See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/349086636

Fuzzy-set Qualitative Comparative Analysis (fsQCA): Guidelines for research

practice in Information Systems and marketing

Article  in  International Journal of Information Management · June 2021

DOI: 10.1016/j.ijinfomgt.2021.102310

CITATIONS READS

14 847

2 authors:

Ilias Pappas Arch G. Woodside

Universitetet i Agder Boston College, USA
105 PUBLICATIONS   2,113 CITATIONS    380 PUBLICATIONS   11,886 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Software Startup Research View project

International Journal of Business & Economics Special Issue View project

All content following this page was uploaded by Ilias Pappas on 09 February 2021.

The user has requested enhancement of the downloaded file.

 International Journal of Information Management 58 (2021) 102310

Contents lists available at ScienceDirect

International Journal of Information Management

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

Technical Note

Fuzzy-set Qualitative Comparative Analysis (fsQCA): Guidelines for

research practice in Information Systems and marketing
Ilias O. Pappas a, b, *, Arch G. Woodside c
a
University of Agder, Campus Kristiansand, Universitetsveien 25, 4630 Kristiansand, Norway
b
Norwegian University of Science and Technology (NTNU), Sem Sælands vei 7-9, 7491, Trondheim, Norway
c
Boston College, 4801 Wigley Road NE, Marietta, GA 30066, USA

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

Keywords: The increasing interest in fuzzy-set Qualitative Comparative Analysis (fsQCA) in Information Systems and
Asymmetric analysis marketing raises the need for a tutorial paper that discusses the basic concepts and principles of the method,
Best practices provide answers to typical questions that editors, reviewers, and authors would have when dealing with a new
Calibration
tool of analysis, and practically guide researchers on how to employ fsQCA. This article helps the reader to gain
Configurations
richer information from their data and understand the importance of avoiding shallow information-from-data
fsQCA
Fuzzy sets reporting. To this end, it proposes a different research paradigm that includes asymmetric, configurational-
Qualitative comparative analysis focused case-outcome theory construction and somewhat precise outcome testing. This article offers a detailed
Set-theoretic methods step-by-step guide on how to employ fsQCA by using as an example an already published study. We analyze the
same dataset and present all the details in each step of the analysis to guide the reader onto how to employ
fsQCA. The article discusses differences between fsQCA and variance-based approaches and compares fsQCA
with those from structured equation modelling. Finally, the article offers a summary of thresholds and guidelines
for practice, along with a discussion on how existing papers that employ variance-based methods are extendable
and complemented through fsQCA.

1. Introduction an outcome of interest. QCA has three main variations: crisp set QCA
(csQCA), multi-value QCA (mvQCA), and fuzzy-set QCA (fsQCA).
“Scientists’ tools are not neutral” (Gigerenzer, 1991, p. 264): both In configurational approaches, the conditions that indicate an
symmetric (e.g., correlation and multiple regression analysis) and outcome, such as user behavior or experience in IS and marketing
asymmetric (i.e., individual case outcome forecasts) data analysis tools research, are regarded as configurations of interrelated structures,
express theoretical stances as well as analytical procedures (Woodside, instead of entities that are examined in isolation (Delery & Doty, 1996;
2019). Qualitative comparative analysis (QCA) is an asymmetric data Fiss, 2007). Analyzing alternative configurations enables systemic and
analysis technique that combines the logic and empirical intensity of holistic views of IS and marketing environments. QCA is useful for
qualitative approaches that are rich in contextual information, with inductive, deductive, and abductive (Park, Fiss, & El Sawy, 2020; Sar-
quantitative methods that deal with large numbers of cases and are more idakis, Angelidou, & Woodside, 2020) reasoning and for remarkably
generalizable (Ragin, 1987) than symmetric theory and tools. This useful theory building, theory elaboration, or theory testing (Greck-
ability for bringing together basic concepts from both qualitative and hamer, Misangyi, & Fiss, 2013; Misangyi et al., 2017). The popularity of
quantitative techniques of analysis differs substantially from traditional QCA and its variations is increasing in different fields such as e-business
methods of quantitative analysis that are often variance-based and (Pappas, Kourouthanassis, Giannakos, & Chrissikopoulos, 2016), social
employ null hypothesis significance testing (NHST). QCA can identify media (Pappas, Papavlasopoulou, Mikalef, & Giannakos, 2020), infor-
logically simplified statements that describe different combinations (or mation systems (Liu, Mezei, Kostakos, & Li, 2017; Park & Mithas, 2020;
configurations) of conditions indicating a specific outcome (Ragin, Park et al., 2020), education (Nistor, Stanciu, Lerche, & Kiel, 2019;
2008b). A configuration is a specific set of causal variables with a syn- Pappas, Giannakos, Jaccheri, & Sampson, 2017), and learning analytics
ergetic nature, that serves as a screen indicating an observed outcome or and multimodal data (Papamitsiou, Economides, Pappas, & Giannakos,

* Corresponding author at: University of Agder, Norway.

E-mail addresses: [email protected], [email protected] (I.O. Pappas), [email protected] (A.G. Woodside).

https://doi.org/10.1016/j.ijinfomgt.2021.102310
Received 10 September 2020; Received in revised form 3 January 2021; Accepted 3 January 2021
0268-4012/© 2021 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

2018; Papamitsiou, Pappas, Sharma, & Giannakos, 2020). At the same (Berg-Schlosser & De Meur, 2009). QCA uses Boolean algebra and
time, standards of good practice in research have been published Boolean minimization algorithms to capture patterns of
(Woodside, 2016a, 2016b) along with more specific ones that offer a multiple-conjunctural causation and to simplify complex data structures
comprehensive and easily accessible “code of good practice” for QCA in a logical and holistic manner Ragin (1987). The use of Boolean
(Schneider & Wagemann, 2010), including textbooks that describe in algebra means that QCA has as input binary data (0 or 1), and uses
detail the method in great detail (Rihoux & Ragin, 2009). logical operations for the procedure, thus it is very important to
QCA studies are designed to combine techniques from qualitative dichotomize the variables in a useful and meaningful manner.
and quantitative approaches, making these studies inherently mixed An extension of QCA is mvQCA, which treats variables as multi-
technique applications (Ordanini, Parasuraman, & Rubera, 2014; Ted- valued instead of dichotomous (Cronqvist, 2004). MvQCA retains the
dlie & Tashakkori, 2009), that take the best attributes from both worlds. idea of performing a synthesis of the dataset and cases with the same
Qualitative inductive reasoning with data being analyzed “by case’’ and value on the outcome variable are explained by a solution, which con-
not “by variable’’ (Ragin, 2000), is combined with quantitative empir- tains combinations of variables that explain a number of cases with the
ical testing, as sufficient and necessary conditions identify outcomes outcome (Cronqvist & Berg-Schlosser, 2009). Since the method was
through statistical methods (Longest & Vaisey, 2008; Ordanini et al., introduced, a discussion emerged on the potential of mvQCA and its
2014). In most cases, QCA are useful in quantitative studies, as it allows usefulness along with csQCA and fsQCA (Thiem, 2013; Vink & Van Vliet,
the researcher to get a deep view of their data through a quantitative 2009; Vink & Vliet, 2013), with its set-theoretic status being unclear.
analysis that has also several characteristics of qualitative analysis. Case MvQCA has remained underutilized (Thiem & Dusa, 2013), compared to
studies focus on describing, explaining, and forecasting, single and the other two variations of QCA (i.e., csQCA and fsQCA).
combinatorial conditional antecedents on outcomes while variable
studies focus on the similarities of variances of two or more variables. A 2.2. FsQCA
“condition” is a point or interval range of antecedent or outcome; a
“variable” characteristic varies. Here are few examples of conditions FsQCA addresses an important limitation of csQCA, the fact that
versus variables: “Male” is a condition; “gender” is a variable. “Swedish” variables are binary, thus restricting the analysis as it cannot fully
is a condition; “nationality” is a variable. “Expert” is a condition; capture the complexity in cases that naturally vary by level or degree
“expertise” is a variable. (Ragin, 2000; Rihoux & Ragin, 2009). This restriction of csQCA is likely
The present tutorial contributes by extending current works as it i) an important reason that QCA has not been widely adopted in multiple
exemplifies the application of fsQCA, ii) argues for the need to perform contexts, including IS and marketing research. FsQCA extends csQCA by
contrarian case analysis, and iii) describes how to perform predictive integrating fuzzy-sets and fuzzy-logic principles with QCA principles
validity of the findings. To this end, we map the recommended steps for (Ragin, 2000; Rihoux & Ragin, 2009), which offers for a more realistic
the above three analyses in two flowcharts that can be found as approach since variables can get all the values within the range of 0–1.
Appendices in this article. Further, we compare fsQCA with PLS-SE and FsQCA is able to overcome several limitations of both csQCA and
discuss conceptual differences among the two methods, and finally we mvQCA, and has received increased attention recently (Fiss, 2011;
summarize some of the frequently used thresholds in fsQCA. This tuto- Ordanini et al., 2014; Pappas et al., 2016; Woodside, 2014), because,
rial focuses on quantitative research and offers a step-by-step guide, in when applied together with complexity theory, it provides the oppor-
an article format on how to employ fsQCA. We focus on fsQCA as this tunity to gain deeper and richer insight into data.
tool can address several shortcomings of symmetric-based analysis.
Additional details are provided in the next section, while we offer sug- 2.3. FsQCA and cluster analysis
gestions on how existing works can be extended by employing fsQCA
based on their findings (Appendix D). Our goal is to make fsQCA easy to Case-based techniques, such as fsQCA and cluster analysis, have been
apply by the scholarly community. Thus, we use as an example a study employed as a way of moving beyond variance-based methods (Cooper
available in the literature (i.e., (Pappas et al., 2016)) and we offer all the & Glaesser, 2011). These two techniques have similarities as they both
details on how to perform the analysis, that are not included in a typical employ multidimensional spaces and often people ask how fsQCA differs
research article. from cluster analysis and why do we need it. A main difference between
This tutorial is structured as follows. First, we present an introduc- the two methods is the kind of research questions they are able to
tion into the different types of QCA and how they differ. Next, Section 3 address (Greckhamer, Furnari, Fiss, & Aguilera, 2018). Specifically,
presents related studies that have performed fsQCA in IS and Marketing cluster analysis answers questions such as which cases are more similar
areas. Section 4 introduces complexity theory and how embracing it to each other, while fsQCA can identify the different configurations that
with fsQCA can take us beyond symmetric tests. Section 5 presents the constitute sufficient and/or necessary conditions for the outcome of
main benefits and limitations of fsQCA. Next, Section 6 presents a interest (Greckhamer et al., 2018; Ordanini et al., 2014). Depending on
detailed step-by-step guide on how to perform fsQCA while giving the the focus of the study the researcher should choose the most appropriate
basic steps that should be followed in the analysis, including screenshots method. Their differences stem from the fact that “QCA addresses the
from the software. For this a dataset from an already published paper is positioning of cases in [multidimensional] spaces via set theoretic operations
used. Section 7 presents a comparison between fsQCA and PLS-SEM while cluster analysis relies on geometric distance measures and concepts of
findings. Section 8 concludes this tutorial. variance minimization” (Cooper & Glaesser, 2011). To this end, prior
studies compare fsQCA with cluster analysis and show how fsQCA can
2. Types of qualitative comparative analysis (QCA) handle causal complexity with fine-grained level data (Fiss, 2011), or
how it can identify more solutions compared to cluster analysis (Orda-
2.1. CsQCA and mvQCA nini et al., 2014). A discussion exists in the literature regarding QCA and
cluster analysis (Greckhamer et al., 2018; Miller, 2018), and both ap-
CsQCA is the first variation of QCA. It is a tool created to deal with proaches have differences making them suitable for different types of
complex sets of binary data (Ragin, 1987). The goal of QCA is to explain studies.
causality in complex real life phenomena through “multi-
ple-conjunctural causation, which refers to “nonlinear, nonadditive, 3. Adoption of fsQCA in relevant studies
non-probabilistic conception that rejects any form of permanent cau-
sality and that stresses equifinality (different paths can lead to the same Configurational approaches are becoming more popular over the
outcome), complex combinations of conditions and diversity” past few years in different areas, with fsQCA playing a large part in this

