You are on page 1of 10

www.ietdl.

org
Published in IET Software
Received on 18th February 2013
Revised on 24th July 2013
Accepted on 23rd August 2013
doi: 10.1049/iet-sen.2013.0029

Special Issue on Empirical Studies in Software


Engineering

ISSN 1751-8806

Externalising tacit knowledge of the systematic


review process
Sandra Camargo Pinto Ferraz Fabbri1, Katia Romero Felizardo2, Fabiano Cutigi Ferrari1,
Elis Cristina Montoro Hernandes1, Fábio Roberto Octaviano3, Elisa Yumi Nakagawa2,
José Carlos Maldonado2
1
Computing Department, Federal University of São Carlos (UFSCar), São Carlos, SP, Brazil
2
Computer Systems Department, University of São Paulo (ICMC/USP), São Carlos, SP, Brazil
3
Computing Department, Federal Institute of São Paulo (IFSP), São Carlos, SP, Brazil
E-mail: fabiano@dc.ufscar.br

Abstract: Systematic Reviews (SRs) have recently intensified in Software Engineering. However, there is a lack of work that
makes explicit how the process to perform SR is in practice. The goal of this paper is externalising the process that reflects
how SRs are currently performed, transforming tacit knowledge into explicit knowledge. We describe the phases and
activities that compose the process and the relationship among them, and explore the iterative characteristic of this process,
focusing on intra- and inter-phase iterations that are necessary to conduct the process appropriately. To achieve the proposed
goal, we devised the process based on practical experience acquired for several years by research groups in Software
Engineering, which include graduate students and researchers who have applied SR. The process has been applied in several
SRs and seems to be effective in keeping the focus of the review at all phases. Moreover, the externalisation of the process
has been useful to help researchers improving the process execution quality. As the process reflects the practice and is
explained in details, it can be used as a guide to better understand the SR process and its details. This shall contribute to
improve all SR process phases, and hence the quality of SR results.

1 Introduction Perhaps one of the factors that can be promoting these


problems is the lack of details in the description of the SR
Systematic review (SR) (also known as systematic literature process. The guidelines used in the SE area [17] were
review) was originally employed in the medical area and is created as an adaptation of several existing guidelines from
recognised as one of the key components to support the other disciplines, particularly medicine, and were described
evidence-based medicine. Inspired by the success in the at a relatively high level of details [18–20]. Although in
medical field, the SR has been addressed within a number 2007 a more comprehensive set of guidelines for
of other subjects, including economics, psychology, social performing SR in SE became available [21], it remained the
science and most health care disciplines [1, 2]. same guidelines used by medical researchers (this was
SRs, since their introduction in the field of software reinforced by several discussions with researches from other
engineering (SE), have gained substantial importance [3, 4], fields [20]).
and have been applied to various topics of interest [5–10]. Staples and Niazi [19] also mentioned that guidelines for
The reasons why more and more SRs are conducted every SR are well established in other disciplines. However, SE
year can be attributed to the advantages of SRs, including researchers have yet to come up with a well understood
reduced likelihood of bias in results and the potential ability consensus about the conduction of SR. Several researchers
to combine data for various quantitative studies such as are working on improving the scientific support for SR in
meta-analysis. SE [3, 15]. It is clear that concepts of evidence-based
Despite its importance, the SR process is still largely practice initially developed in medicine should be adopted
carried out manually, which is both challenging and in SE. However there are difficulties to use the medical
time-consuming. Based on literature, many potentially guidelines for performing SR in the SE context. A key
problematic aspects of the process are identified. issue is to better understand the role of SR in SE [1].
Examples of such problems are: (i) formulation of Kitchenham et al. [2] highlight that adapting the SR
research questions [1]; (ii) conducting searches [1, 11–14]; process to meet the needs of SE is an ongoing process.
(iii) primary study selection [1, 15]; and (iv) quality Based on this scenario, the focus of this paper is to
assessment of primary studies [1, 16]. Therefore the externalise details describing how SR have been carried out
problems and the approaches adopted to tackle these in practice. This is shown in Sections 3 and 4, with the aim
problems vary across studies. of transforming tacit knowledge into explicit knowledge.

