Você está na página 1de 13

Journal of Business Research 69 (2016) 39984010

Contents lists available at ScienceDirect

Journal of Business Research

Estimation issues with PLS and CBSEM: Where the bias lies!
Marko Sarstedt a,d,, Joseph F. Hair b, Christian M. Ringle c,d, Kai O. Thiele c, Siegfried P. Gudergan d
a
Otto-von-Guericke-University Magdeburg, Universittsplatz 2, 39106 Magdeburg, Germany
b
University of South Alabama, Mitchell College of Business, Mobile, AL 36688, USA
c
Hamburg University of Technology (TUHH), Am Schwarzenberg-Campus 4, 21073 Hamburg, Germany
d
University of Newcastle (Australia), University Drive, Callaghan, NSW 2308, Australia

a r t i c l e i n f o a b s t r a c t

Article history: Discussions concerning different structural equation modeling methods draw on an increasing array of concepts
Received 31 March 2016 and related terminology. As a consequence, misconceptions about the meaning of terms such as reective mea-
Received in revised form 30 May 2016 surement and common factor models as well as formative measurement and composite models have emerged.
Accepted 11 June 2016
By distinguishing conceptual variables and their measurement model operationalization from the estimation
Available online 25 June 2016
perspective, we disentangle the confusion between the terminologies and develop a unifying framework. Results
Keywords:
from a simulation study substantiate our conceptual considerations, highlighting the biases that occur when
Common factor models using (1) composite-based partial least squares path modeling to estimate common factor models, and (2) com-
Composite models mon factor-based covariance-based structural equation modeling to estimate composite models. The results
Reective measurement show that the use of PLS is preferable, particularly when it is unknown whether the data's nature is common
Formative measurement factor- or composite-based.
Structural equation modeling 2016 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license
Partial least squares (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction Recently, however, the scholarly community has witnessed a sur-


prising level of acrimony towards PLS. Antonakis, Bendahan, Jacquart,
The extent to which researchers raise an issue is a subtle indicator of and Lalive (2010, p. 1103) allude that there is no use for PLS whatsoev-
its importance. The benets and limitations of partial least squares path er [] thus strongly encourage researchers to abandon it. Other au-
modeling (PLS) is one such issue that scholars have heatedly debated thors similarly suggest that the use of PLS is very difcult to justify
across a variety of disciplines including marketing (e.g., Fornell & (Rnkk & Evermann, 2013, p. 443) or that PLS should not be adopted
Bookstein, 1982; Hair, Sarstedt, Ringle, & Mena, 2012), strategic man- as a tool for psychological research. (Rnkk, McIntosh, & Antonakis,
agement (e.g., Bentler & Huang, 2014; Rigdon, 2012, 2014; Sarstedt, 2015, p. 82). This new harshness climaxed in an editorial from the edi-
Ringle, Henseler, & Hair, 2014), and management information systems tors in chief of the Journal of Operations Management (Guide &
(e.g., Goodhue, Lewis, & Thompson, 2012; Marcoulides & Saunders, Ketokivi, 2015, p. vii) who declared that they were desk rejecting prac-
2006; Ringle, Sarstedt, & Straub, 2012). Such scientic debates are im- tically all PLS-based manuscripts. In a follow-up paper in the very same
portant since they serve as a catalyst that sparks further careful exami- journal, Rnkk, McIntosh, Antonakis, and Edwards (2016, p. 16) echo
nation of a method's properties. Oftentimes, the result is improved this call by suggesting that "the only logical and reasonable action stem-
understanding of the advantages and disadvantages of the focal method, ming from objective consideration of these issues is to discontinue the
but also additional research and methodological advances that stem use of PLS."
from such objective and constructive discussions among scholars. Leaving aside the tone of these and similar statements, which aim at
shutting down any scholarly debate, the question arises why these au-
The authors thank George R. Franke (University of Alabama), Edward E. Rigdon thors cannot nd even a single positive attribute of PLS despite its accep-
(Georgia State University), and the participants of the 2nd International Symposium on tance in scholarly research. In an effort to disentangle these opposing
Partial Least Squares Path Modeling at the University of Seville for their constructive feed-
views, Rigdon (2016) offers an in-depth discussion of PLS and its origins,
back. The authors also thank Jrg Henseler, University of Twente, The Netherlands, for his
support with the replication of Reinartz, Haenlein, and Henseler's study (2009), and with concluding that critics just as proponents of the method frequently offer
the development of the composite model-based data generation approach. Even though incorrect or incomplete rationale for avoiding as well as using PLS. In ad-
this research does not explicitly refer to the use of the statistical software SmartPLS dition, Rigdon (2016) concludes that many misconceptions about PLS
(http://www.smartpls.com), Ringle acknowledges a nancial interest in SmartPLS. have their roots in the method's conceptual underpinnings and particu-
Corresponding author.
E-mail addresses: marko.sarstedt@ovgu.de (M. Sarstedt), joefhair@gmail.com
larly the estimation philosophy it relies on (e.g., Rigdon, 2012).
(J.F. Hair), ringle@tuhh.de (C.M. Ringle), k.thiele@tuhh.de (K.O. Thiele), In fact, when deciding to use PLS, researchersconsciously or
siggi.gudergan@newcastle.edu.au (S.P. Gudergan). unconsciouslyopt for a composite-based approach to structural