2
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

as most studies will prefer fuzzy-set over crisp-set and multi-value QCA imply that the relation between two variables is symmetrical (high
(Thiem & Dusa, 2013). In detail, fsQCA has been employed in infor- perceived usefulness–high intention to use; low perceived use-
mation systems (Fedorowicz, Sawyer, & Tomasino, 2018; Liu et al., fulness–low intention to use). In set theory terms, the presence of
2017), online business and marketing (Pappas et al., 2016; Pappas, perceived usefulness (i.e., a condition) may lead to high intention to use,
2018; Woodside, 2017), consumer psychology (Schmitt, Grawe, & suggesting sufficiency between them. However, high intention to use is
Woodside, 2017), strategy and organizational research (Fiss, 2011; very much likely to exist even when perceived usefulness is absent,
Greckhamer et al., 2018), education (Pappas, Giannakos et al., 2017; suggesting that the presence of perceived usefulness is a sufficient but
Plewa, Ho, Conduit, & Karpen, 2016), data science (Vatrapu, Mukka- unnecessary condition for intention to use a system. Further, in a
mala, Hussain, & Flesch, 2016) and learning analytics (Papamitsiou different context and when other conditions exist (e.g., high perceived
et al., 2018; Sergis, Sampson, & Giannakos, 2018). This tutorial aims to benefits) perceived usefulness may be necessary but insufficient condi-
increase the adoption of fsQCA in IS and marketing studies following the tion for intention to use to occur. Also, sometimes high perceived use-
call for more empirical work in the area (El Sawy, Malhotra, Park, & fulness may lead to high intention to use only when a third condition is
Pavlou, 2010; Fiss, 2011; Woodside, 2014, 2017). present or absent (e.g., high or low/medium perceived ease of use).
FsQCA is useful for both inductive and deductive reasoning for the- As fsQCA is based on fuzzy-sets, the tool enables capturing conditions
ory building, elaboration, and testing (Greckhamer et al., 2018; Park that are (1) sufficient or necessary to explain the outcome and (2)
et al., 2020). Researchers can either explore all the possible solutions insufficient on their own but are necessary parts of solutions that can
that explain the outcome of interest [e.g., (Pappas et al., 2016)] or test explain the result. These are called INUS conditions; insufficient but
specific models and relations (Pappas, 2018), based on theory or pre- necessary part of a condition which is itself unnecessary but sufficient
vious findings. Nonetheless, it is possible to perform both, first to for the result (Mackie, 1965). Such conditions may be present or absent
identify all possible solutions and then test specific propositions (Pap- in a solution, or they may be conditions for which we “do not care”. The
pas, 2018; Pappas et al., 2020). This analysis allows to identify specific “do not care” situation indicates that the outcome may either be present
cases in the sample, who and how many users, that verify the specific or absent and it does not play a role in a specific configuration. Neces-
propositions, as well as indicate other alternative models that can sary and sufficient conditions may be present (or absent) as core and
explain high score of same outcome. With this knowledge, the peripheral elements. Core elements indicate a strong causal relationship
researcher can go back to the cases and use contextual information, not with the outcome, and peripheral elements indicate a weaker relation-
including in the analysis, to further explain and discuss the findings. In ship (Fiss, 2011). Thus, using fsQCA, researchers can identify which
comparison, a typical variance-based analysis would identify a single conditions are indispensable (or not needed) for an outcome to occur,
best solution, thus limiting the results (Woodside, 2013, 2016b). and which combinations of conditions are more (or less) important than
Nonetheless, studies can compare the findings between different data others
analysis techniques to describe how different stories are hidden in the
same dataset, while it is recommended to combine fsQCA with other 5. The benefits of fsQCA – why to use it?
data analysis techniques if possible (Schneider & Wagemann, 2010).
The use of fsQCA can offer several benefits, compared to traditional
4. Complexity and configuration theories methods of analysis. To capture combinations of conditions that are
sufficient for an outcome to occur, fsQCA uses both qualitative and
Relations among variables are naturally complex, sometimes non- quantitative assessments and computes the degree in which a case be-
linear, and sudden changes can cause different results and outcomes longs to a set (Ragin, 2000; Rihoux & Ragin, 2009), thus creating a
(Urry, 2005). Variance-based approaches assume that relations among bridge between qualitative and quantitative methods. FsQCA uses cali-
variables are linear, and one way to overcome this is to examine complex brated measures, as data are transformed into the [0, 1] range. Cali-
phenomena as clusters of interrelated conditions (Woodside, 2017). This bration is common in natural sciences but not so much in social sciences
offers a step towards a holistic and simultaneous understanding of the and can be used to satisfy qualitative researchers’ in interpreting rele-
patterns these conditions create, by employing a configuration theory vant and irrelevant variation as well as quantitative researchers’ in
approach (El Sawy et al., 2010). precisely placing cases relative to one another (Ragin, 2008b; Vis,
As a destination may usually be reached through different routes, an 2012).
outcome may occur through different ways, thus explained by different In line with the focus of this tutorial on quantitative methods, the
combinations of antecedent conditions. Complexity theory and config- main benefits of fsQCA occur when compared to typical variance-based
uration theories have inherent the principle of equifinality, which is the methods and the limitations that the latter have (El Sawy et al., 2010;
premise that multiple combinations of antecedent conditions are equally Liu et al., 2017; Woodside, 2013, 2014). In general, variance-based
effective (Fiss, 2007; Von Bertalanffy, 1968; Woodside, 2014). methods examine variables in a competing environment as they
Numerous factors can influence user experience with any information compute the net effect between variables in a model, while fsQCA fo-
system and, in general, different combinations of these factors can cuses on the complex and asymmetric relations between the outcome of
explain their adoption or use, as well as different levels of the same interest and its antecedents. For example, in typical IS/IT adoption
factors. This means that not all factors (or antecedents) are needed to studies, following the behavioral science paradigm, variables that are
explain adoption or usage, and it is likely that some of them when considered as control variables (e.g., gender, experience) can be part of
combined, can be sufficient to explain high adoption or usage. None- the solutions and combine in ways that explain different parts of the
theless, in some cases, a factor can be indispensable for high adoption or sample. An outcome may be the result of a variety of combinations, and
usage to occur. each combination contributes independently to it. In addition, as we
Configuration theories are based on the principle of causal asymme- seek to design systems that take into account “all users” (i.e., users with
try, based on which a condition (or a combination of conditions) that different requirements), we need methods that enable researchers to
explains the presence of an outcome can be different from the conditions compute multiple solutions for multiple type of users and not the vast
that lead to the absence of the same outcome (Fiss, 2011; Ragin, 2008b). majority explained by the best solution of a regression analysis.
For instance, high perceived usefulness can lead to high intention to use FsQCA is employable on different sample sizes ranging from very
a system, while the low perceived usefulness may not lead to low small (<50 cases) to very large (thousands of cases). The sample size
intention to use a system, typically due to the existence of other con- offers different options to the researcher, either to go back to the cases
ditions. Although such an assumption seems common, when we use and interpret them separately, or identify patterns across many cases
variance-based approaches (e.g., correlation, regression) the findings without returning to the cases (Greckhamer et al., 2013). Further, fsQCA

3
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Fig. 1. Basic steps in fsQCA.

is useful for different types of data (e.g., Likert-scale, clickstreams and seen as outliers. Thus, in fsQCA thus the representativeness of the
multimodal data), as long as the researcher is able to transform them sample does not affect all solutions (Fiss, 2011; Liu et al., 2017), making
into fuzzy sets. Also, fsQCA is combinable with categorical variables (e. it in a way more robust than variance-based methods as it is not sensitive
g., gender) which do not have to be transformed into fuzzy sets. In such to outliers. Testing for contrarian cases and examining the distribution
cases, some variables would be binary (0/1) while some others would of the sample prior to employing fsQCA helps to identify outliers as well
cover all values in the [0,1] range. as to get an indication there are many cases in the sample that are not
Data analysis with fsQCA leads to combinations of independent explained by the main effects (see details on contrarian case analysis in
variables that also include variables that are not identified by typical Section 6.2). Contrarian case analysis is absent typically in research
variance based approaches since the capture main affects only (Wood- articles.
side, 2014). Such variables influence the outcome only for a small FsQCA requires that the researcher has accurate workbench knowl-
number of cases. FsQCA splits the sample into multiple subsets, thus edge both of the examined variables (conditions and outcome) as well as
examining multiple combinations of conditions. Each configuration of the underlying theory and context. Such knowledge is used
represents only a subset of the sample and an outlier will be present in throughout the analysis; (1) data calibration (i.e., transforming variables
only some of the possible solutions. In other words, some solutions are into fuzzy sets), (2) simplifying the multiple solutions, (3) interpreting
likely to explain large parts of the sample, in accordance with a the results. Researchers should make decisions at the different stages
variance-based analysis, while some solutions would explain smaller based on their knowledge that is typical in qualitative analysis. This
parts of the sample as they would include cases that would be typically action is both a limitation and strength of fsQCA. While it can introduce