298 IET Softw., 2013, Vol. 7, Iss. 6, pp. 298–307


& The Institution of Engineering and Technology 2013 doi: 10.1049/iet-sen.2013.0029
www.ietdl.org
Tacit knowledge refers to information that is important to documents one will retrieve; on the other hand, the more
adequately conduct SR, but is not explicit in the guidelines. precise it is, the more relevant documents will be missed.
This tacit knowledge is often crucial for the interpretation Besides this, poorly written abstracts can complicate the
of the explicit knowledge. In this paper, we synthesised study selection process [1, 24–26].
tacit knowledge in a process, named I-SR, which highlights
the need for iterations which naturally occur in an attempt
2.2.2 Sequence against iteration: Currently
to overcome the problems we are faced during the SR
documented, largely adopted SR procedures [17, 18] do not
process. We emphasise that the I-SR process results from
explicitly deal with the iterative nature of the SR process.
the experience obtained by these authors’ research groups in
Nevertheless, previous findings reinforce the cyclic
the context of local research as well as international
characteristics of some phases or subsets of activities within
collaborations. In Section 5, we also summarise the results
a given phase in the SR process. For example, in Zhang
of SRs that have been performed in the last few years with
and Muhammad’s [15] work, many interviewees mentioned
support of the process described in this paper, with special
the study selection and data extraction are done in an
attention to iterations and activities that were externalised in
iterative way, specially when novices are involved in the
our process. A description of related research and
SR. This may even lead to a revision of the inclusion and
conclusions and future work completes this paper (Sections
exclusion criteria and the data extraction form as well. In
6 and 7).
another piece of work, Brereton et al. [1] concluded that
piloting the protocol is essential to find mistakes in the data
2 Background collection and in the synthesis procedures. Obviously,
piloting means that the search will be ‘officially’ executed
2.1 Basic concepts of systematic reviews at some point in the future.
An SR is a ‘means of identifying, evaluating and interpreting
available research relevant to a particular research question, or 2.2.3 Lack of mature guidelines and checklists:
topic area, or phenomenon of interest’ [21]. Since its Staples and Niazi [19] mentioned that guidelines for SR are
introduction in 2004 in the SE field, SRs have been well established in other disciplines, particularly in
increasingly used as a method for conducting SE-related medicine. However, SE researchers have yet to come up
secondary studies [3, 4]. with a well understood consensus about how to conduct
By analogy with medicine, Kitchenham [17] proposed the SR. There are several researchers that are working on
use of SR as a methodology for performing unbiased improving the scientific support for SR in SE [3]. It is clear
aggregation of empirical results in SE. Evidence-based SE that concepts of evidence-based practice initially developed
was first proposed as a means of advancing and improving in medicine should be adopted in SE. However, there are
the discipline of (SE) [17]. SR provides a methodical and difficulties to use the medical guidelines for performing SR
structured process to support the conduction of literature in SE context. A key issue is to better understand the role
reviews [21]. of SR in SE [1]. Kitchenham et al. [2] highlight that
Brereton et al. [1] state that the research question(s) is the adapting the SR process to meet the needs of SE is an
most critical element of an SR, because they are used to: (i) ongoing process.
construct search strings; (ii) determine the data that need to The next sections externalise the iterative process for
be extracted from each primary study; and (iii) constrain conducting SR, transforming tacit, experience-based
the synthesis process. Therefore the research questions are knowledge into explicit one. The process is described
the part of the protocol that should not be changed after the together with a set of checklists to support approval
protocol is reviewed and accepted [1]. activities along the process. We also use examples extracted
from our SR to justify some design decisions made on the
2.2 Major challenges for performing systematic process, in order to make the process explicitly iterative, as
reviews well as to overcome the difficulties discussed in this section.

Several authors have reported and discussed obstacles faced 3 Process overview
by them – or by their research groups – while performing
SRs in the SE field. The major challenges are next The incremental SR process, hereafter called I-SR, is
summarised. compound of four phases, namely: (i) planning, (ii) initial
selection, (iii) final selection and extraction and (4)
2.2.1 Rigour of the search process: Zhang and synthesis. Its overall structure is depicted in Fig. 1. The
Muhammad [15] and Zhang et al. [22] highlight that the process layout is inspired in the evolutionary models for
rigour of the search process is one critical factor that software development. Such models aggregate iteration and
distinguishes SR from traditional, ad hoc literature reviews. evolution in the development process, in which the
However, effective searching is far to be a trivial task. produced software artefacts evolve after each cycle of the
Indeed, many SE researchers [1, 11–14, 16] have lately model. In I-SR, the artefacts evolve after executing each
reported difficulties to search relevant studies. Effective phase similarly. In Fig. 1, phases are each represented as a
searching strongly relies on the quality of the search string; circle that comprises a set of activities that can be done
devising it is a time-consuming and error-prone activity iteratively. The phases are briefly described in the sequence.
[22]. The performance of automated search itself is highly Their details are provided in Section 4.
dependent on the quality of the string, which implies a The planning phase comprises the definition of the SR
continuous refinement process. However, defining a search protocol, a pilot search and the design of a preliminary
string may be a ‘double-edged sword’. As emphasised by version of the data extraction form. It is the first phase of
Boell and Cezec-Kecmanovic [23], on the one hand, the the process and has the ‘define/revise protocol’ activity as
more inclusive is the search string, the more irrelevant its starting point.

IET Softw., 2013, Vol. 7, Iss. 6, pp. 298–307 299


doi: 10.1049/iet-sen.2013.0029 & The Institution of Engineering and Technology 2013
www.ietdl.org

Fig. 1 Process overview

The initial selection phase comprises activities related to update the data sheets because of novel, up-to-date research
the pre-selection of primary studies. Basically, this phase results that make previous ones obsolete.
requires the customisation of the search string for each It is important to highlight that packaging is performed in
target repository, running the searches, pre-selecting the parallel with the whole process execution. Furthermore, the
primary studies and checking the occurrence of new protocol permeates the whole SR lifecycle. Every designed
candidate keywords to enhance the search string. or produced artefact, every decision taken and every
The final selection and extraction phase encompasses procedure performed by the researchers are driven by the
activities related to the final selection of primary studies and protocol. Reporting all these items creates a package that
data extraction. As in initial selection, during this phase enables both process and results audit, as well as SR
some new candidate keywords may emerge, and hence must replication, update and extension.
be checked. Besides that, additional data extraction fields
may be necessary and thus must be considered in a revised 3.1 Supported execution scenarios
version of the data extraction form.
Finally, the synthesis phase addresses the analysis of the The I-SR process is devised to support SR in four different
extracted data, and reporting and publishing the results. scenarios: (i) original SR, (ii) SR replication, (iii) SR
Filtering is required, for example, when it is necessary to extension and (iv) SR update (the last three are called

