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

net/publication/313774114

Mirror, mirror on the wall: a comparative evaluation of composite-based

structural equation modeling methods

Article in Journal of the Academy of Marketing Science · February 2017

DOI: 10.1007/s11747-017-0517-x

CITATIONS READS

2,314 11,049

5 authors, including:

Joe Hair G. Tomas M. Hult

Manchester Metropolitan University Michigan State University
61 PUBLICATIONS 25,727 CITATIONS 293 PUBLICATIONS 84,620 CITATIONS

SEE PROFILE SEE PROFILE

Christian M. Ringle Marko Sarstedt

Hamburg University of Technology Ludwig-Maximilians-Universität in Munich
308 PUBLICATIONS 195,354 CITATIONS 440 PUBLICATIONS 201,295 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Marko Sarstedt on 17 October 2017.

The user has requested enhancement of the downloaded file.

J. of the Acad. Mark. Sci. (2017) 45:616–632
DOI 10.1007/s11747-017-0517-x

ORIGINAL EMPIRICAL RESEARCH

Mirror, mirror on the wall: a comparative evaluation

of composite-based structural equation modeling methods
Joseph F. Hair 1 & G. Tomas M. Hult 2 & Christian M. Ringle 3,4 &
Marko Sarstedt 4,5 & Kai Oliver Thiele 3

Received: 7 July 2016 / Accepted: 12 January 2017 / Published online: 16 February 2017
# Academy of Marketing Science 2017

Abstract Composite-based structural equation modeling on the basis of composite model data, considering a broad
(SEM), and especially partial least squares path modeling range of model constellations. Results of a large-scale simu-
(PLS), has gained increasing dissemination in marketing. To lation study substantiate that PLS and generalized structured
fully exploit the potential of these methods, researchers must component analysis are consistent estimators when the under-
know about their relative performance and the settings that lying population is composite model-based. While both
favor each method’s use. While numerous simulation studies methods outperform sum scores regression in terms of param-
have aimed to evaluate the performance of composite-based eter recovery, PLS achieves slightly greater statistical power.
SEM methods, practically all of them defined populations
using common factor models, thereby assessing the methods Keywords Composite . Generalized structured component
on erroneous grounds. This study is the first to offer a com- analysis . GSCA . Partial least squares . PLS . SEM .
prehensive assessment of composite-based SEM techniques Simulation . Structural equation modeling . Sum scores
regression
John Hulland served as Area Editor for this article.
Electronic supplementary material The online version of this article Introduction
(doi:10.1007/s11747-017-0517-x) contains supplementary material,
which is available to authorized users.
Structural equation modeling (SEM) has become a
* G. Tomas M. Hult quasi-standard with respect to analyzing cause–effect re-
[email protected] lationships between latent variables. Its ability to model
latent variables while simultaneously taking into account
Joseph F. Hair various forms of measurement error makes SEM useful
[email protected] for a plethora of research questions (e.g., Babin et al.
Christian M. Ringle 2008; Steenkamp and Baumgartner 2000), particularly in
[email protected] the marketing field, which typically focuses on examin-
Marko Sarstedt ing unobservable phenomena such as consumer atti-
[email protected] tudes, perceptions, and intentions.
Kai Oliver Thiele To estimate structural equation models, researchers
[email protected] can draw on two main approaches: Factor-based SEM
(Jöreskog 1973) as carried out by software programs
1
University of South Alabama, Mobile, AL, USA such as Amos, EQS, LISREL, or Mplus, and partial
2
Michigan State University, East Lansing, MI, USA least squares path modeling (PLS; Wold 1974) as im-
3 plemented in software programs such as ADANCO,
Hamburg University of Technology (TUHH), Hamburg, Germany
4
PLS-Graph, SmartPLS, or XLSTAT. A review of all
The University of Newcastle, Callaghan, NSW, Australia
empirical studies published in the 30-year period be-
5
Otto-von-Guericke-University Magdeburg, Magdeburg, Germany tween 1986 and 2015 in the Journal of Marketing
J. of the Acad. Mark. Sci. (2017) 45:616–632 617

(JM) and Journal of the Academy of Marketing Science Satisfaction Index (Eklöf and Westlund 2002). The ACSI has
(JAMS)—the top two marketing journals according to become a key performance indicator for companies and gov-
the 2015 Thomson Reuters Journal Citation Report— ernment agencies, as well as entire industries and sectors.
demonstrates the relevance of these two SEM methods Many follow-up studies use ACSI results, for example, to
for applied research. Our search yielded a total of 193 assess the impact of customer satisfaction on market share
studies that used factor-based SEM, while 53 studies (e.g., Rego et al. 2013) or stock returns (e.g., Fornell et al.
applied PLS. Looking at the cumulative number of stud- 2016), furthering our understanding of customer satisfaction’s
ies that appeared between 1986 and 2015 (Fig. 1) value relevance. PLS has also been used extensively in ser-
shows that the use of both SEM approaches has signif- vices research, for example, to analyze the impact of service
icantly increased over time. The review also shows that failures (e.g., Gelbrich 2010; Heidenreich et al. 2015; van der
PLS use has gained momentum relative to factor-based Heijden et al. 2013) or the drivers of service satisfaction
SEM in recent years. While in the early 90s, the ratio (Brady et al. 2012; Dellande et al. 2004; Hennig-Thurau
of factor-based SEM studies to PLS studies was 4.5, et al. 2006). Finally, the technology acceptance model
this ratio decreased to 2 in the period 2011 to 2015. (Burke 2011), arguably one of the most widely used PLS-
Regressing the number of studies on the linear effects based models for studying user adoption in the information
of time yields significant models (factor-based SEM: systems discipline, has been applied to the different elements
F = 62.409, p < 0.01; PLS: F = 25.836, p < 0.01) of the marketing mix (Haenlein and Kaplan 2011), such as
and time effects (factor-based SEM: β = 0.831, salesperson behavior (Sundaram et al. 2007) or product adop-
t = 7.900, p < 0.01; PLS: β = 0.693, t = 5.083, tion (Kaplan et al. 2007).
p < 0.01). For factor-based SEM, a quadratic effect of Wold (1980) created PLS as a complementary approach to
time is negative but not significant (β = −0.139, factor-based SEM that would emphasize prediction while si-
t = −0.314, p > 0.10), indicating that its use has grown multaneously relaxing the demands on data and specification
linearly over time. In contrast, the quadratic time effect of relationships (e.g., Dijkstra 2010; Jöreskog and Wold
for PLS is positive and significant (β = 0.984, 1982). In early writing, researchers noted that PLS estimation
t = 1.817, p < 0.10), indicating its use has accelerated is Bdeliberately approximate^ to factor-based SEM (Hui and
over time. Wold 1982, p. 127), a characteristic that has come to be known
Several of the PLS-based models featured in this review as the PLS bias (e.g., Chin et al. 2003). A number of
have had a lasting effect on the field. For example, the studies have used simulations to demonstrate the alleged
American Customer Satisfaction Index (ACSI; Fornell et al. PLS bias (e.g., Goodhue et al. 2012a; McDonald 1996;
1996) ranks among the most salient models in studying cus- Rönkkö and Evermann 2013), which manifests itself in
tomer satisfaction (e.g., Anderson and Fornell 2000) and has estimates between the constructs and indicators that are
given rise to related indices such as the European Customer higher, while estimates among the constructs are smaller

Fig. 1 Development of factor-

based SEM and PLS applications
in marketing
618 J. of the Acad. Mark. Sci. (2017) 45:616–632

