of
Romanian Fundamental and Applied Scientific Research”, contract number POSDRU/159/1.5/S/140106. This
project is co-financed by European Social Fund through Sectoral Operational Programme for Human Resources
Development 2007-2013. Investing in people!

35
 Quantitative Methods Inquires

BOUNDS TEST APPROACH FOR THE LONG RUN

RELATIONSHIP BETWEEN SHADOW ECONOMY AND
OFFICIAL ECONOMY. AN EMPIRICAL
ANALYSIS FOR ROMANIA1

Adriana AnaMaria DAVIDESCU2

PhD, Lecturer, Department of Statistics and Econometrics,
Bucharest University of Economic Studies,
Researcher, National Scientific Research Institute
for Labour and Social Protection, Bucharest, Romania

E-mail: [email protected]

Abstract
The paper aims to investigate the nature of the relationship between the shadow economy (SE)
and recorded GDP for the case of Romania using Pesaran et al.(2001) bounds tests approach
for cointegration for the period 2000-2010. The size of Romanian shadow economy is
estimated using a revised version of the currency demand approach based on autoregressive
distributed lag (ARDL) approach to cointegration analysis. To investigate the long-run causal
linkages and short-run dynamics between shadow economy and recorded GDP, ARDL
cointegration approach is applied.
The ARDL causality results revealed only the existence of a long-run unidirectional causality
that runs from shadow economy official economy, revealing a negative relationship betwwen
them on long-run. In addition, the CUSUM and CUSUMSQ tests confirm the stability of causal
relationships.

Keywords: shadow economy, currency demand approach, economic development, ARDL

cointegration approach, CUSUM, CUSUMQ tests

1. Introduction

The impact of the shadow economy on overall economic performance was

investigated in various studies (Dell’Anno, 2003; Schneider and Klinglmair, 2004; Schneider,
2005; Dell’Anno, 2008; Halicioglu and Dell’Anno, 2009).
Klinglmair and Schneider (2006),Giles (1997a, b), and Giles et al. (2002) pointed
out that an increase in the size of shadow economy will affect the tax base, leading to lower
official growth. Also, an increased size of SE will be very attractive for workers from official
sector, creating unfair competition between unofficial and official firms (Enste, 2003). Hidden
activities favor corruption and link with criminal activities.
At opposite side, SE can creates positive effects to official economy, creating an
extra-added value that Schneider and Enste(2000) consider that can be spent in the official
economy, estimating that at least two-third of the income earned in unofficial market is
spent in the official economy. Smith (2002) argues that shadow economy have a positive

36
 Quantitative Methods Inquires

effect on employment, helping some individuals that otherwise will be unemployed and so
the unofficial sector may represent a social buffer in the countries with high unemployment
rate. Giles (1997a, 1997b, 1999a) and Giles and Tedds (2000) carried out one of the most
relevant technique-Granger causality approach in New Zealand and Canada, revealing a
significant Granger causality that runs from official economy to unofficial one.
Schneider (2005) quantifies the relationship between SE and official economy,
pointing out that the degree of economic development has relevant implications on both
sectors. The empirical results have pointed out the existence of a negative between SE and
the official economy for developing countries and a positive relationship for industrialized
and transition countries, revealing that SE is pro-cyclical for developing economies and
countercyclical for developed and transition countries.
In this study, I adopted the definition of Schneider (2006) and Schneider et al.
(2010) regarding the shadow economy3 and the subject of the paper do not deal with
typical underground, economic (classical crime) activities, which are all illegal actions that fit
the characteristics of classical crimes like burglary, robbery, drug dealing and also exclude
the informal household economy which consists of all household services and production.
The main empirical results regarding the Romanian shadow economy are obtained
by both national and international studies using different estimation methods and are
presented in table 1.
Tabel 1. The size of Romanian shadow economy (% of official GDP)
Authors Approach Period Size of SE
(min-max)
Albu(2003, 2008, 2010, Discrepancy between 1995- 14.6%-
2011) actual and desired income 2007 22.3%
Institutul Național de Labour input method 1998- 14.5%-
Statistică 2009 23.5%
Johnson(1997, 1998) Physical input method 1990- 18.0%-
1995 28.3%
Lacko(1999) Physical input method 1990- 20.9%-
1995 31.3%
Schneider et al.(1998, DYMIMIC model and 1990- 26.2%-
2000, 2002, 2004, 2005, currency demand approach 2005 37.4%
2006, 2007, 2009)
Schneider et al.(2010) MIMIC model 1999- 34.4%-
2006 36.7%

As Schneider and Enste (2000) stated, no approach is exempt from criticism, the
empirical results being different. So, if according to National Institute of Statistics, the
informal activity represents between 14.5% and 23.5% of official GDP, Schneider et
al.(2010) estimates the size of shadow economy in Romania to overcome the threshold of
35% of official GDP.
The paper aims to investigate the relationship between the size of the shadow
economy (SE) and official economy for the case of Romanian data using bounds test
approach and ARDL causality analysis for quarterly data covering the period 2000-2010. The
size of Romanian shadow economy was estimated using using a revised version of the
currency demand approach based on bounds testing approach to cointegration and error
correction models, developed within an autoregressive distributed lag (ARDL) framework. A
detailed description of the shadow economy estimation is presented in (Davidescu and
Dobre 2013).

37
 Quantitative Methods Inquires

The empirical results of currency demand approach based on ARDL models

emphasizes that there is a general downward trend in the size of the shadow economy as %
of official GDP for the period 2000-2010 with an highlight on two low periods, 2003Q1 and
2008Q4.Thus, the size of the shadow economy as % of official GDP measures approximately
45% at the end of 2000 and achieving the value of 37.4% in the last quarters of the period.
The estimates are in line with the last empirical studies4.
It is important to note that because of its undetectable nature and character, it is
nearly impossible to measure precisely the size of economic activities taking place in the
informal economy of any country in the world, whether developed or less developed. Given
this, any theoretical or empirical inference derived from these results should always be
regarded as an approximation. In the face of these difficulties, the results drawn from these
estimates should be interpreted with due reserve, given the limitations of the methods.
The paper is divided three sections presenting the data, the methodology and the
main econometrical results.

2. The relationship between shadow economy and official economy

for the case of Romania

Official economic situation plays a crucial role in people's decision to work or not in
the informal sector (Bajada and Schneider, 2005; Schneider et al., 2010). In a booming
official economy, people have a lot of opportunities to earn a good salary and even extra
money. This is noi the case of a economy in recession, when people try to compensate the
loss of income from formal economy through involvement in the informal economy.
The shadow economy manifests both positive and negative effects on official
economy. The studies of Frey and Weck-Hannemann (1984), Loayza (1996), Kaufmann si
Kaliberda (1996), Eilat and Zinnes (2000), Schneider and Enste (2000), Ott (2002),
Dell’Anno (2003), Dell’Anno, Gomez and Alañón (2007) and Dell’Anno (2007) argue the
existence of a negative effect of shadow economy on GDP growth, based upon the idea that
unofficial activities, by creating unfair competition, interfere negatively with the market
allocation.
A negative correlation between the size of informal sector and the growth rate of
official real GDP per capita for 14 Latin American countries is also found by Loayza(1996),
while the same conclusion has been drawed be Eilat and Zinnes (2000) in 24 transition
countries, revealing that a one-dollar fall in official GDP was associated with a 31-percent
increase in the size of the SE.
Kaufmann and Kaliberda (1996) estimated that for “every 10 percent cumulative
decline in official GDP, the share of the irregular economy in the overall increases by almost
4 percent” (ibidem, p. 46).The 76 countries survey conducted by Schneider and Enste’s
(2000) pointed out that a growing SE has a negative impact on official GDP growth.
At other side, the shadow economy may manifest an positive impact on GDP
growth, creating markets, increasing financial resources, enhancing entrepreneurship, and
transforming social institutions, economic, legal and necessary capital accumulation (Asea,
1996). The positive realtionship between shadow economy and official one was revealed in
studies such as: Adam and Ginsburgh (1985), Giles (1999), Giles and Tedds (2002), Tedds
(2005), Schneider and Hametner (2007), Chatterjee, Chaudhuri and Schneider (2003),

38
 Quantitative Methods Inquires

Dell’Anno (2008), Bovi and Dell’Anno (2007), Dell’Anno and Halicioglu (2010), Schneider
and Klinglmair (2004) and Brambilla (2008).
Schneider and Enste (2000) considers the informal economy creates additional
value that can be spent in the economy. The informal economy provides employment
opportunities to certain individuals who would otherwise be unemployed and provide
services to low income people who are involved in informal production activities. Thus, it
represents a "social buffer" for countries with high unemployment. Adam and Ginsburg
(1985) found a positive relationship between the growth of the SE and the official economy
under the assumption of low probability of enforcement.
Enste (2003) argues that SE stimulates economic development in transition
countries. He considers the shadow economy as an incentive to develop both the
entrepreneurial spirit and a constraint to limit an excessive growth of the government
activities. Schneider (2003) emphasizes that UE, stimulating higher competition, leads to
more efficient resource allocation on both sides of economy.
Also Dell’anno(2008) has analysed the relationship between unofficial economy
(UE) and official GDP, reaveling a positive correlation is found between unofficial and official
GDP, SE being considered as beneficial to sustain economic growth.
Halicioglu and Dell’Anno(2009) estimated the size of unrecorded economy (SE) of
Turkey over the period 1987-2007 using a revised version of the currency demand
approach and analyzed the relationship between UE and recorded GDP (gross domestic
product) revealing that causality runs from the recorded GDP to the SE.
In Latin American countries, the study of Maloney(1999) revealed empirical
evidence on substantial flow of workers back and forth between formal and informal
employment. Galli and Kucera (2003) assess that “informal employment serves as a
macroeconomic buffer for formal sector employment over the course of business cycles, with
informal employment expanding during downturns and contracting during upturns (ibidem,
p. 17)”.
In 2005, Schneider considers that the effects of SE on the official economic growth
are conditioned to the degree of economic development, revealing a negative relationship
for low-income countries and a positive one in industrialized and transition countries. The
explanation was that in high-income countries citizens are overburdened by taxes and
regulation so that an increasing SE stimulated the official economy as the additional income
earned in the SE was spent in the official sector. On the contrary, for low-income countries,
an increasing SE “erodes the tax base, with the consequence of a lower provision of public
infrastructure and basic public services with the final consequence of lower official economy”
(Schneider, 2005, p. 613).
A valuable paper that traits the relationship between official and unofficial
economy for the ASEAN from 1996 to 2013 is written by Vo and Pham (2014) who finds that
when the official economy is proxied by the GDP growth or the GDP per capita growth, the
unofficial economy negatively contributes to the official economy.

2.1. Data
In the econometrical demarche of the investigation of the relationship between
shadow and official economies, it has been used quarterly data covering the period
2000:Q1 to 2010:Q2.

39
 Quantitative Methods Inquires

The size of Romanian shadow economy as % of official GDP has been obtained
using a revised version of the currency demand approach based on bounds testing approach
to cointegration and error correction models, developed within an autoregressive distributed
lag (ARDL) framework.A detailed description of the shadow economy estimation is presented
in (Davidescu & Dobre, 2013).
The empirical results of currency demand approach based on ARDL models
emphasizes that there is a general downward trend in the size of the shadow economy as %
of official GDP for the period 2000-2010 with an highlight on two low periods, 2003Q1 and
2008Q4.Thus, the size of the shadow economy as % of official GDP measures approximately
45% at the end of 2000 and achieving the value of 37.4% in the last quarters of the period.
The estimates are in line with the last empirical studies5.
The official economy was quantified using real official gross domestic
product(2000=100) expressed in millions RON taken from Tempo database of National
Institute of Statistics.
The graphical evolution of the shadow economy versus official economy reveals the existence
of a negative relationship between variables, intermediate as intensity quantified by a value
of -0.65 of correlation coefficient.

Figure 1. The size of the shadow economy vs. official economy in Romania
Source: Tempo database of National Institute of Statistics

The aim of the paper is to investigate the nature of the relationship between official
economy and the size of the Romanian shadow economy and to identify the direction of
causality between them using ARDL cointegration and causality approach.

2.2. Methodology
The non-stationary analysis is realised using the the unit root tests (The Augmented
Dickey-Fuller (ADF) and Phillips-Perron (PP)). The bounds test approach were applied in
prder to verify the possible relationship between these two variables, having the advantage
that the regressors can have different order of integration.
The models that describe the relationship between these two variales are:
official _ economy t   1  1  SE t   1t (1)
SE t   2   2  official _ economy t   1t (2)

40
 Quantitative Methods Inquires

where: SE t is the size of Romanian shadow economy as % of official GDP obtained through
ARDL models; the official economy is cuantified using is real GDP expressed in prices of
2000; 1 , 2 are constants;  1t ,  2t are the disturbance terms.
The first step in the ARDL approach to cointegration is to estimate the following
relationship using the OLS estimation technique:
m m

t  a0  a1i official
official_ economy t i  a2i SEt i 
_ economy
i1 i0 (3)

t 1  a4  SEt 1  1t
 a3  official_ economy
m m
SEt  b0  b1i SEt i  b2i official_ economyt i 
i 1 i 0 (4)
 b3  SEt 1  b4  official_ economyt 1   2t
where: ∆ is the difference operator; SEt is the size of Romanian shadow economy as % of
official GDP;official economy is expressed using real GDP(2000=1000; 1t and 2t are are the
disturbance terms ; “m” lags.
The first part of equations (3)-(4) with a1i , a 2 i and b1i , b2 i represents the short-run
dynamics of the models and the second part with a 3 , a 4 and b3 ,b4 represent the long-run
phenomenon.
The null hypothesis in the first equation (3) is H 0 : a 3  a 4  0 , which means the
non-existence of a long-run relationship against the alternative H 1 : a 3  a 4  0 meaning
that there is a long-run relationship. In the second equation (4), the null is H 0 : b3  b4  0
against the alternative H 1 : b3  b4  0 which states that we have cointegration. The F tests
for the joint significance of the coefficients on the one period lagged levels of the variables is
compared with the F critical taken from Pesaran6 (2001) or Narayan7 (2005).
Once cointegration is confirmed, we move to the second stage and estimate the
long-run coefficients of the level equations (1)-(2) and the short-run dynamic coefficients via
the following ARDL error correction models8:
m
official _ economyt   0    1i official _ economyt i 
i 1
n
   2i SEt i   3 ECTt 1   1t (5)
i 0
m n
SE t  0   1i SE t i   2 i official _ economy t i 
i 1 i 0

 3 ECTt 1   2 t (6)

where: SE t , official _ economy t are the variables analysed; ∆ is the difference operator
and ECTt-1 is one lag error correction term that maust be negative,  3 , 3 are the
adjustement speed to the equilibrium after a shock. The coefficients  1i ,  2i , 1i ,  2i are the
coefficients for the short-run dynamics of the model’s convergence to equilibrium, and
 1t ,  2t are the error terms. To ascertain the goodness of fit of the ARDL models, diagnostic
and stability tests are conducted. The diagnostic test examines the serial correlation,
functional form, normality, and heteroscedasticity associated with the model. Parameter

41
 Quantitative Methods Inquires

stability is important since unstable parameters can result in model misspecification

(Narayan and Smith, 2004). The stability of parameters is tested using the Cusum and
CusumQ tests.
The third stage includes conducting standard Granger causality tests augmented
with a lagged error-correction term. A statistically significant ECT term implies long-run
causality running from all the explanatory variables towards the dependent variable.
An augmented form of Granger causality test is involved to the error-correction
term and it is formulated in a bi-variate p-th order vector error-correction model (VECM)
which is as follows:
SEt  c1  11p L SEt  12q L official _ economyt   1  ECTt 1   1t (7)
t  c2  Lofficial
official_ economy t  LSEt  2  ECT
t 1   2t
p q
21_ economy 22 (8)
where:
P11 P12 P21 P22
11p L    11p ,i Li , 12p L    12p ,i Li ,  21p L     21p ,i Li  22p L     22p ,i Li
i 1 i 0 i 1 i 0

SE t , official _ economy t are the analysed variables; ∆ denotes the difference operator. L
denotes the lag operator, where (L)∆Yt = ∆Yt-1,  1t and  2t are the disturbance terms.
Iinto a matrix form, the Granger causality looks as follows:
 SEt  c1  p 11p 12p  SEti 
       p   
official_ economyt  c2  i1 21 22 official_ economyt i 
p

(9)
 ECTt 1  1t 
 1  
2 ECTt 1   2t 
where:  is a difference operator, ECT is the error-correction term from ARDL model, c i (i =
1, 2) is constant and i (i = 1, 2) are the disturbance terms. The optimal lag length p is
based on the Akaike Information Criterion. Long-run causality can be revealed through the
significance of the lagged ECTs by t test, while short-run causality is validated by using F-
statistic or Wald test.

2.2. Empirical results

The main goal of the study is to investigate the nature of the relationship between
the shadow economy and the official economy and to identify any possible direction of
causality between them. The analysis of stationarity using Dicley-fuller tst revealed that all
series are integrated on the same order, I(1).
Forthmore, we investigated the possibility of cointegration between the shadow
economy and official one using the bounds tests within the ARDL modeling approach. The
optimal lag length9 required in the bounds test cointegration test has been selected on the
both SBC and AIC Information Criteria.
The lag order selected by AIC in the model in which official economy is the
dependent variable is p  2 if a trend is included and p  4 if not and those selected by
SBC is p  2 irrespective of whether a deterministic trend term is included or not. In view of
the importance of the assumption of serially uncorrelated errors for the validity of the
bounds tests, the lag p  2 has been selected.
In the model in which shadow economy is the dependent variable, the lag order
selected by AIC and SBC is 1, irrespective of whether a deterministic trend term is included
or not.

42
 Quantitative Methods Inquires

A bounds F-test was applied to equation (4) for shadow economy and official
economy to establish a long-run relationship between the variables under the three
scenarios: with restricted deterministic trends (FIV), with unrestricted deterministic trends (FV)
and without deterministic trends (FIII) and with all intercepts unrestricted. The results are
presented in Table 2.

Table 2. The Bounds Test for Co-integration

With Without
Deterministic Trends Deterministic Trend

Variables FIV FV tV FIII tIII Conclusion

Off. economy and SE

Foff ec (off_ec / SE) Ho Accepted
p = 2* -1.65a -
3 -1.30a -
4 -1.57a -1.82a
5 -0.81a -1.44a
6 -1.41a
7 -1.65a
FSE (SE / off_ec) Ho Rejected
p = 1* -5.47c -4.40c
2 -3.57b -2.71b
3 -2.53a -1.73a
4 -2.93a -1.67a

Note: Akaike Information Criterion (AIC) and Schwartz Criteria (SC) were used to select the number of lags
required in the co-integration test. p shows lag levels and * denotes optimum lag selection in each
model as suggested by SBC. FIV represents the F statistic of the model with unrestricted intercept and
restricted trend, FV represents the F statistic of the model with unrestricted intercept and trend, and FIII
represents the F statistic of the model with unrestricted intercept and no trend. tV and tIII are the t ratios
for testing 1Y = 0 in equation (4) and 1X = 0 in Equation (5) respectively with and without
deterministic linear trend. a indicates that the statistic lies below the lower bound, b that it falls within
the lower and upper bounds, and c that it lies above the upper bound(Katircioglu, 2009).

The cointegration test under the bounds framework involves the comparison of the
F and t statistics against the critical values of F and t for ARDL approach presented in table 3
for the three different scenarios.

Table 3. Critical Values for ARDL Modeling Approach

90% level 95% level 99% level
k=1 I (0) I (1) I (0) I (1) I (0) I (1)

FIV 4.05 4.49 4.68 5.15 6.10 6.73

FV 5.59 6.26 6.5 7.30 8.74 9.63
FIII 4.04 4.78 4.94 5.73 6.84 7.84

tV -3.13 -3.63 -3.41 -3.95 -3.96 -4.53

tIII -2.57 -2.91 -2.86 -3.22 -3.43 -3.82

Source: Pesaran(2001) for F-statistics pg.300-301 and for t-ratios pg.303-304.

Note: (1) k10 is the number of independent variables in ARDL models (Erbaykal, 2008), FIV
represents the F statistic of the model with unrestricted intercept and restricted trend, FV
represents the F statistic of the model with unrestricted intercept and trend, and FIII represents
the F statistic of the model with unrestricted intercept and no trend. (2) tV and tIII are the t
ratios for testing a3 = 0 in Equation (4) and b3= 0 in Equation (5) respectively with and
without deterministic linear trend (Katircioglu, 2009).

43
 Quantitative Methods Inquires

Using equations (4)-(5)-each variable is considered as dependent variable in the

calculation of the and t-ratios.
When official economy is the dependent variable, the values of t-ratios for each lag
lies below the lower bound for all lags, revealing that there is not a level official economy
equation, irrespective of trend restrictions. When shadow economy is the dependent
variable, for lag 1 irrespective of trend impositions, the values of t-ratios lies outside the
0.01 critical value bounds, and reject the null hypothesis that there is no level shadow
economy equation.
Overall, the bounds test results support the existence of a mutual long-run
relationship between SE and official economy.
Having cointegrated relationships in bounds tests, the ARDL approach can be now
adopted to estimate the level relationship. On the Akaike Selection Criterion, the selected
ARDL order is 6 for the official economy and 0 for SE without deterministic trend.
The empirical estimates of level relationship for the ARDL error corection
model(lags: 5, 0) revealed that the estimated parameters are statistically significant and the
model shows that official economy have inelastic but negative coefficients. In the long run
period, the long run elasticity (coefficient of offical economy) is statistically significant. (Prob.
=0.00). All five lagged changes in shadow economy are statistically significant, further
justifying the choice of p=5.
The equilibrium correction coefficient is estimated as -0.90 (0.173) which is
reasonably large and highly significant at 1% level. This shows that Romanian shadow
economy coverge to its long run level by 90% by the contribution of official economy. The
intercept is not statistically significant and the lagged coefficients in the short term are
inelastic, but not totally statistically significant.
2
The adjusted R is 0.60 suggesting that such error correction model fit the data
reasonably well. In addition, the computed F-statistics clearly reject the null hypothesis that
all regressors have zero coefficients for all cases. Importantly, the error correction coefficient
carries the expected negative sign and are highly significant in both cases. This helps
reinforce the finding of cointegration.
Finally, we tested the direction of causality within the conditional Granger causality
tests using the ARDL mechanism as a long-run context. The F-statistics for the short-run
causations and the t statistics of ECTs for the long-run causations must be statistically
significant to achieve Granger causality between the shadow economy and official economy.

Table 4. Results of Granger Causality

F-statistics [probability values]
Dependent Variable Official SEt t-stat (prob)
economyt for ECTt-1
Official economyt - () -1.40
[0.18]

SEt () - -1.80*
[0.09]

* denote the rejection of null hypothesis respectively at 0.10 levels.