300 IET Softw., 2013, Vol. 7, Iss. 6, pp. 298–307


& The Institution of Engineering and Technology 2013 doi: 10.1049/iet-sen.2013.0029
www.ietdl.org
‘re-execution’ scenarios. Depending on the current scenario, updating scenario in which new primary studies might be
some activities along the process must be executed in a pre-selected in the initial selection phase. Even though most
particular manner. Owing to this reason, most of the of the search engines enable the user to apply filters related
activities within the planning phase emphasise the revision to publication date (e.g. published after a given year), such
of some artefact, result or procedure. Filtering-relates tasks, filters are generally not fine-grained enough (e.g. published
which are later described in this section and in Section 4, after a particular month of the year) to avoid duplicates.
represent another example of customisable procedures based Therefore, after the researchers query all the target
on the current execution scenario. repositories, they must filter the newly gathered results to,
To illustrate how the process support re-execution in a further step, merge them with the previous round dataset.
scenarios, let us consider an update of an original SR.
During the second interaction (updating), the protocol is
revisited in order to identify required changes to make it 4 Process in details
suitable for the update. For example, the exclusion criteria
require changes to restrict the search only for studies This section details each phase of the I-SR process,
published after the first SR. After running the searches, the emphasising the points where intra-phase and inter-phase
retrieved set of primary studies should be cleaned up to iterations can occur. Checkpoints and associated supporting
avoid overlapping with the previous, original set of studies. checklists are also described for each phase.
In the sequence, the studies are selected (i.e. classified as
selected or discarded [‘Selected’ means the study fulfils the 4.1 Planning phase
inclusion criteria, and ‘discarded’ means the opposite. In
the literature, the terms ‘included’ and ‘excluded’ are also The objective of this phase is planning the SR and establish a
used.]) and all relevant information is extracted. Merging search string that is able to fetch appropriate studies. The plan
(or filtering) tasks are also required to combine previous corresponds to a protocol, for which a template designed by
and newly gathered data, as well as to avoid the presence of Kitchenham and Charters [21] has been used by the SE
obsolete evidence. Section 4 provides details on how to community so far. Two protocol fields deserve special
treat the new studies independently, but still considering the attention since they can generate intra- or inter-phase
whole set of results, and how to aggregate the extracted iterations: keywords and search string. Owing to this, they
data. Related information can also be found elsewhere [27, are drawn as separate steps in Fig. 1. Intra-phase iteration is
28]. assumed to occur because before going to next phase, the
search string must be improved until appropriate studies are
3.2 Intra-phase and inter-phase iteration levels found. Inter-phase iterations triggering the re-execution of
activities of this phase can occur every time a study
The I-SR process is iterative in its essence. Two levels of selection is done. These iterations can be observed in Fig. 1
iteration can be found within and across the phases, (see the connector labelled with ‘A’) and will be explained
hereafter named ‘intra-phase’ iteration and ‘inter-phase’ in Sections 4.2 and 4.3.
iteration, respectively. In Fig. 1, intra-phase iterations are The activities required for the planning phase are:
represented by the arrows placed on the outermost ring of
each phase as well as by the directed edges placed within 1. Define/revise protocol: all fields of the protocol must be
the innermost circles. Inter-phase iterations, on the other filled because they describe the intention of the SR, the way
hand, are represented by the dashed, directed edges that go the SR is carried out, and allow that the SR can be
from one phase to another. replicated by other researchers.
Each directed edge – intra- or inter-phase – is labelled with 2. Revise keywords: the keywords must be revised and
a guard-condition that must hold to activate the associated updated aiming at fetching appropriate studies. As they
transition. As an example, let us consider the ‘approve pilot must be improved until the researcher is satisfied with the
search’ in the planning phase. If the search results are fetched studies, the revision occurs sometimes along the
unsatisfactory (represented by the [not OK] label), the process, particularly inside this phase, as an intra-phase
researchers must re-execute the ‘revise keywords’ activity iteration. Besides, in every study selection activity (be it
and the subsequent ones, until the search results are initial or final), new candidate keywords can be spotted,
approved. Once the results are approved, be it in the first thus triggering an inter-phase iteration.
iteration or not, the researchers can naturally move to the 3. Revise search string: as the keywords are defined or
next activity, which is ‘define/revise preliminary data revised, they must be combined through logical operators,
extraction’ form in this example. generating a search string.
4. Run/revise pilot search: once the search string is defined, it
3.3 Checklists and filters must be submitted to a selected search machine. This machine
is one of the set of sources (i.e. repositories) defined in the
Some activities along the I-SR process bring together protocol, and is used in this phase just to adjust the string.
checklists to help the researcher assess the results obtained 5. Approve pilot search: the studies fetched from the search
after either partial or whole phase executions. For example, executed in previous step must be analysed aiming at
the ‘approve selection’ and ‘approve extraction’ activities in verifying if they are in accordance to the objective of the
the final selection and extraction phase are both supported SR and its main question. As the researchers analyse the
by checklists. Note that a graphical mark is included in results, they can realise that the search string can be
each activity that is supported by a checklist (see the legend improved, triggering the intra-phase iteration, going back to
of Fig. 1). activity 2. This activity is assisted by the checklist shown in
Some other activities in I-SR require data filtering. Their Fig. 2.
goals depend on the ongoing review scenario (i.e. original, 6. Define/revise preliminary data extraction form: based on
replication, extension or update). For example, be a review the SR objective, the researcher is able to define an initial

IET Softw., 2013, Vol. 7, Iss. 6, pp. 298–307 301


doi: 10.1049/iet-sen.2013.0029 & The Institution of Engineering and Technology 2013
www.ietdl.org
identified, they must be considered and will possibly trigger
an inter-phase iteration, causing a re-execution of some
steps of the planning phase.
The activities required for this phase are:

1. Adapt search string: since each search machine uses a


particular syntax for elaborating the search string, the string
defined in the planning phase must be adapted for each of
them. Note that some machines do not accept complex
strings (in terms of composition rules and numbers of
terms), and hence some categories of papers that can be
disregarded by the search string in a given machine cannot
be avoided in other machines. Dieste et al. [13] pinpoint
other important issues about search string construction
considering the most important scientific repositories in
computer science.
2. Query all repositories: after the string is adapted to each
search machine, it must be applied to every repository
Fig. 2 Checklist for pilot search approval aiming at retrieving relevant studies for the research question.