compared to the prespecified values.1 The parameter estimates (e.g., Henseler et al. 2015). Understanding the methods’ perfor-
will approach what has been labeled the Btrue^ parameter values mance in estimating measurement models is crucial to correctly
obtained from factor models when both the number of indicators appreciate their suitability for evaluating measurement model
per construct and sample size increase (Hui and Wold 1982). quality—a fundamental step in model evaluation practice.
While prior simulation studies have aimed to evaluate the Similarly, prior studies have generally analyzed parameter esti-
performance of composite-based PLS, practically all of the sim- mation bias, while neglecting the methods’ sensitivity to Type I
ulations defined populations using factor models in which the and II errors which are, however, fundamental performance fea-
indicator covariances define the nature of the data (for a notable tures of every statistical method (Goodhue et al. 2012a). In light
exception, see Becker et al. 2013a). That is, these studies used of the above and despite the diversity of literature on composite-
factor-based SEM as the benchmark against which the PLS pa- based SEM, it is reasonable to conclude that our understanding
rameters were evaluated with the assumption that they should be of these methods’ actual performance is still rather limited.
the same. In contrast, in composite model populations, the data Addressing these gaps in research, this study is the first to
are defined by means of linear combinations of indicators. offer a comprehensive assessment of PLS and other composite-
Therefore, prior simulation studies universally estimated and based SEM techniques, considering a broad range of model
evaluated PLS models that were incorrectly specified with regard constellations. Specifically, we examine the relative efficacy of
to the population model (Rigdon 2016; Rigdon et al. 2014). PLS, GSCA, and sum scores regression under a variety of con-
Early research warned about the questionable legitimacy of ditions that researchers typically encounter in practice.2 By effi-
such analyses. More than 25 years ago, Lohmöller (1989) and cacy we refer to the ability of composite-based methods to sup-
Schneeweiß (1991) argued that PLS can be seen as a consistent port a researcher’s need to statistically test hypothesized relation-
estimator of parameters as long as researchers emphasize which ships among constructs and indicators in structural and measure-
type of population parameter they attempt to estimate. More ment models (Goodhue et al. 2012a). More specifically, we test
recently, Marcoulides et al. (2012, p. 717) noted Bthe comparison whether these composite-based SEM techniques have different
of PLS to other methods cannot and should not be applied abilities in terms of: (1) producing accurate path estimates, (2)
indiscriminately^ and referred to any evaluation of PLS vis-á- avoiding false positives (Type I errors), and (3) avoiding false
vis factor-based SEM methods as Bcomparing apples with negatives (Type II errors, related to statistical power). Most im-
oranges^ (p. 725). Similarly, in their comparative study on pa- portantly, our approach correctly assesses the methods’ efficacy
rameter recovery of factor-based SEM, PLS, and generalized on the basis of composite model populations, rather than com-
structured component analysis (GSCA; Hwang 2009), Hwang mon factor model populations. As such, this study addresses
et al. (2010, p. 710) note that their common factor model-based numerous calls for further research in this respect (e.g., Chin
data generation approach Bmay have had an unfavorable effect 2010; Hwang et al. 2010; Marcoulides and Chin 2013).
on the performance of partial least squares and generalized struc- Our results show that PLS and GSCA perform very similar
tured component analysis^ and conclude that Bit appears neces- in terms of parameter accuracy, whereas sum scores regression
sary in future studies to investigate whether a particular data does not perform as well when indicator weights differ on the
generation procedure may influence the relative performance of same construct. Further analyses show that PLS achieves higher
the different approaches.^ While these concerns have been ech- power levels than GSCA. The results directly contradict prior
oed by many other authors (e.g., Henseler et al. 2014; McDonald research regarding the performance of PLS and GSCA using
1996; Rigdon 2012; Tenenhaus 2008), some researchers contin- factor model-based populations, where these methods have
ue to adhere to the reflex-like application of factor model popu- been shown to overestimate measurement model parameters
lations to judge the relative performance of PLS (e.g., McIntosh and underestimate structural model parameters (e.g., Barroso
et al. 2014; Rönkkö et al. 2015) and other composite-based tech- et al. 2010; Chin et al. 2003; Reinartz et al. 2009). When using
niques, such as GSCA, and sum scores regression (e.g., Rönkkö the methods to estimate data from composite model-based pop-
and Evermann 2013; Rönkkö et al. 2016). ulations, the direction of parameter bias is reversed and much
Apart from these parameterization issues, previous simulation smaller, approaching zero as sample size increases. Thus, PLS
studies univocally focused on structural model estimates, but and GSCA are consistent estimators of composite-based
neglected the measurement models. This narrow focus is prob- models, whereas this is not the case with sum scores regression.
lematic, particularly since PLS analyses of common factor– In what follows, we first compare key aspects of factor-based
based populations tend to yield higher measurement model esti- and composite-based SEM, deriving recommendations for their
mates, which likely support the measures’ reliability and validity use. Our comparisons focus on PLS since it is regarded as Bthe

2
Our comparison does not consider consistent PLS (PLSc; Dijkstra 2014;
1
Note that researchers frequently distinguish between latent variables/ Dijkstra and Henseler 2015) that corrects the PLS estimates for attenuation
constructs and composites. We use the term latent variable/construct to refer to mimic common factor models. As our objective is to compare composite-
to the entities that represent conceptual variables in a structural equation based SEM techniques on the basis of composite model data, PLSc is not
model. relevant to our study.
J. of the Acad. Mark. Sci. (2017) 45:616–632 619

most fully developed and general system for path analysis with Bprojection to latent structures^ (Wold et al. 2001, p. 110). That
composites^ (McDonald 1996, p. 240). Next, we describe our is, PLS approximates common factor-based reflective measure-
simulation design, followed by the results description, and inter- ment models (Hui and Wold 1982) and will produce biased
pretation. Based on our findings, we derive guidelines for choos- estimates if the common factor model holds—just like factor-
ing among composite-based SEM methods. Finally, we discuss based SEM produces biased estimates when using the method
the implications of our findings for marketing researchers as well to estimate data generated from a composite model (Sarstedt
as our study’s limitations, along with avenues for future research. et al. 2016). Referring to the proxy nature that underlies all mea-
surement (e.g., Cliff 1983; Rigdon 2012), Sarstedt et al. (2016, p.
4002) recently noted that composites produced by PLS Bcan be
Factor-based SEM and PLS
used to measure any type of property to which the focal concept
refers, including attitudes, perceptions, and behavioral intentions.
While factor-based SEM and PLS share the same objective—
(…) As with any type of measurement conceptualization, how-
determining the relationships among constructs and indicators—
ever, researchers need to offer a clear construct definition and
they take different routes to achieve it. Numerous studies have
specify items that closely match this definition—that is, they
contrasted the two approaches, focusing on aspects such as dis-
must share conceptual unity.^ PLS-based composites can also
tributional assumptions, their efficacy for estimating reflective vs.
be applied as a method for dimension reduction, similar to prin-
formative measurement models, and sample size requirements
cipal components analysis, where the aim is to condense the
(e.g., Chin 1998; Hair et al. 2011; Hair et al. 2012a; Henseler
measures so they adequately cover a conceptual variable’s salient
et al. 2009). More recent research, however, emphasizes the in-
features (Dijkstra and Henseler 2011).
terplay between model estimation and treatment of construct
To estimate the model parameters PLS draws on composites
measures as the key distinguishing factor of the two approaches
formed from the indicators and applies a series of ordinary least
(Henseler et al. 2016; Rigdon 2012; Sarstedt et al. 2016).
squares regressions to estimate partial model structures with the
Factor-based SEM initially divides the variance of each indi-
objective of minimizing the error terms (i.e., the residual vari-
cator into two parts: (1) the common variance, which is estimated
ance) of the endogenous constructs. Since PLS does not esti-
from the variance shared with other indicators in the measurement
mate all model relationships simultaneously, the approach en-
model of a construct, and (2) the unique variance, which consists
ables complex models to be estimated with small sample sizes,
of both the specific and the error variance (Bollen 1989; Rigdon
situations in which factor-based SEM often does not converge,
1998). In estimating the model parameters, factor-based SEM
or develops inadmissible solutions (Henseler et al. 2014). This
only draws on the common variance, assuming that the variance
characteristic has greatly contributed to PLS’s popularity but has
of a set of indicators can be perfectly explained by the existence of
also triggered debates among methodologists (e.g., Marcoulides
one unobserved variable (the common factor) and individual ran-
et al. 2012; Rigdon et al. 2014). As Hair et al. (2013, p. 2) note,
dom error (Spearman 1927; Thurstone 1947). This procedure
Bsome researchers abuse this advantage by relying on extremely
conforms to the measurement philosophy underlying reflective
small samples relative to the underlying population^, and BPLS-
measurement models, which, in essence, is why factor-based
SEM has an erroneous reputation for offering special sampling
SEM has limitations when it comes to estimating formatively
capabilities that no other multivariate analysis tool has.^
specified constructs (e.g., Lee and Cadogan 2013).
Factor-based SEM estimates model parameters based on the
Different from factor-based SEM, PLS considers the total
statistical objective of minimizing the discrepancy between the
variance of the indicators in estimating the model (e.g.,
empirical and model-implied covariance matrices. This difference
Tenenhaus et al. 2005). To do so, PLS linearly combines the
serves as a basis for the χ2-based indices, which allow testing a
indicators to form composites, therefore generally conforming
model’s goodness-of-fit. As such, factor-based SEM focuses on
to the measurement philosophy underlying formative mea-
explanatory modeling, which is Bthe use of statistical models for
surement models (Henseler et al. 2016). However, PLS’s des-
testing causal explanations^ (Shmueli 2010, p. 390). Different
ignation as composite-based refers only to the method’s way
from PLS, the construct scores of factor-based SEM need not to
to represent constructs that approximate the conceptual vari-
be known or assumed at any stage of the estimation process
ables from a theoretical model—the method readily accom-
(Jöreskog 1973). A crucial consequence of this indeterminacy is
modates both measurement model types without identification
that the correlation between a common factor and any variable
issues (Hair et al. 2011).
outside the factor model is itself indeterminate—it may be high or
Composites formed by PLS explicitly serve as proxies of the
low, depending on which set of factor scores one chooses
conceptual variables under investigation (e.g., Henseler et al.
(Schönemann and Steiger 1978). As a consequence, factor-
2016; Rigdon 2012).3 Because of this trait, researchers in other
based SEM offers very limited value for predictive modeling
fields such as chemometrics, refer to the acronym PLS as
(Becker et al. 2013a; Evermann and Tate 2016).
3
Note that constructs in factor-based SEM are also proxies for the conceptual PLS on the other hand focuses on prediction, and as such, is
variables under investigation (Rigdon 2012). concerned with generalization (Shmueli et al. 2016), which is the
620 J. of the Acad. Mark. Sci. (2017) 45:616–632