The empirical results reveal the existence of a long-run unidirectional causality that
runs from official economy to shadow economy but in the short run, the lack of F-statistics

44
 Quantitative Methods Inquires

results does not support short-run causations. We have a Granger causality for long-run
period, because the t-statistics for ECT(error correction term) is statistically significant at 10%
levels.
Next, we examine the stability of short-run and long-run coefficients, performing
the CUSUM and CUSUMQ stability tests for the AIC-based error correction models. The tests
applied to the residuals indicate the absence of any instability of the coefficients because the
plots of the CUSUMQ and CUSUM statistic are confirmed within the 5% critical bounds of
parameter stability.

Figure 2. Plots of CUSUM and CUSUMSQ Statistics for Coefficient Stability for the
relationship between shadow economy and official economy

Conclusions

In this paper, we investigated the relationship between official economy and the
size of the Romanian shadow economy using bounds test approach and ARDL causality
analysis for quarterly time series data from 2000-2010. The size of Romanian shadow
economy is estimated using a revised version of the currency demand approach based on
autoregressive distributed lag (ARDL) approach to cointegration analysis. A detailed
description of the estimation process is described in Davidescu and Dobre (2013). The size of
the shadow economy as % of official GDP measures approximately 45% at the end of 2000
and achieving the value of 37.4% in the last quarters of the period.
Cointegration test results does not support any short-run relationship between
official economy and shadow economy but in the long-run official economy have a negative
effect on shadow economy, when it is taken into account a significance level of 10%.
The ARDL causality results revealed the existence of a uni-directional causality that
runs official economy to the shadow economy, but only on long-run. The empirical results
are in line with the studies of Eilat and Zinnes (2000) for 24 transition countries and
Kaufmann and Kaliberda (1996) who estimate a negative impact of official GDP on the size
of the shadow economy, mentioning that a decline in official GDP, will lead to an increase in
the size of the shadow economy.

45
 Quantitative Methods Inquires

References

1. Albu, L.L., Ghizdeanu, I. and Stanica, C. Spatial Distribution of the Informal Economy. A
Theoretical and Empirical Investigation, SCIENZE REGIONALI, FrancoAngeli
Editore, vol. 0(1), 2011, pp. 63-80
2. Davidescu A.A. and Dobre, I. Revisiting the Relationship between U.S. Shadow Economy
and the Level of Unemployment Rate using Bounds Test Approach for
Cointegration and Causality; Economic Computation and Economic Cybernetics
Studies and Research, Vol. 46, No. 2, 2012, pp.91-104
3. Dobre, I. and Davidescu, A. Long-run demand for money and the size of shadow economy
in Romania:An application of ARDL model, Economic Computation and Economic
Cybernetics Studies and Research, Vol. 47, No. 3, 2013, pp. 91-110
4. Eilat, Y. and Zinnes, C. The Evolution of the Shadow Economy in Transition Countries:
Consequences for Economic Growth and Donor Assistance. CAER II Discussion
Paper No. 83, Harvard Institute for International Development. Cambridge, MA, 2000
5. Giles, D.E.A. and Tedds, L.M. Taxes and the Canadian Underground Economy, Canadian Tax
paper 106, Toronto, Canadian Tax Foundation, 2000
6. Katircioglu, S.T. Tourism, Trade and Growth: The Case of Cyprus; Applied Economics, Vol. 41,
No. 21, 2009, pp. 2741-2750
7. Maer-Matei, M.M. Measures Of Occupational Mismatch. SEA-Practical Application of Science,
Vol.5, 2014, pp. 425-430
8. Narayan, P. K. and Smyth, R. The Relationship between the Real Exchange Rate and
Balance of Payments: Empirical Evidence for China from Co-integration and
Causality Testing. Applied Economic Letters, Vol. 11, 2004, pp. 287–291
9. Pesaran, M. H., Shin, Y. and Smith, R. J. Bounds Testing Approaches to the Analysis of Level
Relationships; Journal of Applied Econometrics, Vol. 16, 2001, pp. 289–326
10. Schneider, F. and Buehn, A. Shadow economies and corruption all over the world: revised
estimates for 120 countries, Economics - The Open-Access, Open-Assessment E-
Journal, Kiel Institute for the World Economy, vol. 1, No. 9, 2007, pp. 1-53
11. Schneider, F., Buehn, A. and Montenegro, C. Shadow Economies All over the World: New
Estimates For 162 Countries From 1999 To 2007, Working Papers wp322,
University of Chile, Department of Economics; 2010
12. Strat, V.A. What Happened with the Attractiveness of the Romanian Counties for FDI
during the Period 2001 – 2012?, Journal of Applied Quantitative Methods, Vol. 9,
No. 4, 2014, pp.22-39
13. Vo, D. and Ly, T. Measuring the Shadow Economy in the ASEAN Nations: The MIMIC
Approach, International Journal of Economics and Finance, Vol. 6, No. 10, 2014,
pp.139 – 149
14. Zaman, G., Goschin, Z., Partachi, I. and Herteliu, C. The contribution of labour and capital to
Romania's and Moldova's economic growth, Journal of applied quantitative
methods, 2(1), 2007, pp. 179-185
15. Zaman, G., Goschin, Z., and Herteliu, C. Analysis Of The Correlation Between The Gdp
Evolutions And The Capital And Labor Factors In Romaniam, Romanian Journal
for Economic Forecasting, 2(3), 2005, pp. 5-21
16. Zamfir, A.M., Mocanu, C., Maer-Matei, M.M. and Lungu, E.O. Immigration and Integration
Regimes in EU Countries, Journal of Community Positive Practices, Vol. 14, No. 1,
2014, pp. 104-115
17. *** Quarterly National Accounts database, Eurostat.
18. *** Quarterly Government Finance Statistics database, Eurostat.
19. *** Quarterly Interest Rates database, Eurostat.
20. *** Quarterly Monetary and Financial Statistics database, Eurostat.

46
 Quantitative Methods Inquires

21. *** Tempo database, National Institute of Statistics, www.insse.ro

22. *** Monthly Bulletins of National Bank of Romania, 2000-2010, www.bnr.ro

1
Acknowledgements
This work was supported from the European Social Fund through Sectorial Operational Programme Human
Resources Development 2007–2013, project number POSDRU/ 159/1.5/S/142115, project title “Performance and
Excellence in Postdoctoral Research in Romanian Economics Science Domain”.

2
Adriana AnaMaria DAVIDESCU (ALEXANDRU) has graduated the Faculty of Cybernetics, Statistics and
Economic Informatics in 2006. She holds a PhD diploma in Economics from 2011 and currently she is lecturer within
the Department of Statistics and Econometrics from the Faculty of Cybernetics, Statistics and Economic Informatics
and scientific researcher III within National Scientific Research Institute for Labour and Social Protection. Her main
topics are analysis of informal economy, economic growth, and unemployment and also labour market studies. She
is the author of more than 24 articles in international journals of which 6 in ISI journals, 15 articles published in
volumes of international scientific conferences recognized in the country and abroad. She has also participated to
more than 26 scientific national and international conferences and 4 summer schools. She has a reach experience
in the field of applied statistics and econometrics working in different research projects of which 2 projects in quality
of project manager and various as member of research team.

3
this paper addresses the concept of shadow economy as defined by Schneider (2006) and Schneider et al. (2010)
and does not trait the informal sector.

4
For Schneider et al. (2010), the size of shadow economy in % of official GDP, using the DYMIMIC model is 34.4%
in 2000, 35.4% in 2002, 35.9% in 2004, 36.2% in 2005 and 36.7% in 2006.

5
For Schneider et al. (2010), the size of shadow economy in % of official GDP, using the DYMIMIC model is 34.4%
in 2000, 35.4% in 2002, 35.9% in 2004, 36.2% in 2005 and 36.7% in 2006.

6
Pesaran et al. (2001) have generated critical values using samples of 500 and 1000 observations.

7
Narayan (2005) argued that these critical values are inappropriate in small samples which are the usual case with
annual macroeconomic variables. For this reason, Narayan (2005) provides a set of critical values for samples
ranging from 30 to 80 observations for the usual levels of significance.

8
The Optimal ARDL models are specified on a basis of a set of criteria (Schwarz, Akaike).

9
The maximum duration of lags for both models has been taken as 7.

10
k is the number of repressors for the dependent variable in the ARDL models.

47
 Quantitative Methods Inquires

THE APPLICATION OF GREY SYSTEM THEORY IN PREDICTING

THE NUMBER OF DEATHS OF WOMEN BY COMMITTING
SUICIDE- A CASE STUDY

Kalyan MONDAL1
MSc, Birnagar High School (HS), Birnagar, Ranaghat, , West Bengal, India

E-mail: [email protected]

Surapati PRAMANIK2
PhD, Department Of Mathematics,
Nandalal Ghosh B.T. College, Panpur, West Bengal, India
Corresponding author

E-mail: [email protected]
Abstract:
Sexual harassment, dowry problem, torture, importation of girls, kidnapping, rape and other
social problems are forced a woman to commit suicide. These risk factors include man
dominated social structure, insecurity of woman, unequal priority level of man and woman,
family problems, un-employment etc. Indian constitution offers equal rights for male and
female. So the problem of woman suicide becomes a complicated one that restricts the
development of country and threatens for the parallelism of male-female ratio. Considering the
complexity and uncertainty of the influencing factors on woman suicides, suicide forecasting can
be regarded as a grey system with unknown and known information, so it can be analyzed by
grey system theory. Grey models require only a limited amount of data to estimate the behavior
of unknown systems. In this paper, the original predicted values of woman suicides are
separately obtained by the GM (1, 1) model, the Verhulst model and the GM (2, 1) model. The
results obtained from these models on predicting woman suicide show that the forecasting
accuracy of the GM (1, 1) is better than the Verhulst model and the GM (2, 1) model. Then, the
GM (1, 1) model is proposed to predict woman suicide in Indian context.

Key words: Woman suicide, Grey system theory, GM (1, 1) model, Verhulst model, GM
(2, 1) model, Forecasting

1. Introduction

Indian civilization is one of the greatest civilizations in the world history. Women
suicide is found in the great epic like the Mahabharata and the Ramayana. Committing
suicide is a multidimensional, multifaceted malaise. At present India is a developing nation.
Indian constitution offers the same rights of man and woman. With the development of

48
 Quantitative Methods Inquires

economy overall demands regarding all sphere of life of a woman are increasing day by day.
Now, in urban life, women have to lead first life in order to meet the demand of her family
and other reasons. As a result, they are affected both mentally and physically such as high
blood pressure, high blood sugar, stress, hypertension, mental depression etc. In rural and
urban sections, there have been an increasing number of cases such as sexual harassment,
dowry problem, mental and physical torture, importation of girls, kidnapping, rape, divorce,
love affairs, cancellation or the inability to get married (in accordance with the system of
arranged marriages in India), illegitimate pregnancy, extra-marital affairs, family conflicts,
family problems, illness high expectation, and other unknown problems. These factors are
the main causes behind committing suicide. Many young girls lose their deep love affairs
and take maximum decision of committing suicide [1]3. Eight suicides per day are occurred
due to poverty and dowry dispute [1]. One suicide out of every five suicides was committed
by a housewife [1]. However, the occurrence of woman suicidal cases [1, 2] reflects a rising
tendency as a result of the quick growing of alertness. Though the occurrence of woman
suicidal case is occasional, it can be predicted scientifically based on the related statistical
indexes. Accurate prediction of the woman suicide is important not only for government’s
policy, but also for social organizations that are devoted to deal with woman’s problems.
Grey system theory proposed by Deng [3] in 1982 is a powerful theory for dealing
with partially known and partially unknown information. The concept of the grey system
theory is used in several fields such as rainfall prediction [4], industry [5], business [6] and
geological systems studies [7], environmental studies [8], decision making [9], etc. As an
essential part of grey system theory, grey forecasting models [10] are popularly used in time
-series forecasting because of its simplicity and ability and high precision to characterize an
unknown system by using a few data points [11, 12].
In recent years, the grey system theory has been widely used to forecast in various
fields such as grey prediction model for traffic demand [13], electricity demand [14], and
internet access population [15].
In review of literature, no prediction model for women suicide is still found. In this
paper, the original predicted values of woman suicides are separately obtained by using the
GM (1, 1) model [16], the Verhulst model [17] and the GM (2, 1) model [18]. The results of
these models on predicting woman suicide are compared. Then, the GM (1, 1) model is
proposed to predict woman suicide accidents in Indian context.
Rest of the paper is organized as follows: Section 2 presents mathematical
presentation of three grey prediction models. Section 3 is devoted to present case study in
Indian context. Section 4 presents concluding remarks.

2. Mathematical Presentation Of Prediction Models

2.1. The GM (1, 1) Model [16]

The most commonly used grey forecasting model is GM (1, 1), which indicates that
one variable is employed in the model. The first order differential equation is adopted to
match the data generated by the accumulation generating operation (AGO).
For the algorithm of GM (1, 1), the raw data sequences is presented as follows:
X  x x 
0   x 0  1, 0  2 ,  , 0  n  (1)

49
 Quantitative Methods Inquires

Here n is the total number of modeling data. The AGO formation of X0 1 is
defined as follows:
1 
X 1  x 1, x 1 2 ,  , x 1 n   (2)
Here,
x1 k    j1 x0   j , k = 1,2,. . ., n
k
(3)
The GM (1, 1) model can be formed by establishing a first order differential equation for
X1 k  as follows:
dX1
 a X1  u (4)
dt
Here, the parameters a, u are called the developing coefficient and grey input
respectively.
In practical, the parameters a, u are not calculated directly from the equation (4).
Therefore, the solution of the equation (4) can be obtained by using the least square method
as follows:

x̂
1
k  1   x 0  1  u  eak  u (5)
 â  â
Here â  a , u T  BT B  1 BT Y and
 1
 2 x (1)  x   2
1 1
1

 1 
B   2

x   2  x   3
1 1
1

(6)
  
 1
 2
1   1  
x 3 x 4  1

0  
0  0 
Y N  X 2 , X 3,  , X n   T
(7)
Applying the inverse accumulated generation operation (IAGO), the obtained solution is
presented by:

x̂
0 
k    x 0  1  u  1 e a e a k 1 (8)
 a
1
Here x̂ 1  x 0  1 and k = 2, 3,..., n.

2.2 The Grey Verhulst Model [17]

The Verhulst model [17] was first introduced by a German biologist Pierre Franois
Verhulst. The main purpose of Velhulst model is to restrict the whole development for a real
system. For an initial time sequence,
0   0  
X 0   x 1, x 0  2 ,  , x 0  n  , the initial sequence X is used to construct the
Verhulst model directly as follows:
0 
dX
dt
 u X 0 
a X
0 
  (9)

Here a presents the development coefficient and u denotes the grey action
quantity. The solution of the parameter vector â  a , u T can be obtained by using the least
square method.

50
 Quantitative Methods Inquires

 
Here â  ( A  B) T ( A  B) 1 ( A  B) T Y
and,

    1 (0)  

1 (0)
( x (1)  x (0)
( 2 ))   ( x (1)  x ( 0) (2)) 2 
 2   2  
    
( x (2)  x (3)) 2 
1 (0) 1
A=   2 ( x ( 2)  x (3))  ; B=   2
(0) ( 0 ) ( 0 )

    (10)
     
  ( x ( 0 ) ( n  1)  x ( 0 ) ( n ))  
1
  1 (0) 2 
 2   ( x ( n  1)  x (0)
( n ))  
  2  
 
Y= x ( 0 ) ( 2)  x ( 0 ) (1), x ( 0 ) (3)  x ( 0 ) ( 2), , x ( 0 ) ( n )  x ( 0 ) ( n  1) T (11)
The re-solution of (9) can be presented as follows:
0  ax ( 0 ) (1)
x̂ k  1  k = 0, 1, 2,…, n (12)
ux ( 0) (1)  (a  ux ( 0) (1)) e ak

2.3 The GM (2, 1) Model [18]

The GM (2, 1) model is a single sequence second-order linear dynamic model and
is fitted by differential equations.
Let us assume that an original sequence X(0) be

X(0) = x ( 0 ) (1), x ( 0 ) ( 2),, x ( 0 ) ( n ) . 
A new sequence X(1) is generated by the AGO as follows:

X(1) = x (1) (1), x (1) ( 2),, x (1) ( n ) , here 
X(1)(k) =  kj1 x ( 0 ) ( j) , k = 1, 2,…, n (13)
Now the differential equation of GM (2, 1) model can be presented as follows:
d 2 X (1) dX (1)
 a u (14)
dt 2 dt
aˆ  a, u T  B B  B Y
T 1 T

  x 0  2 1  ( x ( 0 ) ( 2)  x ( 0 ) (1)) 
 0    
 x 3 ( x ( 0 ) (3)  x ( 0 ) ( 2)) 
B
1 
; Y (15)
     
 0    (0) 
 x n  1 ( x ( n )  x ( n  1)) 
(0)

From the equation (14), we have

 u x 0 1  ak u
x̂
1
k  1    e  k  1   x 0 1  u  1  a  (16)
2
a  a  a  a 
 a 
The prediction values of original sequence can be obtained by applying inverse AGO to x̂ 1
as follows:
0 
x̂ k  1  x̂ 1 k  1  x̂ 1 k , here k  1, 2, ..., n (17)

51
 Quantitative Methods Inquires

3. Case Study

In this section, the GM (1, 1) [16], the Verhulst model [17] and the GM (2, 1) [18]
are used for comparison. The woman suicide data [1, 2] in India from 2008 to 2013 is used
to demonstrate the effectiveness and practicability of the models. The data of women suicide
in 2006-2010 is presented to form the three grey prediction models and the data of women
suicide from 2011 to 2013 is used as data set to compare the accuracy of the three
prediction models.
The evaluation criterion is the mean relative percentage error (MRPE), which
measures the percentage of prediction errors. MRPE can be presented as follows:

MRPE 
1 n
n

k 1 x (k )  x̂ (k ) / x (k )
(0) ( 0) (0)
 (18)

In the GM (1, 1) model values of the essential terms are presented as follows:
 44488 1 44750
 44788 1 44825
  
  45193 1 45560
   
B   45780 1 ; Y  46000 ; â  a , u    0.02,  20T (19)
  47601 1  49201
   
  49801 1 50400
  51048 1 51695
  
In the Verhulst model values of the essential terms are presented as follows:
 44488 1979182144   525 
 44788 2005964944  75 
  
  45193 2042407249  735 
   
(A  B)   45780 209580840  Y   440  â  a , u    0.01, 1 / 4428000 T (20)
  47601 2265855201 3201
   
  49801 2480139601 1199
  51048 2605898304 1295
  
In the GM (2, 1) model values of the essential terms are presented as follows:
 44750 1  525 
 44825 1  75 
  
 45560 1  735 
   
B   46000 1 , Y   440  â  a , u    0.21,  9000T (21)
  49201 1 3201
   
  50400 1 1199
  51695 1 1295
  
The real and forecasted values are shown in ‘Table1’ to compare the three model
accuracy and relative error. The corresponding calculated results (the mean error in the
different stage) are shown in Table2.
Table1 indicates that the GM (1, 1) prediction model is smaller than the others by
comparing the relative error. From Table2, it is seen that the MRPE of the GM (1, 1) model,
the Verhulst model and the GM (2, 1) from 2011 to 2013 are 1.360%, 2.007% and 1.503%,

52
 Quantitative Methods Inquires

respectively. The effectiveness and accuracy of GM (1, 1) model is higher than the Verhulst
model and the GM (2, 1) model.

Table1. Model values and prediction error of the woman suicidal case in India
Year Real GM(1,1) Verhulst GM(2,1)
Value Model Error R Model Error R Model Error R
value (%) value (%) value (%)
Model 2006 44225 44225 0 45128 -2.04 44225 0
set up 2007 44750 45098 -0.78 46067 -2.94 46103 -3.02
stage 2008 44825 45989 -2.6 47048 -4.96 46542 -3.83
2009 45560 46898 -2.94 48071 -5.51 47084 -3.35
2010 46000 47825 -3.97 49139 -6.82 47751 -3.81
Post set 2011 49201 48771 0.87 50256 -2.14 48576 1.27
up 2012 50400 49736 1.32 51424 -2.03 49592 1.60
stage 2013 51695 50721 1.88 52649 -1.85 50846 1.64

Table2. Error results for the different prediction models

Stage GM(1,1) Verhulst GM(2,1)
MRPE (%) MRPE (%) MRPE (%)
2008-2010 2.058 4.454 2.802
2011-2013 1.36 2.007 1.503

The comparison of Table1 and Table2 show that the GM (1, 1) model and the GM
(2, 1) model have the better forecasting precision in 2006-2010, but the GM (1, 1)
prediction model offers the lowest post-forecasting errors and it is more suitable to make a
short-term prediction, so the GM (1, 1) model is used to predict women suicide for 2014 and
2015 in India. In Table 3, the comparison between the real values and predicted values
obtained from GM (1, 1) model (see Figure1) for women suicides in India.

Table3. The result of forecasting

Year 2010 2011 2012 2013 2014 2015
Real values 46000 49201 50400 51695 --- ---
GM (1, 1) 47825 48771 49736 50721 51552 52253
Model values

54000 52253
51695 51552
52000 50400 50721
49736
49201
48771
50000
47825
48000
46000
46000
44000
42000
2010 2011 2012 2013 2014 2015

Real values GM (1, 1) Model values

Figure 1. Comparison of real values and predicted values obtained from GM (1, 1) model

53
 Quantitative Methods Inquires

4. Conclusion

Committing suicide is an important issue in social context. This paper demonstrates

how the grey system theory deals with prediction problem with incomplete or unknown
information with large sample. In this paper, we compare the performance of the accuracy
of the three grey forecasting models to predict women suicide in India. This paper
demonstrates that performance of the GM (1, 1) model in prediction is better than the other
two prediction models because it has the merits of both simplicity of application and high
forecasting precision. Therefore, we suggest to using the GM (1, 1) model to predict the
number of suicides in India and other countries for planning and other issues.