† ‘Filtering’: accomplishing this activity may require a


list of data that must be extracted. This list is essential for filter, which is applied while running any SR execution
standardising the reading of the studies. scenario. For example, in a given scenario, search results
7. Approve protocol: it consists in checking if the protocol is obtained from different repositories (e.g. Association for
fully defined and aligned with the SR goals. If the protocol is Computing Machinery (ACM) Digital Library and IEEE
approved, researchers move to the next phase. Otherwise, a eXplore) may overlap and hence must be filtered to
new planning cycle starts and every activity is either remove duplicates; documents such as proceedings
re-executed or revised to achieve protocol adequacy. The foreword must also be discarded. Another example regards
checklist shown in Fig. 3 supports this activity. the re-execution scenarios; in this case, studies that were
processed in a previous execution must be discarded.

4.2 Initial selection phase 3. Perform initial selection: all retrieved studies, from all
repositories, must be selected taking into account their titles
This phase aims to fetch studies in all repositories specified in and abstracts (and possibly other excerpts). If a study can
the protocol. Besides, it includes a first (or initial) selection of be useful for answering the research question, it must be
studies based on the reading of title and abstract (or even other selected; if not, it must be discarded. This decision is based
study excerpts, depending on the protocol definition). This on the inclusion and exclusion criteria defined in the protocol.
phase requires the application of the inclusion and
exclusion criteria defined in the protocol. During the † ‘Filtering’: accomplishing this activity may require a
selection, filters are applied to discard retrieved documents filter, which may be applied while running any SR
not of interest (e.g. studies that are retrieved from more execution scenario. For example, in an SR extension,
than one repository). Besides, if new keywords are after the primary studies are selected, one (or some) of
them may be identified as a new, up-to-date version of
studies which were previously selected. In this case,
according to Kitchenham’s guidelines [17], only the
up-to-date study will populate the dataset. Note that such
situation is likely to occur in any execution scenario, be
it an original SR or a re-execution.

4. Check new keywords: as retrieved studies are read,


different keywords that were not considered in the protocol

Fig. 3 Checklist for protocol approval Fig. 4 Checklist for initial selection approval

302 IET Softw., 2013, Vol. 7, Iss. 6, pp. 298–307


& The Institution of Engineering and Technology 2013 doi: 10.1049/iet-sen.2013.0029
www.ietdl.org
may be identified. This can happen because of different removal of duplicates (i.e. studies retrieved with the
reasons: the researcher can be novice ‘in the research area’, previous version of the search string).
or the keyword can be a new term in that area. Every time 3. Check new data extraction fields and revise data
new keywords are identified, an inter-phase iteration is extraction form: as the studies are read, and the researcher
triggered: the ‘revise keywords’ activity (planning phase) goes deeper into them, new information extraction fields
must be executed, as well as the rest of the activities of that may arise, and the data extraction form must be updated. In
phase. This aims to ensure that the search string and the the ‘refine extraction’ activity, any remaining non-extracted
protocol are consistent with the updated keywords. data will be handled.
5. Approve initial selection: this activity aims to assess the 4. Approve selection: as the SR must be done at least by two
way the initial selection is performed and its results. If the researchers, this approval involves an agreement between
selection is approved researchers move to the next phase. them. Hence, if they have different points of view with
Otherwise, a new initial selection cycle starts to achieve respect to some studies or with respect to some new data
proper selection of studies. To assist this activity, we define extraction fields, then an intra-phase iteration is triggered.
the checklist shown in Fig. 4. This generates a return to ‘perform final selection and
pre-extraction’ and all subsequent activities. The decision on
this iteration takes into account the checklist shown in Fig. 5.
5. Refine extraction: at this activity, the researchers are in
4.3 Final selection and extraction phase accordance with the selected studies and the data extraction
form. Therefore they must completely fill the form for each
The objective of this phase is to read and select all studies selected study, and revise and (possibly) complete the
that are really of interest for the SR. Based on these studies, information extracted along the other activities.
data must be extracted aiming at supporting the synthesis of 6. Approve extraction: this approval aims to check whether
the investigated theme. This phase is compound of six the extracted data are consistent across the full set of
activities and contains intra- and inter-phase iterations as selected studies and unbiased. If not, an intra-phase iteration
follows: is triggered. To assist this activity, we define the checklist
shown in Fig. 6.
1. Perform final selection and pre-extraction: all studies must
be read in full and inclusion and exclusion criteria must be
applied again. Some times, studies that were selected in the
initial selection phase are discarded in this phase. In this
4.4 Synthesis phase
case, the researchers may note that the content is not
relevant for the SR objectives after the full reading. The objective of this phase is summarising the data collected
during the final selection and extraction phase, reporting the
† ‘Filtering’: similarly to the filter described for the process and the results in a clearly way, and publishing
‘perform initial selection’ activity, this filter may be them, providing the interested community with the gathered
applied while running any SR execution scenario. For evidence. This phase is compound of three activities as
example, in a scenario that comprises either an original follows:
SR or a re-execution, when the researchers fully read a
given primary study, they may find out that such study 1. Perform (meta) analysis: the data extracted in the previous
either includes or updates the contributions and results phase should be collected, analysed and summarised
of previous studies published by the same authors. In following the procedures specified and approved in the
this case, the up-to-date study has kept the dataset and protocol. Kitchenham and Charters [21] mentioned that this
the old versions must be removed therefore. synthesis can be descriptive, but advices that quantitative
synthesis (i.e. meta-analysis) should be performed as always
2. Check new keywords: as in the previous phase, new as possible. Cruzes and Dyba [29] present guidelines on
keywords may be identified and the same inter-phase how to conduct descriptive synthesis and Brereton et al. [1]
iteration is triggered, deviating the execution to planning show guidelines of how to conduct meta-analysis in SE.
phase again. Consequently, all the activities of that phase,
as well as the initial selection phase must be re-executed. † Filtering: accomplishing this activity may also require a
Note that the filters present in the process will support the filter, which is mainly intended to be applied during SR
updates and extensions. It comprises the revision of the
data on which (meta) analysis is performed. This is
required in order to keep only valid, non-obsolete
evidence in the data containers (e.g. tables, spreadsheets
and forms).