http://dx.doi.org/10.1016/j.jbusres.2016.06.007
0148-2963/ 2016 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
M. Sarstedt et al. / Journal of Business Research 69 (2016) 39984010 3999

equation modeling (SEM) that linearly combines indicators to form sciences researchmaking PLS the preferred SEM method for most
composite variables (Lohmller, 1989), which serve as proxies for the situations.
concepts under investigation (Rigdon, 2016). This approach is different Based on our ndings, we propose a framework that aligns different
from common factor-based SEM (i.e., covariance-based SEM; CBSEM), measurement and model estimation perspectives. This paper is written
which considers the constructs as common factors that explain the co- with the condence that it will (1) offer researchers a clear roadmap for
variation between their associated indicators. While this distinction the conceptionalization and operationalization of their constructs, (2) pro-
has long been noted (e.g., Jreskog & Wold, 1982; Schneewei, 1991), vide guidance in their choice of the appropriate SEM method, and (3) en-
researchers have traditionally emphasized how PLS is like factor- sure a more balanced perspective concerning recent criticism, which
based SEM but with advantages and disadvantages across different con- largely ignored the common factor vs. composite model distinction.
ditions (Rigdon, 2012, p. 353)see, for example, Hair et al. (2012);
Kaufmann and Gaeckler (2015); Peng and Lai (2012). Only recently 2. Measurement
have scholars started calling for the emancipation of PLS from CBSEM
by acknowledging its status as a purely composite-based method 2.1. Conceptual variables, constructs, and proxies
(e.g., Rigdon, 2012; Sarstedt, Ringle, Henseler et al., 2014). Addressing
this call, Henseler, Hubona, and Ray (2016, p. 3) attempt to provide Irrespective of whether a deductive or an inductive research ap-
an updated view on what PLS actually is and suggest a set of guide- proach is undertaken by social science researchers, at some point in
lines for the interpretation and reporting of results that explicitly con- their search to better understand and explain theory, they deal with
sider the distinction between composite-based SEM and common conceptual variables and theoretical models. A theoretical model re-
factor-based SEM. In their guidelines, the authors note that PLS path ects a set of structural relationships; usually based on a set of equations
models can contain two different forms of construct measurement: fac- connecting conceptual variables that formalize a theory and visually
tor models or composite models and continue by explaining that the represent the relationships (Bollen, 2002). As elements of theoretical
depicted direction of arrows in the measurement model (i.e., reective models, conceptual variables represent broad ideas or thoughts about
or formative) does not necessarily indicate whether PLS estimates a fac- abstract concepts that researchers establish and propose to measure in
tor or composite model (Henseler, Hubona et al., 2016, p. 3). their research (e.g., customer satisfaction).
Such statements leave many applied researchers confused as some Constructs represent conceptual variables in statistical models such as
misperceive the distinction between reective and formative measure- in a structural equation model.1 They are intended to enable empirical
ment specication on the one hand, and common factor and composite testing of hypotheses that concern relationships between conceptual var-
models on the other. The introduction of consistent PLS (PLSc), which iables (Rigdon, 2012) and are conceptually dened in terms of the attri-
Dijkstra and Henseler (2015) developed in an effort to align common bute and the object (e.g., MacKenzie, Podsakoff, & Podsakoff, 2011). The
factor and composite-based SEM methods, further contributed to the attribute denes the general type of property to which the focal concept
confusion. For example, some researchers have started using both PLS refers, such as an attitude (e.g., attitude towards an advertisement), a per-
and PLScwhich assume fundamentally different measurement ception (e.g., perceived ease of use of technology), or behavioral intention
philosophieson the same data without explicitly considering the na- (e.g., purchase intention). The focal object is the entity to which the prop-
ture of the data, model, and the implications of their choice of methods erty is applied. For example, the focus of interest could be a customer's