References
1. Guoa, Z., Song, X. and Ye, J. Verhulst model on time series error corrected for Port
through put forecasting, Journal of East Asia Society of Transport Study, Vol. 6,
2005, pp. 881-891
2. Julong, D. Control problems of grey system, System Control Letters, Vol. 1, No. 5, 1982, pp.
288-294
3. Julong, D., Introduction to grey system theory, Journal of Grey System, Vol. 1, 1989, pp. 1-24
4. Kayacan, E., Kaynak, O. and Ulutas, B. Grey system theory- based models in time series
prediction, Expert Systems with Application, Vol. 37, 2010, pp. 1784-1789
5. Lan, J. and Cheng, H. The grey system and prediction of geological and mineral resources,
Mathematical Geology, Vol. 24, No. 6, 1999, pp. 653-662
6. Lin, Y.H., Chiu, C.C., Lin, Y.J. and Lee, P.C., Rainfall prediction using innovative grey model
with the dynamic index, Journal of Marine Science Technology, Vol. 21, No. 1, 2013,
pp. 63-75
7. Luo, Y. and Che, X., Improvement and application of initial value of non-equidistant GM
(1,1) Model, International Journal of Computer Science Issues, Vol. 10, No. 2, 2013,
pp. 113-118
8. Mohammadi, A., Moradi, L., Talebnejad, A. and Nadaf, A. The use of grey system theory in
predicting the road traffic accident in Fars province in Iran, Australian Journal of
Business Management Research, Vol. 9, 2011, pp. 18-23
9. Niu, W., Zhai, Z., Wang, G., Cheng, J. and Guo, Y., Adaptive multivariable grey prediction
model, Journal of Information Computer Science, Vol. 8, No. 10, 2011, pp. 1801-
1808
10. Pramanik, S. and Mukhopadhyaya, D., Grey relational analysis based intuitionistic fuzzy
multi-criteria group decision-making approach for teacher selection in higher
education, International Jourrnal of Computer Application, Vol. 34, No. 10, 2011, pp.
21-29
11. Quanping, H. and Xiaoyi, Y. Base a EMD-grey model for textile export time series
prediction, International Journal of Data Theory Application, Vol. 6, No. 6, 2013, pp.
29-38
12. Tseng, F.M. and Tzeng, G.H. The Comparison of four kinds of prediction methods: ARIMA,
fuzzy time series, fuzzy regression time series and grey forecasting: an
example of production value forecasting of the mechanical industry in Taiwan,
Journal of Chinese Grey System Association, Vol. 2, No. 2, 1999, pp. 83-98
13. Wu, W.Y. and Chen, S.P., A prediction method using the grey model GM (1, n) combined
with the grey relational analysis: a case study on internet access population
forecast, Applied Mathematics and Computation, Vol. 169, No. 1, 2005, pp. 198-217
14. Yang, C.M., Chen, J.C., Peng, L.P., Yang, J.S. and Chou, C.H., Earthquake-caused landslide
and grey prediction for vegetation recovery, Botanical Bulletin of Academia Sinica,
Vol. 43, 2002, pp. 69-75
15. Yang, Z. and Zhang, Y. Traffic demand forecasting for EGCS with grey theory based multi-
model method, International Journal of Computer Science Issues, Vol. 1, No. 1, 2013,
pp. 61-67

54
 Quantitative Methods Inquires

16. Zhou, P., Ang, B.W. and Poh, K.L. A trigonometric grey prediction approach to forecasting
electricity demand, Energy Conversion Management, Vol. 31, 2006, pp. 2839-2847
17. * * * Accidental deaths and suicides in India, National crime records bureau ministry of home
affairs government of India. R. K. Puram, New Delhi, 2010, http://ncrb.gov.in ,
accessed 11 July 2014
18. * * * Suicides in India, 2012, http:// ncrb.nic.in/CD-ADSI-2012/suicides-11.pdf , accessed 12
July 2014

1
Kalyan Mondal (M.Sc., B. Ed.) passed B. Sc. Honors and M. Sc in Mathematics in 2003 and 2005 respectively
from the University of Calcutta and University of Kalyani. Currently, he is an assistant Teacher of mathematics at
Birnagar High School (HS), Birnagar, Ranaghat, Nadia, Pin Code: 741127, West Bengal, India. He has coauthored
more than six research papers. His field of research interests includes fuzzy goal programming, grey system theory,
intuitionistic Fuzzy decision making and neutrosophic decision making.

2
Dr. Surapati Pramanik (Ph. D., M.SC., M. Ed.) did his B. Sc. and M. Sc. in Mathematics from University of
Kalyani. He received Ph. D. in Mathematics in 2010 from Bengal Engineering and Science University (BESU)
Shibpur, India. He is currently an Assistant Professor of Mathematics at the Nandalal Ghosh B. T. College, Panpur,
P.O.-Narayanpur, West Bengal, India. He has authored/co-authored more than 50 research papers in international
journals; He has published one mathematics method book from Aheli publishers. He coauthored five books for B.
Ed. Courses from Aheli Publisher, Kolkata, India. His research interests include operations research and
optimization, soft computing, grey system theory, neutrosophic decision making, rough sets, mathematics
education, comparative education, international relation.

3
Codification of references within text:
* * * Accidental deaths and suicides in India, National crime records bureau ministry of home affairs
[1] government of India. R. K. Puram, New Delhi, 2010, http://ncrb.gov.in , accessed 11 July 2014
[2] * * * Suicides in India, 2012, http:// ncrb.nic.in/CD-ADSI-2012/suicides-11.pdf , accessed 12 July 2014
[3] Julong, D. Control problems of grey system, System Control Letters, Vol. 1, No. 5, 1982, pp. 288-294
Lin, Y.H., Chiu, C.C., Lin, Y.J. and Lee, P.C., Rainfall prediction using innovative grey model with
[4] the dynamic index, Journal of Marine Science Technology, Vol. 21, No. 1, 2013, pp. 63-75
Tseng, F.M. and Tzeng, G.H. The Comparison of four kinds of prediction methods: ARIMA, fuzzy
time series, fuzzy regression time series and grey forecasting: an example of production value
forecasting of the mechanical industry in Taiwan, Journal of Chinese Grey System Association, Vol. 2,
[5] No. 2, 1999, pp. 83-98
Quanping, H. and Xiaoyi, Y. Base a EMD-grey model for textile export time series prediction,
[6] International Journal of Data Theory Application, Vol. 6, No. 6, 2013, pp. 29-38
Lan, J. and Cheng, H. The grey system and prediction of geological and mineral resources,
[7] Mathematical Geology, Vol. 24, No. 6, 1999, pp. 653-662
Yang, C.M., Chen, J.C., Peng, L.P., Yang, J.S. and Chou, C.H., Earthquake-caused landslide and grey
[8] prediction for vegetation recovery, Botanical Bulletin of Academia Sinica, Vol. 43, 2002, pp. 69-75
Pramanik, S. and Mukhopadhyaya, D., Grey relational analysis based intuitionistic fuzzy multi-
criteria group decision-making approach for teacher selection in higher education, International
[9] Jourrnal of Computer Application, Vol. 34, No. 10, 2011, pp. 21-29
Niu, W., Zhai, Z., Wang, G., Cheng, J. and Guo, Y., Adaptive multivariable grey prediction model,
[10] Journal of Information Computer Science, Vol. 8, No. 10, 2011, pp. 1801-1808
[11] Julong, D., Introduction to grey system theory, Journal of Grey System, Vol. 1, 1989, pp. 1-24
Kayacan, E., Kaynak, O. and Ulutas, B. Grey system theory- based models in time series prediction,
[12] Expert Systems with Application, Vol. 37, 2010, pp. 1784-1789
Yang, Z. and Zhang, Y. Traffic demand forecasting for EGCS with grey theory based multi-model
[13] method, International Journal of Computer Science Issues, Vol. 1, No. 1, 2013, pp. 61-67
Zhou, P., Ang, B.W. and Poh, K.L. A trigonometric grey prediction approach to forecasting
[14] electricity demand, Energy Conversion Management, Vol. 31, 2006, pp. 2839-2847
Wu, W.Y. and Chen, S.P., A prediction method using the grey model GM (1, n) combined with the
grey relational analysis: a case study on internet access population forecast, Applied Mathematics
[15] and Computation, Vol. 169, No. 1, 2005, pp. 198-217
Luo, Y. and Che, X., Improvement and application of initial value of non-equidistant GM (1,1)
[16] Model, International Journal of Computer Science Issues, Vol. 10, No. 2, 2013, pp. 113-118
Guoa, Z., Song, X. and Ye, J. Verhulst model on time series error corrected for Port through put
[17] forecasting, Journal of East Asia Society of Transport Study, Vol. 6, 2005, pp. 881-891
Mohammadi, A., Moradi, L., Talebnejad, A. and Nadaf, A. The use of grey system theory in predicting
the road traffic accident in Fars province in Iran, Australian Journal of Business Management
[18] Research, Vol. 9, 2011, pp. 18-23

55
 Quantitative Methods Inquires

METHODS OF MEASURING CORE INFLATION IN INFLATION

TARGETING COUNTRIES

Ion PARTACHI
PhD, University Professor,
Academy of Economic Studies, Kishinev, Moldova

E-mail: [email protected]

Vitalie MOTELICA
Academy of Economic Studies, Kishinev, Moldova

Abstract:
This article tackles the issue of the transitory effects on growth of prices and the main methods
for the calculation of monetary policy relevant inflation in inflation targeting countries. In order
to have an indicator which would capture the medium term inflation pressures, several
measures of core inflation are considered. Besides that this article mentions the main
advantages and disadvantages of the different measures of medium term inflation pressures.
Most of the central banks which are implementing the inflation targeting regime are using core
inflation indicators based on the exclusion of certain components, however there are several
central banks which are using statistical measures of core inflation such as the trimmed mean
and weighted median. This article also describes the existing core inflation indicators as well as
the main features of the trimmed mean and weighted median for the Republic of Moldova. Even
though the obtained indicators of core inflation have different values, they have similar
trajectories.

Key words: core inflation, statistical measures, trimmed mean, weighted median,
monetary factors

1. Introduction

Within the process of achieving their main objective of price stability, central banks
are usually monitoring, analyzing and forecasting the dynamics of the Consumer Price Index,
even though there are many other indicators which can provide information about the
change of prices such as Producer Price Indices, GDP and consumption deflators etc. The
reason behind this is the fact that the CPI has several advantages such as the fact that it is
known to the public, it is based on the expenditures made by the households and it is
disseminated on a monthly basis just several days after the end of the reference period.
Besides its advantages, CPI has several shortcomings. It doesn’t reveal the growth of prices
which is caused by monetary factors. In other words, some price changes might be

56
 Quantitative Methods Inquires

determined by some sectorial shocks. These have transitory effects on the general price level
and are not considered a part of the inflationary process. In this way, food prices might grow
due to some bad weather conditions which determined a bad harvest or the fuel prices
might grow once the fiscal authority increases the excise taxes. The direct effect of these
changes is temporary and is not a part of the general inflation trend.
A solution for detecting the trend of the inflation in the economy which would
reveal the monetary inflation and which would provide value added to the decision makers
would be the elimination of the high frequency data and keeping just the low frequency
data, in other words calculation of trends. However this procedure would decrease the
timing and the relevancy of the recent information and it wouldn’t be of much use for the
decision makers[2].
Another way to tackle this information is by excluding some volatile components
from the overall index. For example, the food and energy prices are determined to a great
extent by supply shocks caused by weather conditions and the decrease of supply of oil when
some conflicts in Middle East arise. So, by excluding these components from the overall
index, it can be obtained an index which would reflect better the inflation trends in the
economy. However, this method involves some shortcomings as well. There is no certainty
that changes in food prices do not contain useful information on inflation trends in the
economy. The removal could result in the removal of valuable information for decision
makers. Furthermore, there are other factors other than food prices and fuel prices that
could compromise attempts to measure the increase in prices due to monetary factors. The
resulting index does not necessarily have a clear picture of inflationary trends.
In order to overcome these issues, there have been developed new approaches
that are based more on statistical procedures rather than on the characteristics of certain
components, such as the trimmed mean and weighted median. These methods have the
advantage of eliminating of different irregularities in data while still keeping the important
information which was eliminated by traditional approaches.
In this article we will present the main ways to handle the transitory effects on
inflation by investigating what types of core inflation are used by some other central banks in
the region as well as by calculating and analyzing the dynamics of the core inflation
excluding the food, fuel and regulated prices as well as the core inflation determined using
statistical methods such as the trimmed mean and weighted median on CPI data for Republic
of Moldova.
The paper is organized as follows. Section 2 provides a brief literature review
highlighting some important findings of other authors regarding core inflation measures. In
section 3 we present some insight on the data and the main methodology in addressing core
inflation measures in our study. Section 4 presents the main results of our research. The
paper ends with section 5 where conclusions are provided.

2. Literature review

The main shortcomings of the CPI in addressing the monetary policy relevant
inflation was tackled in M.F. Bryan and S. G. Cecchetti paper “Measuring core inflation”
(1993) [2]. They state that it is difficult to measure the monetary inflation as a monetary
phenomenon because of the non-monetary events that can temporarily produce noise in the
data. They list some alternative solutions to this issue such as low-frequency trends or
excluding certain components from the overall CPI index based on the assumptions that they
are most affected by the noise. They also come with some statistical measures of core

57
 Quantitative Methods Inquires

inflation such as median and the trimmed mean. They state that these measures are robust
to the presence of many types of noise. Besides that they tried to evaluate the usefulness of
their proposed measures of core inflation for monetary policy by assessing which of them is
mostly correlated with the money growth. They find that the statistical methods are superior
to CPI in several respects such as the fact that they have higher correlations with past money
growth and provide improved forecasts for future inflation.
M. Silver [9] outlined the many approaches and methods to the measurement of
core inflation and the many approaches to judging the preferred measure. His research
shows that different measures of core inflation yield different results, that is, that choice of
measure matters. Furthermore, he states that different approaches to the choice of measure
yield different results and, even for the same approach to choice, the preferred measure may
differ across countries, and even within a county for different time periods. Choice of
measure should, in principle, be data-driven for each country based on appropriate criteria.
According to his paper, exclusion-based methods are found to be not optimal according to
the criteria selected by the monetary authorities. The choice of the method should be data
driven, so that the methods adopted are tailored to the features of the evolution of that
country’s economy and so that the choice of measures can be justified on an objective,
transparent basis.
After evaluating several candidate series that have been proposed as core
measures of consumer price index (CPI) inflation and personal consumption expenditure
(PCE) inflation for the United States, R. Rich and C. Steindel [8] concluded that policy would
be best served by recognizing that core measures differ in the quality and nature of the
insights they can provide about the dynamics of inflation and to draw from this varied
information for guidance.
M. A. Wynne also reviews various approaches to the measurement of core inflation
that have been pro-posed over the years using the stochastic approach to index numbers as
a unifying framework [12]. According to his paper, there is no theoretical ideal for a
monetary measure of core inflation. He concludes that before choosing a measure of core
inflation one needs to specify what it is one wants the measure for. Depending on the
reason behind it, different methods would be appropriate.
According to B. Meyer and G. Venkatu [6] trimmed-mean inflation statistics
diagnose the most volatile monthly price changes as noise and “trim” them from the price-
change distribution, leaving a clearer inflation signal behind. These measures systematically
remove sources of noise on a monthly basis, rather than ad hoc exclusionary measures such
as the ex food and energy (“core”) CPI. They tried to find whether median CPI was the
appropriate measure of trimmed - mean inflation statistic to use as a measure of underlying
inflation. Besides the symmetric trims, they also tried to use the asymmetric ones. They
conclude that median CPI is generally a better forecaster of future inflation over policy-
relevant time horizons than the headline and core CPI.

3. Data and Methodology

3.1 Data
For calculating different types of core inflation we use the monthly data concerning
change of prices and CPI component weights starting from 2009 until September 2014
available from the statistical authority. The weights of the CPI components are determined by
the statistical authority on a yearly basis based on the Survey concerning income and
expenditures of households in the previous year.

58
 Quantitative Methods Inquires

3.2 Core inflation by exclusion of certain components

Core inflation can be calculated by excluding certain components of the CPI which
are considered to have volatile behavior, are determined by central or local governments or
are subject to frequent supply shocks. In other words, the excluded items are believed to be
beyond the control of the monetary policy. The resulting index can be obtained using the
formulae below:
n m

 wi * iwi   w exj * iw exj

i 1 j 1
CII  (1)
n m

w  w
i 1
i
j 1
ex
j

where:
CII – core inflation index;
wi – the weight of the item in the CPI basket;

iwi – price index of an item in the CPI basket;

w exj – the weight of the item excluded form the CPI basket;

iw exj – price index of an item excluded from the CPI basket;

i – goods and services included in the CPI index;
j – goods and services that are excluded from the CPI index for calculation the CII;
n – number of goods and services that are part of CPI basket;
m – number of goods and services that are excluded from the CPI during calculation of
the CII.

3.3. Statistical measures of core inflation –trimmed mean and weighted median

The trimmed mean of α - percent is calculated by the formula nr. 2:

1
x 
 w x i i , (2)
1 2 iI 
100
where

I  
is  Wi  (1  )
100 100
Wi   j 1 w j
i

w is the weight of the component, x is the monthly price change of the component
In general, the method involves:
 ranking ascending price increases for each period with corresponding weights.
 then, we add the previous weights for each price increase previously ordered.
 next, the monthly increases for which the sum of their weights is less than α will be
excluded,
 the same thing will happen with monthly increases for which the sum of the
corresponding weights is greater than 100-α.

59
 Quantitative Methods Inquires

 the trimmed mean will be calculated then as a weighted average of the remaining
components.
The weighted median is an extreme case of the trimmed mean, so it represents
the growth of the component which is situated in the middle of the increasingly ordered
distribution. Thus, half of the weighted monthly increases are above the weighted median
and half are below. Therefore, the median is calculated according to the previous procedure,
except that it is the first price change whose cumulative weight is greater or equal to 50 %.

3.4. Determining the optimal level of exclusion of information for the trimmed
mean approach
In order to determine the optimal level of exclusion of information from both ends
of the distribution for a given time, several indicators according to the formulae nr. 2 will be
calculated using various exclusion rates (0 percent which is actually CPI until the median -
which excludes all items except the observation in middle of the distribution).For each of
these indicators we will determine the root mean squared error against the trend of inflation
(RMSE , formulae nr. 3). The optimal index is considered the one for which the error is the
smallest. To determine the trend of CPI inflation we will apply the Hondrick-Prescot Filter on
CPI data (lambda =16600) (figure nr. 4).
k

 (T ( )   )
t
2
t

RMSE ( )  t 1 , (3)
k
where
Tt(α)- the index of the trimmed mean α % at time t
 - trend of CPI inflation determined by using the Hondrick Preskot filter on CPI
k- number of observations

4. Results

4.1. Measures of core inflation in other central banks that target inflation
In international practice, usually the objective of monetary policy is price stability
which implies a moderate amount of inflation as measured by CPI. Thus, central banks
assess and communicate the effectiveness of their actions depending on whether the overall
inflation is close to the proposed level in the medium term. However, given the
aforementioned problems on this index, the monetary policy authorities monitor various
measures of core inflation when taking decisions in a timely way and with the desired impact
in order to contain inflation. These are supposed to allow exclusion of transitory effects and
to identify the trends of inflation due to monetary factors. The diversity of the measures of
core inflation can be observed by studying inflation reports published by central banks in
other countries in the region.
Czech Republic switched to inflation targeting regime in December 1997. From
January 2010 the target is 2% ±1 percentage point [5]. The Inflation Report published by the
Czech National Bank denotes that CPI includes Net inflation and regulated prices. The net
inflation is decomposed in food prices, fuel prices and Adjusted net inflation [15].
Central Bank of England adopted the IT strategy in October 1992. The current
target is point target of 2% annual rate of inflation. Central Bank of England, similarly,
practice exclusion method to identify the transitory effects on inflation. Thus, in the inflation
reports published by the bank, CPI inflation is decomposed in food prices, fuel prices, the
prices of services, education, energy and gas prices and the component other. Previously, in

60
 Quantitative Methods Inquires

some older editions of the Bank of England inflation reports one can find the dynamics of
other measures of core inflation such as median and trimmed mean of 15% [13].
The National Bank of Poland adopted the IT regime in 1998. Since 2004 it has a
target of 2.5%±1 percentage point. NBP considers several measures of core inflation in its
reports. In addition to the traditional method of exclusion of some default volatile
components (inflation without the volatile prices, without food and energy prices, without
regulated prices) volatile, it also uses the trimmed mean of 15% [16].
The National Bank of Romania switched to inflation targeting in 2005. Starting
from 2013 it has a 2.5% ±1 percentage point. In the Inflation Reports published by the
National Bank of Romania, there are shown three measures of core inflation. The
component CORE 1 is the difference between total inflation and administered prices.
Component CORE 2 also excludes volatile prices and the component adjusted CORE2 results
from the exclusion of volatile prices, the regulated and the tobacco and alcoholic beverages
from total inflation. The volatile prices components include vegetables, fruits and eggs [14].

4.2. Core inflation measures in Republic of Moldova

In late 2009 the National Bureau of Statistics of the Republic of Moldova adopted
the methodology of the calculation of core inflation index [7]. According to it, core inflation
index is calculated using the method of exclusion. The NBS calculates four measures of core
inflation:
1. Total CPI excluding food and beverages
2. Total CPI excluding products and services with regulated prices
3. Total CPI excluding fuel prices
4. Total CPI excluding food and beverage products and services with regulated prices,
fuel prices.
10,0
9,0
8,0
7,0
6,0
5,0
4,0
3,0
2,0
1,0
0,0
1/10
3/10
5/10
7/10
9/10
11/10
1/11
3/11
5/11
7/11
9/11
11/11
1/12
3/12
5/12
7/12
9/12
11/12
1/13
3/13
5/13
7/13
9/13
11/13
1/14
3/14
5/14
7/14
9/14

CPI Core inflation Inflation target

Figure 1. CPI and core inflation, yoy, %

Although the National Bureau of Statistics publishes several measures of core

inflation, the Inflation Reports published by NBM reveal the dynamics of the core inflation
index that is excluding regulated prices, prices of food and fuel prices.
In 2010 the National Bank started to create the necessary pre-conditions for
implementing the inflation targeting regime. The inflation reports published by the bank
reveal the dynamics of core inflation index which is calculated by excluding from total
inflation prices of food and drinks, regulated prices and fuel prices.