Fig. 5 Checklist for final selection approval Fig. 6 Checklist for extraction approval

IET Softw., 2013, Vol. 7, Iss. 6, pp. 298–307 303


doi: 10.1049/iet-sen.2013.0029 & The Institution of Engineering and Technology 2013
www.ietdl.org
2. Report results: the goal of the report is to make the year of publication – was applied in the search engine,
knowledge acquainted and the evidences found in the the retrieved set of primary studies was overlapping the
primary studies explicit. During the report construction, the previous set of selected studies. This happens because the
researcher should ensure that the research questions are filters by publication date are generally not fine-grained
being answered through clear results. Journals, conference enough (for example, filters do not consider the month of
papers and technical reports are the most common formats publication, thus the whole year is considered). Therefore,
used to report the results. before moving to the next activity, the filter was applied to
3. Publish results: after preparing the report it should be the results in order to remove duplicates.
published to potentially interested parties. A way to select Another example from the ‘AO testing SR’ regards updated
adequate discussion forum for publishing the report is to versions of primary studies. In an updating round, two
take into account the same forums considered by the SR previously selected primary studies became obsolete
selected studies. because of a more recent article published by the same
authors. Such up-to-date article revisited a set of testing
criteria proposed in previous conference papers by the
4.5 Packaging authors. By applying a filter in the ‘perform initial
selection’ activity (initial selection phase), the original
Packing the SR means to group all necessary information for papers were removed from the dataset.
replicating the review and making it auditable. Hence, all During the ‘perform (meta) analysis’ activity (synthesis
artefact generated in all phases should compose such phase), we noted that research results published in a
package, as well as all procedures adopted for conducting primary study subsume previous results published by the
each activity. Any detail and any decision making should same authors. In these pieces of work, the authors
be critically analysed for not becoming tacit knowledge incrementally established a fault taxonomy for AO
causing interference on the review replicability. Owing to programs. In the first SR round, in which the two first
this, packaging is an action that must follow the whole primary studies were selected, the fault types were extracted
process as well as software quality assurance activities must and used to compose a mapping among several fault
follow the software development. Particularly, some taxonomies. In the SR update round, the last publication led
protocol fields aim to guarantee that important procedures to a revision of the map in order to include only the most
(e.g. the initial selection procedure) are registered. Hence, recent descriptions of fault types.
the quality of the protocol is very important since it is an Note that applying filters as in the described examples have
important item of the package. saved time that would otherwise be wasted, and yielded a
The report that contains the results analysis is also part of more precise set of primary studies, since either some
the package, since the transfer of knowledge acquired as a studies were already present in the previous dataset or have
result of the review is one of the goals of the review itself. been updated by the original authors. Although Kitchenham
Therefore, aiming to externalise the importance of [17] states that this overlapping of primary studies must be
packaging, it is an element that ‘involves’ all circles that avoided in an SR, particular situations in which filters
represent each phase of the process, as depicted in Fig. 1. should be applied are not documented in the existing
literature of SR process [17, 18, 28, 31, 32].
5 Evaluation
In this section, we summarise the results of systematic 5.2 Systematic review on architectures for
literature reviews we have performed in the past few years. embedded systems
For each SR herein described, we highlight the activities
described in the I-SR process that were fundamental for A secondary study was conducted to help the comprehension
achieving relevant results. When applicable, we also on how the research community has been investigating and
contrast I-SR with other processes available in the literature. describing software architectures for embedded systems
Antonio et al. [33]. The basis string for running the
5.1 Systematic review on testing of searches was (‘software architecture’ AND reference OR
aspect-oriented software model) AND (‘embedded system’ OR ‘embedded
software’). The surveyed repositories were ACM Digital
We have conducted a SR to characterise the state of the art in Library, IEEE eXplore and Scopus.
testing of aspect-oriented (AO) software [30]. This SR, During the initial selection of primary studies, a new term
hereafter called ‘AO testing SR’, has been updated four was identified – ‘real-time system’ – as frequently used in the
times with support of the I-SR process described in this context of embedded systems. This led to a revision of the
paper. It had a main goal: identifying testing approaches search string, consisting of the addition of the new term,
and fault types researchers and practitioners have been and to the re-execution of the queries in all repositories.
investigating for AO software. Surveyed repositories are The search performed with the revised string retrieved a
IEEE eXplore, ACM Digital Library and others that are total of 573 primary studies. Applying inclusion and
relevant to the computer science area. The final selection exclusion criteria reduced the set to 104 items of interest.
comprises 34 primary studies. Considering these 104 papers, 14 ones were retrieved
As previously mentioned, we have updated the SR four because of the additional term (i.e. real-time system).
times. For the updating rounds, the filters defined in the Note that adding a new term allowed the authors to identify
I-SR process were fundamental to yield a precise dataset on a more comprehensive set of primary studies than the one
which data analysis has been performed. In the updating obtained with the original version of the string. The ‘check
rounds, a filter applied during the ‘query all repositories’ new keywords’ (initial selection phase) of the I-SR process
activity (initial selection phase) helped us to clean up the addresses this addition of new key terms to the search
search results. Even though a search filter – related to the string. Similarly to the filtering activities illustrated in