(e.g., Gelhard & von Delft, 2016). These issues are nicely reected in a re- satisfaction with the products, satisfaction with the services, and satisfac-
cent query by a thoughtful PhD student from the UK who asked one of tion with the prices. In these examples, satisfaction constitutes the attri-
this paper's authors, what is the real difference between reective con- bute, whereas products, services, and prices represent the focal objects.
structs and factor models? What would be a reective composite and Establishing a construct denition also includes determination of the
what would be a common factor? And how is such difference trans- dimensionality that describes the conceptual variable, with each dimen-
ferred to the PLS context in terms of model specication? sion representing a different aspect (e.g., Law, Wong, & Mobley, 1998).
These queries constitute the research questions this paper sheds A conceptual variable is not per se characterized as unidimensional or
light on. By distinguishing measurement model conceptualization multidimensional, let alone two-, three- or four-dimensional (Bollen,
and operationalization from the model estimation perspective, this 2011). Rather it depends on the context-specic denition of the con-
paper disentangles the confusion between reective measurement ceptual variable and the denotation that comes with it. The denotation
and common factor models as well as formative measurement and can, in principle, be innite, since the same conceptual variable can rep-
composite models. More precisely, this paper's aim is to clarify the resent different levels of theoretical abstraction across contexts
interplay between measurement model specication and model es- (Diamantopoulos, 2005; Law & Wong, 1999). Thus, a construct deni-
timation via PLS using different estimation modes (i.e., Mode A vs. tion is subject to the context within which a conceptual variable is ex-
Mode B) and CBSEM. Understanding this interplay is of fundamental amined such that the denition can change from one study to another
importance when deriving measures that suit a specic SEM method, and, accordingly, can differ in terms of dimensionality and the object
or when choosing a specic SEM method that aligns with existing of interest. For example, a customer's satisfaction with the service can
measures or a research objective. Results from a simulation study be broken down into more concrete subdimensions, such as satisfaction
substantiate our conceptual considerations, highlighting the biases with the speed of service, the servicescape, and the staff. The latter di-
that occur when using composite-based PLS to estimate common mension can be differentiated into more concrete subdimensions such
factor models, and common factor-based CBSEM to estimate composite as satisfaction with the friendliness, competence, and outer appearance
models. Specically, our results show that PLS entails practically no bias of the service staff. Each of these aspects can, in principle, be further bro-
when estimating data from a composite model population, regard- ken down into yet more concrete subdimensions (e.g., Rossiter, 2011).
less of the measurement model specication. In contrast, CBSEM Finally, the construct denition also claries how the abstract, concep-
and PLSc estimation of reectively measured constructs when the tual variable relates to measurable, observable quantities. That is, the
data stem from a composite population show severe biases in param- construct denition guides the conceptualization of the measurement
eter estimates, rendering their use inappropriate in these instances.
Further comparisons with common factor model data show that
1
the parameter bias resulting from using an SEM method on discrep- Note that researchers frequently distinguish between latent variables/constructs and
composites, depending on the type of relationship assumed between the latent variable
ant populations is much more severe for CBSEM than for PLS. The real (composite) and its indicators (e.g., MacCallum & Browne, 1993).We use the term latent
bias results when researchers don't know the underlying data popula- variable/construct to refer to the entities that represent conceptual variables in a structural
tion (i.e., common factor or composite)as is widespread in social equation model.
4000 M. Sarstedt et al. / Journal of Business Research 69 (2016) 39984010