4
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

subjective bias into the study, researchers’ own knowledge and under- relations between constructs. Reliability testing, based on the Cronbach
standing of the field and research problem can lead to a richer analysis alpha indicator, showed acceptable indices of internal consistency since
and understanding of the data. In a traditional mixed-methods all constructs exceed the cut-off threshold of 0.70. The AVE for all
approach, a researcher would employ a quantitative study (e.g., constructs ranged between 0.55 and 0.84, all correlations were lower
analyzing questionnaires, clickstreams, log files) and then employ a than 0.80, and square root AVEs for all constructs were larger than their
qualitative analysis (e.g., interviews with key participants) to gain a correlations. The findings in detail for the confirmatory analysis may be
richer understanding of their responses and see connections and patterns found in the original paper (Pappas et al., 2016).
that are by nature hard to capture. FsQCA is not intended to measure the
unique contribution of each variable to the overall observed data; 6.2. Contrarian case analysis
instead, its objective is to identify complex solutions and combinations
of independent variables. Contrarian case analysis is performed outside fsQCA, but we present
Recently, fsQCA has been used to analyze quantitative data (e.g., it here because it can serve as an easy and quick way to examine how
(Pappas et al., 2016; Vatrapu et al., 2016; Woodside, 2017)), while many cases in our sample are not explained by main effects, and thus
calibrating qualitative data into fuzzy-sets is also possible (Basurto & they would not be included in the outcome of a typical variance-based
Speer, 2012; Henik, 2015). Also, QCA has been employed in approach (e.g., correlation or regression analysis) (Woodside, 2014,
mixed-method studies (Cairns, Wistow, & Bambra, 2017) or to bring 2016a). Existing studies perform a contrarian case analysis (Pappas
together quantitative and qualitative data (Kane, Lewis, Williams, & et al., 2016), however, despite its usefulness, many studies employing
Kahwati, 2014) in different fields. FsQCA provides a novel way of fsQCA do not report tests for contrarian cases. Indeed, when examining
analyzing current datasets, can bring together quantitative or qualita- main relations between two variables, we typically test if a variable
tive data, and broaden our methodological approaches and data ana- positively or negatively affects another variable, which means that most
lyses. Despite its benefits, configurational analysis has limitations that cases in a sample verify this relationship. However, it is likely that the
need to be considered when employing fsQCA (Liu et al., 2017; Mendel opposite relationship exists for some of the cases in the sample. The
& Korjani, 2012; Woodside, 2014). Furthermore, best practices have existence of such cases can be identified through a contrarian case
been proposed and should be taken into account when employing QCA analysis (Woodside, 2014), since contrarian cases occur regardless of the
(Greckhamer et al., 2018; Schneider & Wagemann, 2010; Woodside, significance of the main effects.
2016b). It should be noted that while fsQCA offers increased flexibility To perform a contrarian case analysis, first, the sample needed to be
to the researcher when it comes to data analysis, and it can be used divided in order to investigate the relationships among the examined
either for exploratory or confirmatory purposes, researchers should not variables. To do so we split the same by using quintiles (i.e., dividing the
employ it in a mechanistic way simply following existing guidelines and sample into five equal groups). Other splitting methods of continuous
best practices. Conscious decisions should be made during the analysis, variables, such as median split, should be avoided because this may lead
which should be reported to show the validity of the study and enable to a reduction of statistical power as well as to false results when the
replicability. variables are correlated (Fitzsimons, 2008). Next, we performed
cross-tabulations across the quintiles, which crosstabs allows us to
6. How to use fsQCA – Step by step example of employing fsQCA compute the degree of association between the variables, which suggests
in a typical e-commerce study a dependence between the two variables and describes main effects
between them. The result for any two variables is a 5 × 5 table that
For those IS and marketing researchers who are not familiar with presents all combinations for all of the cases in the sample between the
fsQCA, this section presents a sample step-by-step analysis of a typical two variables (Fig. B4) as shown in Appendix B. The top left and bottom
dataset that uses the technique discussed in this article. For this tutorial right cases represent the main effects (e.g., degree of association), while
we use the data and findings from a recently published study (Pappas the bottom left and top right represent cases not explained by the main
et al., 2016). We present more details in every step of the process, as well effects. The latter are contrarian cases existing in our sample.
as the guidelines on how to perform fsQCA. Finally, we have visualized All details on how to perform a contrarian case analysis are given in
the recommended steps for employing fsQCA using fsQCA software in Appendix B. Next, the results for the contrarian case analysis for all
Fig. 1. variables are presented in Appendix C, as it appears on the original study
of Pappas et al. (2016). The findings show the existence of various re-
6.1. Basic information about the study from the prior published paper lationships between the variables, separate from the main effect, sup-
porting the need to perform a configurational analysis.
To offer the necessary context for the study, in this section, we offer
the basic information regarding the goal and research design. The study 6.3. Calibration
examined cognitive and affective perceptions as antecedents of online
shopping behavior in personalized e-commerce environments. We used 6.3.1. Data treatment
a typical a snowball sampling methodology to recruit participants and The most important step in fsQCA is data calibration. Most types of
controlled for respondents’ previous experience with both online data can be used (e.g., survey responses, clickstreams, user performance
shopping and personalized services. Eventually, the sample comprises data, and physiological data). When a variable or construct is measured
582 individuals with experience in online shopping and personalized with multiple items, we need to compute one value per construct that
services. We collected data through a questionnaire built with measures will be used as input in fsQCA. In other words, for each case (row) in our
that have been adopted from the literature. Appendix A (as presented in dataset we need one value for every construct (column). The simplest
the original study) lists construct definitions, the questionnaire items way is to compute the mean of all the items in order to come up with one
used to measure each construct, along with descriptive statistics and single value per case (such as when testing correlations test). Nonethe-
loadings. less, there are more ways to do this while taking into account the indi-
Typical with similar quantitative studies, first we evaluate constructs vidual effect of each item on the construct itself (DiStefano, Zhu, &
for reliability and validity. This is a step that should be always per- Mindrila, 2009).
formed when it is appropriate, and it is not directly related with the Furthermore, fsQCA does not test for construct reliability and val-
fsQCA analysis as it depends on the type of variables that are used in the idity as these tests refer to the measures and not the method of analysis.
study. Construct reliability and validify, as the name implies, refer to the If the constructs used in a study need to be tested for their reliability and
construct itself and not to the method of analysis used to examine validity, then this is done prior to fsQCA analysis, following the

5
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Fig. 2. Compute thresholds using percentiles.

traditional methods, and must be reported accordingly. that a case with a fuzzy membership score of 1 is a full member of a fuzzy
In fsQCA, different from traditional methods, instead of working set (fully in the set), and a case with a membership score of 0 is a full
with probabilities data are transformed from ordinal or interval scales non-member of the set (fully out of the set). A membership score of 0.5 is
into degrees of membership in the target set, which shows if and how exactly in the middle, thus a case would be both a member of the fuzzy
much a case belongs into a specific set. “In essence, a fuzzy membership set and a non-member, and is therefore a member of what is known as
score attaches a truth value, not a probability, to a statement” (Ragin, the intermediate set. The intermediate-set point is the value where there
2008a). For example, the variable intention to purchase can be coded as is maximum ambiguity as to whether a case is more in or more out of the
“high intention to purchase”, and we will be looking for the presence or target set.
absence of the condition high intention to purchase (“intention to pur- Data calibration may be either direct or indirect. In the direct cali-
chase” is the variable; “high intention to purchase” is a condition). bration the researcher needs to choose exactly three qualitative break-
Similarly, we code the rest of the variables. The method computes the points, which define the level of membership in the fuzzy set for each
presence of a condition or its opposite (i.e., negation). The negation of a case (fully in, intermediate, fully out). In the indirect method, the
condition is referred in the literature as the absence of a condition, and measurements need to be rescaled based on qualitative assessments. The
the two terms have been used interchangeably based on how the absence researcher may choose to calibrate a measure differently, depending on
is computed (Fiss, 2011; Pappas, 2018; Ragin, 2008b). The term absence what one is investigating. Either method may be chosen, depending on
has been also used to describe when the condition is irrelevant in a researcher’s substantive knowledge of both data and underlying theory
configuration (Nagy et al., 2017; Woodside, 2017), similar to the “do not (Rihoux & Ragin, 2009). The direct method is recommended and is more
care” term that is also often used in the literature (Fiss, 2011; Pappas common, in which the researcher sets three values corresponding to
et al., 2016). This distinction is not often addressed or clarified (Pappas, full-set membership, full-set non-membership, and intermediate-set
2018), thus we suggest researchers to clearly define these terms in future membership. This can lead to more rigorous studies which are easier
works to avoid misunderstandings. to be replicated and validated, since it is clearer on how the thresholds
have been chosen.
6.3.2. Transform data into fuzzy-sets
In fsQCA we need to calibrate our variables to form fuzzy sets with 6.3.3. Choosing thresholds for direct calibration
their values ranging from 0 to 1 (Ragin, 2008b). Consider a fuzzy set as a To calibrate the data, we can choose the values 0.95, 0.50, 0.05 as
group, then the values from 0 to 1 define if and at what amount a case the three thresholds (or breakpoints), which will transform the data into
belongs to this group. The fact that all values range from 0 to 1 means the log-odds metric with all values being between 0 and 1. We do not use

Fig. 3. The dataset after calibration.

6
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

the researcher to select the three thresholds. For example, if the data do
not have a normal distribution but instead are skewed then the 80 %, 50
%, and 20 % can be set as the thresholds for full-set membership,
intermediate-set membership, and full-set non-membership, respec-
tively (Pappas, Mikalef, Giannakos, & Pavlou, 2017). In any case, the
choice of thresholds should be justified and reported accordingly, along
with a table that presents the original values that correspond to each
threshold (Fig. 2).
Especially in the case of the widely used seven-point Likert scales
(1=Not at all, 7=Very much), previous studies suggest that the values of
6, 4, and 2 can be used as thresholds (Ordanini et al., 2014; Pappas et al.,
2016). Similarly, for a five-point Likert scale the thresholds could be 4,3,
and 2. For the example presented in Fig. 2, a seven-point Likert scale was
used. We note that, for most variables, the percentiles give the same
Fig. 4. Selecting variables and outcome for fsQCA. values as if we had chosen directly the values of 6,4, and 2. However,
this is not the case for all variables. In detail, Weak_Negative emotions
exactly 1 and 0 as breakpoints because the two membership scores have the lowest scores overall, and the 95 %, 50 %, and 5 % are the
would correspond to positive and negative infinity, respectively, for the values 4.33, 2.00, and 1.00. Since this is a construct measured with a
log of the odds (Ragin, 2008a). To find which values in our dataset seven-point Likert scale, if we use the 95 % it means that users with
correspond to the 0.95, 0.50, and 0.05 we use percentiles. The percen- scores 4.33 or higher fully belong to the set, which is high values of weak
tiles allow the calibration of any measure regardless of its original negative emotions. However, this would be an inaccurate representation
values. In detail, we can compute the 95 %, 50 %, and 5 % of our of those cases, as users that choose 4 or 5 are nearer to the neutral point
measures and use these values as the three thresholds in fsQCA software. rather than to the higher point in the scale. Thus, using the thresholds 6,
This can be done easily, for example in SPSS using the “Percentiles” 4, 2 provides a more accurate representation of our sample.
function (Frequencies > Statistics > Percentiles) and choosing the
desired percentages (Fig. 2). Nonetheless, the thresholds should not be 6.3.4. Calibrating the data in fsQCA software
chosen mechanically but the researcher should understand what it Once we have decided the thresholds, we proceed to the data cali-
means for example that 5 % defines being fully out of the set. This may bration in fsQCA software (version 3.0 is used here) (Ragin & Davey,
mean that the threshold should be changed or adjusted, since it is up to 2016). The dataset file needs to be in “comma-separated values” (.csv)

Fig. 5. Truth table in fsQCA.

7
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Fig. 6. A sorted truth table in fsQCA based on raw consistency after removing combinations with low frequency.

format to be able to open it in fsQCA. Calibration is performed by using for each possible combination), while several lines will have a frequency
the Calibrate function of the software, which takes as input the variable of zero, which means that none of the cases in the sample are explained
that will be calibrated and the three breakpoints (from the highest to the by them. As the number of variables in an analysis increases, the number
lowest values) (Fig. 3). It should be noted that the researchers may use of possible configurations increases exponentially (2k), thus the more
other software for the calibration part, and that it is not mandatory to variables the more combinations are likely to have a frequency of zero.
calibrate all values following a logistic function; instead, other mem- Including more variables in the analysis might benefit by having also a
bership functions (linear or non-linear) may be used (Mendel & Korjani, larger sample, which is common in typical quantitative analysis (e.g.,
2012). Besides, the fsQCA software a package for R exists as well (Thiem MRA, SEM).
& Dusa, 2013). Next, the truth table needs sorting by frequency and consistency
In fsQCA, the cases that are exactly on 0.5 are dropped from the (Ragin, 2008b). This is done using the options in the Sort menu. Since
analysis which makes it difficult to analyze the conditions that are set frequency describes how many cases in the sample are explained by a
exactly on 0.5 (i.e., intermediate-set membership) (Ragin, 2008b). To configuration, to ensure that a minimum number of cases is obtained for
overcome this, Fiss (2011) suggests adding a constant of 0.001 to the the assessment of the relationships, a frequency threshold is set (i.e.,
causal conditions below full membership scores of 1. This can be done by column number) (Fig. 5). A higher frequency threshold means that each
adding 0.001 in all conditions after the calibration has been performed. configuration refers to more cases in the sample, but as a result will
Once all variables have been calibrated the dataset includes both reduce the percentage (i.e., coverage) of the sample that is explained by
versions of each variable (Fig. 3) and we can proceed to the next step, the retained configurations. On the other hand, a lower frequency
which is running the fuzzy-set algorithm and the generation of the truth threshold increases the coverage of the sample, although each combi-
table. nation refers to fewer cases in the sample. For samples larger than 150
In order to run the fsQCA algorithm choose “Analyze > Truth Table cases the frequency threshold may be set at 3 (or higher), while for
Algorithm”, and at this point the researcher has to select the variables smaller samples the threshold may be set at 2 (Fiss, 2011; Ragin, 2008b).
that will be included in the analysis (Figure 4). In detail, the causal As our sample is 582, the threshold is set at 3, and all combinations with
conditions are the independent variables, and the outcome is the smaller frequency are removed from further analysis.
dependent variable. At this point, the researcher may choose to either Once removing configurations with low frequency, the truth table
compute the presence of outcome or negation of the outcome. By should be sorted by “raw consistency” (Fig. 6) At this point a consistency
clicking OK, fsQCA produces the truth table. threshold should be set, with the minimum recommended value being
The truth table computes all possible configurations (or combina- 0.75 (Rihoux & Ragin, 2009). A first indicator for choosing the consis-
tions) that may occur, providing 2k rows, with k representing the tency threshold is to identify natural breaking points in the consistency
number of outcome predictors, and each row representing every values that have been obtained, however this is not absolute. In detail, in
possible combination (Fig. 5). When computing all possible configura- Fig. 6 we note the lowest consistency values of 0.862958, 0.859605,
tions, the frequency is also presented (i.e., the number of observations 0.839190, 0.781378, 0.695669. These values indicate that both

8
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

versa (Dul, 2016).