304 IET Softw., 2013, Vol. 7, Iss. 6, pp. 298–307


& The Institution of Engineering and Technology 2013 doi: 10.1049/iet-sen.2013.0029
www.ietdl.org
Section 5.1, this activity was not clearly defined in the field was added to the data extraction form: ‘Which process
existing literature of SR process. phase (or activity) is supported by the technique?’, related
to ‘check new data extraction fields and revise data
5.3 Systematic review on the use of visual data extraction form activity’. As mentioned before, although
mining (VDM) to support systematic reviews Kitchenham [17] mentions that the data extraction is an
iterative activity, the activity flow designed for the I-SR
The aim of this SR, hereafter called VDM-SR, was to process allowed us to improve the data extraction form to
understand of the use of VDM techniques to support the SR collect new pieces of information and to yield a more
process, examining the types of techniques which may be complete data synthesis.
applied in each activity [34]. More specifically, the research
question was: ‘What evidence is there of VDM techniques
being applied to help in the SR process?’.
During the ‘perform final selection and pre-extraction’ 6 Related work
activity (final selection and extraction phase in I-SR),
information necessary was collected using the data Kitchenham [17] proposed a three-phase process for SR in
extraction form, which was created and approved in the SE. In the planning phase, the need for a new SR is
planning phase. We collected information related to VDM identified, so the researchers define the SR objectives and
techniques and phases/activities of the SR process to which create the protocol. Conducting the review phase is divided
VDM was applied. Although reading the studies and into two stages: (i) the ‘selection execution’ which
collecting data, we found out that most of the papers were encompasses the acquisition and analysis of primary studies
not related to SE. Consequently, we decided to collect and (ii) ‘information extraction’, in which relevant
additional information from the studies. In the ‘check new information is extracted from the included studies. Finally,
data extraction fields and revise data extraction form’ (final in reporting the review phase, the results of the primary
selection and extraction phase), we added a new field – studies that meet the SR purpose are summarised; the
‘topic area’ – to the original data extraction form. synthesis of results can be either qualitative or quantitative,
Consequently, the extraction was not approved and an that latter based on statistical analyses. Unlike the
intra-phase iteration was necessary (from ‘approve Kitchenham’s process, our I-SR process externalises the
extraction’ to ‘refine extraction’, in the final selection and intra- and inter-phase iterations that compose the SR process.
extraction phase). Kitchenham [17] also proposed guidelines to perform SR
We reviewed 20 papers and our results indicate a scarcity in SE. In addition, Biolchini et al. [18] developed a
of research on the use of VDM to help with conducting template protocol which facilitates SR planning and
SRs in the SE domain. However, most of the studies (16 of execution in SE. The I-SR process introduces five
20) have been conducted in the field of medicine and they checkpoints which are not described in those previous
revealed that the activities of data extraction and data reports. One checkpoint was established to guarantee that
synthesis have more VDM support than other activities. the search string is suitable. The pilot search must be
Although Kitchenham [17] mentions that the data evaluated and if problems are found, the researchers must
extraction is an iterative activity, she does not exemplify return to revise keywords. Similarly, if problems regarding
the consequences of the insertion or absence of fields in the protocol, selection of primary studies, extraction forms,
the data extraction form. The VDM-SR shows that the extraction procedures are found, the I-SR processor and
information acquired with the new field topic area was some of its phases and/or activities may be re-executed.
essential to conclude that there is a scarcity of research in In a previous work [28], we adapted the SR process
SE area and to signal an opportunity for future research. proposed by Biolchini et al. [18] to allow us to perform SR
updates. The adaptation concerned the update planning and
5.4 Systematic review of experimental studies in results filtering and merging. The filters proposed for the
software inspection I-SR process presented in this paper were inspired by
the filters proposed in our previous work. In spite of this,
The main goal of this study was to identify in literature which the I-SR includes several other activities and iterations not
software inspection processes were investigated through included in the mentioned processes [18, 28]. Examples are
experimental studies. We surveyed four repositories: the identification of new keywords, which implies a return
SCOPUS, IEEE eXplore, ACM Digital Library and Web of to the planning phase, the approval related activities and its
Science. The final selection comprises 79 primary studies. associated intra-phase iterations.
During the ‘perform final selection and pre-extraction’ Molléri et al. [32] showed the SR process proposed by
activity (final selection and extraction phase in I-SR), we Kitchenham [17] using the business process model notation.
collected necessary information to answer the research The authors detailed the activities that compose the process,
questions using the data extraction form. In the planning the produced artefacts, participants involved and the
phase, we expected that the techniques mentioned in the communication mechanisms. The process was formalised
studies would be related to reading techniques, which are through a group of elements, such as events, gateways and
commonly used to identify defects in artefacts. However, communication flows. The representation created by Molléri
during the pre-extraction activity, we noted that authors of et al. can be used to guide the development of
primary studies mention techniques to support other computational tools to support the SR process. Similarly,
inspection process phases or experimental process phases our work also represents the SR process. However, our
(for instance: defects classification, defects estimation and focus is not to represent the process in a formal notation,
selection of inspectors). but highlighting the iterative sequence flows of the
We figured out this information should be collected to activities. Therefore we adopted arrows, which are useful
answer the research question accordingly. Hence, a new symbols to represent iterations along the process.