models, which entails deciding whether to measure a construct reec- Fig. 1 shows a reective measurement model for a latent variable Y1,
tively or formatively. measured with four indicators x1, x2, x3, and x4 as well as the conceptual
Constructs are not just theoretical concepts under a different name variable the construct seeks to represent, illustrated by a triangle in the
as implied by commonly used denitions of this term (e.g., Bollen, upper part of the gure (Rigdon, 2012). Reective indicators, also re-
2002; Pedhazur & Pedhazur Schmelkin, 1991), but representations of ferred to as effect indicators, can be viewed as a representative sample
conceptual variables in a statistical model. Importantly, constructs do of all the possible items available within the conceptual domain of the
not represent conceptual variables perfectly since any concept and construct (Nunnally & Bernstein, 1994). Since a reective measurement
any construct denition has some degree of ambiguity associated with model dictates that all items reect the same construct, indicators asso-
it (e.g., Gilliam & Voss, 2013). In addition, constructs stem from data ciated with a particular construct should be highly correlated with each
and therefore share the data's idiosyncrasies (Cliff, 1983; MacCallum, other (Edwards & Bagozzi, 2000). In addition, individual items should
Browne, & Cai, 2007), which further detach them from the concepts be interchangeable, and any single item can generally be left out with-
they intend to represent. In this context, Michell (2013, p. 20) notes out changing the meaning of the construct, as long as the construct
that constructs are contrived in a way that is detached from the actual has sufcient reliability (Jarvis et al., 2003). The fact that the relationship
structure of testing phenomena and held in place by an array of quanti- goes from the construct to its indicators implies that if the evaluation of
tative methods, such as factor analysis, which gratuitously presume the latent trait changes (e.g., because of a change in the standard of com-
quantitative structure rather than infer it from the relevant phenomena parison), all indicators will change simultaneously (e.g., Diamantopoulos
(). Similarly MacCallum et al. (2007, p. 153) state that factor analyt- & Winklhofer, 2001).
ical procedures such as CBSEM cannot fully represent the undoubtedly The second approach is formative measurement. In a formative mea-
large number of minor common factors that inuence measured vari- surement model the indicators form the construct by means of linear
ables and account in part for their intercorrelations. There are many combinations (Fig. 1). A change in an indicator's value due to, for exam-
other sources of error in such models. At best, a factor analysis model ple, a change in a respondent's assessment of the trait being captured by
is an approximation of real-world phenomena. Against this back- the indicator, changes the value of the construct. That is, variation in
ground, Rigdon (2012, pp. 343344) concludes that constructs should the indicators precedes variation in the latent variable (Borsboom,
rather be viewed as something created from the empirical data Mellenbergh, & van Heerden, 2003, p. 208), which means that, by de-
which is intended to enable empirical testing of propositions regarding nition, constructs with a formative measurement model are inextricably
the concept. That is, all measures of conceptual variables are approxi- tied to their measures (Diamantopoulos, 2006). Besides the difference
mations of or proxies for conceptual variables, independent from how in the relationship between indicator(s) and construct, formative
they were derived (e.g., Wickens, 1972). Thus, irrespective of the quality measurement models do not require correlated indicators. In practi-
with which a conceptual variable is theoretically substantiated and op- cal applications, however, indicators in formative measurement
erationally dened and the rigor that encompasses measurement models may be highly correlated, yielding satisfactory levels in reli-
model development, any measurement in structural equation models ability and validity statistics whose use, from a conceptual perspec-
produces only proxies for latent variables (Rigdon, 2012). This assess- tive, should be restricted to reective measurement models (Hair
ment is in line with the proliferation of all sorts of instruments that et al., 2012).
claim to measure essentially the same construct, albeit often with little Despite these clear conceptual differences, deciding whether to
chance to convert one instrument's measures into any other specify measurement models reectively or formatively is not clear-
instrument's measures (Salzberger, Sarstedt, & Diamantopoulos, cut in practice, as constructs do not inherently follow a reective or for-
2016). For example, business research and practice has brought forward mative measurement logic (e.g., Baxter, 2009). Rather, the researcher
a multitude of measurement instruments for corporate reputation, has the exibility to conceptualize a measurement model based on the
which rest on the same denition of the concept but differ fundamental- construct denition the researcher species. As Baxter (2009, p. 1377)
ly in terms of their underlying conceptualizations and measurement notes, there are often quite different possibilities for conceptualization
items (e.g., Sarstedt, Wilczynski, & Melewar, 2013). of what might at rst sight appear to be the same construct and, most
importantly, there may be quite distinct lines of enquiry underlying
2.2. Measurement model conceptualization and operationalization the multiple possible conceptualizations. Consider, for example, the
concept of perceived switching costs. Jones, Mothersbaugh, and Beatty
Based on the construct denition, the next step is to specify a mea- (2000, p. 262) dene perceived switching costs as consumer percep-
surement model, which expresses how to measure the construct by tions of the time, money, and effort associated with changing service
means of a set of indicators (e.g., Jarvis, MacKenzie, & Podsakoff, 2003; providers. Their measurement approach in the context of banking ser-
MacKenzie, 2003). Generally, there are two broad ways to conceptualize vices draws on three items, which constitute reections or conse-
measurement models (Coltman, Devinney, Midgley, & Venaik, 2008; quences of perceived switching costs (In general it would be a hassle
Diamantopoulos & Winklhofer, 2001), which entail fundamentally dif- changing banks, It would take a lot of time and effort changing
ferent approaches to generating items (e.g., Churchill, 1979; banks, and For me, the costs in time, money, and effort to switch
Diamantopoulos & Winklhofer, 2001; MacKenzie et al., 2011). The rst banks are high). Hence, the authors implicitly assume that there is a
approach is referred to as reective measurement. In a reective mea- concept of perceived switching costs, which can be manifested by que-
surement model the indicators are considered to be error-prone mani- rying a set of (e.g., three) items. Barroso and Picn (2012, p. 532), on the
festations of an underlying construct with relationships going from other hand, consider perceived switching costs as a latent aggregate
the construct to its indicators (Bollen, 1989). The relationship between construct that is expressed as an algebraic composition of its different
an observed and an unobserved variable is usually modeled as dimensions. These authors identify a set of six dimensions (benet
expressed in the following equation: loss costs, personal relationship loss costs, economic risks costs, evalua-
tion costs, set-up costs, and monetary loss costs), which represent cer-
tain specic characteristics, each covering an independent part of the
x l  Y e; 1
perceived switching costs concept. As such, Barroso and Picn's
(2012) construct denition of perceived switching costs follows a for-
where x is the observed indicator variable, Y is the latent variable, the mative measurement model logic. Of course, the underlying items can
loading l is a regression coefcient quantifying the strength of the rela- be empirically correlated, and perhaps causally related, but they are
tionship between x and Y, and e represents the random measurement not actually exchangeable in the way the reective measurement
error. model conceptualization assumes they are (Rigdon et al., 2011).
M. Sarstedt et al. / Journal of Business Research 69 (2016) 39984010 4001