The last step while working on the truth table is to decide if each
combination (i.e., line of the table) explains the outcome or not. Using
the consistency thresholds, the researcher has to insert the value of 1 or
0 in the column with the outcome variable. Choosing 1 or 0 defines if a
combination explains the outcome or not. Once this is complete, the
researcher may proceed to obtain solution sets (command: Standard
Analyses).
Next, the researcher may decide whether a single independent var-
iable should be present or absent at all times in the chosen configura-
tions (Fig. 7), contributing to the intermediate solution (to be explained
below). Unless otherwise needed (e.g., based on the theory or litera-
ture), it is suggested to choose “Present or Absent” in order to obtain all
Fig. 7. Setting specific causal conditions as present of absent for the interme- the possible configurations.
diate solution.

0.781378 and 0.839190 may be breaking points and potential frequency 6.4. Obtaining the configurations/solutions
thresholds. Thus, the researcher needs to decide which is the appropriate
threshold and justify this choice. To aid the researcher, fsQCA software FsQCA computes three solution, namely complex solution, parsi-
calculates PRI consistency, which stands for “Proportional Reduction in monious solution, and intermediate solution. Here, “solution” refers to a
Inconsistency” and is an alternate measure of the consistency of subset combination of configurations that is supported by a high number of
relations in social research, and only relevant to fuzzy sets. PRI consis- cases, where the rule “the combination leads to the outcome” is
tency is used to avoid simultaneous subset relations of configurations in consistent.
both the outcome and the absence of the outcome (i.e., negation). PRI The complex solution presents all the possible combinations of con-
consistency scores should be high and close to raw consistency scores (e. ditions when traditional logical operations are applied (Fig. 8). In gen-
g., 0.7), while configurations with PRI scores below 0.5 indicate sig- eral, because the number of identified configurations can be very large,
nificant inconsistency (Greckhamer et al., 2018). Thus, a PRI consis- the number of complex solutions can be very large and these may
tency threshold should also be used. include configurations with several terms, making the interpretation of
Finally, fsQCA software calculates SYM consistency (i.e., Symmetric the solutions rather difficult and in most cases impractical. For this
Consistency), which was developed for fuzzy-sets and can be used when reason, they are simplified further into parsimonious and intermediate
the researcher examines both the presence and negation of the outcome solution sets.
and wants to use the same consistency standard for both analyses (i.e., The parsimonious solution set is a simplified version of the complex
presence and its negation). In general, most papers do not present the solution, based on simplifying assumptions, and presents the most
truth table in their analysis, however presenting this can increase the important conditions which cannot be left out from any solution (Fig. 9).
validity of the findings and strengthen the rigorousness of the process. It These are called “core conditions” (Fiss, 2011) and are identified auto-
should be noted that a low consistency threshold leads to the identifi- matically by fsQCA. The major difference between parsimonious and
cation of more necessary conditions, reducing type II errors (i.e., false complex solution is that the complex solution excludes counterfactual
negatives), but increasing type I errors. (i.e., false positives), and vice cases, involving limited simplification, while the parsimonious solution

Fig. 8. The complex solution.

9
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Fig. 9. The parsimonious solution.

Fig. 10. The intermediate solution.

includes any counterfactual combination that can contribute to a logi- variables should be considered as only present, only absent, or either, in
cally simpler solution. explaining the outcome. By default, either present or absent is
Finally, the intermediate solution is obtained when performing computed. Any decisions made regarding the connection between each
counterfactual analysis on the complex and parsimonious solutions causal condition and the outcome need to be based on theoretical or
including only theoretically plausible counterfactuals (Liu et al., 2017; substantive knowledge (Fiss, 2011). The intermediate solution is part of
Ragin, 2008b) (Fig. 10). The intermediate solution uses a subset of those the complex solution and includes the parsimonious solution. While core
simplifying assumptions used to compute the parsimonious solution, conditions appear in both parsimonious and intermediate solutions, the
which should be consistent with theoretical and empirical knowledge. conditions that are eliminated in the parsimonious solution and appear
Based on previous knowledge the researcher may choose if one of the only in the intermediate solution are called “peripheral conditions” (Fiss,

10
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Table 1
FsQCA findings.

2011). In other words, since the intermediate solution presents both core indicated with a black circle (●), the absence/negation with a crossed-
and peripheral conditions, and peripheral conditions are removed from out circle (⊗), and the “do not care” condition with a blank space (Fiss,
the parsimonious solution, an easy way to identify the core conditions is 2011). As we mentioned earlier (see Section 6.3.1), the negation of a
to examine the parsimonious solution since it does not include periph- condition is referred in the literature also as absence, and the two terms
eral conditions. Also, the parsimonious solution is typically smaller than have been used interchangeably (Pappas, 2018). The distinction be-
the intermediate. However, it is possible that they could be exactly the tween core and peripheral is made by using large and small circles,
same, meaning that no elaboration is useful beyond the parsimonious respectively. The researcher needs to present the overall solution con-
solution. By including additional conditions in the solution, we increase sistency and the overall solution coverage. The overall coverage de-
the complexity in favor of increased consistency. Comparing Figs. 9 and scribes the extent to which the outcome of interest may be explained by
10, we see that the intermediate solution has a higher consistency than the configurations, and is comparable with the R-square reported on
the parsimonious. A more detailed and mathematically justified regression-based methods (Woodside, 2013). In our example, the results
description of the steps in counterfactual analysis is available in Mendel indicate an overall solution coverage of 0.84, which suggests that a
and Korjani (2012). substantial proportion of the outcome is covered by the nine solutions.
The fsQCA findings, as reported in the original article, are readable
6.5. Interpreting and presenting the solutions as follows. For high purchase intentions to occur, solutions 1–3 reflect
combinations of the presence and absence of cognitive with affective
FsQCA software provides all three solutions every time. Complex and perceptions. Quality of personalization and strongly positive emotions
parsimonious solutions are computed regardless of any simplifying as- are core constructs, pointing out the importance of these factors. In
sumptions employed by the researcher (e.g., choosing the presence or detail, the combination of high quality of personalization with strongly
absence/negation of a variables) while the intermediate solution de- positive emotions towards personalized services, with the absence of
pends on these assumptions. While the intermediate solution includes message quality and both types of negative emotions, lead to high
both core and peripheral conditions, we need an easy way to make the purchase intentions, regardless of the level of benefits of personalization
distinction that will help us interpret and present the solutions in a and weakly positive emotions (solution 1). To this end, when all
better manner. cognitive perceptions are present, in order to achieve high purchase
Combining the parsimonious and intermediate solutions can offer a intentions, they may be combined with either (i) strongly positive
more detailed and aggregated view of the findings (Fiss, 2011). To do emotions, with the absence of weakly positive and negative emotions,
this, the researcher can identify and mark the conditions of the parsi- and regardless of strongly negative emotions (solution 2), or (ii) all types
monious solution that also appear in the intermediate solution. This of emotions (solution 3), or (iii) with the absence of negative emotions
practically will lead to an intermediate solution which has highlighted and regardless of positive ones (solution 4). With the absence of all
the core conditions, clearly presenting all core and peripheral condi- emotions, high purchase intentions may be achieved with either high
tions, allowing for a better interpretation of the findings. Frequently, we personalization and message quality, regardless of their benefits (solu-
may have situations where more than one core condition co-occurs in a tion 5), or just by high personalization benefits regardless of its quality
given case. Hypothetically, if we have a parsimonious solution of A + BC (solution 7). Solution 6 combines personalization quality and strongly
+ BD and an intermediate solution of AcD + BCE + ABF + ABCDf, we negative emotions, with the absence of message quality along with the
report AcD + BCE + ABF + ABCDf, with bold characters indicating core rest of emotions. Personalization benefits play a minor role in this so-
conditions. Nonetheless, the researcher may choose to present only the lution. On the other hand, in solution 8, benefits are an important (core)
parsimonious solution and focus only on the core conditions that cannot factor which combined with message quality, and the presence of pos-
be left out of any solution. itive emotions only lead to high purchase intentions. Finally, the same
Next, to improve the presentation of the findings we can transform outcome may be achieved by the presence of weakly positive emotions
the solutions from fsQCA output (Figs. 9 and 10) into a table that is combined with the absence of all other emotions and all cognitive per-
easier to read (Table 1). Typically, the presence of a condition is ceptions (solution 9).

11
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

a proposition can be the following: Customers having high quality of

personalization, high message quality, and high strongly positive emotions
will have high intention to purchase online. Thus, the model we need to
compute is the presence of quality of personalization, message quality and
positive emotions. Indeed, the model should be seen as one variable, thus
we use the option Compute from the Variable menu. The function fuz-
zyand(x,..,) is used, which takes as input all the variables that are present
(Fig. 11). Also, here we can test for the negation of a specific condition if
it is required by a proposition. In this case, first we use the fuzzynot(x)
for every variable that is negated, which computes the negation (1-x) of
a variable (fuzzy set). Then we use the fyzzyand function which takes as
input all the variables that are present in each configuration and the new
variables that occurred as the outcome of the fuzzynot(x) function. The
fuzzyand(x,…,) function returns a minimum of two variables (fuzzy
sets).
Finally, the new variable (model) is plotted against the outcome of
interest using the XY Plot option from the fsQCA menu (Graphs – XY
Plot) (Fig. 12). Consistency and coverage values are presented here. The
findings show that the proposition is partially supported. Indeed, it does
not correspond to a specific solution identified by fsQCA, instead it al-
Fig. 11. Computing a specific proposition (model).
lows the identification of specific cases, or persons, (who and how many)
that will have high or low/medium intentions depending on specific
6.6. Testing for specific propositions antecedent conditions (if they are high or low/medium). Asymmetric
analysis indicates that high scores on the model (i.e., configuration)
After obtaining all the possible solutions that can explain the usually occur for high scores on the outcome condition, making the
outcome of interest with fsQCA, we can also test for specific propositions model useful for the researchers. However, this model does not predict
and examine for how many cases in the sample these propositions hold all cases with high scores on the outcome, as other models exist that
true (Pappas, 2018; Pappas et al., 2020). In addition, we can identify predict high scores of the same outcome (i.e., the upper left side in the
specifically which cases are these in the sample, if such information is plot). Models with consistency above 0.80 are useful and can serve
relevant. This is performed by computing the specific configuration in theory advancement (Woodside, 2017).
fsQCA, thus creating a model, and plotting it against the outcome of
interest.
The first step is to compute the configuration as described by the
proposition and transform into a model in fsQCA software. For example,

Fig. 12. Plotting a specific proposition (Proposition 1; high quality of personalization, high message quality, and high strongly positive emotions will have high intention
to purchase.