61
 Quantitative Methods Inquires

Food prices Regulated prices

Fuel prices Core inflation

32,8%
37,2%

6,0%

24,1%

Figure 2. CPI structure in Moldova (2014)

Since 2010, the annual evolution of the above-mentioned core inflation indicator
was characterized by a significantly lower volatility than that of the overall inflation (figure
1). At the same time it oscillated closer to the medium-term inflation target of 5.0 percent ±
1.5 percentage points, its average being 4.7 percent in the period. Core inflation has seen
an upward trend with business recovery after the crisis of 2009. In the end of 2011, along
with the slowdown in economic activity, it reversed its previous trend decreasing from 6.8
percent in September to 3.6 percent in October 2012. Starting from the beginning of 2013,
core inflation recorded a slightly upward trend increasing up to a value of 5.8 percent in
September 2014, driven mostly by monetary policy measures undertaken to prevent annual
inflation from leaving the target range.
40,0

30,0

20,0

10,0

0,0

-10,0

-20,0
12/10

4/11

8/11

12/11

4/12

8/12

12/12

4/13

8/13

12/13

4/14

8/14

Volatile components Rest of food prices

Figure 3. Food prices, yoy, %

However, the core inflation measure presented above has several drawbacks. It
excludes, in addition to regulated prices and fuel prices, completely the component food and
drinks. As a result, the core inflation component has a weight of about 32.8 percent of the

62
 Quantitative Methods Inquires

total CPI (in 2014) (figure 2) which is quite low compared with core inflation indexes
monitored in other countries with similar regimes. So, according to the above mentioned
procedure more than two thirds of the information is excluded from core inflation measure
analyzed in the Inflation Report
The exclusion of the food prices from the core inflation is usually justified by the fact
that they have a high volatility driven largely by supply-side factors and not the medium-
term inflation trend. However, in the structure of the food price component can be identified
very volatile subcomponents such as prices of vegetables, fruits, eggs which dynamics is
indeed mostly caused by transitory effects. The structure of the food prices also contains
some less volatile components which includes most of the processed foods and are not as
sensitive to agro-meteorological conditions. These might be largely influenced by aggregate
demand (see figure 3). Thus, they might present useful information about medium-term
inflation trends in the economy and might not be excluded from the measure of core
inflation. Also, in this way the share core inflation in CPI structure would significantly
increase.

4.3. The trimmed mean and weighted median measure for CPI data from Moldova
Given the fact that most core inflation indicators published by the statistical
authority in Moldova are calculated by the method of exclusion of pre-determined
components and at the moment there is not an alternative core inflation index, next we will
provide the trimmed mean and weighted median for the consumer prices in Moldova based
on the formulae nr. 2.

20

15

10

0
2001M01

2002M01

2003M01

2004M01

2005M01

2006M01

2007M01

2008M01

2009M01

2010M01

2011M01

2012M01

2013M01

2014M1

-5

Trend of inflation using HP filter

Annual CPI inflation

Figure 4. CPI inflation and the trend of inflation

The information on the RMSE suggests that the optimal measure of the trimmed
mean, i.e. the closest to the trend of inflation is the one for which 10 % from each end is
truncated. This means that at the upper and at the lower end of the distribution we will
exclude 10 percent of observations on price changes (figure no. 5).
The annual dynamics of the trimmed mean (figure 6) is slightly different from the
annual growth rate of core inflation calculated by the method of exclusion. Thus, although in
early 2010 they had similar values, the trimmed mean recorded a faster increase in the first
quarter of 2010 which determined a higher trajectory than that of the traditional core

63
 Quantitative Methods Inquires

inflation index. suggesting higher inflationary pressures form the aggregate demand
compare to the second indicator.

3,5

3,0

2,5

2,0
RMSE

1,5

1,0
optimal level of trimmed mean (10 %)
0,5

0,0
1%

4%

7%

10%

13%

16%

19%

22%

25%

28%

31%

34%

37%

40%

43%

46%

49%
The level of the trimmed mean

Figure 5. RMSE

In the first quarter of 2011 the trimmed mean experienced a pronounced

downward evolution, while the traditional core inflation had a stable dynamics. After this
episode, by the end of 2011 both indicators had similar increasing dynamics signaling
pressures from increasing demand on prices. However, the overall path of the trimmed
mean was lower than that of the core inflation calculated by the exclusion method( aprox. 1
penrcentage point). In 2012, both indicators had a downward trajectory due to decrease in
the economic activity and the difference between core inflation calculated by the truncated
mean method and calculated by the method of exclusion had been maintained. In late
2012, both above mentioned indicators started a moderate increasing path which lasted till
the end of the sample (3rd quarter 2014). However the difference between the two recorded
a slight increase.

8,0

7,0

6,0

5,0

4,0

3,0

2,0

1,0

0,0
apr..10

apr..11

apr..12

apr..13

apr..14
ian..10

iul..10

oct..10

ian..11

iul..11

oct..11

ian..12

iul..12

oct..12

ian..13

iul..13

oct..13

ian..14

iul..14

weighted median trimmed mean (20 %) core inflation (exclusion method)

Figure 6. Core inflation measures using exclusion method and statistical methods

64
 Quantitative Methods Inquires

In case of the weighted median (figure 6) in 2010, it had a similar pattern to that of
core inflation calculated by the method of exclusion. After significant reduction in first
quarter of 2011, and a more modest increase to the end of the year, the weighted median
trajectory was significantly lower than that of the other indicator. Towards the end of 2011,
the difference between the two measures of core inflation was about. 4.0 percentage points.
This difference, however, decrease in 2012 and early 2013 to approx. 2.5 percentage points.
The weigted median started a moderate upward trend in early 2013 similar to the core
inflation calculated by the exclusion method and similar to the trimmed and by the end of
the 3rd quarter 2014 it reached 2 %. The basic idea of the 2 alternative core inflation
indicators is that they suggest lower pressures on inflation coming from the aggregate
demand compare to the traditional core inflation measure.

5. Conclusion

This article tackles some of the main issues the policymakers face when monitoring
the price dynamics within the inflation targeting regime. The Consumer Price index, besides
important information on medium term inflation trends, might still contain information
determined by transitory effects or some measurement errors. Therefore, it is of high interest
to have a so called core inflation index that would be useful for taking the right decisions to
contain inflation within the medium term target.
The measures of core inflation calculated by the method of exclusion of certain pre-
determined components whose dynamics is mostly driven by external factors, by the
decisions of authorities or which exhibited a very volatile behavior over the history are more
commonly used by central banks implementing inflation targeting regime. However, these
indicators have several drawbacks and the most important of them were mentioned within
the article such as the fact that these methods might exclude important information from the
CPI data. The index which is left after the exclusion can be in the end a small part of the
initial CPI index. Furthermore, this index can also still include some components which are
not a part of the medium term inflation process and are not relevant for policy makers. In
this way it can sometimes provide an inaccurate view on inflationary pressures caused by
monetary factors.
This article suggests that there are alternative methods to the traditional exclusion
procedures, such as statistical methods for determining core inflation, the trimmed mean
and weighted median. According to these measures, the excluded components differ in each
period, and their exclusion criterion is determined by certain statistical properties, in this case
how far the respective component is from the central tendency in a certain period, and does
not contain any economic reasoning.
Given the fact that the inflation reports published by the National Bank of Moldova
denote the dynamics of a core inflation index which is determined by the exclusion of the
food prices, fuel prices and regulated prices form the CPI index, the trimmed mean and the
weighted median of inflation might be an important additional source of information for
policymakers in Republic of Moldova. Even though these indicators of core inflation have
different values, they have similar trajectories over the sample analyzed in the article
compare to the traditional core inflation indexes calculated for Moldova .
As a conclusion, the trimmed mean and the weighted median for Moldova could
present useful information about inflationary trends that might be missed by traditional core

65
 Quantitative Methods Inquires

inflation measures and should be considered as an additional source of information to guide

decision making in the process of keeping the overall inflation in the inflation target band.

Bibliography

1. Bike, L. and Stracca, L. A persistence weighted measure of core inflation in the euro-area,
European Central Bank, 2007
2. Bryan, F. Stephen, G. and Cecchetti, M. Measuring Core Inflation, 1993
3. Cutler, J. Core inflation in UK, Bank of England, 2001
4. Griffiths, D. Core inflation measures produced in New Zealand, 2009
5. Hammond, G. State of the art of inflation targeting, Handbook no. 29, Centre for Central
Banking Studies, Bank of England, 2012
6. http://en.wikipedia.org/wiki/Truncated_mean
7. http://www.statistica.md/public/files/Metadate/alte/Metodologia_Inflatia_de_baza.pdf
8. Rich, R. and Steindel, C. A review of Core inflation and an evaluation of its measures,
Federal Reserve Bank of New York, 2005
9. Silver, M. Core inflation measures and statistical issues in choosing among them, 2006
10. Zaman, G., Goschin, Z., and Herteliu, C. Analysis Of The Correlation Between The Gdp
Evolutions And The Capital And Labor Factors In Romaniam, Romanian Journal
for Economic Forecasting, 2(3), 2005, pp. 5-21
11. Zaman, G., Goschin, Z., Partachi, I. and Herteliu, C. The contribution of labour and capital to
Romania's and Moldova's economic growth, Journal of applied quantitative
methods, 2(1), 2007, pp. 179-185
12. Wynne, M.A. Core Inflation: A review of Some Concepptual Issues, 1999
13. www.bankofengland.co.uk
14. www.bnr.ro
15. www.cnb.cz
16. www.nbp.pl

66
 Quantitative Methods Inquires

CONNECTIONS BETWEEN WILL TO EMIGRATE AND

ATTACHMENT THEORY – A DATA MINING APPROACH1

Angel-Alex HAISAN2
PhD, Faculty of Economics and Business Administration,
Babes-Bolyai University, Cluj-Napoca, Romania

E-mail: [email protected]

Vasile Paul BRESFELEAN3

PhD, Lecturer, Faculty of Economics and Business Administration,
Babes-Bolyai University, Cluj-Napoca, Romania

E-mail: [email protected]

Abstract:
Many studies have been carried out in the last years in Romania, due to the scale of the
phenomenon, on the factors that are responsible for the migration decision of people and the
consequences of this act.
The family is being generally accepted as the nucleus of a society and that’s why is very
important to create a medium where it can develop and grow in a healthy way. Family is one of
the 12 domains of our main study direction, quality of life and it still is the first options when it
comes to individual’s support.
A new aspect that should be brought into attention should be represented by the recent findings
in the emigration research domain, which show that there is a significant relationship between
the will to emigrate and an unresolved attachment status.
So, we would like to see if indeed the emigration decision is based on financial motivations, as
we’ve discovered so far, or it also has ties with other indicators.

Key words: family, attachment theory, emigration, data mining, teachers

1. Introduction

Many studies have been carried out in the last years in Romania, due to the scale of
the phenomenon, on the factors that are responsible for the migration decision of people
and the consequences of this act. We can encounter in the research literature, as major
domains of research, different approaches, such as economical [1,18], sociological [28, 40],
educational [41, 22] or psychological [46, 37] directions, each of them having their own sub
directions that most of the times intersect each other, thus needing an interdisciplinary
approach.
Although much has been written regarding the migration process of Romanians in
general, very few studies approach the problem on a smaller scale, such as socio-
professionals categories. We can talk about the brain-drain problem here, which like all

67
 Quantitative Methods Inquires

major concepts, has been discussed in the beginning without analyzing a certain category
and only in the last couple of years it began to know a more focused approach, the attention
of the research community being mainly drawn by the migration of specialists from health
care and IT. As a consequence to that, in one of our past studies we have tried to approach
the migration subject from a different category’s point of view, namely teachers [21]. We’ve
tried to discover, by employing data mining techniques, the motivations that reside behind
the emigration decision of this category based on their marital status and found out that the
economical factor is of great importance for the married without children and unmarried
ones while for the ones that are married with children the family comes first.
The family is being generally accepted as the nucleus of a society and that’s why is
very important to create a medium where it can develop and grow in a healthy way. Also,
family is one of the 12 domains of our main study direction, quality of life and it still is the
first options when it comes to individual’s support [15]. Regardless if we are talking about
close or enlarged family, its members offer comfort to the individual whether we are
considering its moral, financial or other needs. Another aspect that should be brought into
the equation should be that recent findings in the emigration research domain show that
there is a significant relationship between the will to emigrate and an unresolved attachment
status [47]. Attachment theory refers to the way a human responds within relationships after
he has been heart, separated from its loved ones or threatened, this way being formed in
early life, before we can talk, based on how reliable, responsive and understanding our
caregiver is [10]. Having these in mind, we are set to analyze the will to emigrate of our
subjects by dividing them into groups depending on the vital status of their parents.

2. Problem Formulation

The necessary data for conducting the following study has been extracted from the
general data base generated by our main research direction, quality of life of pre-university
professors from Cluj-Napoca. This direction, being one that covers all domains of one’s life,
generates a very large quantity of data and includes aspects about close and enlarged
family. The questionnaire, especially developed for this research, using EQLS’s approach
with 12 domains: health, job, income, education, family, social involvement, housing,
environment, infrastructure, personal safety, leisure and life satisfaction, has been
distributed to all physical education professors from pre-university schools from Cluj-
Napoca, the most important city in the Nord-West development region, second largest in
Romania and at the same time one of the largest university centers in the country. After
centralizing the results, we’ve counted 105 valid answers from a total of 149 potential
respondents, thus having a response rate of 70,46%.
After a first analysis we’ve identified several indicators as potential material for
future studies, because of their unusual high values. Among these indicators the will to
emigrate was by far the one that registered the highest values, 38% of the respondents that
offered valid answers saying they would like to emigrate and if this wasn’t concerning
enough, 22% of these stated that they would emigrate anywhere. As a consequence to this
we’ve started to analyze this particular indicator in relation with others that we’ve considered
they could influence it. First on our list was to see how this indicator varies depending on the
marital status of our respondents, so in relation with their close family, because a lot of
issues could come from this decision, including relational problems with their life partner, or

68
 Quantitative Methods Inquires

even worse psychological traumatizing experiences or sentimental deprivation for children

[21]. We then continued with its analysis in relation with the financial status of our
respondents, starting from the premises that the decision to emigrate has strong ties with the
way they cope with daily financial needs [24]. In the present paper, after revising the
research literature, we would like to return to the family indicator but from another
perspective, enlarged family, more precisely how the emigration indicator varies among our
subjects in relation with the vital status of their parents, mostly because all of our subjects
declared that have a good and very good relation with their families, family support being
an important part of the quality of life assessment and also because it seems that there are
strong ties between the will to emigrate and attachment problems, that occur in early life
mainly because the caregiver, usually the parents, don’t respond well to the needs of the
child. So we would like to see if indeed the emigration decision is based on financial
motivations, as we’ve discovered so far, or it also has ties with other indicators.
As a result, we were set to analyze which could be the indicators that had the most
influence in the decision to emigrate, how these decisions differentiate based on the vital
status of our respondent’s parents and if there’s a connection with the attachment theory. In
order to achieve this, we’ve split the study group into four categories by taking into
consideration their answers to their parent’s vital status:
1. Their mother and father are alive – coded as “mom_yes_dad_yes” for
compatibility with the utilized software;
2. Their mother is alive and their father is dead – coded as “mom_yes_dad_no” for
compatibility with the utilized software;
3. Their mother is dead and their father is alive – coded as “mom_no_dad_yes” for
compatibility with the utilized software;
4. Their mother and father are dead – coded as “mom_no_dad_no” for compatibility
with the utilized software.
The fifth category, “NA”, comprised of the ones that haven’t answered to the
income question, was excluded because it didn’t bear any relevance.

3. State of the Art Research

3.1. Emigration and family

Confucius said “The strength of a nation derives from the integrity of the family”.
So, in accordance with this statement if we were to change something in our country we
should begin by promoting and supporting the concept of a united family.
Kent Hoffman, which is one of the founders of a parental training program based
on the theory of attachment, stated at one of his seminars held in Romania that the dynamic
of family from ex communist countries was very affected and distorted by this doctrine, but at
the same time those hard times managed to unite it and strengthen the connections
between its members.
If communism is a thing of the past, at least on paper, our society is confronted
nowadays with new challenges. One of the toughest it seems to be the impossibility to offer
young capable people a stable and healthy environment where they can develop, start a
family and live a decent life. Recent studies have surfaced worrying results such as the fact
that the majority of the ones that chose to emigrate belong to the age interval 20-35, which
represents basically the ideal fertile period and more so, over 60% of these are females [42].

69
 Quantitative Methods Inquires

These findings can explain the low fertility rates that Romania faces in present. Also, another
problem could be represented by the fact that a large number of highly qualified workers,
doctors, nurses, engineers, informaticians, choose to emigrate [35, 39].
Having in mind these drawing attention findings, some researchers tried to identify
the reasons behind peoples will to emigrate and found that most of them are seeking a
better salary and a higher standard of living [4]. Although several studies have shown that
there are some benefits, mostly economical, for the families that have one or both adults
members implicated in an emigration process [3], the vast majority of the research literature
states the same conclusion: long term effects on family are devastating and it usually begins
with the alienation of its members and ends with scission, the most affected being the
children which can develop unhealthy psychological behaviors [13, 38, 6]. If adults can, at
some point, get over the pain caused by a separation, children on the other hand will
remain with scars that will be transmitted to further generations, this being the reason why it
is so important to take care of our children within the family unit.
The ideal situation, towards which we should aspire, would be to raise our children
in a happy stable family, because it has been scientifically proven that happy raised children
manage to do better in life on all areas when they benefit of the care of both parents [27],
without thinking about the money. Happiness is free and paradoxically at the same time
expensive. It can be achieved for free by doing small things that cost you noting, like a kind
word, offering a helping hand, moral support or by simply being there, or you can work hard
to earn it, although with a different outcome, as researchers from the Warwick University
have discovered [50].

3.2. Data mining

Data mining is defined as the “nontrivial process of extracting valid, previously
unknown, comprehensible and useful information from large databases” [36]. Fayyad et al.,
[16] identified that for the knowledge to be discovered through this technique some
preparatory steps need to be fulfilled: data cleaning, data reduction, data transformation,
data mining and pattern evaluation.
Classical data mining techniques were mostly used to collect data and later became
a tool for analyzing large quantity of data [2]. In the last years, its role has extended beyond
initial borders, today being encountered in almost all fields, such as marketing, banking,
medicine, astronomy, education, sociology, etc., thus becoming an important tool in decision
making processes. Its utilization basically flourished due to the fact that almost every life
domain became data-intensive [48].
In the research literature we have found many studies that use data mining
techniques covering various fields: security technology [33], manufacturing [26], banking
[14], management [45], sport [34], medicine [29], transport [49], etc., but very few that used
data mining techniques upon traditional education. Instead, we could observe that this
method is nowadays extensively used in e-learning [25, 43, 11, 52, 31, 32]. We have
managed in some of our past studies to associate it with traditional education by successfully
establishing raw connections between various indicators concerning life of pre-university
schools teachers. This offered us unique inside perspectives with a direct impact on the
national educational system [21, 23, 24].
The data mining method most used in our research was classification learning and
allowed us to automatically learn models [51]. We have based our approach towards

70
 Quantitative Methods Inquires

classification on decision trees, mainly due to the fact that they can operate under
supervision by being provided with the actual outcome for each of the training examples.
The models have been used for scanning the data in order to generate trees and make
predictions.
The instances are classified by decision trees based on their feature values, each
node being a feature in an instance to be classified and each branch a value the node can
obtain [30]. The main advantages of this method are that it creates models that are easy to
understand and missing values within the data don’t affect them [5], although due to the fact
that it only permits single dependent variable it can create certain restrictions [44].
For the present classification learning experiment, J48 and J48graft methods
(developed from C4.5 algorithm) have been employed, with the help of the very popular
Weka 3 open source GNU software for machine learning [51].

3.3. Attachment theory

John Bowlby, who is considered to be the father of attachment theory, defined it as
“one specific and circumscribed aspect of the relationship between a child and caregiver that
is involved with making the child safe, secure and protected” [7].
The emergence of this theory dates back to the late ‘40’s, when Bowlby began with
the help of James Robertson to observe hospitalized and institutionalized children that have
been separated from their parents[12]. After some studies he concluded, based on empirical
evidence, that a small child in order to grow up mentally healthy ”should experience a
warm, intimate, and continuous relationship with his mother, or permanent mother
substitute, in which both find satisfaction and enjoyment” [9]. Also, he emphasized on the
role of social networks and economy as factors which influence well developed functioning
relationships between mother and child stating in one of his books that “children are
absolutely dependent on their parents for sustenance, so in all hut the most primitive
communities, are parents, especially their mothers, dependent on a greater society for
economic provision. If a community values its children it must cherish their parents” [17].
Unfortunately the negative outcome of this behavior has long term repercussions
on the individual because “the initial relationship between self and others serves as
blueprints for all future relationships” [8].
So, returning to the context of our study, the vast majority of our respondents lived
most of their life or have been raised for the first years of their life in a communist society.
Family in the communism period although was proclaimed in the official ideology as “basic
cell of the society”, wasn’t just a simple propaganda but a justification of the intervention of
the state in private space, in order to gain control by destroying its traditional values [19].
The extensive character of the communist economy, in the context of a forced
industrialization, required growth of human workforce. So, along with the exploitation of the
rural workforce, came the concept of women emancipation, which in order to function
required for women to be relieved of family duties [19].
If we analyze these statements with the ones in the previous paragraph we start to
see that some connections begin to form. A lot of our subjects want to emigrate and not all
of them have financial reasons behind, as we’ve discovered in our previous studies. Children
with attachment problems have a tendency to run away from home [20] and as long as this
wish of alienation persists even after reaching adult age and no other reasons such as
financial exist, we can consider emigration a form of “run away”. In this regard, a recent

71
 Quantitative Methods Inquires

study has analyzed from the attachment’s theory perspective the Dutch and Belgian
immigrants from California and found a significant relationship between unresolved
attachment status and being an immigrant [47].

4. Results

4.1. Mother alive, father alive

We proceed by analyzing the first category, where both parents of our respondents
are alive. This is the most numerous group with a number of 51 subjects, from which 29
wish to emigrate, the average age being 33.29. After employing data mining techniques
we’ve obtain the following decision tree, which is graphical represented in figure number 1.
no (3.0)
<=0
no_vacations

= basic_necessities >0
yes

no (7.0)

yes (6.0)
<=4
income no_room_
=
>4
no (1.0)

=great effort basic

=NA
no (4.0)
no (0.0)

=div
no (1.0)

marital_stat =married with

=unmarried yes (1.0)

=married without

no (0.0)

=unmarried_with

no (0.0)

=NA

no (0.0)

Figure 1. J48 decision tree based on “mom_yes_dad_yes” group

72
 Quantitative Methods Inquires

As we can see, the main indicators that influence the will to emigrate for this group are of
financial type. So, the respondents that are most likely to emigrate are of 3 types: a) the
ones that have an income that only covers their basic necessities and managed to spend
more than 0 vacations in the last five years in a resort from Romania or abroad with 19
persons. These were initially divided, by their answers to the will to emigrate, in 7 that didn’t
want and 12 that did. So, our program, as we can see from the bellow graph, identified 4
persons from the ones that initially stated they don’t wish to emigrate as in fact being
potential candidates for emigration; b) the ones that consider that their income doesn’t cover
even their basic necessities and have an apartment with less than 4 rooms with 7 persons. In
this group’s case our program kept the initial distribution of the respondents based on their
questionnaire answers with 6 that would wish to emigrate and 1 not; c) the ones that didn’t
respond to the income indicator and are unmarried with 1 person for which our program
hasn’t found any other connections.
Concerning direct connections with attachment issues, our program didn’t
identified any, because indicators that could suggest such a thing like “help from parents” or
“members of living unit”, weren’t taken into consideration. So, in order to find some
connections we’ve dug deeper and isolated for each of the 3 groups the ones that have been
identified by the program as potential candidates for emigration and found out that: for the
a) group 4 persons from total 16 ones have a special cohabitation situation, in the sense that
they are still living with their parents and more than that, all of them are unmarried and
have ages very close and over 30, so it’s possible that these particular subjects could
experiment attachment problems. The help that they are receiving from parents consists only
in food; for the b) and c) groups we could not find any connections.