Table 1 Comparison between factor-based SEM and PLS-SEM

Criterion Factor-based SEM PLS Sample references

Model estimation
Estimation process Estimates model parameters Estimates model parameters so that Jöreskog (1973); Jöreskog and
so that the discrepancy between the explained variance of the Wold (1982); Tenenhaus
the estimated and sample endogenous constructs / indicators et al. (2005)
covariance matrices is minimized is being maximized
Convergence Defines convergence as the Defines convergence as the point at Jöreskog (1973);
increase / decrease in the function which no substantial difference Tenenhaus et al. (2005)
value beyond a certain threshold occurs in the model estimations
from one iteration to the next
Estimation objective Explanatory modeling Predictive modeling Rigdon (2012); Shmueli (2010);
Shmueli et al. (2016)
Treatment of construct Treats constructs as common Treats constructs as composites–the Henseler et al. (2016);
measures factors–only common variance total variance is used to estimate Rigdon (2012);
is used to estimate model model parameters Sarstedt et al. (2016)
parameters
Latent variable scores Indeterminate Explicitly Estimated Becker et al. (2013a);
McDonald (1996);
Schönemann and Steiger (1978)
Model specification
Model complexity Requires models of small to Accommodates all kinds of model Hair et al. (2012a)
moderate complexity complexity including large models
with many constructs and
indicators
Measurement model Handles reflectively specified Handles reflectively and formatively Diamantopoulos and Riefler
specification constructs; limitations in handling specified constructs (2011);
formatively specified constructs Sarstedt et al. (2016)
Data
Distributional assumptions Standard maximum likelihood Non-parametric; makes no Cassel et al. (1999); Reinartz et al.
estimation requires multivariate distributional assumptions (2009); Rigdon (2016)
normality but numerous robust
estimators exist
Sample Size Generally requires relatively Produces parameter estimates with Goodhue et al. (2012a);
high sample sizes to produce small sample sizes Reinartz et al. (2009);
(robust) parameter estimates Rigdon (2016)
Fit Measure Offers goodness-of-fit statistics Several goodness-of-fit criteria Henseler et al. (2016);
have been proposed but their Henseler et al. (2014);
adequacy and interpretation Henseler and Sarstedt (2013)
remains subject for future research

ability to predict sample data, or preferably out-of-sample data because these indices are not optimized in the estimation
(e.g., Shmueli 2010). Complementing the range of metrics procedure.^
for assessing a model’s predictive relevance (e.g., Hair Factor-based SEM and PLS are not interchangeable but
et al. 2017), researchers have advocated several PLS-based rather complementary, a fact that was stressed by the method’s
goodness-of-fit indices such as the standardized root mean originators (Jöreskog and Wold 1982). Researchers need to
square residual (SRMR) or the root mean square residual apply the SEM approach that best suits their research objec-
covariance. However, literature casts doubt on whether mea- tive, measurement properties, and model setup. Table 1 sum-
sured fit—as understood in a factor-based SEM context—is a marizes the differences between factor-based SEM and PLS
relevant concept for PLS (Hair et al. 2017; Lohmöller 1989; along with sample references discussing each criterion.
Rigdon 2012). Different from factor-based SEM, the dis-
crepancy between the empirical and the model-implied co-
variance matrices, which serves as the basis for fit indices Simulation design and model estimation
such as SRMR, is a byproduct of the PLS algorithm and not
explicitly minimized. Correspondingly, Lohmöller (1989, p. An extensive body of literature provides the technical underpin-
222) long ago noted that these goodness-of-fit measures nings of PLS (e.g., Lohmöller 1989; Tenenhaus et al. 2005; Wold
Bshould not be used for a decision about the fit of the model, 1982) and GSCA (e.g., Henseler 2012; Hwang 2009; Hwang
J. of the Acad. Mark. Sci. (2017) 45:616–632 621

Fig. 2 Simulation model

et al. 2010). GSCA is an alternative method for path analysis & The number of indicators per latent variable: 2, 4, 6, 8 (4
with composites. The method replaces factors by exact linear factor levels)
combinations of observed variables, employs a least squares cri- & The standardized indicator weights for different numbers of
terion to estimate model parameters, and retains the advantages indicators per latent variable: equal, unequal (2 factor levels)4
of PLS (e.g., less restrictive distributional assumptions and no & The data distribution: normal, non-normal (i.e., diff normal),
improper solutions). Model estimation using sum scores regres- and extremely non-normal (i.e., log normal) (3 factor levels)5
sion (sum scores) is similar to PLS, but differs in that the sum & The sample size: 100, 250, 500, 1000, 10,000 (5 factor levels)
scores approach assumes equal (typically unit) indicator weights,
whereas weightings in PLS represent the partialized effect of the This is a full factorial design with 4·2·3·5 = 120 factor level
indicators on their corresponding construct, and thus control the combinations. To obtain stable average outcomes for our anal-
individual effects of all the other construct indicators (e.g., yses, we conducted 300 replications of each factor-level com-
Goodhue et al. 2012b; Tenenhaus 2008). We do not elaborate bination resulting in the generation of 120·300 = 36,000
further on the methodological underpinnings of these techniques, datasets. For the three methods under investigation (i.e.,
but refer to the relevant literature. PLS, GSCA, and sum scores), this simulation study draws
Our simulation study considers the path model shown in on a total number of 108,000 model estimations.
Fig. 2, which Reinartz et al. (2009) also used in their compre- For the data generation, we use the pre-specified measurement
hensive simulation study on the comparison of PLS and factor- model (Table A1) and structural model (Fig. 2) coefficients to
based SEM. The model mirrors the structure of the ACSI model determine the indicators’ population correlation matrix (Becker
(e.g., Fornell et al. 1996) and the European Customer et al. 2013a; Ringle et al. 2014; Schuberth et al. 2016). Its
Satisfaction Index model (e.g., Eklöf and Westlund 2002) Cholesky decomposition allows us to extract the indicator data
whose estimations routinely draw on PLS. We also chose this for a pre-specified number of observations and the desired data
type of model, because it reflects the typical degree of complex- distribution (see the Online Appendix for further information).
ity used in composite-based structural equation modeling in the For the data generation and the (partial) regression model estima-
marketing discipline (Hair et al. 2012a). tions by means of the sum scores approach, we use the statistical
For the simulation study, we selected low (i.e., 0.15; γ2, γ3, software R (R Core Team 2016). PLS estimations draw on the
β6), medium (i.e., 0.30; β3), and high (i.e., 0.50; γ1, γ6, β1, β2, semPLS (Monecke and Leisch 2012) package, while the GSCA
β4, β5) pre-specified values of standardized path coefficients estimations rely on the ASGSCA (Romdhani et al. 2014) pack-
(Fig. 2). As a result of corresponding calls in the literature age. Moreover, the R software’s snowfall package (Knaus 2013)
(e.g., Marcoulides et al. 2012a), we extended the simulation allowed us exercising parallel computing on several hundred pro-
study by adding a construct (ξ2) with two null paths (γ4 and cessors of a high-performance computing cluster.
γ5) to the original model.
Our choice of manipulated factors and their factor levels fol-
lows prior research (e.g., Becker et al. 2013b; Chin et al. 2003; Results
Henseler 2012; Hwang et al. 2010; Vilares and Coelho 2013).
These conditions compare well with those seen in the marketing In order to draw a comprehensive picture of each approach’s
field as evidenced in prior reviews of PLS use (e.g., Hair et al. efficacy for estimating path models with composite model
2012a). Specifically, we manipulate the following factors: data, analyses of our results address the following aspects
related to the methods’ performance: (1) parameter accuracy,
(2) the direction of the estimation bias, and (3) statistical
4 5
Table A1 in the Online Appendix shows the indicator weights for different For further details about the non-normal data, see the additional information
numbers of indicators. on the data generation in the Online Appendix.
622 J. of the Acad. Mark. Sci. (2017) 45:616–632

power and Type I errors. Prior to these analyses, we examined contrary, the sample size has a significant bearing on the
whether the candidate solutions for the iterations of each results—MAE values drop considerably as sample size in-
method tend to get closer to the desired solution (i.e., conver- creases. When comparing PLS and GSCA, we find that
gence), which is a key area of concern when using iterative GSCA always yields slightly lower MAE values across all
algorithms (Henseler 2010). All methods examined in this simulation conditions. These differences become smaller
study (i.e., PLS, GSCA, and sum scores) converge across all with higher sample sizes.
factor level combinations, providing support for their ability to Complementing our assessment of the methods’ pa-
provide proper solutions. rameter accuracy in the measurement models, the fol-
lowing analyses address their accuracy when estimating
the relationships between latent variables in the structur-
Parameter accuracy
al model. The presentation of MAE values for the struc-
tural model in Fig. 4 follows the same principle as in
In order to be useful, an SEM method should produce
Fig. 3 with columns and rows showing the MAE values
parameter estimates that are very similar to the pre-
for different types of weights and numbers of indicators.
specified parameters of the artificially generated
The x-axis maps the sample size, while the y-axis shows
datasets. The mean absolute error (MAE; e.g., Hulland
each method’s MAE values for the structural model
et al. 2010) is a commonly used quantity to measure
(MAE SM). An MAE value for the structural model
how close estimates and pre-specified parameters are.
of, for example, 0.05 indicates that the absolute devia-
The MAE is defined as
tion between estimated and pre-specified path coeffi-
1 t ^ cients is on average 0.05 units. Different from our pre-
MAE ¼ ∑ θj −θ j j; ð1Þ
t j¼1 vious assessment, the charts also display sum scores
since this approach entails an explicit estimation of
where t equals the number of parameters, θj is the pre- structural model relationships.
specified parameter j, and θ j is the parameter estimate Similar to the MAE results of the measurement models, we
in any replication. Thus, the lower the MAE the higher find that the MAE values of the structural model only margin-
a method’s parameter accuracy. An MAE for the mea- ally vary across different numbers of indicators per measure-
surement models of, for example, 0.05 indicates that the ment model and the type of weights (i.e., equal vs. unequal).
absolute deviation between estimated and pre-specified The MAE values produced by PLS and GSCA are almost
weights is on average 0.05 units. identical, except for small sample sizes of 100 where PLS is
Figure 3 displays the MAE results for the measurement slightly more accurate. As sample size increases, parameter
models in a series of eight charts.6 The left column of charts accuracy generally improves for all three methods, but this
shows the results for unequal weights, whereas the right col- improvement is less pronounced for the sum scores approach
umn shows those for equal weights. The four rows display the compared to PLS and GSCA.
results for two, four, six, and eight indicators per measurement Further contrasting the methods’ performance shows that
model. Each chart in Fig. 3 illustrates the PLS and GSCA PLS and GSCA generally have higher parameter accuracy
results for different sample sizes (i.e., 100, 250, 500, 1000, than the sum scores approach when indicator weights are un-
and 10,000) on the x-axis, and the MAE values of the mea- equal. The differences between PLS and GSCA on the one
surement models (MAE MM) on the y-axis. When using the hand and the sum scores approach on the other are particularly
sum scores approach, no measurement model estimation oc- pronounced for larger sample sizes; for example, with four
curs as this approach draws on equal weights by design. Thus, indicators and 10,000 observations the sum scores approach’s
assessing sum scores’ performance in measurement models is MAE values are four times as high as those of PLS and
not meaningful. GSCA. Only in two situations with unequal weights
When comparing the charts in Fig. 3 across rows and (when measurement models have six or eight indicators
columns, we find that the MAE values vary only marginally and 100 observations), does the sum scores approach
depending on the number of indicators per measurement perform slightly better than PLS and GSCA. As sample
model and the type of weights (i.e., equal vs. unequal). sizes increase, however, its performance deteriorates rel-
MAE values are lower for fewer indicators per measurement ative to the other approaches.
model but the differences are not pronounced. On the Not surprisingly, a different picture emerges when indicator
weights are equal. In this situation, the sum scores approach
performs better or as well as PLS and GSCA across all simu-
6 lation conditions. The differences between the methods are
As the analyses show only marginal differences between normal and non-
normal data, the following results presentations use the joint outcomes of the marginal, however, especially for sample sizes of 250 and
different data distribution types considered in this simulation study. higher, independent of the number of indicators. Only for
J. of the Acad. Mark. Sci. (2017) 45:616–632 623