12
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Fig. 13. Basic steps in performing predictive validity.

because achieving only good model fit does not necessarily mean that
Table 2
the model offers good predictions. We present here how to perform
Solutions (models) from the subsample.
predictive validity testing in fsQCA (Pappas et al., 2016). This process is
Models from subsample Raw Unique Consistency visualized in Fig. 13.
coverage Coverage
In order to test for predictive validity, the first step is to randomly
BP●MQ●~SP●~SN●~WP●~WN 0.532 0.044 0.877 divide the sample into a subsample and a holdout sample, and run the
QP●BP●~MQ●~SP●~WP●~WN 0.309 0.027 0.853
same analysis for both samples, as was described in the previous sec-
QP●BP●~MQ●~SP●~WP●~WN 0.309 0.027 0.853
QP●BP●SP●~SN●WP●~WN 0.349 0.011 0.957
tions. The second step is to run the fsQCA for the subsample, and then
QP●BP●MQ●SP●SN●WP●WN 0.125 0.018 0.946 the findings obtained should be tested against the holdout sample.
QP●BP●MQ●~SN●~WP●~WN 0.559 0.015 0.895 Testing for predictive validity including hold-out samples is always
QP●BP●MQ●SP●~SN●~WN 0.459 0.012 0.949 possible and doing so substantially increases the added value for both
Overall solution consistency 0.869
empirical positivistic and interpretative case studies (Woodside, 2010).
Overall solution coverage 0.791
Table 2 presents the solutions from the subsample.
QP; Quality of Personalization, BP; Benefits of Personalization, MQ; Message After obtaining the findings from the subsample, the researcher must
Quality, SP; Strongly Positive Emotions, WP; Weakly Positive Emotions, SN; use the holdout sample to proceed with predictive validity testing. From
Strongly Negative Emotions, WN; Weakly Negative. ●; Logical conjunction
the findings of the subsample, each solution that contains the various
(AND), ~; Negation (NOT).
combinations of present and absent variables, needs modelling as one
variable following a similar procedure as described in Section 6.6 (using
6.7. Testing for predictive validity functions fuzzyand(x,…,) fuzzynot(x). Instead of computing a prede-
signed proposition we compute every solution from the findings from
Testing our solutions (models) for predictive validity is important. the testing with the subsample. Finally, the new variable is plotted
Predictive validity shows how well the model predicts the dependent against the outcome of interest using the holdout sample (Fig. 14).
variable in additional samples (Gigerenzer & Brighton, 2009; Gigerenzer Consistency and coverage values are presented here, which should not
& Gaissmaier, 2011; Woodside, 2014). Predictive validity is important contradict the consistency and coverage of the solution. The numbers

13
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Fig. 14. Fuzzy-Plot of Model 1 (from Table 2) using data from the holdout sample.

This section presents findings from a PLS-SEM analysis on the same

constructs and dataset in order to compare them with the fsQCA findings
that are presented earlier in the paper. Here, PLS-SEM, CB-SEM, or MRA
would lead to similar results as all independent variables are set as
predictors of one dependent variable. In detail, Fig. 15 shows the find-
ings from PLS-SEM analysis, which was performed using SmartPLS
software, with five out of seven relations being significant. Quality of
personalization and benefits of personalization have the strongest in-
fluence on intention to purchase, suggesting that they are the most
important factors in this model. Also, message quality and weakly pos-
itive emotions have a weaker positive effect on intention to purchase,
while weakly negative emotions have a negative effect in intention to
purchase. Strongly positive and negative emotions do not influence
intention to purchase. Finally, the model explains 57 % of the variance
Fig. 15. Findings from PLS-SEM analysis. of intention to purchase (R2 = 0.57).
Overall, a typical paper could report and explain these findings,
below the “Plot” button show set-theoretic consistency scores (Ragin, connect to previous studies and also discuss some unexpected results
2018). If one of these two numbers indicates high consistency, the other such as the non-significant effect of strongly positive emotions.
can be interpreted as a coverage score. In our example, 0.892873 in- Furthermore, on average the explained variance of 57 % is considered as
dicates high consistency, while 0.46756 indicates the coverage. These acceptable and a good result. However, the question remains on how we
calculations indicate that the data are largely consistent (89 %) with the can explain the rest of the sample (or its variance), as well as on how we
argument that Model 1 is a subset of intention to purchase and its can get more details on why we had unexpected results, such as the weak
coverage of intention to purchase is 47 %. That is, Model 1 accounts for effect of message quality on intention to purchase or the non-significant
47 % of the sum of the memberships in intention to purchase. effect of positive emotions. Different models may explain better our
findings, as consumers’ emotions when using personalized services may
7. When to use fsQCA in preference to variance-based also have either a mediating (Pappas, Kourouthanassis, Giannakos, &
approaches Chrissikopoulos, 2014) or moderating (Pappas, Kourouthanassis, Gian-
nakos, & Chrissikopoulos, 2017) role. Nonetheless, the purpose of this
FsQCA can be complementary to the traditional variance-based ap- analysis is to demonstrate how a variance-based approach differs in
proaches and the researchers can decide which method to use in their practice with the configurational approach. Since these methods can
study, depending on various factors since fsQCA can help overcome complement each other, the researchers may perform both of them in a
some limitations of variance-based approaches. Researchers can paper if that is suitable [e.g., (Fang, Shao, & Wen, 2016; Mikalef &
perform fsQCA to determine if causal asymmetry exists in their datasets Pateli, 2017)].
and if their findings are subject to causal equifinality or asymmetry (Fiss, Looking the findings from the fsQCA (Table 1) and from the PLS-SEM
2011). (Fig. 15), we can make several observations, highlighting how the two

14
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Table 3 Table 4
Findings using fsQCA and PLS-SEM on the same sample. Conceptual differences between fsQCA and RBMs.
FsQCA findings PLS-SEM findings FsQCA Regression based models (RBMs)

Quality of personalization is present in 6 Relationships between variables can be Relations are assumed to be
out of 9 solutions that explain either symmetric or asymmetric, non- symmetrical and linear.
Quality of personalization is an
intention to purchase linear.
important predictor of intention to
This can also be read as: High quality of The conditions that explain high levels of We focus on identifying determinants
purchase
personalization explains intention to an outcome are not exact opposites of that explain high levels of an outcome,
purchase in 6 out of 9 solutions. the ones that explain low levels of an assuming that the exact opposite will
Benefits of personalization are present in Benefits of personalization are an outcome (causal asymmetry) lead to low levels of the same outcome.
5 out of 9 solutions that explain important predictor of intention to Allows for case-based modelling to Focuses on the unique contribution of a
intention to purchase purchase identify localized effects and also to variable while holding constant the
Cognitive perceptions are more return to the cases after the analysis for values of all other variables in the
All cognitive perceptions influence
frequently present than absent when deeper understanding of the results and equation.
intention to purchase
explaining intention to purchase the sample.
Affective perceptions are more We identify multiple solutions as We identify a single best solution
frequently absent or negated than Weakly positive and negative emotions algorithms that are sufficient or (unifinality).
present when explaining intention to influence intention to purchase necessary to explain the same outcome
purchase (equifinality).
Strongly positive and negative emotions We identify the INUS condition, that is We report on the most important
can be either present or negated in insufficient but necessary part of a conditions that explain an outcome.
Strongly positive and negative emotions
explaining intention to purchase condition which is itself unnecessary We can examine direct, indirect, and
have no effect on intention to purchase
depending on how they combine with but sufficient for the result. moderating effects
the other factors. We explain both main effects and Findings explain cases that are
Message quality and weakly negative contrarian cases represented by main effects.
emotions can be either present or Message quality and weakly negative FsQCA transforms variables into scales RBMs use probabilities to compute
negated in explaining intention to emotions have weak effects (0.1) on and gives them a truth value that estimates that provide the single best
purchase depending on how they intention to purchase. defines their membership in the set. solution.
combine with the other factors. FsQCA computes the coverage, which
RBMs, using probabilities, estimate the
Different levels of weakly negative based on the truth value that they
variance of the model which is
emotions can contribute to explaining receive, is the actual number of cases
presented with the R2
intention to purchase. Weakly Weakly negative emotions have a explained by the solution.
negative emotions are present in one negative effect on intention to purchase Solutions are computed based on the
solution, and not present in eight coverage value and the existence of
Acceptance or rejection is based on the
solutions. cases in the sample that explain a
effect size and the p-value is used to
Different levels of weakly positive configuration, and consistency is used
determine its significance.
emotions may contribute to explaining to define the strength of a relationship,
intention to purchase. Weakly positive Weakly positive emotions have a supported by empirical evidence.
emotions are present in three positive effect on intention to purchase Variables compete with each other to
Variables work together and combine
solutions, but not present in four explain the largest percentage of
with each other to explain the outcome
solutions. variance.
All cognitive perceptions and emotions Not all cognitive perceptions and We develop propositions that can
can be present at the same time and emotions influence intention to examine the role of combinations of We develop hypotheses that depict the
explain intention to purchase purchase. variables in explaining the outcome and relation between 2 variables.
Positive and negative emotions can Weakly positive and weakly negative allow case identification in the model.
coexist in explaining intention to emotions may coexist in explaining Hypothesis of positive or negative
purchase. intention to purchase. effects influences the findings and the
Counter-hypothesized relations are
The presence of (weakly positive) acceptance or rejection of hypotheses.
expected and can be explained.
emotions can lead to high intention to Weakly negative emotions have a It is difficult to explain counter
purchase even when cognitive negative effect on intention to purchase hypothesized relations.
perceptions are not present.
The multiple solutions may refer to The single solution is the “best” solution
different types of users, with different that explains the largest percentage of information that fsQCA provides.
perceptions and needs. variance.
Overall, the findings from PLS-SEM show that quality of personali-
The results indicate an overall solution
coverage of .84, which suggests that a
The results indicate an R2 of 0.57, which zation, message quality, benefits of personalization, weakly positive
means that 57% of the variance of emotions, and weakly negative emotions are determinants of intention
substantial proportion of intention to
intention to purchase is explained by the
purchase is covered by the nine
model
to purchase. FsQCA reveals similar findings, but more importantly it also
solutions identifies conditions that are (1) sufficient or necessary to explain the
outcome and (2) insufficient on their own but are necessary parts of
solutions that can explain the result (INUS conditions; insufficient but
necessary part of a condition which is itself unnecessary but sufficient
methods are complementary, but more importantly how we are able to
for the result
get more insight into the dataset by employing fsQCA. Table 3 presents a
The literature on fsQCA is increasing and more researchers employ
detailed summary of how we can read the findings from the two
configurational analysis to get a deeper understanding of the phenom-
methods. The interpretation of the findings in fsQCA is an elaborate
ena under examination. However, as the two methods can complement
process due to the very rich information that the analysis provides. This
each other, it means that they look at a phenomenon from a different
means that researchers can choose on which part of the findings to focus
point of view. For more details on best practices on QCA we direct the
more depending on the context as well as their knowledge on the topic.
reader to existing works (Greckhamer et al., 2018; Schneider & Wage-
In detail, the “PLS-SEM findings” column presents the typical way of
mann, 2010; Woodside, 2016b). In Table 4, we summarize some con-
reading the findings from such a model. Based on that, the “FsQCA
ceptual differences between fsQCA and typical RBMs, related to key
findings” column presents what can be considered as the equivalent
assumptions between the two methods, as well as how the analysis is
finding in configurational analysis. It must be noted that an overlap may
performed and how the findings are presented and interpreted.
occur between the fsQCA findings, which is done here on purpose to
Furthermore, Table 5 summarizes some of the frequently used
highlight the difference with the PLS-SEM findings as well as the rich
thresholds during the analysis. In fsQCA, the value of 1 defines the full-