4.2. Mother alive, father dead

We’ve continued by analyzing the second category, in which case only the mother
of our respondents is alive. This group has a number of 25 subjects with an average age of
45,96. As we’ve discovered in our previous studies, once the age increases the will to
emigrate decreases and this is also true for this group, because only 5 persons want to
emigrate. Another possible explanation for the low number of persons that wish to emigrate
would be that only their mother is alive and they choose to stay around and help. After
running the program on the data, we’ve identified that for this group the indicators that
count the most, when talking about emigration, are the ones related to work and family. So,
the most likely to emigrate belong to 4 categories: a) the ones that have a second job and
evaluate the educational system as being one of a poor quality with 2 respondents; b) the
ones that have a second job and didn’t respond to the question regarding the quality of the
educational system with one person; c) the ones that don’t have a second job and live with
their life partner and parents with one person; d) the ones that don’t have a second job and
live with their child with one person.
As we can see from the graphic bellow, our program discovered some direct
connections between one of the indicators that could signalize attachment problems
“members of living unit” and will to emigrate. After the isolation of these two persons, we’ve
observed something very interesting: the one that doesn’t have a second job and lives with
his life partner and parents answered negative to the will to emigrate question, but the
program, as we can see, indentified him as a potential emigrant; and the one that declared
he lives only with his children, his marital status is married, although he had the divorced or

73
 Quantitative Methods Inquires

widow options to choose from, so we would be inclined to think that these two persons have
some problems of attachment.
no (1.0)

=neither bad nor good

yes (2.0)
=bad
edu_eval
= good
no (2.0)

=NA
yes (1.0)
= yes

no (1.0)

nd =husband/wife parents
2 _work_plac no (8.0)
= NA no (0.0)

=
no (1.0)

=no =child parents

no (0.0)

=husband/wife child parents

yes (1.0)

memb_home =husband/wife child grandpar

=alone no (0.0)

=child

yes (1.0)

=husband/wife child

no (2.0)

=parents
no (3.0)

=husband/wife
no (0.0)
=NA
no (2.0)

Figure 2. J48 decision tree based on “mom_yes_dad_no” group

In order to find connections between attachment theory and will to emigrate for the
other 3 persons belonging to the a) and b) groups, we’ve proceeded as in the case of the

74
 Quantitative Methods Inquires

precedent category and isolated them. The findings were interesting, because one of two
persons that were identified by the program to belong to the a) group presents similar
characteristics as the ones from the a) group in “mother alive, father alive” category. He’s
unmarried with the age of 31 and lives with his parents, in this case only with his mother.
The help received from parent consists in durable goods.

4.3. Mother dead, father alive

The third category is represented by the ones that declared their mother is dead
and their father is alive and it is the smallest of our study with only 3 respondents. Average
age continues to increase reaching 50,3. Our program has identified for this category as
principal indicator in the decision to emigrate gender, factors that could be related to
attachment problems not being directly taken into consideration. Although it doesn’t bare
much relevance, because of the small number of respondents, we would like to proceed with
a small analysis. So, as we can see from the bellow graph, if they are females they do not
want to emigrate and if they are males they do. After isolating the one that wants to
emigrate we also couldn’t find any relations to the indicators that could signal an attachment
problem.
no (2.0)

=F

gender

=M

yes (1.0)

Figure 3. J48 decision tree based on “mom_no_dad_yes” group

4.4. Mother dead, father dead

Last group of our study is represented by the ones that declared that none of their
parents are alive. It has 22 respondents and an average age of 57,72. For this group the
indicators that count the most in the emigration decision, as we can see from the bellow
graph, are ones of financial type. Initially 4 respondents from this group stated they would
like to emigrate, but our program has identified, based on their answers to the others
indicator, that one of them actually wouldn’t. The most likely to emigrate belong to 2
categories: a) the ones that own an apartment and land with 2 persons; and b) the ones that
own an apartment and a car and their income doesn’t even cover their basic necessities with
one person. Due to the fact that both parents of the subjects included in this category are
dead it will be irrelevant to pursue a connection between will to emigrate and attachment
theory.

75
 Quantitative Methods Inquires

no_vacations
no

=apt
no (2.0)
no (7.0)

=basic_necessities

=apt_land_auto no (2.0)
=all confort
prop_goods income
=apt auto
=lower than basic
=apt_land yes (1.0)

=great_effort_basic
yes
no (4.0)

=auto

no (1.0)

Figure 4. J48 decision tree based on “mom_no_dad_no” group

5. Conclusions

Although the relevance of this study can be easily contested, mainly due to the low
number of respondents and the presence of not so many indicators related to the
attachment theory, our goal wasn’t necessarily to obtain hard evidence but to establish a
conceptual framework and method for future studies.
Having this in mind we declare ourselves satisfied with the results and can affirm
that data mining methods helped us on one hand to directly identify indicators that count for
our respondents in the decision to emigrate based on the vital status of their parents and on
the other hand to narrow down possible special cases for further analysis of the indicators
that could denote problems in relation with attachment theory. In fact, we consider this
method very promising because it managed to find in such a small number of respondents
some cases that can be correlated to our main hypothesis.
The four categories chosen are in accordance with the findings of the studies
conducted by the National Institute of Statistics which state that females benefit of a higher
life expectancy then males and it is demonstrated, in our study, by the number of
respondents for the first three categories. So, the very low number of respondents for the
“mother alive, father dead” category would be somehow justified.
We continue by synthesizing the results given by the use of the data mining
method. For the first category “mother alive, father alive” the main reasons for emigrating
would be of financial type. For the second category “mother alive, father dead”, the reasons
change and refer to work and family. For the third “mother alive, father dead” category

76
 Quantitative Methods Inquires

although the number of respondents was very low the indicator that counted in the decision
to emigrate or not was gender. For the last category “mother dead, father dead”, as if it was
a circle of life, the reasons return to financial ones.
We conclude with the findings related to the attachment theory. As a general rule,
where our program didn’t find any direct connections, we’ve proceeded with an individual
analysis of the ones that were identified by the program with a wish to emigrate. For the first
category “mother alive, father alive” no direct connections were found, so we’ve isolated the
respondents and identified four possible drawing attention cases, mainly because of their
cohabitation status, age and marital status. For the second category “mother alive, father
dead” our program found direct connections through the indicator “members of living unit”.
Interestingly enough is the fact that this particular respondent, found by the program, has a
very similar situation with the ones identified by us in our analysis of the first group, he’s 31
years old, lives with his mother and is unmarried. For the third category “mother alive, father
dead” neither the program nor us could find any connections, we think mainly due to the
low number of respondents. The last category “mother dead, father dead” didn’t bear any
relevance in this case.

References

1. Ailenei, D. and Bunea, D. Labour market flexibility in terms of internal migration, Annals of
University of Oradea, 2010, pp. 153-158
2. Andronie, M. and Crisan, D. Commercially Available Data Mining Tools used in the
Economic Environment, Database Systems Journal, 2010, pp. 45-54
3. Antman, F. The Impact of Migration on Family Left Behind, Bonn: IZA, 2012
4. Asociatia Nationala a Birourilor de Consiliere pentru Cetateni. Romanii si migratia fortei de
munca in Uniunea Europeana, Bucharest, ANBCC, 2005
5. Berson, A, Smith, S. and Thearling, K. Building data mining applications for CRM, USA:
McGraw Hill, 2000
6. Botezat, A. and Pfeiffer, F. The Impact of Parents Migration on the Well-Being of Children
Left Behind – Initial Evidence from Romania, ZEW - Centre for European Economic
Research Discussion Paper, 2014
7. Bowlby, J. Attachment and loss: retrospect and prospect, American Journal of
Orthopsychiatry, 1982
8. Bowlby, J. Attachment theory, separation anxiety and mourning, American Handbook of
Psychiatry, 1975, pp. 292-309
9. Bowlby, J. Maternal care and mental health, World Helath Organization, 1951
10. Bowlby, J. O baza de siguranta. Aplicatii clinice ale teoriei atasamentului, Bucharest, Ed.
Trei, 2011
11. Bresfelean, V.P. Implicatii ale tehnologiilor informatice asupra managementului
institutiilor universitare, Cluj-Napoca, Risoprint, 2008
12. Bretherton, I. The Origins of Attachment theory: John Bowlby and Mary Ainsworth,
Developmental Psychology, 1992, pp. 757-775
13. Ciuperca, N. http://ciupercaniculina.blogspot.ro/2009/05/efectele-emigrarii-asupra-
familiei.html
14. Divsalar, M., Roodsaz, H., Vahdatinia, F., Norouzzadeh, G. and Behrooz, A. A Robust Data-
Mining Approach to Bankruptcy Prediction, Journal of Forecasting, 2012, pp. 504–
523
15. European Foundation for the Improvement of Living and Working Conditions. EURLife,
http://www.eurofound.europa.eu/areas/qualityoflife/eurlife/index.php.

77
 Quantitative Methods Inquires

16. Fayyad, U., Piatetsky-Shapiro, G. and Smyth, P. From Data Mining to Knowledge Discovery in
Databases, American Association for Artificial Intelligence, 1996, pp. 37-54
17. Feig, C. The Confluence of Attachment Style, Perceived Social Support, and Role
Attainment in Women Experiencing Postpartum Mood Disorders, Dissertations,
2012
18. Frunza, R., Maha, L.G. and Mursa, C.G. Reasons and Effects of the Romanian Labour Force
Migration in European Union Countries, CES Working Papers, 2009, pp. 37-62
19. Ghebrea, G. Regim social-politic si viata privata. Familia si politica familiala in Romania,
Bucharest, Ed. Universitatii, 2000
20. Golder, S., Gillmore, M., Spieker, S. and Morrison, D. Substance Use, Related Problem
Behaviors and Adult Attachment in a Sample of High Risk Older Adolescent
Women, Journal of Child and Family Studies , 2005, pp. 181-193
21. Haisan, A.-A. Disfunctionalitati in sistemul educational national, Cluj-Napoca, Presa
Universitara Clujeana, 2013
22. Haisan, A.-A. Life quality of physical education and sport teachers, The 6TH International
Conference “Perspectives in the Science of Human Movement”, Cluj-Napoca, Faculty
of Physical Education and Sport - UBB, 2012, pp. 143-148
23. Haisan, A.-A., and Bresfelean, V.P. A Data Mining Examination on the Romanian
Educational System - teachers Viewpoint, International Journal of Mathematical
Models and Methods in Applied Sciences, Vol. 7, no. 3, 2013, pp. 277-285
24. Haisan, A.-A., and Bresfelean, V.P. A Data Mining Survey on the Factors that Influence
Emigration Decisions among Romanian Teachers based on their Incomes,
Journal of Social and Economic Statistics, 2014, pp. 38-52
25. Hanna, M. Data mining in the e-learning, Campus-Wide Information Systems, 2004
26. Harding, J.A., Shahbaz, M. and Kusiak, A. Data mining in manufacturing: A review, Journal
of Manufacturing Science and Engineering - ASME, 2006, pp. 969-976
27. Institute for Family Studies. Strong Families, Sustainable Societies, http://ifstudies.org/strong-
families-sustainable-societies/
28. Ionescu, I. The migration of Romanians in European Community, Review of Research and
Social Intervention, 2008, pp. 23-35
29. Kharat, A. Data mining in radiology, The Indian journal of radiology & imaging, 2014
30. Kotsiantis, S.B., Zaharakis, I.D. and Pintelas, P.E. Machine learning: a review of classification
and combining techniques, Artificial Intelligence Review, 2006, pp. 159-190
31. Lile, A. Analyzing E-learning systems using educational data mining techniques,
Mediterranean Journal of Social Sciences, 2011, pp. 403-419
32. Maiorana, F, Mongioj, A. and Vacc, M. A Data Mining E-learning Tool: Description and Case
Study, Proceedings of the World Congress on Engineering . London: WCE, 2012
33. McCarthy, J. Security and data mining, InfoWorld, 2003, pp. 48-51
34. McKeever, S. Sports data mining, Hanover, Informs, 2012
35. Moldoveanu, R, and Chiricescu, A. http://www.evz.ro/detalii http://www.evz.ro/detalii/stiri/cei-
10000-de-medici-emigrati-va-saluta-de-peste-mari-si-tari-954499.html.
36. Olaru, C., and Wehenkel, L. Data mining, Computer Applications in Power IEEE, 1999, pp. 19-
25
37. Paduraru, M. E. Romania - Emigration's Impact on Families and Children, Journal of
Community Positive Practices, 2014, pp. 27-36
38. Pescaru, M. Consecințele migrației familiei contemporane asupra creșterii și educării
copiilor, Reconstruind Socialul. Riscuri si solidaritati noi - Prima Conferinta
Internaţionala a Societatii Sociologilor din Romania, Cluj-Napoca, Facultatea de
Sociologie si Asistenta Sociala, 2010, pp. 1-12
39. Petcana, A. M. Financiar - O treime dintre romani vor sa plece din Romania,
http://www.gandul.info/financiar/aproape-doua-treimi-dintre-romani-cumpara-doar-
strictul-necesar-de-alimente-10754023

78
 Quantitative Methods Inquires

40. Petrescu, R. M., Zgura, I.D. and Bac, D.P. Descriptive Analysis Of The International
Migration Phenomenon In Romania Between 1991 And 2008, Annals of Faculty
of Economics, 2011, pp. 288-294
41. Pociovalisteanu, D.-M. Migration for Education Nowadays, Annals of the „Constantin
Brancusi” University of Targu Jiu, 2012, pp. 82-87
42. Pritulescu, R. http://www.adevarul.es/stiri/social, http://www.adevarul.es/stiri/social/cum-
dispare-populatia-romaniei
43. Romero, C. and Ventura, S. Educational Data Mining: A Review of the State of the Art, IEEE
Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews, Vol.
40, no. 6, 2010, pp. 601-618
44. Shah, S, Roy, R. and Tiwari, A. Technology Selection for Human Behaviour Modelling in
Contact Centres, Decision Engineering Report Series, Cranfield University, 2006
45. Shen, C. Data Mining the Data Processing Technologies for Inventory Management,
Journal of computers, 2011
46. Timofti, I. C. Romanians' attitudes towards emigration, Journal of Psychological and
Educational Research, 2011, pp. 117-123
47. van Ecke, Y., Chope, R. and Emmelkamp, P. Immigrants and attachment status: Research
findings with Dutch and Belgian immigrants in California, Social Behavior and
Personality, 2005, pp. 657-674
48. Venkatadri, M. and Reddy, C. A Review on Data mining from Past to the Future, International
Journal of Computer Applications, 2011, pp. 19-22
49. Wallander, J, and Makitalo, M. Data mining in rail transport delay chain analysis,
International Journal of Shipping and Transport Logistics, 2012, pp. 269-285
50. Wilson, C. and Oswald, A. How Does Marriage Affect Physical and Psychological Health? A
Survey of the Longitudinal Evidence, Bonn: IZA, 2005
51. Witten, I., Eibe, F. and Hall, M. Data Mining: Practical Machine Learning Tools and
Techniques, Burlington, Morgan Kaufmann, 2011
52. Zhou, M. Data Mining and Student e-Learning Profiles, E-Business and E-Government (ICEE).
Guangzhou: IEEE, 2010, pp. 5405-5408

Appendices

Appendix 1 - Mother alive, father alive

=== Run information ===

Scheme: weka.classifiers.trees.J48 -C 0.25 -M 1 -A

Relation: mother alive, father alive
Instances: 51
Attributes: 18
marital_stat
age
sex
no_child
support_parent
no_room_apt
prop_goods
memb_home
marriage
achivment
income
will_emigr
no_vacations
2nd_work_place
soc_traject
edu_eval
fin_retrib

79
 Quantitative Methods Inquires

prof_eval
Test mode: evaluate on training data

=== Classifier model (full training set) ===

J48 pruned tree

------------------
income = basic_necessities
| no_vacations <= 0: no (3.0)
| no_vacations > 0: yes (16.0/4.0)
income = great_effort_basic: yes (16.0/6.0)
income = all_confort: no (7.0)
income = lower_than_basic
| no_room_apt <= 4: yes (6.0)
| no_room_apt > 4: no (1.0)
income = NA
| marital_stat = div: no (0.0)
| marital_stat = married_with: no (1.0)
| marital_stat = unmarried: yes (1.0)
| marital_stat = married_without: no (0.0)
| marital_stat = NA: no (0.0)
| marital_stat = unmarried_with: no (0.0)

Number of Leaves : 12
Size of the tree : 16
Time taken to build model: 0.01 seconds

=== Evaluation on training set ===

=== Summary ===

Correctly Classified Instances 41 80.3922 %

Incorrectly Classified Instances 10 19.6078 %
Kappa statistic 0.5771
Mean absolute error 0.3315
Root mean squared error 0.3809
Relative absolute error 67.5203 %
Root relative squared error 76.9048 %
Total Number of Instances 51

=== Detailed Accuracy By Class ===

TP Rate FP Rate Precision Recall F-Measure ROC Area Class

0.545 0 1 0.545 0.706 0.846 no
1 0.455 0.744 1 0.853 0.846 yes
Weighted Avg. 0.804 0.258 0.854 0.804 0.79 0.846

=== Confusion Matrix ===

a b <-- classified as
12 10 | a = no
0 29 | b = yes

Appendix 2 - Mother alive, father dead

=== Run information ===

Scheme: weka.classifiers.trees.J48 -U -M 2
Relation: mother alive father dead
Instances: 25
Attributes: 18
marital_stat
age
sex
no_child
support_parent
no_room_apt
prop_goods
memb_home
marriage
achivment

80
 Quantitative Methods Inquires

income
will_emigr
no_vacations
2nd_work_place
soc_traject
edu_eval
fin_retrib
prof_eval
Test mode: evaluate on training data

=== Classifier model (full training set) ===

J48 unpruned tree

------------------
2nd_work_place = yes
| edu_eval = good: no (2.0)
| edu_eval = neither_bad_nor_good: no (1.0)
| edu_eval = bad: yes (2.0)
| edu_eval = NA: yes (1.0)
2nd_work_place = no
| memb_home = husband/wife: no (0.0)
| memb_home = NA: no (2.0)
| memb_home = parents: no (3.0)
| memb_home = husband/wife_child: no (2.0)
| memb_home = husband/wife_parents: no (1.0)
| memb_home = child: yes (1.0)
| memb_home = alone: no (0.0)
| memb_home = husband/wife_child_grandparents: yes (1.0)
| memb_home = husband/wife_child_parents: no (0.0)
| memb_home = child_parents: no (1.0)
memb_home = husband/wife_nephew: no (0.0)
2nd_work_place = NA: no (8.0)

Number of Leaves: 16
Size of the tree: 19
Time taken to build model: 0 seconds

=== Evaluation on training set ===

=== Summary ===

Correctly Classified Instances 25 100 %

Incorrectly Classified Instances 0 0 %
Kappa statistic 1
Mean absolute error 0
Root mean squared error 0
Relative absolute error 0 %
Root relative squared error 0 %
Total Number of Instances 25

=== Detailed Accuracy By Class ===

TP Rate FP Rate Precision Recall F-Measure ROC Area Class

1 0 1 1 1 1 no
1 0 1 1 1 1 yes
Weighted Avg. 1 0 1 1 1 1

=== Confusion Matrix ===

a b <-- classified as
20 0 | a = no
0 5 | b = yes

Appendix 3 - Mother dead, father alive

=== Run information ===

Scheme: weka.classifiers.trees.J48 -U -M 1
Relation: mother dead father alive
Instances: 3
Attributes: 18

81
 Quantitative Methods Inquires

marital_stat
age
sex
no_child
support_parent
no_room_apt
prop_goods
memb_home
marriage
achivment
income
will_emigr
no_vacations
2nd_work_place
soc_traject
edu_eval
fin_retrib
prof_eval
Test mode:evaluate on training data

=== Classifier model (full training set) ===

J48 unpruned tree

------------------
sex = F: no (2.0)
sex = M: yes (1.0)

Number of Leaves : 2
Size of the tree : 3
Time taken to build model: 0 seconds

=== Evaluation on training set ===

=== Summary ===

Correctly Classified Instances 3 100 %

Incorrectly Classified Instances 0 0 %
Kappa statistic 1
Mean absolute error 0
Root mean squared error 0
Relative absolute error 0 %
Root relative squared error 0 %
Total Number of Instances 3

=== Detailed Accuracy By Class ===

TP Rate FP Rate Precision Recall F-Measure ROC Area Class

1 0 1 1 1 1 no
1 0 1 1 1 1 yes
Weighted Avg. 1 0 1 1 1 1

=== Confusion Matrix ===

a b <-- classified as
2 0 | a = no
0 1 | b = yes

Appendix 4 - Mother dead, father dead

=== Run information ===

Scheme: weka.classifiers.trees.J48 -U -M 2
Relation: mother dead father dead
Instances: 22
Attributes: 18
marital_stat
age
sex
no_child
support_parent
no_room_apt

82
 Quantitative Methods Inquires

prop_goods
memb_home
marriage
achivment
income
will_emigr
no_vacations
2nd_work_place
soc_traject
edu_eval
fin_retrib
prof_eval
Test mode: evaluate on training data

=== Classifier model (full training set) ===

J48 unpruned tree

------------------
prop_goods = apt: no (3.0/1.0)
prop_goods = apt_auto
| income = basic_necessities: no (2.0)
| income = great_effort_basic: no (4.0)
| income = all_confort: no (2.0)
| income = lower_than_basic: yes (1.0)
prop_goods = apt_land_auto: no (7.0)
prop_goods = apt_land: yes (2.0)
prop_goods = auto: no (1.0)

Number of Leaves : 8
Size of the tree : 10
Time taken to build model: 0 seconds

=== Evaluation on training set ===

=== Summary ===

Correctly Classified Instances 21 95.4545 %

Incorrectly Classified Instances 1 4.5455 %
Kappa statistic 0.8308
Mean absolute error 0.0606
Root mean squared error 0.1741
Relative absolute error 19.2771 %
Root relative squared error 45.0273 %
Total Number of Instances 22

=== Detailed Accuracy By Class ===

TP Rate FP Rate Precision Recall F-Measure ROC Area Class

0.75 0 1 0.75 0.857 0.986 yes
1 0.25 0.947 1 0.973 0.986 no
Weighted Avg. 0.955 0.205 0.957 0.955 0.952 0.986

=== Confusion Matrix ===

a b <-- classified as
3 1 | a = yes
0 18 | b = no

1
Acknowledgements
This paper has been financially supported within the project entitled “Horizon 2020 - Doctoral and Postdoctoral
Studies: Promoting the National Interest through Excellence, Competitiveness and Responsibility in the Field of
Romanian Fundamental and Applied Scientific Research”, contract number POSDRU/159/1.5/S/140106. This
project is co-financed by European Social Fund through Sectoral Operational Programme for Human Resources
Development 2007-2013. Investing in people!