Fig. 3 Mean absolute error (MAE) in the measurement models (MM). 10,000) on the x-axis, and the MAE of the measurement models (MM) on
Note: Charts in the left (right) column show the results for unequal (equal) the y-axis. An MAE for the measurement models of, for example, 0.05
indicator weights; the four rows of charts show the results for 2, 4, 6, and indicates that the absolute deviation between estimated and pre-specified
8 indicators per measurement model. Each chart illustrates the results of weights is on average 0.05 units
PLS and GSCA for different sample sizes (i.e., 100, 250, 500, 1000, and
624 J. of the Acad. Mark. Sci. (2017) 45:616–632

Fig. 4 Mean absolute error (MAE) in the structural model 1000, and 10,000) on the x-axis, and the MAE of the structural
(SM). Note: Charts in the left (right) column show the results model (SM) on the y-axis. An MAE for the structural model
for unequal (equal) indicator weights; the four rows of charts of, for example, 0.05 indicates that the absolute deviation be-
show the results for 2, 4, 6, and 8 indicators per measurement tween estimated and pre-specified path coefficients is on aver-
model. Each chart illustrates the results of PLS, sum scores, age 0.05 units
and GSCA for different sample sizes (i.e., 100, 250, 500,
J. of the Acad. Mark. Sci. (2017) 45:616–632 625

100 observations and when measurement models have six or parameter values. Our assessment of each method’s direction
eight indicators, differences between the sum scores approach of the estimation bias builds on the mean error (ME), which is
and the other two methods are more pronounced. defined analogous to the MAE (Eq. 1), but not in terms of an
To further our assessment of parameter accuracy, and in absolute value:
order to facilitate the comparison of the results with those of
1 t ^
Reinartz et al. (2009), we also computed the mean absolute ME ¼ ∑ θj −θ j Þ: ð3Þ
relative error (MARE) defined as t j¼1

1 t θ^j −θ j Here, t equals the number of parameters, θj is the pre-

MARE ¼ ∑ ð2Þ specified parameter j, and θ j is the parameter estimate in any
t j¼1 θ j
replication. The goal is to obtain a low ME close to zero,
where t equals the number of parameters, θj is the pre- which substantiates that a systematic underestimation or over-
specified parameter j, and θ j is the parameter estimate in any estimation is not an issue. A substantial positive (negative)
replication. This equation is similar to the MAE (Eq. 1) in that ME confirms that the estimated parameter values are system-
the MARE represents the MAE relative to the pre-specified atically larger (smaller) than the pre-specified values of this
parameter values. A MARE value of, for example, 0.10 indi- simulation study. An ME for the measurement models of, for
cates that the pre-specified coefficients were missed by 10% example, −0.05 implies that the estimated values are on aver-
on average. Thus, if a pre-specified parameter has a relatively age 0.05 units lower than the pre-specified parameters.
low value, a comparatively small MAE entails a relatively Analyzing the ME of the weights estimates in the measure-
rather large MARE.7 ment models (Fig. A3 in the Online Appendix), we find that
The findings resulting from the MARE assessment parallel PLS and GSCA perform very well with practically no system-
those from the MAE analysis.8 While the types of weights and atic bias across most simulation conditions. Both methods
number of indicators have a limited impact, the MARE im- show slight underestimation tendencies, which are more pro-
proves considerably for increasing sample sizes in both, the nounced when sample sizes are small (i.e., 100), particularly
measurement and structural models (Figures A1 and A2 in the for six and eight indicator models, independent of whether the
Online Appendix). For example, for 100 observations the indicator weights are equal or not. Even in these settings,
methods’ MARE values in the structural model (Fig. A2) are however, the ME values never exceed 0.025 units, which is
around 0.25, and decrease to 0.15 for 250 observations, 0.10 very low in absolute terms. Thus, as sample size increases the
for 500 and 1000 observations, and finally, less than 0.05 for ME diminishes, approaching zero for 10,000 observations.
10,000 observations—except for the sum scores approach With regard to the ME in the structural model (Fig. A4 in
when indicator weights are unequal, particularly for large the Online Appendix), PLS and GSCA show slight overesti-
sample sizes. Most notably, the MARE values in Reinartz mation tendencies. The overestimation is more pronounced
et al. (2009) show a similar pattern, but at a considerably when sample sizes are small (i.e., 100), particularly for six
higher level. For example, regardless of the number of indica- and eight indicator models. But with larger sample sizes, the
tors, MARE values in Reinartz et al. (2009) never drop below magnitude of the overestimation becomes trivial, especially
0.10, even for a sample size of 10,000. This difference under- for sample sizes of 250 and higher.
lines the importance of aligning the data generation procedure PLS and GSCA clearly outperform the sum scores ap-
with the assumptions underlying the composite-based SEM proach when indicator weights are unequal. The sum scores
methods such as PLS. approach always underestimates the path coefficients, partic-
ularly when the measurement models only have two or four
Direction of the estimation bias indicators. In this situation, the sum scores approach’s ME
values peak at 0.046 units. The pronounced underestimation
The previous analysis indicated the degree of estimation bias. tendency for the sum scores approach when indicator weights
The next step is to evaluate the direction of this bias. By doing are unequal, however, translates into a slight overestimation
so, we evaluate whether the methods exhibit a systematic un- tendency when indicator weights are equal. In this situation,
derestimation or overestimation tendency of the pre-specified the approach always yields ME values equal to or lower than
PLS and GSCA.
7
For example, for the condition with 500 observations, two indicators with
equal weights of 0.625, PLS yields a MAE value of 0.05814 in the measure-
ment models, which translates into an MARE value of 0.093. On the contrary, Statistical power and Type I errors
a very similar MAE value of 0.06029 for the condition with 500 observations,
eight indicators with equal weights of 0.25 translates into a MARE value of
The next analyses addresses the question of how well the
0.241.
8
Note that the MARE is not defined for the two null paths γ4 and γ5 (Fig. 2). methods reveal relationships with low (β = 0.15), medium
Hence, we did not include these two paths in the MARE computations. (β = 0.30), and high (β = 0.50) population effect sizes. This
626 J. of the Acad. Mark. Sci. (2017) 45:616–632