IET Softw., 2013, Vol. 7, Iss. 6, pp. 298–307 305


doi: 10.1049/iet-sen.2013.0029 & The Institution of Engineering and Technology 2013
www.ietdl.org
7 Conclusions 2 Kitchenham, B., Brereton, P., Turner, M., et al.: ‘Refining the systematic
literature review process – two participant-observer case studies’, Empir.
Softw. Eng. (ESE), 2010, 15, (6), pp. 618–653
SR is an important component to identify and aggregate 3 Zhang, H., Muhammad, A.: ‘An empirical investigation of systematic
research evidence related to a specific research theme. reviews in software engineering’. Proc. Fifth Int. Symp. Empirical
Despite their relevance, there is a lack of work that explains Software Engineering and Measurement (ESEM), 2011, pp. 1–10
in details the SR process. Owing to this fact, this paper 4 Petersen, K., Nauman, B.: ‘Identifying strategies for study selection in
transformed tacit knowledge into explicit knowledge, systematic reviews and maps’. Proc. Fifth Int. Symp. Empirical
Software Engineering and Measurement (ESEM), 2011, pp. 1–10
externalising the way that the SR process has been 5 Mendes, E.: ‘A systematic review of web engineering research’. Proc.
conducted in practice. Fourth Int. Symp. Empirical Software Engineering (ISESE), 2005,
Examples of factors externalised in the I-SR process, pp. 498–507
described in this paper, were the iterations imposed by the 6 Hannay, J.E., Sjøberg, D.I.K., Dybå, T.: ‘A systematic review of theory
use in software engineering experiments’, IEEE Trans. Softw. Eng.,
SR process, which are essential for maintaining the SR 2007, 33, (2), pp. 87–10
quality. These iterations can occur at intra- and inter-phase 7 MacDonell, S.G., Shepperd, M.J.: ‘Comparing local and global software
levels, and are triggered, in general, by unforeseen effort estimation models – reflections on a systematic review’. Proc. First
situations during the planning phase. Examples are the Int. Symp. Empirical Software Engineering and Measurement (ESEM),
identification of new keywords for search strings or new 2007, pp. 401–409
8 Sulayman, M., Mendes, E.: ‘A systematic literature review of software
data extraction fields, which can be identified as the process improvement in small and medium web companies’.
researchers go deeply into the investigated theme (after the Advances in Software Engineering, vol. 59 of Communications in
initial selection of papers or data extraction activities). Computer and Information Science, 2009, pp. 1–8
Frequently, they impose updates in the protocol, which is 9 Dieste, O., Juristo, N.: ‘Systematic review and aggregation of empirical
available during all tasks of each phase, as showed in the studies on elicitation techniques’, IEEE Trans. Softw. Eng., 2011, 37,
(2), pp. 283–304
I-SR process diagram (Fig. 1). Although iterative activities 10 Salleh, N., Mendes, E., Grundy, J.: ‘Empirical studies of pair
were already mentioned by some authors [17, 18, 28, 31, programming for CS/SE teaching in higher education: a systematic
32], the I-SR process explicitly presents the moments when literature review’, IEEE Trans. Softw. Eng., 2011, 37, (4), pp. 509–525
these iterations happen and should be considered during the 11 Dieste, O., Padua, A.: ‘Developing search strategies for detecting
execution of the SR process. Note that although the relevant experiments for systematic reviews’. Proc. First Int. Symp.
Empirical Software Engineering and Measurement (ESEM), 2007,
inter-phase and intra-phase iterations have been observed in pp. 215–224
practice in the studies described in Section 5, the process 12 Bailey, J., Zhang, C., Budgen, D., Turner, M., Charters, S.: ‘Search
should be applied by other researchers aiming to obtain engine overlaps: do they agree or disagree?’. Proc. Second Int.
quantitative evidence of its adequacy. Workshop on Realising Evidence-Based Software Engineering
(REBSE), 2007, pp. 1–6
We highlighted that besides the protocol, the packaging is 13 Dieste, O., Griman, A., Juristo, N.: ‘Developing search strategies for
also present all the time, evincing that all information or detecting relevant experiments’, Empir. Softw. Eng., 2009, 14, (5),
artefact generated during the review are important for pp. 513–539
allowing review audit and replication. Particularly, there are 14 Kitchenham, B., Brereton, P., Turner, M., et al.: ‘The impact of limited
some protocol fields that must register essential procedures search procedures for systematic literature reviews – a
participant-observer case study’. Proc. Third Int. Symp. Empirical
for making the review replicable, thus not leaving important Software Engineering and Measurement (ESEM), 2009, pp. 336–345
information as tacit knowledge. 15 Zhang, H., Muhammad, A.: ‘Systematic reviews in software
Note that systematic mappings (SMs) [31] follow a process engineering: an empirical investigation’, Inf. Softw. Technol., 2013,
that is similar to an SR process. Therefore the process here 55, (7), pp. 1341–1354
presented can be easily adapted for SM conduction. In the 16 Riaz, M., Sulayman, N., Salleh, M., Mendes, E.: ‘Experiences
conducting systematic reviews from Novices’ perspective’. Proc. 14th
context of our research groups, SM has been executed Int. Conf. Evaluation and Assessment in Software Engineering
through such process and some examples used in this paper (EASE), BCS-eWiC, 2010, pp. 1–10
came from them. 17 Kitchenham, B.: ‘Procedures for performing systematic reviews, joint
As future work, we intend to apply a survey to identify SE technical report TR/SE-0401 (Keele) – 0400011 T.1 (NICTA)’.
Software Engineering Group – Department of Computer Science –
researchers’ experience in applying SR. This shall enable us Keele University and Empirical Software Engineering – National ICT
to enhance the quality of the I-SR process. Finally, we Australia Ltd., 2004
consider that this paper can be used as a guide for novice 18 Biolchini, J., Mian, P., Natali, A., Travassos, G.: ‘Systematic review in
researchers as well as for researchers who intent to software engineering: relevance and utility’. Technical report, PESC/-
understand the SR process deeply. COPPE/UFRJ, 2005
19 Staples, M., Niazi, M.: ‘Experiences using systematic review
guidelines’, J. Syst. Softw., 2007, 80, (1), pp. 1425–1437
8 Acknowledgments 20 Stapic, S., LÓpez, E.G., Cabot, A.G., Ortega, d.M., Strahonja, L.V.:
‘Performing systematic literature review in software engineering’.
Central European Conf. Information and Intelligent Systems (CECIIS’
The authors thank the financial support received from 12), 2012, pp. 441–447
FAPESP (grant 2012/02524-4), INEP (Observatório da 21 Kitchenham, B., Charters, S.: ‘Guidelines for performing systematic
Educação, project 3280) and CAPES, Brazil. The authors literature reviews in software engineering’. Technical report, EBSE
also acknowledge the support granted by CNPq and 2007-001, Keele University and Durham University, UK, 2007
22 Zhang, H., Muhammad, P., Tell, A.B.: ‘Identifying relevant studies in
FAPESP, Brazil, to the INCT-SEC (National Institute of software engineering’, Inf. Softw. Technol., 2011, 53, (1), pp. 625–637
Science and Technology – Critical Embedded Systems, 23 Boell, S., Cezec-Kecmanovic, D.: ‘Are systematic reviews better, less
Brazil, grants 573963/2008-8 and 08/57870-9). biased and of higher quality?’. Proc. 19th European Conf. Information
Systems (ECIS), Helsinki, Finland, 2011
24 Dybå, T., Dingsøyr, T., Hanssen, G.: ‘Applying systematic reviews to
9 References diverse study types: an experience report’. Proc. First Int. Symp.
Empirical Software Engineering and Measurement (ESEM), 2007,
1 Brereton, P., Kitchenham, B., Budgen, D., Turner, M., Khalil, M.: pp. 225–234
‘Lessons from applying the systematic literature review process within 25 Kitchenham, B., Brereton, O., Owen, S., Butcher, J., Jefferies, C.:
the software engineering domain’, J. Syst. Softw., 2007, 80, (4), ‘Length and readability of structured software engineering abstracts’,
pp. 571–583 IET Softw., 2008, 2, (1), pp. 37–45