2
Mr. Haisan Angel-Alex (born in Piatra Neamt, 18.04.1983) has a PhD in sociology and currently is a postdoctoral
researcher at the Institute of National Economy, where he is conducting an interdisciplinary research on the quality
of life of pre-universitary teachers from Romania, under the title "Reality or myth? Geographical positioning and its
effects on socio-professional categories".

83
 Quantitative Methods Inquires

3
Mr. Bresfelean Vasile Paul (born in Bistrita, 3.12.1979) is a Lecturer PhD at Babes-Bolyai University of Cluj-
Napoca, Romania. His educational background includes: PhD in Cybernetics and Economic Statistics (Babes-Bolyai
University of Cluj-Napoca, Romania, 2009), Master of Science in Business Information Systems (Babes-Bolyai
University of Cluj-Napoca, Romania, 2004), Bachelor of Economics (Babes-Bolyai University of Cluj-Napoca,
Romania, 2003).

84
 Quantitative Methods Inquires

A STUDY OF SURVIVAL MODELLING IN DIALYSIS PATIENTS

APPLYING DIFFERENT STATISTICAL TOOLS1

Diana-Silvia ZILISTEANU
MD, PhD, Assistant Professor,
Carol Davila University of Medicine and Pharmacy, Bucharest, Romania

E-mail: [email protected]

Ion Radu ZILISTEANU

PhD, Bucharest University of Economic Studies, Romania

Mihai VOICULESCU
MD, PhD, Professor,
Carol Davila University of Medicine and Pharmacy, Bucharest, Romania

Abstract:
The aim of this study is to develop a model for survival probabilities of incident dialysis patients
based on demographic, clinical and biological characteristics. We used statistical methods,
mainly Cox regression, and 2 statistical tools: SPSS version 21 and Excel. During the first stage,
data were analysed using SPSS software, edition 21. We performed survival analysis using Cox
proportional hazard regression test, to assess the relationship of explanatory variables with
survival time. Second stage of analysis was performed using Excel computations based on the
results provided by Cox analysis. Starting from the basal risk curve, and applying the coefficients
derived from the Cox regression analysis, the hazard curve was calculated for any combination
of values for the variables included in the equation. Based on these elements, we constructed an
Excel model for survival simulation.

Key words: statistics, SPSS, survival analysis, Cox proportional hazard regression,
mathematical model

1. Introduction

Utility of mathematical algorithms applied for assessing the risk of future negative
health outcomes emerged from the Framingham Risk Score, which was first developed based
on data obtained from the Framingham Heart Study in order to estimate the 10-year risk of
developing coronary heart disease. Because risk scores give an indication of the likely
benefits of prevention, they are useful for both the individual patient and for the clinician in

85
 Quantitative Methods Inquires

helping decide whether lifestyle modification and preventive medical treatment, and for
patient education.
Morbidity and mortality in patients with chronic kidney disease included in the
replacement of renal function program is influenced by a number of factors related both to
the patient (age, sex, renal disease and comorbidities) and the "quantity and quality" of
nephrological care in the predialysis period. O’Hare et all (2005) describe the impact of age
on the mortality risk in chronic renal failure.

1.1. Working hypothesis

The aim of this study is to develop a model that can estimate the probability of
survival of incident dialysis patients based on demographic, clinical and biological
characteristics recorded at the time of dialysis initiation. Upon this model, we can make
recommendations for optimization and planning care for a patient with chronic renal failure
before and after inclusion in chronic dialysis.
The working hypothesis of the study is that late referral to the nephrologist
adversely affect the survival of patients with chronic kidney disease included in the dialysis
program. In this paper, we propose to assess the impact that can be attributed to nephrology
referral on mortality in incident dialysis patients, analyzing the effect of this factor,
controlling for other factors that may influence survival of these patients.

1.2. Material and method

Patients with chronic kidney disease, hospitalized in Nephrology Department,
Fundeni Clinical Institute, aged over 18 years, incident in the dialysis program between
January 1st, 2007 and July 1st, 2012, were included in this analysis. Follow-up period ended
on September 1st, 2014. The main indicator of the evolution of patients was survival from the
time of inclusion in the dialysis program.
For the study patients, we recorded the following data:
- Demographical data: date of birth, gender, age at initiation of dialysis;
- Etiology of renal disease: hypertensive nephropathy, diabetic nephropathy, tubulo-
interstitial nephropathy, primitive or secondary glomerular disease, genetic diseases
(including autosomal dominant polycystic disease, Alport syndrome), systemic vasculitis
(systemic lupus erythematosus, ANCA vasculitis, Henoch Schonlein etc), multiple myeloma
or amyloidosis; cases where etiology was unknown were also recorded;
- Type of dialysis (hemodialysis or peritoneal dialysis) and access method used for renal
replacement (central venous or arteriovenous fistula for hemodialysis or peritoneal dialysis
catheter);
- Nephrology monitoring interval in months from the time the patient was evaluated for the
first time in a nephrology service until entry into dialysis;
- Clinical manifestations (hyperhydration status, presence of pericarditis, heart failure,
arrhythmias, pleural effusion, pulmonary infections, neurological, digestive manifestations,
bleeding syndrome);
- Biological parameters recorded at the time of dialysis initiation: glomerular filtration rate
estimated using the CKD-EPI formula (ml/min/1.73m2), hemoglobin (g/dL), leucocytes
(count/mmc), platelets (count/mmc), sideremia (mcg/dl), serum ferritin (ng/mL), serum
sodium (mmol/ L), potassium levels (mmol/L) total serum calcium (mg/dL) serum phosphate

86
 Quantitative Methods Inquires

(mg/dL), an intact parathyroid hormone PTH (pg/mL), serum albumin (g/dL), blood pH and
serum bicarbonate concentrations (mmol/L).

2. Statistical analysis

The study was performed on a total of 430 patients included in dialysis,

hospitalized between January 2007 and July 2012. Survival data were collected until
September 2014.

2.1. Identification of survival indicators

During the first stage, data were analyzed using SPSS software, edition 21. As
shown by Kleinbaum and Klein (2005), we performed survival analysis using Cox
proportional hazard regression test, to assess the relationship of explanatory variables to
survival time. Cases were considered censored (value status = 0) if the patient was alive or
lost to follow-up during the study period, while deceased patients were considered cases that
met the study goal (value status = 1), as explained by Sedgwick (2011). We applied a Cox
regression sequence using different control variables to identify indicators that significantly
affect survival, gradually eliminating the variables for which we did not obtained significant
values.
The final model for survival (Table 1) included the following variables that
significantly influence survival: age (p <0.0001), heart failure (p = 0.001), bleeding
syndrome (p = 0.003), diagnosis of multiple myeloma / amyloidosis (p <0.0001) serum
albumin (p <0.0001). Although we did not obtain statistically significant values for the
variable coefficient logarithm-referral logR (p = 0.252), we included this variable in the final
model in order to shape the effect of the nephrological monitoring on survival of incident
dialysis patients.

Table 1. The final Cox regression analysis including the identified variables influencing
significantly the survival: age, heart failure, hemorrhagic syndrome, serum
albumin, etiology of multiple myeloma/amyloidosis, and length of referral period
Survival indicators B SE Wald df Sig. Exp(B)
Age 0,040 0,007 34,541 1 0,000 1,041
Heart failure 0,647 0,196 10,855 1 0,001 1,910
Hemorrhagic syndrome 0,694 0,230 9,089 1 0,003 2,001
Serum albumin -0,676 0,158 18,267 1 0,000 0,508
Multiple myeloma/amyloidosis 1,845 0,290 40,569 1 0,000 6,328
logR -0,061 0,053 1,314 1 0,252 0,941

By applying Cox regression analysis, we developed the basal risk curve, which
expresses the probability of death at a certain time for patients who survived until that time.
This probability is not constant over time, permanently changing depending on the time the
analysis is done. Based on the risk curve, survival curve is calculated directly by the
arithmetic operation, which does not require other parameters.
To model mathematically the chances of survival at a certain time, we used the
above results obtained by determining statistically significant variables Cox analysis, to
create a survival function.
The relationship between the survival curve S(t) and the cumulative hazard curve
CumH is exponential, and is given by the ecuation:

87
 Quantitative Methods Inquires

S (t )  exp(CumH )
The hazard function h(t) for a given combination of characteristics (values of
explanatory variables) is the product of:
- The basic hazard function (the baseline hazard), h0 (t),
- Exponential of linear sum of the products of the values of explanatory variables (x1, x2,
..., xn) and the corresponding coefficients (β1, β2, ..., βn).
Thus, the hazard function becomes:
h(t )  h0 (t )  exp(1 x1   2 x 2  ...   n x n )
Using SPSS analysis, we determined:
- The basal cumulative hazard curve;
- Coefficients for each explanatory variable, β1, β2, ..., βn.
Based on that information, one can calculate the cumulative curve hazard function
and the survival curve for every combination of values of the explanatory variables.
Hazard or survival curves are well estimated by the following type of equations:
hˆ(t )  a  t b  ln( hˆ(t ))  ln( a )  b ln(t )
This is visible when we exemplify the hazard or survival curve through a logarithmic
scale chart. In such a graph, a linear relationship between the natural logarithm of the
hazard curve, ln(h(t)), natural logarithm of the scale of the time, ln(t) shows the relationship
expressed in the equation above.
The use of equations for computing the values on survival curve corresponding to a
certain value on time axis and to a certain combination of explicative values, even if
introduces some errors towards using the values obtained from data, has the advantage to
highlight the main characteristics of survival curve and allows the focus on them, and the
visual comparison between curves corresponding to different cathegories of patients are
eased by eliminating non-essential variations.

2.2. Development of survival model

Second stage of analysis was performed using Excel computations based on the
results provided by Cox analysis.
Using Cox analysis, we identified the following explanatory variables, which are
statistically significant:
- X1 = patient age (in years);
- X2 = age of referral to the nephrologist (months x 10);
- X3 = presence of heart failure (categorical variable, which can take values of 0 or 1,
signifying the absence or presence of disease);
- X4 = presence of bleeding syndrome (categorical variable, which can take values of 0
or 1, signifying the absence or presence of disease);
- X5 = serum albumin (10 x g / dl);
- X6 = multiple myeloma / amyloidosis (categorical variable, which can take values of 0
or 1, signifying the absence or presence of disease).
Also, the βi coefficients obtaind from Cox regression are:
β1 = 0.040;
β2 = -0.061;
β3 = 0.647;
β4 = 0.694;

88
 Quantitative Methods Inquires

β5 = -0.068;
β6 = 1.845.
Empirical survival curves can be described by mathematical functions that are more
flexible, because they require knowledge of a small number of parameters; derived
mathematical function well approximated survival curve which was empirically obtained (r2 =
98%) (Figure 1). For this reason, we further applied the mathematical function for building
survival curves.

Empirically
obtained
survival curve

Survival curve
obtained by
mathematical
function

Figure 1. Correspondence between empirically obtained survival curve and curve obtained
by the mathematical function

Starting from the basal risk curve, and applying the coefficients derived from the
Cox regression analysis, the hazard curve can be calculated for any combination of values
for the variables included in the equation.
Based on these elements, we constructed an Excel model for survival calculation
(Figure 2), with the following advantages over SPSS output:
- Is more flexible than SPSS output, which can only express the curves for categorical
variables, but not modeling for continuous variables;
- Allows a higher resolution analysis of the relationship between survival and variables of
interest; thus, can analyze the impact of small changes in clinical and biological indicators
on survival.

Figure 2. Excel model for survival estimation

89
 Quantitative Methods Inquires

The Excel model for survival estimation allows different simulations based on
combinations of values of included variables, even for hypothetical cases that have no
counterpart in the database of patients in the study group. Thus, one can choose different
values of continuous variables included (age, length of reference, serum albumin level),
while selecting various combinations of categorical variables indicating the
presence/absence of a diagnosis of multiple myeloma/amyloidosis and of the uremic
complications (heart failure, bleeding syndrome).
Based on the selected combinations of values, Excel model estimates the survival
chance calculated as a percentage, for a certain survival threshold.
We assessed whether increasing the length of nephrological care before dialysis, by
earlier referral, can improve the survival handicap given by the presence of heart failure or
bleeding syndrome after initiation of dialysis (Figure 3).

Figure 3. Percentage of initial survival difference between patients who have a certain status
(heart failure or hemorrhagic syndrome) versus those without that clinical condition,
which can be recovered by earlier referral to nephrologist. It is considered that at
baseline (reference length = 0.1 months) there is a survival difference of 100%
between the two categories

Considering that, for the referral length of 0.1 months, the difference in survival is
100%, we found the following:
- For patients with heart failure, the difference in survival drops to 94% if was referred at
1 month before dialysis, to 89% for the referral vintage of 6 months, at to 87% for referral
of 12 months;
- For patients with bleeding syndrome, the difference in survival drops to 96% for 1
month referral, to 92% for the 6 months referral, and becomes 90% for 12 months
referral vintage.
Therefore, we can say that for a patient with heart failure syndrome or bleeding
syndrome at the time of initiation of dialysis, his chances of survival improve more so as he
was referred earlier to the nephrologist.

90
 Quantitative Methods Inquires

3. Discussions and conclusions

The problem of late referral to nephrologist and initiation of renal replacement

therapy in the emergency situation is extremely serious, considering that in Romania the
number of dialysis patients incident has gradually increased in recent years, exceeding the
number of 3000 in 2011 (from 1933 in 2007 to 3161 in 2011), as reported by Romanian
Renal Registry. Medical care of these patients requires significant human and material costs,
while being associated with a high mortality rate in short, medium and long term, as shown
by Van Biesen (1999), Obialo (2005) and Black (2010). This justifies the need for coherent
health policies related to chronic kidney disease, as shown by the report published by Levey
and colleagues (2009). According to Vassalotti (2010) and McCullough (2011), a successful
program has been promoted during the last years in United Stated, proving that a
community-based screening approach can address disparities in chronic kidney disease.
This study emerged from the necessity of estimating the risk of future negative
health outcomes for patients with chronic kidney disease included in the replacement of
renal function program, based on influences by a number of factors related both to the
patient (age, sex, renal disease and comorbidities) and the length of nephrological care in
the predialysis period. Our results are similar with other of studies in the literature. Thus,
Khan et all (2005) showed that consistent nephrology care may be more important than
previously thought, especially because the frequency and severity of uremic complications
increase as patients approach dialysis. This was supported also by Jones et al (2006), who
showed the different decline in kidney function before and after nephrology referral and the
effect on survival in moderate to advanced chronic kidney disease.
The mathematical model we developed is based on survival data in our group.
Based on this model, we demonstrated that early referral can contribute to the partial
recovery of handicap given by the unfavorable profile of a patient. This model we have
developed, by estimating the chance of survival in patients enrolled in chronic renal dialysis
program, could become a useful tool for scoring the severity of clinical and biological status
in chronic renal patients. Future research will focus on expanding the patients’ database in
order to create a better approximation of survival chances based on cited parameters.
However, the utility of such mathematical model can be extended beyond the study
in which was originally designed. This model can be considered a template for further
survival analysis in different patients’ categories, using diverse indicators and variables. Of
great interest to the medical field would be the creation of modular software that can be
used independently by each physician as a tool for tailored estimation of the risk score for an
individual patient, by applying specific characteristics of each subject.

References

1. Annual Report 2011 of Romanian Renal Registry. Ministry of Health - Hospital Nephrology
"Dr Carol Davila", Bucharest, Romania, 2012, available at
http://www.srnefro.ro/media/RRR.Raport.2011.pdf
2. Black, C., Sharma, P., Scotland, G., McCullough, K., McGurn, D., Robertson, L., and Smith, C.
Early referral strategies for management of people with markers of renal
disease: a systematic review of the evidence of clinical effectiveness, cost-
effectiveness and economic analysis, NIHR Health Technology Assessment
programme: Executive Summaries, Vol. 14, No. 21, 2010

91
 Quantitative Methods Inquires

3. Jones, C., Roderick, P., Harris, S. and Rogerson, M. Decline in kidney function before and
after nephrology referral and the effect on survival in moderate to advanced
chronic kidney disease, Nephrology Dialysis Transplantation, Vol. 21, 2006, pp.
2133-2143
4. Khan, S., Xue, J., Kazmi, H., Gilbertson, D., Obrador, G., Pereira, B. and Collins, A. Does
predialysis nephrology care influence patient survival after initiation of
dialysis? Kidney International, Vol. 67, 2005, pp.1038–1046
5. Kleinbaum, D.G and Klein, M. Survival Analysis. A Self-Learning Text. 2nd ed. New York,
Springer, 2005
6. Levey, A. S., Schoolwerth, A. C., Burrows, N. R., Williams, D. E., Stith, K. R. and McClellan, W.
Comprehensive public health strategies for preventing the development,
progression, and complications of CKD: report of an expert panel convened by
the Centers for Disease Control and Prevention, American Journal of Kidney
Diseases, Vol. 53, No. 3, 2009, pp. 522-535
7. McCullough, P. A., Vassalotti, J. A., Collins, A. J., Chen, S. C. and Bakris, G. L. National Kidney
Foundation's Kidney Early Evaluation Program (KEEP) annual data report 2009:
executive summary, American Journal of Kidney Diseases, Vol. 55, No. 3, 2010, S1-
S3
8. Obialo, C. I., Ofili, E. O., Quarshie, A. and Martin, P. C. Ultralate referral and presentation
for renal replacement therapy: socioeconomic implications, American journal of
kidney diseases, Vol. 46, No. 5, 2005, pp. 881-886
9. O’Hare, A. M., Bertenthal, D., Covinsky, K. E., Landefeld, C. S., Sen, S., Mehta, K. and Walter, L.
C. Mortality risk stratification in chronic kidney disease: one size for all ages?,
Journal of the American Society of Nephrology, Vol. 17, No. 3, 2006, pp. 846-853
10. Sedgwick, P. Survival (time to event) data: censored observations, British Medical Journal;
Vol. 343, 2011, d4816
11. Sedgwick, P. Cox proportional hazards regression, British Medical Journal, Vol. 347, 2013,
f4919
12. Van Biesen, W., De Vecchi, A., Dombros, N., Dratwa, M., Gokal, R., LaGreca, G.,and Lameire, N.
The referral pattern of end-stage renal disease patients and the initiation of
dialysis: a European perspective, Peritoneal dialysis international, Vol. 19, Suppl. 2,
1999, pp. S273-S275
13. Vassalotti, J.A., Li, S., McCullough, P.A. and Bakris G.L. Kidney early evaluation program: a
community-based screening approach to address disparities in chronic kidney
disease. Seminars in Nephrology, Vol. 30, 2010, pp. 66–73
14. Zilisteanu, D.S. Late nephrology referral and impact on morbidity and mortality of
patients with chronic renal disease, Ph.D. Thesis, Carol Davila University of
Medicine and Pharmacy, Bucharest, 2013

1
Acknowledgment
This paper was co-financed from the European Social Fund, through the Sectorial Operational Programme Human
Resources Development 2007-2013, project number POSDRU/159/1.5/S/138907 "Excellence in scientific
interdisciplinary research, doctoral and postdoctoral, in the economic, social and medical fields - EXCELIS",
coordinator The Bucharest University of Economic Studies.

92
 Quantitative Methods Inquires

THE REDISTRIBUTIVE EFFECT OF THE ROMANIAN TAX-

BENEFIT SYSTEM: A MICROSIMULATION APPROACH1

Eva MILITARU
Postdoctoral fellow, Bucharest University of Economic Studies, Romania
Researcher, National Research Institute for Labour and Social Protection, Romania

E-mail: [email protected]

Abstract:
This paper attempts to investigate the income distribution of Romanian households, focusing on
the role of the tax-benefit system in income redistribution. We evaluate the redistributive effect
by estimating income inequality changes due to tax-benefit components. We use EU-SILC
microdata and the EUROMOD microsimulation model to simulate income components. The
results point out that income inequality is considerably reduced through the tax-benefit system,
as a great deal of income is redistributed among households. The analysis of the income
components that contribute to inequality reduction emphasizes that pensions, personal income
taxes and social benefits are in favour of inequality reduction, while social contributions act the
opposite way. Our results are sensitive to social and fiscal policy changes.

Key words: Income distribution, Redistributive effect, Tax-Benefit System, Income

inequality, Microsimulation

1. Introduction

The aim of this paper is the investigation of household income distribution in

Romania, with focus on the role of the tax-benefit system in income redistribution. The
evolution of income distribution in Romania has encountered many changes during recent
years. We remark average household income growth before the economic crisis (up to
2008), income decline during crisis and slight recovery during the most recent years (2012-
2013). It seems that the poorer households have benefited more from pre-crisis positive
economic developments and have lost lower proportions of their incomes during the
economic crisis, as compared to higher income households. The unequal changes along the
income distribution have shaped a more equal income distribution in 2013 compared to
2007. Besides the economic developments which had a direct influence on household
income levels, the changes that took place in the tax-benefit system (as social and fiscal
policy response to crisis) have a serious impact on household income developments.
The paper attempts to evaluate the redistributive effect of the Romanian tax-benefit
system, by estimating income inequality changes due to tax-benefit components. It covers
the period between 2007 and 2013. The results indicate that half of the income inequality
before taxes and transfers is reduced through the tax-benefit system. We find out that the
economic crisis has led to the decline of income inequalities, as richer households have lost

93
 Quantitative Methods Inquires

more of their market incomes; but income inequality has dropped also due to tax-benefit
system changes that were adopted in order to cope with the emerging situation. We use the
EUROMOD microsimulation model in order to split the household income into income
components (i.e. social benefits: pensions, means-tested benefits, non means-tested
benefits, etc.; taxes: personal income tax, social insurance contributions) and assess the role
of each of these components in the redistribution of income. The redistribution of income
through the tax-benefit system is evaluated by calculating an indicator derived from the Gini
coefficient of pre and post social transfers and taxes. We concentrate on the income
components which are responsible for the differences between the two measures, by
decomposing the Gini coefficient by income source.
The rest of the paper is organized as follows. We continue with a brief overview of
the general framework with respect to recent empirical findings concerning the evolution of
income distribution in Romania. Then, we focus on the description of methodology, data
and indicators used. The following section summarizes the most important findings
concerning the estimation of the redistributive effect of the tax-benefit system in Romania.
The paper ends with some concluding remarks.