assessment is particularly important to applied researchers, Even though a number of simulation studies have focused on
because any failure to correctly depict such relationships as evaluating the performance of composite-based SEM tech-
significant may lead to erroneous conclusions with adverse niques, most notably PLS, practically all of these studies began
consequences for theory building and testing (Cohen 1988). by assuming that the common factor model represents the truth
To test for each method’s statistical power, we calculated the (Becker et al. 2013a; Rigdon 2016). Hence, most previous sim-
relative occurrence of significant paths found by each method ulations evaluated models that were incorrectly specified rela-
and averaged the results with paths of identical population tive to the population and, thereby, introduced a substantial
effect size (Reinartz et al. 2009). For example, in terms of research design bias. Addressing corresponding calls in prior
low effect sizes, the analysis considers the model parameters research (Chin 2010; Hwang et al. 2010; Marcoulides and Chin
β6, γ2, and γ3 (Fig. 2). 2013; Marcoulides et al. 2012), this study is the first to provide
The results show that PLS performs best in terms of statis- a systematic assessment of and comparison between PLS,
tical power, followed by GSCA and the sum scores approach GSCA, and the sum scores approach based on composite model
(Table A2 in the Online Appendix). For low effect sizes and data that is consistent with the functional principles of the
small sample sizes, PLS yields power values of approximately estimators.
0.70, whereas the sum scores approach and GSCA have pow- Our results show that PLS and GSCA are consistent esti-
er values of 0.60 and 0.40, respectively. The differences be- mators of measurement and structural model coefficients
come less pronounced as sample sizes increase and ultimately when the underlying population is composite model-based—
diminish for 1000 observations or more. This pattern of results approaching the pre-specified values as sample sizes increase.
also holds for medium and high effect sizes, with the only However, both methods show slight tendencies to underesti-
difference being that the methods’ power values already align mate measurement model parameters and small overestima-
at a sample size of 250. tion tendencies of structural model parameters. In fact, in ab-
Mistakes in the interpretation of results and false conclu- solute terms the parameter biases in measurement models are
sions can also occur if the methods erroneously render a null quite pronounced for a small sample size of 100, yielding
relationship as significant (i.e., Type I error; false positives). MAE values above 0.10, regardless of the number of indica-
For this reason, we analyze the null relationships (γ4 and γ5) in tors and weights pattern (i.e., equal or unequal). In the extreme
the structural model (Fig. 2) and determine the probability of case (a condition with 100 observations and eight indicators
rendering these relationships significant—hence, low proba- with unequal weights at relatively low levels), PLS and
bilities indicate small Type I errors. We find that all three GSCA show relative deviations of measurement model
methods have very low Type I error rates of less than 10% parameter estimates (as expressed through MARE) of
(Table A2). This especially holds for PLS, which has the low- 60% and more. For larger sample sizes, however, the
est error rates in most of the factor level constellations. relative deviations decrease substantially, dropping be-
low 10% for 10,000 observations.
Compared to the measurement models, structural model
Discussion and conclusions MAE values are lower when the sample size is small (i.e.,
100 observations), with values of around 0.075. For 250 ob-
Summary of results servations, structural model MAE values drop below 0.05,
diminishing further as sample size increases. Similarly, the
Along with the increasing prominence of composite-based relative deviations of parameter estimates (i.e., the MARE
SEM methods, particularly in the marketing field (Hair et al. values) in the structural model quickly decrease for larger
2012a; Henseler et al. 2009), researchers have started calling for sample sizes. As sample size increases to 10,000, the relative
their emancipation from factor-based SEM methods. As deviation falls below 5%, providing support for the consisten-
Rigdon (2012, p. 353) notes with regard to PLS, Binstead of cy of the PLS and GSCA methods. This finding does not hold
emphasizing how it is ‘like’ factor-based SEM but with advan- for the sum scores approach, however, which exhibits MARE
tages and disadvantages across different conditions, PLS path values of 7.5% and higher for large sample sizes when indi-
modeling should celebrate its status as a purely composite- cator weights are unequal. As unequal indicator weights are
based method.^ To address this call and further the emancipa- more the rule than the exception in applied research, these
tion of composite-based SEM from its factor-based sibling, results cast substantial doubt on prior research that called for
researchers require a more nuanced understanding of these using sum scores over PLS or GSCA, but in making this
methods’ performances based on strong simulation results recommendation relied on factor model-based data (Rönkkö
(Henseler et al. 2014; Sarstedt et al. 2016). However, reviewing and Evermann 2013).
prior simulation studies gives rise to substantial concern. Similarly, our results demonstrate that prior allegations
Any simulation study begins by specifying a population—a suggesting that PLS is not suitable for disclosing null effects,
true state from which data are sampled (Paxton et al. 2001). without formally testing this claim (Rönkkö and Evermann
J. of the Acad. Mark. Sci. (2017) 45:616–632 627

2013; Rönkkö et al. 2016), or drawing on factor model-based estimating factor model-based populations. For example, in
data (Rönkkö et al. 2015) are without merit. Our study reveals their study, factor-based SEM’s parameter estimates deviate
that model estimation using PLS, GSCA, and sum scores between 30% and 65% for small sample sizes (Fig. 2 in
yields low Type I error rates of around 5%, regardless of Reinartz et al. 2009), depending on the number of indicators
sample size and measurement model set-up. Additionally, in the measurement models, while in this study deviations pro-
when statistical power is considered, our results indicate that duced by PLS are around 25% (Fig. A2). Clearly, when using
PLS should be the preferred method vis-à-vis GSCA. While PLS on correctly specified populations in simulation studies,
both methods achieve power values close to 1.00 for sample the biases are smaller than when using factor-based SEM on
sizes of 500 or more, PLS excels at small sample sizes, par- correctly specified populations.
ticularly when effect sizes are low. In this situation, PLS’s Our results also cast doubt on the widely held belief that
power is approximately 75% higher than that of GSCA. PLS is universally applicable for small sample size situations
(e.g., Chin and Newsted 1999; Haenlein and Kaplan 2004;
Implications for marketing researchers Hair et al. 2011). While this study is not the first to question
PLS’s performance in this regard (Goodhue et al. 2012a;
None of the composite-based SEM approaches dominates in Rönkkö and Evermann 2013), it is the first to do so on the
all criteria. Instead, each approach has certain strengths and grounds of an approriately specified (i.e., composite model)
weaknesses, which make it suitable when the focus is on pa- population. However, this conclusion primarily holds for the
rameter accuracy in the measurement models, the structural measurement models, where the relative deviations of param-
model, or on statistical power. Against this background, Fig. 5 eter estimates are pronounced. PLS’s performance is much
visualizes the recommendations to marketing researchers that less affected in the structural model where the relative devia-
result from the simulation study. tions for small sample sizes are comparable to those produced
When the aim is to maximize parameter recovery accuracy, by factor-based SEM for sample sizes of 250 to 500 when
researchers should draw on PLS or GSCA. The latter method estimating factor model-based data (Reinartz et al. 2009).
has slight advantages in terms of estimating measurement The implications of this aspect of our analysis are twofold.
models, but PLS offers higher accuracy in the estimation of First, these results give confidence that the ACSI model
structural model parameters. Only when sample size is small (Fornell et al. 1996) or the technology acceptance model
(i.e., 100) and measurement models are complex (i.e., 6 or (Davis 1989) and its extensions (Venkatesh et al. 2003), which
more indicators), the sum scores approach performs favorably have been validated using hundreds or thousands of observa-
in terms of structural model parameter accuracy while having tions, offer valid results on both measurement and structural
the same Type I error rate and statistical power as PLS. In this model levels (Anderson et al. 2004). Researchers frequently
situation, the bias in indicator weights due to sampling vari- interpret the latent variable scores from these models and use
ability is higher than the bias resulting from the assignment of the scores to run between-model mean-comparisons. For ex-
equal weights. The lower average weights that occur when ample, drawing on latent variables scores from PLS analyses,
using many indicators amplify this effect. Except for this spe- Hult et al. (2017) recently used ACSI scores to investigate the
cific condition, however, our results advise against the use of extent to which managers’ perceptions of the levels and
the sum scores approach. When the primary focus is disclos- drivers of their customers’ satisfaction and loyalty align with
ing the significance of structural model effects, researchers that of their actual customers. Similarly, a multitude of studies
should choose PLS in light of the method’s high statistical on the marketing-finance interface draws on ACSI scores de-
power compared to the sum scores approach and GSCA. rived from PLS-based indicator weighting (e.g., Angulo-Ruiz
Our results in terms of parameter recovery directly contradict et al. 2014; Fornell et al. 2016; Habel and Klarmann 2015; Lee
previous findings regarding the performance of composite- et al. 2015) to offer evidence for the value relevance of cus-
based PLS in factor model-based populations, where the meth- tomer satisfaction and related marketing asset variables. Our
od has been shown to overestimate measurement model param- results regarding the parameter accuracy of indicator weights
eters while underestimating structural model parameters (e.g., provide support for the validity of such analyses, given that
Barroso et al. 2010; Chin et al. 2003; Reinartz et al. 2009). In the sample size is sufficiently large, as it is the case with the
composite model-based populations, the direction of parameter original ACSI data. However, for small sample sizes (e.g.,
bias is reversed and much smaller. For example, PLS’s relative 100), indicator weights estimates show pronounced biases that
deviations in structural model estimates are always smaller than may compromise the validity of latent variable scores.
those obtained when using the method to estimate factor model- Second, caution needs to be exercised when interpreting
based data, as in Reinartz et al.’s (2009) study, which used a PLS or GSCA results on an item level when sample size is
highly similar simulation design. Interestingly, the very same small since the biases produced in this situation potentially
holds when comparing PLS’s performance in this study with cast doubt on any prioritization on the grounds of indicator
that of factor-based SEM in Reinartz et al.’s (2009) study when weights. This result is particularly relevant for studies
628 J. of the Acad. Mark. Sci. (2017) 45:616–632