306 IET Softw., 2013, Vol. 7, Iss. 6, pp. 298–307


& The Institution of Engineering and Technology 2013 doi: 10.1049/iet-sen.2013.0029
www.ietdl.org
26 Dybå, T., Dingsøyr, T.: ‘Strength of evidence in systematic reviews ComputaÇão, Universidade de São Paulo (ICMC/USP), São Carlos/
in software engineering’. Proc. Second ACM-IEEE Int. Symp. SP – Brasil, 2010
Empirical Software Engineering and Measurement (ESEM), 2008, 31 Petersen, K., Feldt, R., Mujtaba, S., Mattsson, M.: ‘Systematic mapping
pp. 178–187 studies in software engineering’. Proc. 12th Int. Conf. Evaluation and
27 Dieste, O., López, M., Ramos, F.: ‘Formalizing a systematic review Assessment in Software Engineering (EASE), BCS-eWiC, University
updating process’. Proc. Sixth Conf. Software Engineering Research, of Bari, Italy, 2008, pp. 1–10
Management and Applications (SERA), 2008, pp. 143–150 32 Molléri, J.S., Silva, L.E., Benitti, F.B.V.: ‘Proposal of an automated
28 Ferrari, F., Maldonado, J.: ‘Experimenting with a multi-iteration approach to support the systematic review of literature process’. Proc.
systematic review in software engineering’. Proc. Fifth Experimental 25th Conf. Software Engineering & Knowledge Engineering (SEKE),
Software Engineering Latin America Workshop (ESELAW), Salvador, 2013, pp. 488–493
Brazil, 2008, pp. 1–10 33 Antonio, E.A., Ferrari, F.C., Fabbri, S.C.P.F.: ‘A systematic mapping of
29 Cruzes, D., Dyba, T.: ‘Recommended steps for thematic synthesis in architectures for embedded software’. Proc. Second Brazilian Conf.
software engineering’. Proc. Fifth Int. Symp. Empirical Software Critical Embedded Systems (CBSEC), 2012, pp. 18–23
Engineering and Measurement (ESEM), 2011, pp. 275–284 34 Felizardo, K., MacDonell, S., Mendes, E., Maldonado, J.: ‘A systematic
30 Ferrari, F.C.: ‘A contribution to the fault-based testing of aspect-oriented mapping on the use of visual data mining to support the conduct of
software’. PhD thesis, Instituto de Ciências Matemáticas e de systematic literature reviews’, J. Softw. (JSW J.), 2011, 7, (2), pp. 450–461

IET Softw., 2013, Vol. 7, Iss. 6, pp. 298–307 307


doi: 10.1049/iet-sen.2013.0029 & The Institution of Engineering and Technology 2013

You might also like