2. General framework

During the last decades there has been a great interest in the measurement of the
redistributive effect of social benefits and fiscal systems and of the contribution of each
income component to redistribution. Starting with Kakwani (1977a, 1977b) who has laid the
foundations for the measurement of income redistribution through the difference between
the pre and post taxes and transfers Gini coefficients, a large strand of the literature in this
field has focused on theoretical issues regarding measurement, but also on effective
assessment of income redistribution through the tax-benefit system.
We mention as follows several recent relevant studies dealing with household
income distribution in Romania. Most of these studies were focusing on the estimation and
explanation of income inequalities, and very few are concerned with the effects of the tax-
benefit system on income distribution.
One of the most relevant studies concerning the estimation and analysis of income
inequalities can be attributed to Molnar (2010) who has decomposed income inequality by
groups of main household characteristics. Her results show that the most important elements
driving income inequalities between groups of households are education and labour market
status. A decomposition exercise has been employed also by Zamfir et al. (2008) who have
investigated the impact of remittances sent by Romanians working abroad on income
inequalities between and within urban and rural areas. Their results show that remittances
have driven the decline of income inequalities both between and within rural and urban
areas. Dachin and Mosora (2012) have studied the inequalities driven by the regional
distribution of household income and shown that the most relevant factors driving the
unequal distribution of income by regions are the employment structure by economic
activities and the prevalence of subsistence agriculture.
Concerning the effects of the country’s economic development on income
distribution, we mention Militaru and Stroe (2010) who have investigated the income
dynamics in Romania between 2000 and 2007 using a growth incidence curve approach.
Their findings clearly show that the economic growth has been pro-poor, meaning that the

94
 Quantitative Methods Inquires

average income growth of poor households has been more substantial than the income
growth of the rest of the households. Both households from rural and urban areas have
been affected by crisis, but not equally. This is an issue addressed by Dachin and Sercin
(2012) who concluded that household income in rural areas is less affected by crisis,
compared to household income in the urban area. This can be explained by the different
structure of household income by income sources between the two areas, the consumption
from own-resources and the prevalence of informal income in the rural area as well.
The effectiveness of social policies in reducing income inequalities has been
investigated by Precupetu (2013) who has focused on income inequalities in Romania after
1990. The concern on the tax-benefit system’s effect on income distribution in Romania is
very recent though. For example, Voinea and Mihaescu (2009) have measured the changes
in the income distribution due to the income tax reform that took place in 2005, shifting
from a progressive to a flat rate personal income tax and showed that only the richest 20%
are clear winners of this reform. Avram et al. (2012) in their study on the distributional
impact of fiscal consolidation measures taken in Romania (and other eight EU countries)
during the recent economic crisis have shown that richer households have lost higher
proportions of their incomes than poorer households, as a result of the above-mentioned
measures. A similar analysis has been carried out by De Agostini et al. (2014), but they have
measured the effects of all changes in the tax-benefit system (not limited to fiscal
consolidation). They have concluded that in Romania the changes in the tax-benefit system
were progressive, in the sense that their distributional effects were mainly beneficial for the
bottom of the income distribution. Avram, Levy et al. (2014) have studied the redistributive
effect of the tax-benefit system and found out that in Romania, unlike in most of the EU
countries, social contributions increase income inequalities mostly due to higher limits set on
contribution base.

3. Methodological issues

3.1. Methodology and data

We base our analysis on microdata from the European Union Survey on Income
and Living Conditions (EU-SILC). The data is collected annually and it is nationally
representative for the Romanian population. We use data collected during the 2008 and
2010 surveys, the income reference years being 2007, respectively 2009. Using updating
factors by detailed income components (i.e. change in the average value of an income
component between the year of the data and the current/ policy year), we adjust the value
of the income variables from 2007 to 2008 and from 2009 to 2010-2013. Other variables
(demographic, household size and composition, labour market variables) are kept constant
to the survey years. We estimate the direct, static effect of the tax-benefit system on
household income distribution.
We make use of a tax-benefit microsimulation instrument, namely the tax-benefit
microsimulation model EUROMOD. The model comprises the Romanian tax-benefit policy
rules for 2007-2013 and is built on EU-SILC data. The model can simulate the entitlement to
cash social benefits (i.e. in-kind benefits are not taken into account) and tax and social
contribution liabilities. The implemented tax-benefit policy rules are those in place at the
middle of the year (i.e. the 30th of June), being assumed that no changes have occurred
during the rest of the year. In the interpretation of results from microsimulations using

95
 Quantitative Methods Inquires

EUROMOD, one should bear in mind this very important issue. In our case, several relevant
changes in the Romanian tax-benefit system took place during the second half of the year,
they being effectively implemented in EUROMOD in the following year’s rules. The model
assumes 100% benefit take-up (exception in the case of the minimum guaranteed income)
and no tax evasion. Asset tests that condition the entitlement for means-tested benefits are
not simulated due to the lack of adequate information (EUROMOD Country Report:
Romania, 2007-2009, 2009-2010, and 2009-2013).
The household disposable income is calculated as the sum between the original
income (i.e. market or gross income) and the social transfers, minus direct taxes. The social
transfers (benefits) are split into three categories: pensions, means-tested benefits (i.e.
beneficiaries have to comply with some eligibility criteria regarding income levels below a
threshold, often differentiated by household size, number of children, etc.; the beneficiaries
may also be subject to asset tests for some of the benefits), non means-tested benefits (such
as the state allowance for children, etc.). The category of direct taxes includes the flat-rate
personal income tax (together with the tax allowance) and the social insurance contributions
paid by the employees, self-employed (and pensioners) in order to cover the risks of
retirement, sickness, unemployment, work-accidents, etc. The household size and age
structure is taken into account by using the modified OECD equivalence scale. Thus,
household income is adjusted and each household member is assigned the same amount of
income.

3.2. Indicators
In order to measure the redistributive effect of the tax-benefit system in Romania,
we use the common approach proposed by Kakwani (1977a, 1977b), who has suggested
the assessment of the size of the income redistribution (RE) through the social benefit and tax
system by the difference between the Gini coefficients of pre-fiscal income (no social benefits
and taxes) (GX) and post-fiscal income (GN):

(1)
The Gini coefficient measures the income inequality by the area between the Lorenz
curve and the equality line. A progressive tax-benefit system moves the Lorenz curve towards
the equality line; therefore the income inequality will be lower in this case. The redistributive
effect is larger for greater average tax rates and greater progressivity. Atkinson (1980) and
Plotnick (1981) pointed out that the tax-benefit system induces, besides the movement of the
Lorenz curve, the re-ranking of individuals/ units, which can be measured by the difference
between the Gini and the concentration coefficient of post taxes and transfers income. A few
years later, Kakwani (1984) has decomposed the redistributive effect into vertical
(progressivity) and re-ranking terms. In other words, the redistributive effect is reduced by
the changes in the new ranking of individuals/ households which occurred in the post- tax
and transfers system (see formula (2) below):

(2)
k AP
where, V is the Kakwani vertical effect and R the Atkinson-Plotnick index of re-ranking.
The vertical effect can be computed as below:

(3)

96
 Quantitative Methods Inquires

where tx is the average tax rate and PTk is the progressivity of the tax-benefit system (named
the Kakwani index of progressivity).
The re-ranking effect is the difference between the Gini coefficient of post-taxes
and transfers (GN) and the concentration coefficient of post-tax and transfers income (DNx):

(4)
We estimate the redistributive effect of the Romanian tax-benefit system between
2007 and 2013 and then, we decompose the effect into vertical and re-ranking effect. The
Gini coefficient is decomposed by income source in order to estimate the contribution of
each income component to income inequality, following the approach described in Lerman
and Yitzhaki (1985) and in Stark, Taylor and Yitzhaki (1986), which allows the calculation of
the impact that a marginal change in a particular income source will have on inequality. The
influence of an income component on total income inequality depends on the importance of
the income source with respect to the total income (Sk), the extent of equality/ inequality in
the distribution of that income source (Gk) and on the correlation of the income source with
the total income distribution (Rk) (see formula (5)).

(5)
Using the above decomposition we estimate the effect that 1% change in income
from source k will have on total income inequality, as:

(6)
This approach concerning the measurement and decomposition of the redistributive
effect of the tax-benefit system has been most recently used by Verbist and Figari (2014),
and the decomposition of inequality by income source has been employed by Avram et al.
(2014). These papers follow a comparative framework and analyse groups of EU countries.
The contribution of our paper is that the methodology is applied on Romania, for the period
between 2007 and 2013 and is focused on the dynamics of the redistributive effect,
explaining the impact of certain changes that took place in the tax-benefit system on the size
of the redistributive effect.

4. Main findings

4.1. Income distribution and the structure of the tax-benefit system, 2007-2013
Between 2007 and 2013, the household income dynamics has been strongly
influenced by the economic downturn which became visible in Romania by the end of 2009.
The average household income (real income, adjusted with the consumer prices index,
reference 2007) has dropped in 2009 by approximately 5%. The negative developments of
household incomes continued during the next two years, but the pace of decline was
smoother than in 2009 (see Fig. 1).

97
 Quantitative Methods Inquires

Figure 1. Annual percentage change in the average household disposable income, by

quintile groups, %
Source: own calculations using EU-SILC, EUROMOD ver. G.1.0
Note: incomes are adjusted with the consumer prices index, reference year 2007; quintiles are
constructed based on the equivalised household disposable income.

However, the developments were uneven along the income distribution (by quintile
groups, each quintile comprises 20% of the population), the middle and the upper quintiles
have benefited more from the economic growth in 2008, but also lost more during the crisis
than the bottom quintiles, who have managed to preserve their levels of income from one
year to another (except for the year 2010). This is mostly due to important changes in the
tax-benefit system, the so-called “austerity measures” aiming fiscal consolidation, but also
helping the worse off population. The fiscal policy changes that took place in 2010 and 2011
seem to have had a positive impact on household disposable incomes, while some of the
changes in the social benefit system had a positive impact on household disposable income
(i.e. changes in the means-tested benefits) and others a negative effect (i.e. the decrease of
the unemployment benefit and the changes in the rules for the child raising allowance).
Overall, the changes in the tax-benefit system seem to be progressive, as the bottom of the
income distribution is advantaged in terms of income losses.
The Romanian tax-benefit system’s largest component is public pensions. The
pensions’ share in the average household disposable income accounts for around 23-28%,
slightly changing with the years. The other social benefits, either means-tested or not, do not
exceed 8% of the household disposable income. The direct taxes, which consist of personal
income tax and social insurance contributions account for almost 30% of the household
disposable income. The social insurance contributions are designed to cover contingencies
such as old-age, sickness, unemployment, work accidents, etc. and are paid by employees
and self-employed. Additionally, pensioners with pension levels exceeding a statutory
threshold pay the health insurance contribution.
As it can be seen in the figure bellow (Fig.2), the structure of the tax-benefit system
has not changed considerably between 2007 and 2013. We notice though an increased

98
 Quantitative Methods Inquires

share of social benefits in 2010. This can be explained by the increase of the income
eligibility thresholds for some means-tested benefits (i.e. minimum social pension, social
assistance benefit). There was also a decline in the share of social contributions after 2011,
most likely as a result of the introduction of an upper ceiling to the social insurance
contribution of employees and self-employed, and the introduction of lower limits to health
insurance contribution for all population (active population and pensioners). In 2011, the
share of social benefits has contracted, consequence of the following changes: decrease of
the unemployment benefit, maximum threshold set for the child raising benefit and the
policy rules were changed, increase of the child raising incentive, the allowance for the new-
born children was abolished, the income thresholds and the amounts of the means-tested
family benefit and the means-tested heating benefit have been changed. We should note
that some of the above mentioned changes took place during the second half of the year
2010, being part of the austerity measures, but according to EUROMOD rules, they are
implemented in 2011 (see the previous section on methodology of the paper).

Figure 2. Structure of the tax-benefit system, % of household

disposable income, 2007-2013
Source: own calculations using EU-SILC, EUROMOD ver. G.1.0

It is important to mention that the structure of the tax benefit-system varies a lot by
quintile groups constructed based on the equivalised household disposable income.
Naturally, the bottom quintile (1st quintile) relies on means-tested benefits to a much greater
extent than the other parts of the income distribution. On the other hand, the upper quintiles
(4th and 5th quintiles) are paying a higher proportion of their disposable income as personal
income taxes and social insurance contributions (see Fig. 3). This picture points towards a
progressive tax-benefit system, where poorer households benefit more from social transfers

99
 Quantitative Methods Inquires

and the richer pay more taxes, but the size of the redistribution is to be treated in the next
sub-section.

Figure 3. Structure of the tax-benefit system, by quintile groups,

% of household disposable income, 2013
Source: own calculations using EU-SILC, EUROMOD ver. G.1.0

4.2. Redistribution of income through the tax-benefit system

We have measured the redistributive effect of the tax-benefit system as a whole, by
the difference between the Gini coefficient of pre and post taxes and transfers, as described
in detail in the section on methodological issues. The results are presented in the figure
below (see Fig. 4). More than half of the income inequality before taxes and transfers (i.e.
original or market income) is reduced through the tax-benefit system. It seems that the
economic crisis has led to the decline of income inequalities, as richer households have lost
more of their market income, but also due to the tax-benefit system changes that were
adopted in order to cope with the crisis. Thus, the redistributive effect of the tax-benefit
system was highest in 2011 and lowest in 2009.
We have decomposed the redistributive effect into vertical effect and re-ranking
effect, the idea behind being that the vertical effect is actually reduced by the re-ranking of
individuals that has occurred in the post-tax and transfers system. As it can be seen in the
figure below (Fig. 4), the re-ranking effect resulted from the redistribution of income lowers
the total redistributive effect by approximately 40%. Only in 2009 and 2010, the re-ranking
effect has exceeded 50% of the total redistributive effect. Nevertheless, the dynamics of the
redistributive effect is strongly driven by the vertical equity term.

100
 Quantitative Methods Inquires

Figure 4. Redistributive effect of the tax-benefit system, 2007-2013

Source: own calculations using EU-SILC, EUROMOD ver. G.1.0

The decomposition of the Gini coefficient by income source shows that pension
income is the most important driver for inequality reduction from all income components.
This is because pensions are the largest component of the tax-benefit system and are more
equally distributed among the whole population than other income components (except for
the personal income tax and the social insurance contributions). The means-tested benefits
contribute to the reduction of income inequality due to their negative correlation with the
distribution of total income, as the lower part of the income distribution benefits more from
these transfers.
The personal income tax is decreasing income inequality due to its distribution and
strong negative correlation with total income distribution. However, the size of the effect is
lowered by the nature of the tax rate, this being a flat-rate tax. The non means-tested
benefits have lower impact on income inequality. As expected, their share in total household
disposable income is the lowest. The social insurance contributions have acted in the sense
of inequality reduction after 2011, as a result of several important changes that took place in
2011 in the social insurance system. On one hand, an upper ceiling was introduced for the
social insurance contribution paid by employees and self-employed which could have
increased income inequality, but this was counterbalanced by the introduction of lower limits
to health insurance contribution in the case of pensioners, thus the overall effect being in
favour of inequality reduction. The dynamics of the marginal effect of personal income tax
shows a decline in the contribution of the income tax to income inequality reduction.
During the first years of economic crisis (2009-2010), the means-tested benefits
have strongly acted as to decrease income inequalities.

101
 Quantitative Methods Inquires

5. Conclusions

The paper has attempted to study the income distribution of Romanian households,
concentrating on the structure of the tax-benefit system and on the effects on the income
redistribution of income components and of the system as a whole.
Our analysis covers the period between 2007 and 2013 and is based on annual
nationally representative microdata from the European Union Survey on Income and Living
Conditions (EU-SILC). We use the tax-benefit microsimulation model EUROMOD in order to
simulate the components of the tax-benefit system. In order to measure the redistributive
effect of the tax-benefit system in Romania, we use the approach proposed by Kakwani
(1977a, 1977b) and we assess the size of the income redistribution through the social
benefit and tax system by the difference between the Gini coefficients of pre-fiscal and post-
fiscal income. We decompose the redistributive effect into vertical and re-ranking effect. In
order to establish the contribution of each income component to income inequality, we
decompose the Gini coefficient by income source, following the approach described in
Lerman and Yitzhaki (1985) and in Stark, Taylor and Yitzhaki (1986), which allows the
calculation of the impact on income inequality of a marginal change in a particular income
source.
Our results show that between 2007 and 2013, the household income dynamics
has been strongly influenced by the economic downturn which became visible in Romania by
the end of 2009. The average household income has dropped in 2009 and the negative
developments have continued for the next two years, but the pace of decline was smoother
than in 2009. Starting from 2012, we notice a slight increase in the average level of
household income. Though, the developments were unequal along the income distribution,
the middle and the upper quintiles have benefited more from the economic growth in 2008,
but also lost more during the crisis than the bottom quintiles, who generally have managed
to preserve their levels of income. This latter result is mostly due to important changes that
took place in the tax-benefit system.
With respect to income redistribution, the results indicate that income inequality
before taxes and transfers is reduced to half through the tax-benefit system. During the
economic crisis, richer households have lost more of their market income. This is reflected in
the reduction of the original income inequality (before taxes and transfers). Additionally, the
role of the tax-benefit system was considerable in income inequality reduction, due to
changes that were adopted in order to cope with the economic crisis. The decomposition of
the redistributive effect into vertical effect and re-ranking effect shows that the redistributive
effect is reduced by the re-ranking of households that has occurred in the post-tax and
transfers system. The re-ranking of households lowers the total redistributive effect by
approximately 40%. In 2009 and 2010, the re-ranking effect has exceeded 50% of the total
redistributive effect. However, the dynamics of the redistributive effect is mainly driven by the
vertical equity term.
The decomposition of the Gini coefficient by income source has shown that
pensions, which account for the larger part of the tax-benefit system, play the most
important part in income redistribution, while social insurance contributions increase income
inequalities (especially before 2011). The personal income tax is redistributive, though its
effect is not substantial due to its flat-rate. Means-tested and non means-tested social

102
 Quantitative Methods Inquires

benefits are conducive to income inequality reduction. During the economic crisis, the
means-tested benefits have been the most influential on decreasing income inequality.

References

1. Atkinson, A. Horizontal Inequity and the Distribution of the Tax Burden. In Aaron,
H. and Boskin, M. eds. “The Economics of Taxation”, Washington D.C. The
Brookings Institution, 1980
2. Avram, S., Figari, F., Leventi, C., Levy, H., Navicke, J., Matsaganis, M., Militaru, E.,
Paulus, A., Rastrigina, O. and Sutherland, H. The distributional effects of
fiscal consolidation in nine EU countries. Research Note 1/2012 of the
European Observatory on the Social Situation and Demography, European
Commission, 2012
3. Callan, T., Leventi, C., Levy, H., Matsaganis, M., Paulus, A. and Sutherland, H. The
distributional effects of austerity measures: a comparison of six EU
countries. Research Note 2/2011 of the European Observatory on the Social
Situation and Demography, European Commission, 2011
4. Dachin, A. and Mosora, L.C. Influence factors of regional household income
disparities in Romania, in Journal of Social and Economic Statistics, no.1,
vol. 1, 2012.
5. Dachin, A. and Sercin, A. Effects of the economic crisis on rural household incomes
in Romania, Paper presented at the 3rd International Symposium "Agrarian
Economy and Rural Development - realities and perspectives for Romania",
Bucharest, Romania, 2012
6. Davidescu (Alexandru), A. A. Estimating the size of romanian shadow economy a
labour approach, Journal of Social and Economic Statistics, vol. 3, no. 3,
2014, pp. 25-37
7. De Agostini, P., Paulus, A., Sutherland, H. and Tasseva, I. The effect of tax-benefit
changes on income distribution in EU countries since the beginning of
the economic crisis. Research note 03/2013 of the Social Situation Monitor,
European Commission, 2013
8. Immervoll, H., Levy, H., Lietz, C., Mantovani, D., O’Donoghue, C., Sutherland, H. and
Verbist, G. Household incomes and redistribution in the European
Union: quantifying the equalizing properties of taxes and benefits in
Papadimitriou, D. B. (Ed.) “The Distributional Effects of Government Spending
and Taxation”, London, Palgrave Macmillan, 2006
9. Kakwani, N. Measurement of tax progressivity: an international comparison, in
Economic Journal, Vol. 87, 1977, pp. 71-80
10. Kakwani, N. Application of Lorenz Curves in Economic Analysis, in Econometrica,
Vol. 45, No. 3, 1977, pp. 719-727
11. Molnar, M. Income distribution in Romania, MPRA Paper No. 30062, 2010
12. Militaru, E. and Stroe, C. Poverty and Income Growth: Measuring Pro-Poor
Growth in the Case of Romania, Proceedings of the 11th WSEAS
Mathematics and Computers in Science Engineering, WSEAS Press, 2010
13. Precupetu, I. and Precupetu, M. Growing inequalities and their impacts in
Romania. Country Report, GINI Growing inequalities’ impact, 2013

103
 Quantitative Methods Inquires

14. Plotnick, R. A Measure of Horizontal Inequity, in Review of Economics and

Statistics, Vol. 63, 1981, pp. 283-288
15. Reynolds, M. and Smolensky, E. Public Expenditures, Taxes, and the Distribution of
Income: The United States, New York, Academic Press, 1977
16. Sutherland, H. EUROMOD: the tax-benefit microsimulation model for the
European Union. In Gupta, A. and Harding, A. (Eds.) “Modelling our future:
population ageing, health and aged care”, International Symposia in Economic
Theory and Econometrics Vol. 16, Elsevier, Amsterdam, 2007, pp. 483-488
17. Stroe, C., Militaru, E., Avram, S. and Cojanu, S. EUROMOD Country Report: Romania
2007-2010, ISER, University of Essex, Colchester, 2012
18. Verbist, G. Redistributive effect and progressivity of taxes: an international
comparison across the EU using EUROMOD, EUROMOD Working Paper
Series: EM5/04, 2004
19. Voinea, L. and Mihaescu, F. The impact of the flat tax reform on inequality – the
case of Romania, Romanian Journal of Economic Forecasting, Vol. 4, 2009
20. Zamfir, A.M., Mocanu, C., Militaru, E. and Pirciog, S. Impact of Remittances on
Income Inequalities in Romania, In Schuerkens, U. (Eds.) “Globalization
and Transformations of Social Inequality”, Routledge Taylor& Francis Group,
New York, London, 2010, pp. 58-75

1
Acknowledgment
This paper was co-financed from the European Social Fund, through the Sectorial Operational Programme Human
Resources Development 2007-2013, project number POSDRU/159/1.5/S/138907 "Excellence in scientific
interdisciplinary research, doctoral and postdoctoral, in the economic, social and medical fields -EXCELIS",
coordinator The Bucharest University of Economic Studies.