15
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Table 5 set membership, 0 defines the full set non-membership and 0.5 the in-
A summary of thresholds used in fsQCA. termediate membership. Because fsQCA uses log-odds, thus not capable
Threshold Sources Examples of computing memberships that are exactly 0 or 1, the membership
thresholds are set as follows; 0.95, 0.50, 0.05. Nonetheless, the re-
95 % - Full set
membership searchers can choose different thresholds depending on the background,
(Fiss, 2011;
Any type of data 50 % - context and previous knowledge. Finally, to explicate how existing
Ragin, 2008b;
(including Intermediate
Rihoux & Ragin,
studies may benefit by applying fsQCA, in Appendix D we summarize
Likert Scales) membership their main outcomes and suggest possible extensions through fsQCA
2009)
5 % - Full set non-
membership
based on their findings.
6 - Full set
membership 8. Conclusion
4 - Intermediate (Ordanini et al., (Pappas et al.,
7-point Likert Scale
membership 2014) 2016, 2020)
The choice of an appropriate method of analysis matters and such
2 - Full set non-
membership choices are defined by our research questions and research objectives.
Any type of data (Pappas, Mikalef Nonetheless, various limitations may influence these choices such as
80 % - Full set
skewed to the
membership
et al., 2017); sample size limitations, availability of tools, or sometimes our own
right or left (e.g., Likert Scale knowledge of specific tools and methods. Often, we tend to keep using
Likert scale, 50 % - (Papamitsiou
clickstreams, user Intermediate et al., 2020);
the same tools and methods of analysis because we have gained a certain
performance, membership Clickstreams, amount of expertise along with the level of convenience it brings. While
physiological 20 % - Full set non- multimodal data, gaining expertise on a tool or topic is much desired, we may avoid using
data) membership user performance new tools and methods because they require extra effort and resources to
Overall solution
> 0.75 minimum get past their learning curve, something that we may often miss as ac-
consistency
> 0.80 suggested ademics. Methodology defines how we study a phenomenon and how we
Can be higher think about it (Bagozzi, 2007). While quantitative and qualitative ap-
Identify natural proaches have their strengths and weaknesses, employing mixed method
breaking points in approaches can offer significantly deeper insight into our datasets, and
the consistency
values
by extension the phenomena we study. FsQCA is tool that combines
Test multiple aspects of both quantitative and qualitative approaches in one analysis,
values of thus bridging the quantitative-qualitative gap that exists in most fields.
consistency Similar to Gefen, Straub, and Boudreau (2000) seminal study on how
(Dul, 2016; Fiss,
threshold for the
2011; Mendel & to conduct SEM and regression in research, our goal is to offer an easy to
same analysis and
assess the
Korjani, 2012; follow guide that describes step-by-step how to employ fsQCA in typical
Ragin, 2008b;
Raw consistency difference between studies in information management and marketing. This tutorial can
Rihoux & Ragin,
the results.
2009)
also be followed by studies in other fields in which similar methods are
Low consistency used.
threshold leads to
more necessary
Looking at the literature, while the number of articles employing
conditions, fsQCA is increasing, it is evident that some venues (journals or confer-
reducing type II ences) are more frequented than others. This mainly happens because
errors (i.e., false reviewers (and editors) with more expertise are available in these
negatives), but
venues, thus making it possible to better understand the method and its
increasing type I
errors (i.e., false implications. This creates a natural cluster with researchers employing
positives), and vice fsQCA looking for venues that have already published articles employ-
versa. ing the method. Nonetheless, more and more journals and conferences
Close to “Raw will accept such articles showing that the field is evolving and interest
consistency” (Greckhamer
PRI consistency
(~0.70) et al., 2018)
towards less popular methods is increasing. Editors and reviewers can
> 0.50 minimum direct authors towards employing fsQCA and taking a look at their data
No specific from a different point of view. It must be clear that fsQCA is not a so-
threshold lution to all problems and is not always appropriate. The authors should
For Small-N (Greckhamer
Coverage justify the reasons for employing fsQCA and always follow the recom-
sample size high et al., 2013)
coverage is mended thresholds and guidelines, in order to offer meaningful and
expected. valid results. FsQCA should not be employing blindly in a mechanical
<50: Small-N
(Greckhamer
manner. Based on our own personal experience, editors and reviewers
>50: Large-N will disagree or will not be able to decide about the appropriateness of
et al., 2013;
Sample Size No restriction on
Vatrapu et al., even the methodological soundness of a manuscript. As with the tutorial
how big the sample
2016) on SEM 20 years ago more knowledgeable opinions should be weighted
can be
<60 (Capatina, more heavily than those of less understanding (Gefen et al., 2000). Our
Micu, Micu, goal with this tutorial is to help researchers better understand fsQCA,
1 for Small-N Bouzaabia, &
either they act as authors, or as editors and reviewers, and offer a
Bouzaabia,
2018) practical guide that can answer many questions for those less familiar
500+ (Pappas with QCA.
Frequency 2 or 3 Large-N
et al., 2016)
3000+ (Schmitt, CRediT authorship contribution statement
It is recommended
Grawe, &
keeping about 80
Woodside, 2017;
% of the cases in Ilias O. Pappas: Conceptualization, Methodology, Formal analysis,
Sergis et al.,
the analysis.
2018) Investigation, Writing - original draft. Arch G. Woodside: Methodol-
ogy, Writing - review & editing.

16
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Acknowledgements Appendix B. Contrarian case analysis

The authors would like to thank the Editor of IJIM and the anony- We create quintiles [i.e., dividing the sample into five equal groups]
mous reviewers for their helpful developmental comments on the work by ranking the cases. An easy way to do this in SPSS in using the “Rank
presented in this manuscript. The authors express their appreciation to Cases” function of SPSS. In detail, on menu “Transform”, select “Rank
Katja Bley for her critical reading and thoughtful suggestions during Cases…”. Next, under “Rank Types…” uncheck “Rank” and select the
writing the final version of manuscript. We thank Peer Fiss for offering “Ntiles” option and set it to 5 (i.e., quintile) (Fig. B1). Then select the
his insight while developing this tutorial article, as well as Michail variables of interest and press OK to create the Ntiles.
Giannakos, Panos Kourouthanassis, Patrick Mikalef for helping in earlier This leads to the creation of new variables in the dataset (Fig. B2).
phases of this project. Finally, we thank the workshop participants at The quintiles can be also computed by using the percentiles within the
Swansea University, American College of Greece, and University of descriptive statistics function of SPSS; however, the Rank Cases option is
Agder for their useful feedback and insightful suggestions. This project simpler
has received funding from the European Union’s Horizon 2020 research Next, we performed cross-tabulations across the quintiles, using the
and innovation programme, under the Marie Sklodowska-Curie Grant SPSS “Crosstabs” function (Analyze > Descriptive Statistics > Cross-
Agreements Nos. 751550. tabs), between every independent variable and the dependent variable
(Fig. B3). The crosstabs allow us to compute the degree of association
Appendix A. Construct and scale items between the variables, which suggests a dependence between the two
variables and describes main effects between them.
Table A1 The result for any two variables is a 5 × 5 table that presents all
combinations for all of the cases in the sample between the two variables
(Fig. B4). The top left and bottom right cases represent the main effects
(e.g., degree of association), while the bottom left and top right

Table A1
Scale items with mean, standard deviation and standardized loading.
Construct and scale items Mean S.D. Loading

Quality of Personalization (CA ¼ 0.89)

How well and efficiently are perceived the offered services, to fulfill customers’ needs
1. Online vendors can provide me with personalized deals/ads tailored to my activity context. 4.6 1.44 0.84
2. Online vendors can provide me with more relevant promotional information tailored to my preferences or personal interests. 4.6 1.36 0.88
3. Online vendors can provide me with the kind of deals/ads that I might like. 4.5 1.32 0.84
Message Quality (CA ¼ 0.93)
Customer’s general perception of the accuracy and completeness of Website information as it relates to products and transactions, when using personalized services.
1. Personalized services provide correct information about items or services I want to purchase. 4.3 1.30 0.74
2. Overall, I think personalized services provide useful information. 4.5 1.32 0.80
3. Personalized services provide timely information on an item/service. 4.5 1.29 0.76
4. Personalized services provide sufficient information when I try to make an online purchase. 4.3 1.33 0.77
5. I am satisfied with the information that personalized services provide. 4.6 1.40 0.84
6. Overall, the information personalized services provide is of high quality. 4.4 1.43 0.89
7. Personalized services provide timely information on an item/service. 4.3 1.31 0.86
Benefits of Personalization (CA ¼ 0.90)
Customer’s belief about the extent to which he or she will become better off from the online transaction with a certain Website, when using personalized services.
1. I think the use of personalized services is convenient. 4.9 1.43 0.85
2. I can save money by using personalized services. 4.7 1.60 0.76
3. I can save time by using personalized services. 5.2 1.57 0.86
4. Using personalized services enables me to accomplish a shopping task more quickly than using traditional methods. 5.9 1.57 0.84
5. Using personalized services increases my productivity in shopping (e.g., make purchase decisions or find product information within the shortest time 4.8 1.57 0.74
frame).
Intention to Purchase (CA ¼ 0.90)
Customer’s intention to shop online based on personalized services.
1. In the future I intend to continue shopping online based on personalized services. 4.7 1.47 0.93
2. My general intention to buy online based on personalized services is very high. 4.4 1.54 0.94
3. I will shop online in the future based on personalized services. 4.4 1.42 0.87

Emotions

Measuring customer’s emotions, based on valence and control, when using personalized services.

Mean SD Loading Mean SD Loading

Strongly Positive (CA ¼ 0.91) Weakly Positive (CA ¼ 0.89)

1. Pleasure 3.7 1.73 0.95 1. Contentment 3.7 1.73 0.93
2. Joy 3.5 1.71 0.93 2. Admiration 3.1 1.73 0.76
3. Pride 2.8 1.63 0.74 3. Love 2.4 1.57 0.64
4. Amusement 3.8 1.57 0.75 4. Relief 3.0 1.79 0.75
5. Interest 4.21 1.51 0.68
Strongly Negative (CA ¼ 0.83) Weakly Negative (CA¼0.89)
1. Anger 2.99 1.71 0.72 1. Disappointment 2.72 1.65 0.68
2. Hate 2.66 1.45 0.85 2. Shame 2.12 1.48 0.87
3. Contempt 3.09 1.83 .65 3. Regret 2.47 1.61 0.80
4. Disgust 2.15 1.39 0.86 4. Guilt 2.04 1.33 0.84
5. Fear 2.79 1.72 0.57 5. Sadness 1.93 1.32 0.77
6. Compassion 2.12 1.38 0.69

CA = Cronbach’s alpha.

17
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Fig. B1. Create the quintiles using the Rank Cases function in SPSS.

Fig. B2. Quintiles as new variables in the dataset.

Fig. B3. Performing crosstabs in SPSS to find the main effects and contrarian cases.

Fig. B4. Cross-tabulation and degree of association for message quality and intention to purchase.

18
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

represent contrarian cases, that are the ones not explained by the main
effects. Results for the contrarian case analysis for all variables are
presented in Appendix C, as it appears on the original study of Pappas
et al. (2016). The findings show the existence of various relationships
between the variables, separate from the main effect, supporting the
need to perform a configurational analysis.

Appendix C. Results of contrarian case analysis

Table C1

Table C1
Results from the contrarian case analysis.