Fig. 5 Guidelines for choosing

among composite-based SEM
methods

implementing formative measurement models, which contin- quarter of all models. In our review of PLS use in JM and
ue to feature prominently in marketing research (e.g., Ranjan JAMS between 1986 and 2015, the majority of studies
and Read 2016; Rubera et al. 2016; Wolter and Cronin 2016). (64.15%) incorporated at least one model with a sample size
Similarly, findings from an importance-performance map of less than 250; seven of these 53 PLS studies (13.21%) had
analyses (Ringle and Sarstedt 2016), which contrast each at least one model with less than 100 observations, with Green
item’s total effect on a target construct (importance) with the et al. (1995) using a sample size as low as 39. Different from
average latent variable scores (performance) are prone to bi- factor-based SEM, PLS provides measurement model esti-
ased indicator weights with small sample sizes. Considering mates even when the sample size is very small. But authors,
the prominence of composite-based SEM when sample sizes reviewers and editors should question the value of those
are small, this caveat holds for many studies. For example, estimates in small sample size situations and rather focus on
Hair et al.’s (2012b) review of PLS use in strategic manage- the structural model outcomes. However, when the aim is to
ment shows that more than half of all models estimated with interpret measurement model results, we concur with Rigdon
PLS draw on sample sizes of 100 or less. In other fields such (2016, p. 600) who recently noted that Bwith respect to both
as marketing (Hair et al. 2012a), management information composite-based and factor-based approaches to SEM, if sam-
systems (Ringle et al. 2012), and supply chain management ple size is small, the best course is to get more data.^ In B-to-B
(Kaufmann and Gaeckler 2015), this is the case for about a research, for example, large samples often are not available. In
J. of the Acad. Mark. Sci. (2017) 45:616–632 629

such situations, when populations are small and/or data is thereby broadening our understanding of the methods’ perfor-
difficult to obtain, the application of PLS with smaller samples mance in composite model-based populations.
denotes a viable attempt to advance knowledge in these areas. Finally, while the parameter bias that PLS produces when
estimating common factor models has been extensively debat-
Future research avenues ed in the literature (e.g., Goodhue et al. 2012a; Reinartz et al.
2009; Rönkkö and Evermann 2013), the bias that factor-based
Our results in terms of parameter recovery in the measurement SEM produces when estimating composite models has not yet
models reinforce Rigdon’s (2012, p. 353) call that methodol- been explored in depth. Initial results show that this bias can
ogists Bshould work to complete and validate a purely be substantial (Sarstedt et al. 2016), but future research should
composite-based approach to evaluating modeling results.^ aim at evaluating in greater detail how data and model con-
Using composite-based SEM instead of factor-based SEM stellations affect factor-based SEM’s parameter accuracy in
does not mean rejecting rigor. Instead, it means defining rigor this context. Understanding this bias would help researchers
in composite terms (Rigdon 2014, 2016). With regard to PLS, to clarify the consequences of mistakenly using factor-based
such an evaluation approach should take the method’s predic- SEM on composite model-based populations. This endeavor
tion orientation into account (Shmueli et al. 2016). Further would be particularly fruitful since researchers increasingly
advancing the set of prediction-oriented evaluation criteria is oppose adherence to the common factor model in SEM anal-
a particularly promising area of future research (see Becker yses (e.g., Rigdon 2012; Schönemann and Wang 1972;
et al. 2013a; Evermann and Tate 2016; Sarstedt et al. 2014; Treiblmaier et al. 2011). For example, among 72 articles pub-
Schubring et al. 2016). lished during 2012 in what Atinc et al. (2012) consider the
With respect to the recovery of composite model parame- four leading management journals (Academy of Management
ters, subsequent studies should extend the simulation design Journal, Journal of Applied Psychology, Journal of
by considering more complex model structures, such as hier- Management, and Strategic Management Journal) that tested
archical component models (e.g., Becker et al. 2012), interac- one or more common factor model(s), fewer than 10%
tion terms (e.g., Henseler and Fassott 2010), mediating effects contained a common factor model that did not have to be
(Hair et al. 2017), and nonlinear effects (e.g., Rigdon et al. rejected. While these results do not necessarily imply that
2010). While the use of these modeling elements has recently the composite model predominantly holds in practice, under-
become more en vogue, little is known about the efficacy of standing the consequences of using factor-based SEM on such
different composite-based SEM methods when estimating models would contribute to broadening our understanding of
such model structures. Follow-up simulation studies should this method type, which is still the mainstay for analyzing
also complement our design, which focused on parameter ac- cause–effect models in marketing.
curacy as well as Type I and II errors, by considering the
methods’ predictive power and different weighting modes
(i.e., correlation versus regression weights; Becker et al. Acknowledgements Earlier versions of the manuscript have been pre-
2013a). Future research should also include a broader set of sented at the 2015 Academy of Marketing Science Annual Conference
composite-based SEM methods such as regularized general- held in Denver, Colorado, and the 2nd International Symposium on
Partial Least Squares Path Modeling: The Conference for PLS Users
ized canonical correlation analysis (Tenenhaus and Tenenhaus held in Seville, 2015. The authors would like to thank Jan-Michael
2011), best fitting proper indices (Dijkstra and Henseler Becker, University of Cologne, Jörg Henseler, University of Twente,
2011), and the extended PLS algorithm, which Lohmöller and Rainer Schlittgen, University of Hamburg, for their support and help-
(1989) recommended to overcome some of the original PLS ful comments when developing the simulation study and its data genera-
tion in order to improve earlier versions of the manuscript. Even though
method’s restrictions in terms of model specifications. For this research does not explicitly refer to the use of the statistical software
example, the extended PLS algorithm allows for imposing SmartPLS (http://www.smartpls.com), Ringle acknowledges a financial
restrictions on model parameters (e.g., the orthogonalization interest in SmartPLS.
of latent variables to circumvent collinearity issues) and for
assigning an indicator to multiple constructs. Researchers
have largely overlooked the extended PLS algorithm; conse- References
quently, its performance has not yet been examined across a
broader set of data and model constellations. Anderson, E. W., & Fornell, C. G. (2000). Foundations of the American
Complementing our empirical perspective, future studies customer satisfaction index. Total Quality Management, 11(7), 869–882.
should aim at providing analytical support for the methods’ Anderson, E. W., Fornell, C. G., & Mazvancheryl, S. K. (2004).
consistency in estimating measurement and structural model Customer satisfaction and shareholder value. Journal of
Marketing, 68(4), 172–185.
parameters. Providing such an analytical rationale would com- Angulo-Ruiz, F., Donthu, N., Prior, D., & Rialp, J. (2014). The financial
plement prior research on the consistency of PLS estimates in contribution of customer-oriented marketing capability. Journal of
common factor model populations (Hui and Wold 1982), the Academy of Marketing Science, 42(4), 380–399.
630 J. of the Acad. Mark. Sci. (2017) 45:616–632