104
 Quantitative Methods Inquires

SURVEY REGARDING RESISTANCE TO CHANGE IN

ROMANIAN INNOVATIVE SMEs FROM IT SECTOR1

Eduard Gabriel CEPTUREANU

PhD, Assistant Professor,
Bucharest University of Economic Studies, Romania

E-mail: [email protected]

Abstract:
Unfortunately, few changes predominantly generate positive effects involving major effort and
costs are often not far short of expectations. Why efforts to implement the changes result in
failure or do not match the expected results? We will try to formulate a response based on
achieving an investigation on a sample of 819 SMEs innovative IT Romanian order: (i) identify
the types of resistance to change prevailing in the analyzed companies; (ii) identify change
management tools used to reduce resistance to change; (iii) proliferation substantiate future
directions of change management in Romanian.

Key words: change management, innovative SMEs, resistance to change

1. Introduction

Resistance to change issue is based on a set of logical reasons arising from the
third law of Newton's dynamics that every movement always meets resistance forces. To
overcome resistance to change we must answer at least two questions:
• what are the causes of resistance to change?
and
• how to work on these causes to eliminate or substantially reduce?
Before we attempt to answer these questions, we consider useful to present the
opinion of Rick Maurer, author of "Beyound The Wall Of Resistance". According to it, the
base resistance are two sets of elements that represent two distinct levels:
• Level 1- such information-logical, visible, relatively easier to see and countered;
• Level 2- personal and emotional, that often people do not flaunt it, to be discovered,
evaluated and addressed specific means.
This postulate is reflected by P. Senge showing the life cycle of change through a
curve that unrealized potential for growth and development due to resistance to change
manifested in various forms:
American researcher I. Ansoff [1] notes that resistance to change is "a
multiaspectual phenomenon generating unexpected obstacles in the process of
organizational change and instability thus introducing unexpected efforts in the process. At
the same time, is an expression of irrational behavior of organization members who refuse
to recognize the new dimensions of reality and ignore the logical arguments. "

105
 Quantitative Methods Inquires

Unfulfilled potential

Level
of
develo Factors who
pement influence real limits
of developement
(resistance to
change,
organisational
potential etc.)

Life cycle

Time
Figure 1. Diagram of life cycle for processes of change [17]

Based on the statements, we can consider that (Ceptureanu, 2012):

 resistance organization is a permanent phenomenon generated by the
tendency of a system to maintain a relatively stable equilibrium inside and
organizational change is perceived as a destabilizing phenomenon;
 resistance to change should not be seen solely as a negative reaction
because, given the appearance of objectivity creates the prerequisites necessary
to test the viability of new ideas;
 although resistance to change is objective and has a legitimate source is
subjective element of the system (organization) - the man who has a major
importance in the development activity, fulfilling both the role of "organizer" (by
behavior, initiative, incentives) and the "destabilizing". given that the source of
opposition to change a form subjective element of the system, as the source of
this phenomenon objectively analyzing subjective reasons such as: fear of the
new and inertia are presented in several forms. why people are hostile to
something new are very different and are limited to the following associations:
property damage, loss of current status, new responsibilities, limitation of rights,
liquidation function, increase the volume and complexity of work, loss of moral
advantages (status, authority, power), replacing old methods of work, formal and
informal relationships, feelings of incompetence for new tasks, functions.
This particular form of organizational behavior - resistance can occur in two forms:
 Active: when the manager hears, sees, understands why a negative feedback
and take steps to change it;
 Passive (hidden / masked) when nobody open disagreement, but no changes
are not implemented (no resistance or "deformation").
Even transformations routine daily occurring in the coordination of any business,
such as launching new products, forming interactions or new systems are often accompanied
by tension, disagreement, stress - in other words by resistance . If this is the case of reduced
scale transformations, we can imagine how hard it is to achieve major changes involving

106
 Quantitative Methods Inquires

changes in formal and informal structures, such as: restructuring the organization, merger,
managerial reengineering, culture change etc [16].
An analytical research conducteb by Ovidiu Nicolescu [13] identify the most
frequent sources of resistance to change, which refers both to those directly involved in
changing and changing context. In Figure 2 we present the main sources of resistance to
change:

Figure 2. The main sources of resistance to change by employees

Without specifically insist further explained briefly what is each potential source of
resistance to change:
 personal convenience is a factor that is found in a certain proportion to each
person. At the level of each of us is manifested with a certain intensity tend to save
available forces, not always use them to make something new, mulţumindune with

107
 Quantitative Methods Inquires

what we have, with the current situation, even if not the best or favorable for us. The
expression of this situation is devoted to "anything goes".
 individual habits. Over time, each person has formed certain habits, resulting from
the specific personality and background conditions involved. There is a tendency not
to give up our habits and organizational changes that are involved always affects
some of our habits.
 the fear of the unknown. No matter how strong a person psychologically, how
much confidence in itself and in those around him, changing and its promoters,
always appears a sense of anxiety and fear. The stronger it is, the resistance to
change is more intense.
 own economic interests. Sometimes expected changes may cause a decrease in
meeting our economic interests in the organization - salary, bonus, bonuses, access
to machinery spaces protocol etc. Such situations are strong motivations for the
persons concerned to oppose, to "resist" change.
 lack of confidence in change and / or those who promote it. Whenever a
person involved in the change process does not trust those who promote or does not
believe in its success, it will manifest itself, consciously or unconsciously, a certain
resistance. To prepare people for change and promoting its prestige and possessing
the ability to exchange helps eliminate the inhibitor of change.
 the risks involved in change. When a person certain risks associated with the
expected change in personal, group or organization, even if its promoters trusts and
the end result, he will show some restraint or opposition to engage in change.
 loss of power and / or reducing personal prestige. Such motivation to resist the
change applies particularly to managers and specialists, people in formal or informal
power and prestige are intrinsic components of their work. Naturally, when I see that
the change envisaged will diminish their power and prestige, they will be tempted to
block this change.
 incompetence. Organizational change always causes changes in different
proportions in employee tasks and how to do. In situations where employees do not
have the knowledge to achieve them, it is likely that these changes seek to avoid or
to reduce as much.
 disrputions on networking system Disruption of the person within the
organization. Each employee is integrated in a micro office in the organization,
being in some work and personal relationships with other people. When the
employee is satisfied with it, and the change will affect the relational context and
position within it, it will tend not to get involved and do not favor this change.
 different perceptions of change. Presentation by the managers of change that will
achieve is not always perceived in the manner intended by them. The employees
who develop different perceptions of the objectives, content, implications and effects
of change, is likely not generate the same motivation for change sometimes occur
even motivations antischimbare generating passivity or even resistance to their
implementation.
 conservative personality. A proportion of the population in any country, is
characterized by native tendencies to avoid new, the lock, excessive cantonându the
past and present. The ability to take risks, tolerance for ambiguity inherent in
innovation, resistance to stress are reduced. Employees who fall into this category -

108
 Quantitative Methods Inquires

and they are not few - will always tend to block change, or at least not to get
involved in their operationalization. They must apply special treatment, especially for
strategic change, large-scale.
 inadeqacy of change forces. As noted, in any organization there are forces that
resist change generated by previous factors. Countering their organizational level is
done by generating forces that promote and encourage change, higher premiums. If
not done this superiority, perceived by employees and other stakeholders, their
resistance to change will be more intense.
 lack of leadership. Multiple internal sources, intrinsic resistance to change, you
have listed, can be removed and / or substantially diminished when those people
show their impact on a strong leader influential promoter of innovations consistently.
Whenever there is such a leader, employees will manifest insufficient responsiveness,
passivity and even resistance to expected changes. The leader is a driving force for
successful change.
 organisational culture. Although it is an external factor in relation to persons
involved in changing organizational culture strongly influences their attitude towards
change. Companies that possess organizational culture focused on innovation,
effort, team spirit, obtaining performance from employees will induce a favorable
attitude change, thus diminishing the explicit and implicit resistance to change.
Naturally, this factors are not exhaustive, but only a selection of the most intense
and frequent, occurring in firms in general, including those in Romania (Ceptureanu,
2010).Resistance to change is a natural psychological reaction caused by the action of any of
the factors listed above. People always need a certain level of stability and safety, and the
change involves a new situation of uncertainty that causes a feeling of uncertainty and
therefore it is likely that employees feel vulnerable in several respects (risk taking,
committing mistakes, s. a.).

2. Distrust
7. Acceptance

Self-
estee 6.Search of direction
m
level
3. Depression
5. New experiments

1. Immobilization
(shock, fear) 4.Acceptance of reality (disclaimer)

Time
Figure 3. Changes suffered by self-esteem during transitions [5]
Note: 1,2,3, - negative reactions to change; 4- neutral reaction; 5,6,7 - positive reactions to change

Very few people are prepared to give up ideas for your loved obvious risks. Difficult to give
up something very specific for the human being and it happens because it seems quite

109
 Quantitative Methods Inquires

dangerous to give a firm foothold and you head into the unknown (Ceptureanu, 2012).
Every instinct of human logic, emotion of self-preservation and oppose this action extremely
risky. From the point of view of psychology whose criteria do not necessarily reflect those of
logic, these events are easier to understand. The vast majority of people under risk losing the
flexibility of thinking. Preventing and resolving resistance to change depends on the ability to
understand the reactions of individuals in such situations vary depending on a variety of
criteria: mentality, character and culture. Thus, some want new and are pleased
transformation, while others feel fear and exhibit resistance to loss of the status quo. It is
possible that ambivalence to get more complex aspects: people may welcome the change
and at the same time, to show resistance to its implementation (Masssa, 2008).
When reacting to a significant change in people, according to L. Clarke follows a
predictable pattern of response - was called "transition curve" (Figure 3.) Showing an
individual's reaction to change in a period of time.
As we see, the beginning of the transition process that involves changes are
negative aspects related to the perception of change, followed by adjustment period, which
lasts differently to different people, depending on the individual flexibility.
According to the American consultant J. Kotter [9] differ tangled emotions that
occur change as anger, pessimism, arrogance, pride, cynicism, panic, fatigue, distrust, anger
and emotions that help to achieve that change: positive trust, optimism, results orientation,
satisfaction the positive results achieved, incentives, concern, excitement, hope. Also the
author emphasizes, in particular, the need to act on emotions cause people to change itself
and later to change things can change. Emphasizing and arguing prevailing social aspects of
change for a successful outcome researcher suggests the following method of working with
people: SEE -> FEEL -> CHANGE, i.e. employees must be shown opportunities and threats
in a convincing manner and particularly the EU would achieve it aware of the need for
change and actually achieve it (Lester, 2001).

Massive
Change is considered
nenecessary
Change is considered
Accept unfeasible
Acceptance
ance of
of change
change

Level of Low
acceptance
Narrow Average Large
Dimensions of change

Figure 4. Dependence between dimensions of change and levels of acceptance [8]

110
 Quantitative Methods Inquires

According to G. Johns [8] in general, there are two reasons that "justify" change
(Figure 4):
1. The change is not necessary because there is only a small discrepancy between the
condition of the moment and the ideal state of the organization;
2. The change cannot be achieved because there is a discrepancy between the present
state too large of things or requirements.
As we see in the figure as the size change (their size or depth) is higher, the change
is more disagreeable, and the same reaction when forming its dimensions are small. Indeed,
it is hard to convince and to convince others that "good is the enemy of good" and that
perfection has no limits (Prusak, 2007). The middle is the only approved: the magnitude of
change coincides with the needs, desires for change and the potential that we can use.

2. Change management research on innovative Romanian SMEs

2.1. Sample size and structure of SMEs

To analyse the trends, motives and peculiarities of change management in ITC
innovative Romanian SMEs, we use a survey database that was collected by Romanian
National Trade Registration Office- main legal entity with function of keeping the register of
trade. The survey targeted SMEs, defined as enterprises with 1-249 employees, and also
large companies and was implemented by means of computer-assisted telephone
interviewing. Data collection was done over a 2 month period during September- October
2014. To reliably identify trends only respondents with long tenure and representing
enterprises that systematically innovate and implement change, were selected. The survey
therefore started with screening questions. Respondents first indicated if their company had
developed implementation of change management processes and at least one innovation in
the past year. This could either be a product, process-, organizational- or marketing-related
innovation as defined by the Oslo manual (a set of integral guidelines for the collection of
innovation data, see OECD, 2005). Secondly, respondents had to be involved at least in one
implementation of change management process during the last 5 years. In this way, the
screening ensured that respondents all represented SMEs with systematic efforts in change
and they were in a position to adequately judge if and how change processes had developed
over the past years. The sample was represented only by representatives of ITC domain
(generate by difficulties to identify innovative SMEs on Romanian economy) and
disproportionally stratified across four size classes (0–9, 10-49, 50-249 employees) (official
EU classification of SMEs) and > 250 employees. Enterprises with less than 10 employees
(micro-enterprises) were not excluded since they generally have limited identifiable
innovation activities and this population usually contains many start-ups who are very
innovative in order to survive on the market. Interviewers explicitly asked for those who were
responsible for implementation of change, i.e. small business owners, general managers or
staff managing new business development activities.
Distribution by Romanian counties
No. Counties Number of companies
1 Alba 1
2 Arad 3
3 Argeș 4
4 Bacău 1
5 Bihor 16

111
 Quantitative Methods Inquires

Bistrița-
6 Năsăud 2
7 Botoșani 2
8 Brașov 32
9 București 435
10 Buzău 2
11 Călărași 1
12 Cluj 83
13 Constanța 12
14 Covasna 2
15 Dolj 10
16 Galați 5
17 Giurgiu 2
18 Gorj 2
19 Harghita 2
20 Hunedoara 3
21 Iași 39
22 Ilfov 20
23 Maramureș 7
24 Mureș 13
25 Neamț 4
26 Olt 1
27 Prahova 19
28 Satu Mare 7
29 Sibiu 25
30 Suceava 3
31 Timiș 54
32 Tulcea 1
33 Vaslui 2
34 Vâlcea 4
Total 819

Given the age of SMEs (Figure 1), most of the companies that were the subject of
research were older the 10 years (47%), followed by enterprises between 6-10 years (33%)
and those established in the last 5 years (20%).

0-5 years 6-10 years over 10 years

161 273 385
20% 33% 47%

Analyzed companies by age

20%
47% 0‐5 years
6‐10 years
33%
over 10 years

Source: own research

112
 Quantitative Methods Inquires

Considering the size of the organizations, as shown in Figure 2, small enterprises

represents 50% of the SMEs surveyed, microenterprises account for 27% and midsize
companies have a rate of 19% .We also consider a sample of 4% of large companies in
order to simulate accurately the conditions of Romanian economy.

Dimension of analyzed companies (employee 
criteria)

50‐249 over 250 
0‐9 employees
employees
employees
19% 4% 27%

10‐249 
employees
50%

0‐9 employees 10‐249 employees 50‐249 employees over 250 employees

0-9 employees 10-249 employees 50-249 employees over 250 employees

223 408 154 34
27% 50% 19% 4%
Source: own research

As regards the legal form of SMEs, 99% of companies are private companies
limited by shares and 1% public limited companies. See Figure 3.

Sample delimited by type of 
business
Public limited 
Private 
company
company 
1%
limited by 
shares
99%

Source: own research

Type of business
Private company limited by shares Public limited company
808 11
99% 1%

Given the NACE codes, the structure of companies is as follows: 54.9% of

companies NACE code principal- 6201 (Activities to develop custom software (software-
oriented client), 20.9% CAEN 6202 (consultancy activities information technology), 1.2% -

113
 Quantitative Methods Inquires

NACE 6203 (management activities (administration and operation) of calculation), 9.9% -

NACE code 6209 (Other information technology service activities), 10,9% - NACE 6311
(data processing, hosting and related activities), 1,2% - NACE 6312 (activities of web portals)
and 6391 and 1% mainly operate on CAEN code 6399 and 6391 (Other information service
activities ).

SAMPLE STRUCTURE BY NACE CODES
500

400

300

200

100

0
NACE  NACE 
6201 6202 NACE  NACE 
6203 6209 NACE  NACE 
6311 NACE 
6312 NACE 
6391
6399

Figure 5. Sample structure by NACE code

Source: own research

2.2. Information about the change processes in investigated companies

Table 1. Survey variables
Variables (partial Findings
approach)
Correlation between 74,5% of respondents agree with the statement that the change provides
change and survival of better conditions for survival of the company in the medium and long term
the organization
The level of involvement Regarding organizational structures involved in the change process we
of organizational emphasis the role of Sales Department (33,3%), followed by R&D
subdivisions in the Department (31,9%) and Production Department (18,8%). Unfortunately,
process of change Management Dept. is ranked 4th.
Perception of Changes results on the market is reflected especially into creation of a
organizational changes product or service (39,2%), use of new resource
on the market (e.g. knowledge builders, T managers etc.)-34,3% or using an old
idea/product/service into a new manner (13,95)
Determinants of change The determinants of change- new ideas are represented by higher-level
managers (63,61%), changing interests of owners (59,7%), liquidity crises
and success crises (58,36%).58,24% of respondents believe that the
process of organizational change cannot be controlled completely vs.
41,75% believe that it is possible to direct organizational change.
Areas affected by the The areas highly affected by the change are represented by new products/
change services (55,31%) human resources (51,52%), organizational structure
(49,08%).
The types of change used Proactive change represent roughly 54,21% from answer opposed to
in companies analysed reactive changes-45,78%
The techniques used to 52% of respondents use techniques such as restructuration in crisis
implement organizational conditions (50,54%), managerial reengineering of BPM instruments
changes (46,15%) and organizational development (28,69%)
The success of Negative results during implementation were obtained in roughly

114
 Quantitative Methods Inquires

implemented changes 62,39% of analysed companies, while only 22,83% of respondents

were fully satisfied with the results
Role of subjects of change Only 25, 07% of respondents mentioned that mid-level managers played
the role of strategists, 57,14% were implementators and 17,79 were
passive subjects of change.
Measuring resistance to 72,28% of mid-level and high level managers have positive reactions to
change the categories of change, the remaining 39,82% saw the change as a threat
employees
Manifestations of Unfortunately 74,72% of employees show an active resistance to change
resistance to change
Frequency of using tactics Reducing resistance to change was obtained negotiation with employees
to reduce resistance to reluctant to change (21,5%), Staff training (21,2%), Providing information
change -actions of senior needed for the adaptation of change (12,85%), Managers personal
managers on change involvement in change management (18,8%), Stimulation and support in
adapting change (14,2%), Rotation posts (6,5%) and Job enrichment (5%)
Source: own research

3. Conclusions

Generally, considering the results, we find out that that:

• Resistance to change was and is a problem that faced all the organizations
investigated, and attempts to reduce resistance to change problematic went to all;
• Conduct direct actions change (implementation plan) so was a difficult for
domestic enterprises;
• Achieve quick results is only possible if it was developed a good plan of action
coupled situational management practices in situations when there were
"surprises" that it was not possible to foresee at the planning stage;
• Strengthening the change in corporate culture is an intangible result is sometimes
very difficult to get him and requiring time. Respondents recognized that this
requirement has been ignored in the past unconscious, lack of knowledge of
change management;
• Assess the results of implementing change can be achieved easily by comparing
staff to plan, analysing external and internal sources of information taking into
account the social implications of changes completed;
• A distinction is made between strategic and operational change;
• Use models to stimulate and clarify thinking about change and impacts;
• Pursue the technical, cultural, etc. - Are interdependent;
• Attention is paid to transition management, and not the final aspects of change;
• Strategies are not filled with procedures, tactics;
• Preparatory measures (changing organizational culture and conducting training
with employees) are vital Success is the approach;

References

1. Ansoff, I. and McDonell, E. Strategic Management, Palgrave Macmillan Publishing

House, 2007
2. Ceptureanu, E.G. and Ceptureanu, S.I. Model of organizational change by
reengineering in Romanian companies, IBIMA Conference, 2012

115
 Quantitative Methods Inquires

3. Ceptureanu, S.I. Knowledge management model for Romanian companies, Review

of International Comparative Management, Vol.11, No. 1, 2010
4. Ceptureanu, S.I., Ceptureanu, E.G., Tudorache, A. and Zgubea, F. Knowledge Based
Economy Assessment in Romania, Economia. Seria Management, 2012
5. Clarke, L. Managementul schimbarii: ghid practic privind producerea,
mentinerea si controlul schimbarilor intr-o organizatie, Bucharest, Teora
Publishing House, 2002
6. Ileanu, B. V., Isaic-Maniu, A. and Herteliu, C. Intellectual capital components as
causes of regional disparities. A case study in Romania, Romanian
Journal of Regional Science, Vol. 3, No. 2, 2009, pp. 39-53
7. Isaic-Maniu, A. an Herteliu, C. Ethnic and religious groups in Romania-Educational
(co) incidences, Journal for the Study of Religions and Ideologies, No. 12,
2005, pp. 68-75
8. Johns, G. Comportament organizational, Bucharest, Economica Publishing House,
1998
9. Kotter, J.P. Leading Change, Harvard Business Press, 1996
10. Lester M. Innovation and Knowledge Management, The Long View, Creativity and
Innovation Management, no. 3, 2001
11. Massa, S. and Testa, S. Innovation and SMEs: Misaligned perspectives and goals
among entrepreneurs, academics, and policy makers, Technovation, Vol.
28, 2008, pp. 93–407
12. Maurer, R. Beyond the wall of resistance, Palgrave Macmillan Publishing House, 2010
13. Nicolescu, O. and Nicolescu, C. Tranzitia organizationala si rezistenta la
schimbari, Revista Economie teoretica si aplicata, no. 7, 2006
14. Prusak L. and Matson E. Knowledge Management and Organisational Learning,
Oxford University Press, Oxford, 2007
15. Raducanu, A. M., Feraru, V., Herteliu, C. and Anghelescu, R. Assessment of The
Prevalence of Dental Fear and its Causes Among Children and
Adolescents Attending a Department of Paediatric Dentistry in
Bucharest, OHDMBSC, Vol. 8, No. 1, 2009, pp. 42-49
16. Sabau, G. Schimbarile organizationale produse prin reingineria proceselor
economice, Bucharest, ASE Publishing House, 2000
17. Senge, P. The Dance of Change: The Challenges to Sustaining Momentum in
Learning Organizations, Crown Business, 1999

1
Acknowledgements
This work was cofinanced from the European Social Fund through Sectoral Operational Programme Human
Resources Development 2007-2013, project number POSDRU/159/1.5/S/142115 „Performance and excellence in
doctoral and postdoctoral research in Romanian economics science domain”

116