19
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Table D1
Examples of recent studies that perform variance-based analysis (e.g., MRA, SEM) along with suggestions for possible extensions with fsQCA.
Source Topic Description Main findings Possible extension with fsQCA

Anticipated quality, configured trust, and Examine if the dimensions of anticipated

Citizen-government interaction via social media. approved adaptation as important factors in quality and configured trust, are indispensable
Aladwani and Dwivedi citizen-government interaction. factors to achieve approved adaptation.
(2018) Anticipated governmental social media quality
The article proposes and tests a new model in a Identify which of these dimensions are
influences configured trust, which in turn
quest for a SocioCitizenry theory. sufficient to explain high approved adaptation.
influences the extent of approved adaptation.
System quality, user satisfaction and habit Examine if satisfaction is a necessary factor for
positively influenced by intention to continue high continuance usage intention of e-filing or
using e-filing. if it needs to be combined with other factors.
Identify if combinations of system quality,
Satisfaction is the most important factor having
satisfaction, and habit are sufficient to explain
Veeramootoo, Nunkoo, Continuance usage intention of e-filing in e- the stronger impact on the outcome.
the outcome.
and Dwivedi (2018) government
Identify how information quality, service
Information quality, service quality and quality, and perceived risk play a role for a
perceived risk do not significantly predict smaller part of the sample or if they help to
continuance usage intention of e-filing. formulate solutions that are different from the
main effects.
The factors representing both planned (i.e.,
Examine how the absence of planned behavior
perceived social recognition and altruistic
and absence of unplanned behavior can explain
motivation) and unplanned behavior (i.e.,
political content sharing behavior as well as its
Hossain, Dwivedi, extroversion and impulsiveness) affect people’s
negation (i.e., not sharing).
Chan, Standing, and Sharing Political Content in Online Social Media political content sharing behavior.
Olanrewaju (2018) Complement the role of collective opinion as
Collective opinion moderates planned behavior, moderator by identifying the combinations in
but not unplanned behavior. which it is necessary to be present or absent for
people to share political content.
SNSs participation motivations significantly
Examine how participation, brand trust, brand
affect customer participation, which in turn
loyalty can combine to explain brand co-
positively influences brand trust, brand loyalty
creation for different customers depending on
and consequently resulted in branding co-
their gender and age.
Kamboj, Sarmah, creation.
Customer participation in social media brand
Gupta, and Dwivedi Examine how brand loyalty in social media
communities applying the S-O-R framework. Gender influences brand trust and brand co-
(2018) brand communities’ changes when brand trust
creation, while age influences brand loyalty
is present and when it is absent.
Brand trust mediates the relationship between
customer participation and brand loyalty in
social media brand communities.
Test the same antecedents for both positive and
Administrative conflict, political biasness and negated attitude and compare. Conditions that
professional growth have significant effects on explain the presence of the outcome are not
negative attitude. necessarily mirror opposites of those explaining
the absence of the outcome.
Dwivedi, Shareef, Include insecurity and corruption as they may
Exploratory study in involvement in emergency Impact of insecurity and corruption is non-
Mukerji, Rana, and play a role in a small part of the sample (not
supply chain for disaster management significant on attitude.
Kapoor (2018) explain by the main model).
Insecurity and Corruption constructs were
removed from the model to improve the model
Examine if negative attitude is a necessary
the fit.
factor for behavioral intentions.
Negative attitude has a very strong effect on
behavioral intentions.
Identify how the antecedents form different
UMEGA outperforms all other models for e-
solutions that explain attitude and intention
government
separately and compare.
Examine if attitude is a necessary factor for high
Government context should be taken into intention, and if not, what other combinations
Propose and test the unified model of electronic
Dwivedi et al. (2017) account. are sufficient to explain intention that do not
government adoption (UMEGA)
include attitude.
Since the explanatory power is so large
UMEGA is simpler to use and has a better complement by additional solutions for the
explanatory power than the UTAUT. small part of the sample that is not explained by
this model.
Identify how the antecedents form different
solutions that are sufficient to explain attitude
Propose a revised alternative theoretical model and intention separately and compare.
Attitude was central to behavioral intentions and
for explaining the acceptance and use of IS and IT Examine if attitude is a necessary factor for high
usage behaviors, partially mediated the effects of
Dwivedi et al. (2019) innovations using a combination of meta-analysis intention, and if not what other combinations
exogenous constructs on behavioral intentions,
and structural equation modelling (MASEM) are sufficient to explain intention that do not
and had a direct influence on usage behaviors.
techniques. include attitude.
Test specific propositions coming from MASEM
and identify specific cases in the sample.
Performance expectancy, effort expectancy, Since the model performs better than other
Adoption of emerging electronic government social influence, facilitating conditions, anxiety acceptance models, identify if any of the
Rana et al. (2017)
(eGov) applications in India are important predictors of attitude and antecedents is a necessary condition for the
behavioral intention. outcome to occur.
(continued on next page)

20
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Table D1 (continued )
Source Topic Description Main findings Possible extension with fsQCA

Identify combinations that explain a different

The model performs better compare to previous
(smaller) part of the sample that is not
technology acceptance models.
represented in the main model.
It revealed that consumer segmentation and Examine if consumer segmentation and target
target marketing is the most effective way to marketing are necessary in order to effectively
communicate with consumers through communicate with consumers through
Shareef, Dwivedi,
Examine the effects of promotional marketing promotional marketing conducted by the mobile promotional marketing conducted by the
Kumar, and Kumar
through SMS. phone SMS. mobile phone SMS.
(2017)
It also suggested that this promotional marketing Identify different combinations of consumer
is valuable only for highly reputable vendors/ segmentation and target marketing that are
retailers. sufficient to explain the outcome.
All predictors (Performance Expectancy, Effort
Expectancy, Trust, Hedonic Motivation, Price Examine if all predictors (except SI) of
Value) of behavioral intention to adopt mobile behavioral intention are necessary to be present
The study examines the factors influencing banking were found significant except for social in order to explain adoption of mobile banking.
Alalwan et al. (2017) behavioral intentions and adoption of Mobile influence.
banking by Jordanian bank customers. Identify combinations that explain other part of
The variance explained by the model in the sample, not described in the main model
behavioral intention was found as 65 %. that explains 65 % of the variance of behavioral
intention.
This study addresses whether a service delivery
channel, based on mobile phones and provided Identify which factors are necessary or
Citizens are quite satisfied with SMSs used as a
through a short messaging service (SMS) can be sufficient to develop citizen perceptions of high
service delivery channel by the public service
included in public administration to meet citizen value and effectiveness of this new service
domain.
requirements, regarding their perceptions of high delivery channel.
value and effectiveness.
Shareef, Dwivedi,
If public service providers can effectively Produce solutions that profile citizens based on
Kumar, and Kumar
segment the market based on time, location, and their country and identify which factors are
(2016)
requirements, and can deliver the preferred necessary or sufficient for high positive
Also, the study examines if culture plays a role on
message to concerned users with relevant and attitudes towards this service for the different
developing a positive attitude toward this service
information that is easy to access and process, cultures.
delivery channel.
citizens will regard this service delivery channel The relatively small sample per country allows
as effective and satisfactory, and as competent as to go back to the cases and explain behavior
its private counterparts. using contextual information
Identify how the antecedents form different
All are verified. solutions that explain attitude and intention
Rana, Dwivedi, separately and compare.
Adoption of online public grievance redressal
Williams, and Examine if attitude is a necessary factor for high
system in India
Weerakkody (2016) The model performs better compared to previous intention, and if not what other combinations
technology acceptance models. are sufficient to explain intention that do not
include attitude.
Performance expectancy, social influence, Identify which of the proposed antecedents are
innovativeness, and perceived risk influence sufficient or necessary for high behavioral
behavioral intentions. intentions.
Make propositions based on the SEM findings
Adoption of Remote Mobile Payments (MP) in the Effort expectancy and trust do not influence
and test them to identify for how many cases in
UK. Extending the Unified Theory of Acceptance behavioral intentions.
Slade, Dwivedi, Piercy, the sample (and which ones) they hold true.
and Use of Technology (UTAUT), with more
and Williams (2015) Segment consumers based on their knowledge
consumer-related constructs (Innovativeness,
of MP, identify who they are and use knowledge
Risk, and Trust) for non-users of these services.
Knowledge of MP moderates the effects of that is not included in the model to explain their
antecedents of behavioral intention behavior. The latter could lead to qualitative
studies with specific consumers (e.g., via
interviews).

Appendix D. Possible extension of recent studies with fsQCA Cairns, J., Wistow, J., & Bambra, C. (2017). Making the case for qualitative comparative
analysis in geographical research: A case study of health resilience. Area, 49(3),
369–376.
Table D1 Capatina, A., Micu, A., Micu, A. E., Bouzaabia, R., & Bouzaabia, O. (2018). Country-
based comparison of accommodation brands in social media: An fsQCA approach.
References Journal of Business Research, 89, 235–242.
Cooper, B., & Glaesser, J. (2011). Using case-based approaches to analyse large datasets:
A comparison of Ragin’s fsQCA and fuzzy cluster analysis. International Journal of
Aladwani, A. M., & Dwivedi, Y. K. (2018). Towards a theory of SocioCitizenry: Quality Social Research Methodology, 14(1), 31–48.
anticipation, trust configuration, and approved adaptation of governmental social Cronqvist, L. (2004). Presentation of TOSMANA. Adding multi-value variables and visual
media. International Journal of Information Management, 43, 261–272. aids to QCA. COMPASSS: WP,20.
Alalwan, A. A., Dwivedi, Y. K., & Rana, N. P. (2017). Factors influencing adoption of Cronqvist, L., & Berg-Schlosser, D. (2009). Multi-value QCA (mvQCA). Configurational
mobile banking by Jordanian bank customers: Extending UTAUT2 with trust. Comparative Methods: Qualitative Comparative Analysis (QCA) and Related Techniques,
International Journal of Information Management, 37(3), 99–110. 51, 69–86.
Bagozzi, R. P. (2007). The legacy of the technology acceptance model and a proposal for Delery, J. E., & Doty, D. H. (1996). Modes of theorizing in strategic human resource
a paradigm shift. Journal of the Association for Information Systems, 8(4), 3. management: Tests of universalistic, contingency, and configurational performance
Basurto, X., & Speer, J. (2012). Structuring the calibration of qualitative data as sets for predictions. The Academy of Management Journal, 39(4), 802–835.
qualitative comparative analysis (QCA). Field Methods, 24(2), 155–174. DiStefano, C., Zhu, M., & Mindrila, D. (2009). Understanding and using factor scores:
Berg-Schlosser, D., & De Meur, G. (2009). Comparative research design: Case and Considerations for the applied researcher. Practical Assessment, Research, and
variable selection. In B. Rihoux, & C. C. Ragin (Eds.), Configurational comparative Evaluation, 14(1), 20.
methods. Qualitative comparative analysis (QCA) and related techniques (pp. 19–32). Dul, J. (2016). Identifying single necessary conditions with NCA and fsQCA. Journal of
Thousand Oaks and London: Sage. Business Research, 69(4), 1516–1523.