Atinc, G., Simmering, M. J., & Kroll, M. J. (2012). Control variable use Dijkstra, T. K. (2010). Latent variables and indices: Herman Wold’s basic
and reporting in macro and micro management research. design and partial least squares. In V. Esposito Vinzi, W. W. Chin, J.
Organizational Research Methods, 15(1), 57–74. Henseler, & H. Wang (Eds.), Handbook of partial least squares:
Babin, B. J., Hair, J. F., & Boles, J. S. (2008). Publishing research in concepts, methods and applications (pp. 23–46). Berlin: Springer.
marketing journals using structural equation modeling. Journal of Dijkstra, T. K. (2014). PLS' Janus face – response to professor Rigdon's
Marketing Theory and Practice, 16(4), 279–285. ‘rethinking partial least squares modeling: in praise of simple
Barroso, C., Carrión, G. C., & Roldán, J. L. (2010). Applying maximum methods’. Long Range Planning, 47(3), 146–153.
likelihood and PLS on different sample sizes: studies on Dijkstra, T. K., & Henseler, J. (2011). Linear indices in nonlinear struc-
SERVQUAL model and employee behavior model. In V. Esposito tural equation models: best fitting proper indices and other compos-
Vinzi, W. W. Chin, J. Henseler, & H. Wang (Eds.), Handbook of ites. Quality & Quantity, 45(6), 1505–1518.
partial least squares: concepts, methods and applications (pp. 427– Dijkstra, T. K., & Henseler, J. (2015). Consistent partial least squares path
447). Berlin: Springer. modeling. MIS Quarterly, 39(2), 297–316.
Becker, J. M., Klein, K., & Wetzels, M. (2012). Hierarchical latent vari- Eklöf, J. A., & Westlund, A. H. (2002). The pan-European customer
able models in PLS-SEM: guidelines for using reflective-formative satisfaction index program: current work and the way ahead. Total
type models. Long Range Planning, 45(5–6), 359–394. Quality Management, 13(8), 1099–1106.
Becker, J. M., Rai, A., & Rigdon, E. E. (2013a). Predictive validity and Evermann, J., & Tate, M. (2016). Assessing the predictive performance of
formative measurement in structural equation modeling: Embracing structural equation model estimators. Journal of Business Research,
practical relevance. In 2013a Proceedings of the international con- 69(10), 4565–4582.
ference on information systems. Milan. Fornell, C. G., Johnson, M. D., Anderson, E. W., Cha, J., & Bryant, B. E.
Becker, J. M., Rai, A., Ringle, C. M., & Völckner, F. (2013b). (1996). The American customer satisfaction index: nature, purpose,
Discovering unobserved heterogeneity in structural equation models and findings. Journal of Marketing, 60(4), 7–18.
to avert validity threats. MIS Quarterly, 37(3), 665–694. Fornell, C., Morgeson, F. V., & Hult, G. T. M. (2016). Stock returns on
Bollen, K. A. (1989). Structural equations with latent variables. New customer satisfaction do beat the market: gauging the effect of a
York: Wiley. marketing intangible. Journal of Marketing, 80(5), 92–107.
Brady, M. K., Voorhees, C. M., & Brusco, M. J. (2012). Service Gelbrich, K. (2010). Anger, frustration, and helplessness after service
sweethearting: its antecedents and customer consequences. Journal failure: coping strategies and effective informational support.
of Marketing, 76(2), 81–98. Journal of the Academy of Marketing Science, 38(5), 567–585.
Burke, S. J. (2011). Competitive positioning strength: market measure-
Goodhue, D. L., Lewis, W., & Thompson, R. (2012a). Does PLS have
ment. Journal of Strategic Marketing, 19(5), 421–428.
advantages for small sample size or non-normal data? MIS
Cassel, C., Hackl, P., & Westlund, A. H. (1999). Robustness of partial
Quarterly, 36(3), 981–1001.
least-squares method for estimating latent variable quality structures.
Goodhue, D. L., Lewis, W., & Thompson, R. (2012b). Comparing PLS to
Journal of Applied Statistics, 26(4), 435–446.
regression and LISREL: a response to Marcoulides, Chin, and
Chin, W. W. (1998). The partial least squares approach to structural equa-
Saunders. MIS Quarterly, 36(3), 703–716.
tion modeling. In G. A. Marcoulides (Ed.), Modern methods for
Green, D. H., Donald, W. B., & Ryans, A. B. (1995). Entry strategy and
business research (pp. 295–358). Mahwah: Erlbaum.
long-term performance: conceptualization and empirical examina-
Chin, W. W. (2010). Bootstrap cross-validation indices for PLS path
tion. Journal of Marketing, 59(4), 1–16.
model assessment. In V. Esposito Vinzi, W. W. Chin, J. Henseler,
& H. Wang (Eds.), Handbook of partial least squares (pp. 83–97). Habel, J., & Klarmann, M. (2015). Customer reactions to downsizing:
Berlin: Springer. when and how is satisfaction affected? Journal of the Academy of
Chin, W. W., & Newsted, P. R. (1999). Structural equation modeling Marketing Science, 43(6), 768–789.
analysis with small samples using partial least squares. In R. H. Haenlein, M., & Kaplan, A. M. (2004). A beginner's guide to partial least
Hoyle (Ed.), Statistical strategies for small sample research (pp. squares analysis. Understanding Statistics, 3(4), 283–297.
307–341). Thousand Oaks, CA: Sage. Haenlein, M., & Kaplan, A. M. (2011). The influence of observed het-
Chin, W. W., Marcolin, B. L., & Newsted, P. R. (2003). A partial least erogeneity on path coefficient significance: technology acceptance
squares latent variable modeling approach for measuring interaction within the marketing discipline. Journal of Marketing Theory and
effects: results from a Monte Carlo simulation study and an Practice, 19(2), 153–168.
electronic-mail emotion/adoption study. Information Systems Hair, J. F., Ringle, C. M., & Sarstedt, M. (2011). PLS-SEM: indeed a silver
Research, 14(2), 189–217. bullet. Journal of Marketing Theory and Practice, 19(2), 139–151.
Cliff, N. (1983). Some cautions concerning the application of causal model- Hair, J. F., Sarstedt, M., Ringle, C. M., & Mena, J. A. (2012a). An
ing methods. Multivariate Behavioral Research, 18(1), 115–126. assessment of the use of partial least squares structural equation
Cohen, J. (1988). Statistical power analysis for the behavioural sciences modeling in marketing research. Journal of the Academy of
(2nd ed.). Hillsdale: Lawrence Erlbaum Associates. Marketing Science, 40(3), 414–433.
Core Team, R. (2016). R: a language and environment for sta- Hair, J. F., Sarstedt, M., Pieper, T. M., & Ringle, C. M. (2012b). The use of
tistical computing. Vienna, Austria: R Foundation for partial least squares structural equation modeling in strategic manage-
Statistical Computing. ment research: a review of past practices and recommendations for
Davis, F. D. (1989). Perceived usefulness, perceived ease of use, future applications. Long Range Planning, 45(5–6), 320–340.
and user acceptance of information technology. MIS Hair, J. F., Ringle, C. M., & Sarstedt, M. (2013). Partial least squares
Quarterly, 13(3), 319–340. structural equation modeling: rigorous applications, better results
Dellande, S., Gilly, M. C., & Graham, J. L. (2004). Gaining compliance and higher acceptance. Long Range Planning, 46(1–2), 1–12.
and losing weight: the role of the service provider in health care Hair, J. F., Hult, G. T. M., Ringle, C. M., & Sarstedt, M. (2017). A primer
services. Journal of Marketing, 68(3), 78–91. on partial least squares structural equation modeling (PLS-SEM)
Diamantopoulos, A., & Riefler, P. (2011). Using formative measures in (2nd ed.). Thousand Oaks, CA: Sage.
international marketing models: a cautionary tale using consumer Heidenreich, S., Wittkowski, K., Handrich, M., & Falk, T. (2015). The
animosity as an example. In M. Sarstedt, M. Schwaiger, & C. R. dark side of customer co-creation: exploring the consequences of
Taylor (Eds.), Advances in international marketing (Vol. 22, pp. 11– failed co-created services. Journal of the Academy of Marketing
30). Bingley: Emerald. Science, 43(3), 279–296.
J. of the Acad. Mark. Sci. (2017) 45:616–632 631

Hennig-Thurau, T., Groth, M., Paul, M., & Gremler, D. D. (2006). Are all Knaus, J. (2013). R package snowfall: Easier cluster computing (version:
smiles created equal? How emotional contagion and emotional labor 1.84–6). cran.r-project.org/web/packages/snowfall/.
affect service relationships. Journal of Marketing, 70(3), 58–73. Lee, N., & Cadogan, J. W. (2013). Problems with formative and higher-order
Henseler, J. (2010). On the convergence of the partial least squares path reflective variables. Journal of Business Research, 66(2), 242–247.
modeling algorithm. Computational Statistics, 25(1), 107–120. Lohmöller, J.-B. (1989). Latent variable path modeling with partial least
Henseler, J. (2012). Why generalized structured component analysis is squares. Heidelberg: Physica.
not universally preferable to structural equation modeling. Journal Marcoulides, G. A., & Chin, W. W. (2013). You write, but others read:
of the Academy of Marketing Science, 40(3), 402–413. common methodological misunderstandings in PLS and related
Henseler, J., & Fassott, G. (2010). Testing moderating effects in PLS path methods. In H. Abdi, W. W. Chin, V. Esposito Vinzi, G.
models: an illustration of available procedures. In V. Esposito Vinzi, Russolillo, & L. Trinchera (Eds.), New perspectives in partial least
W. W. Chin, J. Henseler, & H. Wang (Eds.), Handbook of partial squares and related methods (pp. 31–64). New York, NJ: Springer.
least squares: concepts, methods and applications (pp. 713–735). Marcoulides, G. A., Chin, W. W., & Saunders, C. (2012). When imprecise
Berlin: Springer. statistical statements become problematic: a response to Goodhue,
Henseler, J., & Sarstedt, M. (2013). Goodness-of-fit indices for Lewis, and Thompson. MIS Quarterly, 36(3), 717–728.
partial least squares path modeling. Computational McDonald, R. P. (1996). Path analysis with composite variables.
Statistics, 28(2), 565–580. Multivariate Behavioral Research, 31(2), 239–270.
Henseler, J., Ringle, C. M., & Sinkovics, R. R. (2009). The use of partial McIntosh, C. N., Edwards, J. R., & Antonakis, J. (2014). Reflections on
least squares path modeling in international marketing. In R. R. partial least squares path modeling. Organizational Research
Sinkovics & P. N. Ghauri (Eds.), Advances in international Methods, 17(2), 210–251.
marketing (Vol. 20, pp. 277–320). Bingley: Emerald. Monecke, A., & Leisch, F. (2012). semPLS: structural equation modeling
Henseler, J., Dijkstra, T. K., Sarstedt, M., Ringle, C. M., Diamantopoulos, using partial least squares. Journal of Statistical Software, 48(3), 1–32.
A., Straub, D. W., Ketchen, D. J., Hair, J. F., Hult, G. T. M., & Paxton, P., Curran, P. J., Bollen, K. A., Kirby, J., & Chen, F. (2001).
Calantone, R. J. (2014). Common beliefs and reality about partial Monte Carlo experiments: design and implementation. Structural
least squares: comments on Rönkkö & Evermann (2013). Equation Modeling, 8(2), 287–312.
Organizational Research Methods, 17(2), 182–209. Ranjan, K. R., & Read, S. (2016). Value co-creation: concept and mea-
Henseler, J., Ringle, C. M., & Sarstedt, M. (2015). A new criterion for surement. Journal of the Academy of Marketing Science, 44(3),
assessing discriminant validity in variance-based structural equation 290–315.
modeling. Journal of the Academy of Marketing Science, 43(1),
Rego, L. L., Morgan, N. A., & Fornell, C. (2013). Reexamining the
115–135.
market share–customer satisfaction relationship. Journal of
Henseler, J., Hubona, G. S., & Ray, P. A. (2016). Using PLS path model-
Marketing, 77(5), 1–20.
ing in new technology research: updated guidelines. Industrial
Reinartz, W. J., Haenlein, M., & Henseler, J. (2009). An empirical compar-
Management & Data Systems, 116(1), 1–19.
ison of the efficacy of covariance-based and variance-based SEM.
Hui, B. S., & Wold, H. O. A. (1982). Consistency and consistency at large
International Journal of Research in Marketing, 26(4), 332–344.
of partial least squares estimates. In K. G. Jöreskog & H. O. A. Wold
Rigdon, E. E. (1998). Structural equation modeling. In G. A. Marcoulides
(Eds.), Systems under indirect observation, part II (pp. 119–130).
(Ed.), Modern methods for business research (pp. 251–294).
Amsterdam: North Holland.
Mahwah: Erlbaum.
Hulland, J., Ryan, M. J., & Rayner, R. K. (2010). Modeling customer
satisfaction: a comparative performance evaluation of covariance Rigdon, E. E. (2012). Rethinking partial least squares path modeling: in
structure analysis versus partial least squares. In V. Esposito Vinzi, praise of simple methods. Long Range Planning, 45(5–6), 341–358.
W. W. Chin, J. Henseler, & H. Wang (Eds.), Handbook of partial Rigdon, E. E. (2014). Rethinking partial least squares path modeling: break-
least squares: concepts, methods and applications (pp. 307–325). ing chains and forging ahead. Long Range Planning, 47(3), 161–167.
Berlin: Springer. Rigdon, E. E. (2016). Choosing PLS path modeling as analytical method
Hult, G. T. M., Morgeson III, F. V., Morgan, N. A., Mithas, S., & Fornell, in European management research: a realist perspective. European
C. (2017). Do managers know what their customers think and why? Management Journal, 34(6), 598–605.
Journal of the Academy of Marketing Science, 45(1), 37–54. Rigdon, E. E., Ringle, C. M., & Sarstedt, M. (2010). Structural modeling of
Hwang, H. (2009). Regularized generalized structured component analy- heterogeneous data with partial least squares. In N. K. Malhotra (Ed.),
sis. Psychometrika, 74(3), 517–530. Review of marketing research (pp. 255–296). Armonk: Sharpe.
Hwang, H., Malhotra, N. K., Kim, Y., Tomiuk, M. A., & Hong, S. (2010). Rigdon, E. E., Becker, J.-M., Rai, A., Ringle, C. M., Diamantopoulos, A.,
A comparative study on parameter recovery of three approaches to Karahanna, E., Straub, D. W., & Dijkstra, T. K. (2014). Conflating
structural equation modeling. Journal of Marketing Research, 47(4), antecedents and formative indicators: a comment on Aguirre-Urreta
699–712. and Marakas. Information Systems Research, 25(4), 780–784.
Jöreskog, K. G. (1973). A general method for estimating a linear struc- Ringle, C. M., & Sarstedt, M. (2016). Gain more insight from your PLS-
tural equation system. In A. S. Goldberger & O. D. Duncan (Eds.), SEM results: the importance-performance map analysis. Industrial
Structural equation models in the social sciences (pp. 255–284). Management & Data Systems, 116(9), 1865–1886.
New York, NJ: Seminar Press. Ringle, C. M., Sarstedt, S., & Straub, D. W. (2012). A critical look at the
Jöreskog, K. G., & Wold, H. O. A. (1982). The ML and PLS techniques for use of PLS-SEM in MIS Quarterly. MIS Quarterly, 36(1), iii–xiv.
modeling with latent variables: historical and comparative aspects. In Ringle, C. M., Sarstedt, M., & Schlittgen, R. (2014). Genetic algorithm
H. O. A. Wold & K. G. Jöreskog (Eds.), Systems under indirect segmentation in partial least squares structural equation modeling.
observation, part I (pp. 263–270). Amsterdam: North-Holland. OR Spectrum, 36(1), 251–276.
Kaplan, A. M., Schoder, D., & Haenlein, M. (2007). Factors influencing Romdhani, H., Grinek, S., Hwang, H., & Labbe, A. (2014). R package
the adoption of mass customization: the impact of base category ASGSCA: Association studies for multiple SNPs and multiple traits
consumption frequency and need satisfaction. Journal of Product using generalized structured equation models (Version 1.4.0),
Innovation Management, 24(2), 101–116. http://bioconductor.org/packages/ASGSCA/.
Kaufmann, L., & Gaeckler, J. (2015). A structured review of partial least Rönkkö, M., & Evermann, J. (2013). A critical examination of common
squares in supply chain management research. Journal of beliefs about partial least squares path modeling. Organizational
Purchasing and Supply Management, 21(4), 259–272. Research Methods, 16(3), 425–448.
632 J. of the Acad. Mark. Sci. (2017) 45:616–632