21
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Dwivedi, Y. K., Rana, N. P., Janssen, M., Lal, B., Williams, M. D., & Clement, M. (2017). Papamitsiou, Z., Pappas, I. O., Sharma, K., & Giannakos, M. (2020). Utilizing multimodal
An empirical validation of a unified model of electronic government adoption data through fsQCA to explain engagement in adaptive learning. IEEE Transactions on
(UMEGA). Government Information Quarterly, 34(2), 211–230. Learning Technologies, 13(4), 689–703. https://doi.org/10.1109/TLT.2020.3020499.
Dwivedi, Y. K., Rana, N. P., Jeyaraj, A., Clement, M., & Williams, M. D. (2019). Re- Pappas, I. O. (2018). User experience in personalized online shopping: A fuzzy-set
examining the unified theory of acceptance and use of technology (UTAUT): analysis. European Journal of Marketing, 52(7/8), 1679–1703. https://doi.org/
Towards a revised theoretical model. Information Systems Frontiers, 21(3), 719–734. 10.1108/EJM-10-2017-0707.
Dwivedi, Y. K., Shareef, M. A., Mukerji, B., Rana, N. P., & Kapoor, K. K. (2018). Pappas, I. O., Kourouthanassis, P. E., Giannakos, M. N., & Chrissikopoulos, V. (2014).
Involvement in emergency supply chain for disaster management: A cognitive Shiny happy people buying: The role of emotions on personalized e-shopping.
dissonance perspective. International Journal of Production Research, 56(21), Electronic Markets, 24(3), 193–206.
6758–6773. Pappas, I. O., Kourouthanassis, P. E., Giannakos, M. N., & Chrissikopoulos, V. (2016).
El Sawy, O. A., Malhotra, A., Park, Y., & Pavlou, P. A. (2010). Research commentary- Explaining online shopping behavior with fsQCA: The role of cognitive and affective
seeking the configurations of digital ecodynamics: It takes three to tango. Information perceptions. Journal of Business Research, 69(2), 794–803.
Systems Research, 21(4), 835–848. Pappas, I. O., Papavlasopoulou, S., Mikalef, P., & Giannakos, M. N. (2020). Identifying
Fang, J., Shao, Y., & Wen, C. (2016). Transactional quality, relational quality, and the combinations of motivations and emotions for creating satisfied users in SNSs:
consumer e-loyalty: Evidence from SEM and fsQCA. International Journal of An fsQCA approach. International Journal of Information Management, 53, Article
Information Management, 36(6), 1205–1217. 102128.
Fedorowicz, J., Sawyer, S., & Tomasino, A. (2018). Governance configurations for inter- Pappas, I. O., Giannakos, M. N., Jaccheri, L., & Sampson, D. G. (2017). Assessing student
organizational coordination: A study of public safety networks. Journal of behavior in computer science education with an fsQCA approach: The role of gains
Information Technology, 33(4), 326–344. and barriers. ACM Transactions on Computing Education (TOCE), 17(2). Article No.
Fiss, P. C. (2007). A set-theoretic approach to organizational configurations. The 10.
Academy of Management Review, 32(4), 1180–1198. Pappas, I. O., Kourouthanassis, P. E., Giannakos, M. N., & Chrissikopoulos, V. (2017).
Fiss, P. C. (2011). Building better causal theories: A fuzzy set approach to typologies in Sense and sensibility in personalized e-commerce: How emotions rebalance the
organization research. The Academy of Management Journal, 54(2), 393–420. purchase intentions of persuaded customers. Psychology & Marketing, 34(10),
Fitzsimons, G. J. (2008). Death to dichotomizing. The Journal of Consumer Research, 35 972–986.
(1), 5–8. Pappas, I. O., Mikalef, P., Giannakos, M. N., & Pavlou, P. A. (2017). Value co-creation
Gefen, D., Straub, D., & Boudreau, M.-C. (2000). Structural equation modeling and and trust in social commerce: An fsQCA approach. Paper Presented at the 25th
regression: Guidelines for research practice. Communications of the Association for European Conference on Information Systems (ECIS).
Information Systems, 4(1), 7. Park, Y., & Mithas, S. (2020). Organized complexity of digital business strategy: A
Gigerenzer, G. (1991). From tools to theories: A heuristic of discovery in cognitive configurational perspective. MIS Quarterly, 44(1).
psychology. Psychological Review, 98(2), 254. Park, Y., Fiss, P., & El Sawy, O. A. (2020). Theorizing the multiplicity of digital
Gigerenzer, G., & Brighton, H. (2009). Homo heuristicus: Why biased minds make better phenomena: The ecology of configurations, causal recipes, and guidelines for
inferences. Topics in Cognitive Science, 1(1), 107–143. applying QCA. MIS Quarterly, 44(4), 1493–1520.
Gigerenzer, G., & Gaissmaier, W. (2011). Heuristic decision making. Annual Review of Plewa, C., Ho, J., Conduit, J., & Karpen, I. O. (2016). Reputation in higher education: A
Psychology, 62, 451–482. fuzzy set analysis of resource configurations. Journal of Business Research, 69(8),
Greckhamer, T., Furnari, S., Fiss, P. C., & Aguilera, R. V. (2018). Studying configurations 3087–3095.
with qualitative comparative analysis: Best practices in strategy and organization Ragin, C. C. (1987). The comparative method: Moving beyond qualitative and quantitative
research. Strategic Organization, 18(3), 482–495. strategies. Univ of California Press.
Greckhamer, T., Misangyi, V. F., & Fiss, P. C. (2013). The two QCAs: From a small-N to a Ragin, C. C. (2000). Fuzzy-set social science. University of Chicago Press.
large-N set theoretic approach. In P. C. Fiss, B. Cambré, & A. Marx (Eds.), Ragin, C. C. (2018). User’s guide to Fuzzy-Set/Qualitative comparative analysis 3.0. Irvine,
Configurational theory and methods in organizational research (pp. 49–75). Emerald California: Department of Sociology, University of California.
Group Publishing Limited. Ragin, C. C. (2008a). Measurement versus calibration: A set-theoretic approach. In
Henik, E. (2015). Understanding whistle-blowing: A set-theoretic approach. Journal of J. M. Box-Steffensmeier, H. E. Brady, & D. Collier (Eds.), The Oxford handbook of
Business Research, 68(2), 442–450. political methodology (pp. 174–198).
Hossain, M. A., Dwivedi, Y. K., Chan, C., Standing, C., & Olanrewaju, A.-S. (2018). Ragin, C. C. (2008b). Redesigning social inquiry: Fuzzy sets and beyond (Vol. 240). Wiley
Sharing political content in online social media: A planned and unplanned behaviour Online Library.
approach. Information Systems Frontiers, 20(3), 485–501. Ragin, C. C., & Davey, S. (2016). fs/QCA [Computer Programme], version 3.0. Irvine, CA:
Kamboj, S., Sarmah, B., Gupta, S., & Dwivedi, Y. (2018). Examining branding co-creation University of California.
in brand communities on social media: Applying the paradigm of Stimulus- Rana, N. P., Dwivedi, Y. K., Lal, B., Williams, M. D., & Clement, M. (2017). Citizens’
Organism-Response. International Journal of Information Management, 39, 169–185. adoption of an electronic government system: Towards a unified view. Information
Kane, H., Lewis, M. A., Williams, P. A., & Kahwati, L. C. (2014). Using qualitative Systems Frontiers, 19(3), 549–568.
comparative analysis to understand and quantify translation and implementation. Rana, N. P., Dwivedi, Y. K., Williams, M. D., & Weerakkody, V. (2016). Adoption of
Translational Behavioral Medicine, 4(2), 201–208. online public grievance redressal system in India: Toward developing a unified view.
Liu, Y., Mezei, J., Kostakos, V., & Li, H. (2017). Applying configurational analysis to IS Computers in Human Behavior, 59, 265–282.
behavioural research: A methodological alternative for modelling combinatorial Rihoux, B., & Ragin, C. C. (2009). Configurational comparative methods: Qualitative
complexities. Information Systems Journal, 27(1), 59–89. comparative analysis (QCA) and related techniques (Vol. 51). Thousand Oaks, CA: Sage
Longest, K. C., & Vaisey, S. (2008). Fuzzy: A program for performing qualitative Publications.
comparative analyses (QCA) in Stata. The Stata Journal, 8(1), 79–104. Saridakis, C., Angelidou, S., & Woodside, A. G. (2020). What type of CSR engagement
Mackie, J. I. (1965). Causes and conditions. American Philosophical Quarterly, 2(4), suits my firm best? Evidence from an abductively-derived typology. Journal of
245–264. Business Research, 108, 174–187.
Mendel, J. M., & Korjani, M. M. (2012). Charles Ragin’s fuzzy set qualitative comparative Schmitt, A. K., Grawe, A., & Woodside, A. G. (2017). Illustrating the power of fsQCA in
analysis (fsQCA) used for linguistic summarizations. Information Sciences, 202, 1–23. explaining paradoxical consumer environmental orientations. Psychology &
Mikalef, P., & Pateli, A. (2017). Information technology-enabled dynamic capabilities Marketing, 34(3), 323–334.
and their indirect effect on competitive performance: Findings from PLS-SEM and Schneider, C. Q., & Wagemann, C. (2010). Standards of good practice in qualitative
fsQCA. Journal of Business Research, 70, 1–16. comparative analysis (QCA) and fuzzy-sets. Comparative Sociology, 9(3), 397–418.
Miller, D. (2018). Challenging trends in configuration research: Where are the Sergis, S., Sampson, D. G., & Giannakos, M. N. (2018). Supporting school leadership
configurations? Strategic Organization, 16(4), 453–469. decision making with holistic school analytics: Bridging the qualitative-quantitative
Misangyi, V. F., Greckhamer, T., Furnari, S., Fiss, P. C., Crilly, D., & Aguilera, R. (2017). divide using fuzzy-set qualitative comparative analysis. Computers in Human
Embracing causal complexity: The emergence of a neo-configurational perspective. Behavior, 89, 355–366.
Journal of Management, 43(1), 255–282. Shareef, M. A., Dwivedi, Y. K., Kumar, V., & Kumar, U. (2016). Reformation of public
Nagy, G., Megehee, C. M., Woodside, A. G., Laukkanen, T., Hirvonen, S., & Reijonen, H. service to meet citizens’ needs as customers: Evaluating SMS as an alternative service
(2017). Achieving requisite variety in modeling firms’ strategy heterogeneities: delivery channel. Computers in Human Behavior, 61, 255–270.
Explaining paradoxical firm-market performances. Industrial Marketing Management, Shareef, M. A., Dwivedi, Y. K., Kumar, V., & Kumar, U. (2017). Content design of
65, 100–128. advertisement for consumer exposure: Mobile marketing through short messaging
Nistor, N., Stanciu, D., Lerche, T., & Kiel, E. (2019). “I am fine with any technology, as service. International Journal of Information Management, 37(4), 257–268.
long as it doesn’t make trouble, so that I can concentrate on my study”: A case study Slade, E. L., Dwivedi, Y. K., Piercy, N. C., & Williams, M. D. (2015). Modeling consumers’
of university students’ attitude strength related to educational technology adoption intentions of remote mobile payments in the United Kingdom: Extending
acceptance. British Journal of Educational Technology, 50(5), 2557–2571. UTAUT with innovativeness, risk, and trust. Psychology & Marketing, 32(8), 860–873.
Ordanini, A., Parasuraman, A., & Rubera, G. (2014). When the recipe is more important Teddlie, C., & Tashakkori, A. (2009). Foundations of mixed methods research: Integrating
than the ingredients a Qualitative Comparative Analysis (QCA) of service innovation quantitative and qualitative approaches in the social and behavioral sciences. Sage.
configurations. Journal of Service Research, 17(2), 134–149. https://doi.org/ Thiem, A. (2013). Clearly crisp, and not fuzzy: A reassessment of the (putative) pitfalls of
10.1177/1094670513513337. multi-value QCA. Field Methods, 25(2), 197–207.
Papamitsiou, Z., Economides, A. A., Pappas, I. O., & Giannakos, M. N. (2018). Explaining Thiem, A., & Dusa, A. (2013). QCA: A package for qualitative comparative analysis. The R
learning performance using response-time, self-regulation and satisfaction from Journal, 5(1), 87–97.
content: An fsQCA approach. Paper Presented at the Proceedings of the 8th International Urry, J. (2005). The complexity turn. Theory, Culture & Society, 22(5), 1.
Conference on Learning Analytics and Knowledge. Vatrapu, R., Mukkamala, R. R., Hussain, A., & Flesch, B. (2016). Social set analysis: A set
theoretical approach to big data analytics. IEEE Access, 4, 2542–2571.

22
I.O. Pappas and A.G. Woodside International Journal of Information Management 58 (2021) 102310

Veeramootoo, N., Nunkoo, R., & Dwivedi, Y. K. (2018). What determines success of an e- Woodside, A. G. (2013). Moving beyond multiple regression analysis to algorithms:
government service? Validation of an integrative model of e-filing continuance Calling for adoption of a paradigm shift from symmetric to asymmetric thinking in
usage. Government Information Quarterly, 35(2), 161–174. data analysis and crafting theory. Journal of Business Research, 66(4), 463–472.
Vink, M. P., & Van Vliet, O. (2009). Not quite crisp, not yet fuzzy? Assessing the Woodside, A. G. (2014). Embrace• perform• model: Complexity theory, contrarian case
potentials and pitfalls of multi-value QCA. Field Methods, 21(3), 265–289. analysis, and multiple realities. Journal of Business Research, 67(12), 2495–2503.
Vink, M. P., & Vliet, O.v. (2013). Potentials and pitfalls of multi-value QCA: Response to Woodside, A. G. (2017). The complexity turn: Cultural, management, and marketing
Thiem. Field Methods, 25(2), 208–213. applications. Berlin: Springer.
Vis, B. (2012). The comparative advantages of fsQCA and regression analysis for Woodside, A. G. (2019). Accurate case-outcome modeling in economics, psychology, and
moderately large-N analyses. Sociological Methods & Research, 41(1), 168–198. marketing. Psychology & Marketing, 36(11), 1046–1061.
Von Bertalanffy, L. (1968). General system theory: Foundations, development, applications. Woodside, A. G. (2016a). Bad to good: Achieving high quality and impact in your research.
New York: Braziller. Emerald Group Publishing.
Woodside, A. G. (2010). Bridging the chasm between survey and case study research: Woodside, A. G. (2016b). The good practices manifesto: Overcoming bad practices pervasive
Research methods for achieving generalization, accuracy, and complexity. Industrial in current research in business. Elsevier.
Marketing Management, 39(1), 64–75.

23

View publication stats