Rönkkö, M., McIntosh, C. N., & Antonakis, J. (2015). On the adoption of Tenenhaus, M. (2008). Component-based structural equation modelling.
partial least squares in psychological research: caveat emptor. Total Quality Management & Business Excellence, 19(7–8), 871–886.
Personality and Individual Differences, 87, 76–84. Tenenhaus, A., & Tenenhaus, M. (2011). Regularized generalized canon-
Rönkkö, M., McIntosh, C. N., Antonakis, J., & Edwards, J. R. (2016). Partial ical correlation analysis. Psychometrika, 76(2), 257–284.
least squares path modeling: time for some serious second thoughts. Tenenhaus, M., Esposito Vinzi, V., Chatelin, Y.-M., & Lauro, C. (2005).
Journal of Operations Management, 47-48(November), 9–27. PLS path modeling. Computational Statistics & Data Analysis,
Rubera, G., Chandrasekaran, D., & Ordanini, A. (2016). Open innova- 48(1), 159–205.
tion, product portfolio innovativeness and firm performance: the Thurstone, L, L. (1947). Multiple factor analysis. Chicago: The
dual role of new product development capabilities. Journal of the University of Chicago Press.
Academy of Marketing Science, 44(2), 166–184. Treiblmaier, H., Bentler, P. M., & Mair, P. (2011). Formative constructs
Sarstedt, M., Ringle, C. M., Henseler, J., & Hair, J. F. (2014). On the implemented via common factors. Structural Equation Modeling: A
emancipation of PLS-SEM: a commentary on Rigdon (2012). Long Multidisciplinary Journal, 18(1), 1–17.
Range Planning, 47(3), 154–160. van der Heijden, G. A. H., Schepers, J. J. L., Nijssen, E. J., & Ordanini, A.
Sarstedt, M., Hair, J. F., Ringle, C. M., Thiele, K. O., & Gudergan, S. P. (2013). Don’t just fix it, make it better! Using frontline service
(2016). Estimation issues with PLS and CBSEM: where the bias employees to improve recovery performance. Journal of the
lies! Journal of Business Research, 69(10), 3998–4010. Academy of Marketing Science, 41(5), 515–530.
Schneeweiß, H. (1991). Models with latent variables: LISREL versus Venkatesh, V., Morris, M. G., Davis, G. B., & Davis, F. D. (2003). User
PLS. Statistica Neerlandica, 45(2), 145–157. acceptance of information technology: toward a unified view. MIS
Quarterly, 27(3), 425–478.
Schönemann, P. H., & Steiger, J. H. (1978). On the validity of indeterminate
Vilares, M. J., & Coelho, P. S. (2013). Likelihood and PLS estimators for
factor scores. Bulletin of the Psychonomic Society, 12(4), 287–290.
structural equation modeling: an assessment of sample size, skew-
Schönemann, P. H., & Wang, M.-M. (1972). Some new results on factor
ness and model misspecification effects. In J. Lita da Silva, F.
indeterminacy. Psychometrika, 37(1), 61–91.
Caeiro, I. Natário, & C. A. Braumann (Eds.), Advances in regres-
Schuberth, F., Henseler, J., & Dijkstra, T. K. (2016). Partial least squares
sion, survival analysis, extreme values, Markov processes and other
path modeling using ordinal categorical indicators. Quality &
statistical applications (pp. 11–33). Berlin: Springer.
Quantity, forthcoming.
Wold, H. O. A. (1974). Causal flows with latent variables: partings of
Schubring, S., Lorscheid, I., Meyer, M., & Ringle, C. M. (2016). The PLS ways in the light of NIPALS modelling. European Economic
agent: predictive modeling with PLS-SEM and agent-based simula- Review, 5(1), 67–86.
tion. Journal of Business Research, 69(10), 4604–4612. Wold, H. O. A. (1980). Model construction and evaluation when theoret-
Shmueli, G. (2010). To explain or to predict? Statistical Science, 25(3), ical knowledge is scarce: theory and application of PLS. In J.
289–310. Kmenta & J. B. Ramsey (Eds.), Evaluation of econometric models
Shmueli, G., Ray, D., Manuel, J., Estrada, V., & Chatla, S. B. (2016). The (pp. 47–74). New York, NJ: Academic Press.
elephant in the room: evaluating the predictive performance of PLS Wold, H. O. A. (1982). Soft modeling: the basic design and some exten-
models. Journal of Business Research, 69(10), 4552–4564. sions. In K. G. Jöreskog & H. O. A. Wold (Eds.), Systems under
Spearman, C. (1927). The abilities of man. London: MacMillan. indirect observations, part II (pp. 1–54). Amsterdam: North-Holland.
Steenkamp, J.-B. E. M., & Baumgartner, H. (2000). On the use of struc- Wold, S., Sjöström, M., & Eriksson, L. (2001). PLS-regression: a basic
tural equation models for marketing modeling. International tool of chemometrics. Chemometrics and Intelligent Laboratory
Journal of Research in Marketing, 17(2/3), 195–202. Systems, 58(2), 109–130.
Sundaram, S., Schwarz, A., Jones, E., & Chin, W. W. (2007). Technology Wolter, J. S., & Cronin, J. J. (2016). Re-conceptualizing cognitive and
use on the front line: how information technology enhances individ- affective customer-company identification: the role of self-motives
ual performance. Journal of the Academy of Marketing Science, and different customer-based outcomes. Journal of the Academy of
35(1), 101–112. Marketing Science, 44(3), 397–413.

View publication stats