Fig. 1. Measurement model conceptualization and operationalization.

That is, their correlation is not because the construct of perceived While this characteristic is fully comprehensible from a model estima-
switching costs is assumed to be their common cause. There are many tion perspective (see Diamantopoulos, 2006), from a measurement per-
more examples of constructs that carry the same label but which spective, there is no reason to assume that the sources of error that have
rely on different (i.e., reective vs. formative) measurement model traditionally been associated with reective indicators do not apply to
conceptualizationssee, for example, Albers (2010), Baxter (2009), causal indicators. For example, why would the use of double-barreled
and Chang, Franke, and Lee (2016). items or of suggestive item wordings trigger error in a reective indica-
Further contributing to the difculties of deciding on the measure- tor but not in a causal indicator? The following equation represents a
ment perspective is the fact that there is not one type of formative mea- measurement model comprised of causal indicators, where wi indicates
surement modelas had been implied in the early works on formative the contribution of xi (i = 1, , I) to Y, and z is an error term associated
measurement (e.g., Diamantopoulos & Winklhofer, 2001) and the use with Y:
of formative measurement models in statistical analysis (e.g., Hair,
I
Ringle, & Sarstedt, 2011). Rather, two types of indicators exist in forma- Y i1 wi  xi z 2
tive measurement models: causal indicators and composite indicators
(Bollen, 2011; Bollen & Bauldry, 2011). Models with causal indicators The other type of indicators, referred to as composite indicators,
follow a realist approach to measurement, which acknowledges that closely resembles that of causal indicators except for one aspect. In con-
under any denition of a conceptual variable, there is a true value but trast to constructs measured with causal indicators, constructs mea-
this can never be measured with complete accuracy (e.g., Grace & sured with composite indicators do not have an error term (see
Bollen, 2008). Therefore, the indicators should have conceptual unity construct Y3 in Fig. 1). This distinction has an important implication
in that all the indicators correspond to the researcher's denition of for the conceptualization of formative measurement models (Henseler
the concept (Bollen & Diamantopoulos, 2016). Breadth of coverage of et al., 2014) because composite indicators operate as contributors to a
the domain is extremely important to ensure that the domain of content construct rather than truly causing it (Bollen, 2011; Bollen &
is adequately captured: Omitting important indicators implies omitting Bauldry, 2011). They form the composite representing the construct in
a part of the conceptual variable that the construct represents full by means of linear combinations. Therefore, a formative construct
(e.g., Bollen & Lennox, 1991). measured with composite indicators does not have an error term
Since causal indicators are expected to cover all aspects of the con- (i.e., the error term is set to zero). As with causal indicators, composite
tent domain (Bollen & Bauldry, 2011), constructs measured with causal indicators are assumed to be error free. The following equation illus-
indicators (Y2 in Fig. 1) have an error term (z in Fig. 1). This error term trates a measurement model with composite indicators, where Y is a lin-
captures all the other causes of the construct not included in the ear combination of indicators xi, each weighted by an indicator weight
model (Diamantopoulos, 2006). Or as Diamantopoulos, Rieer, and wi (Bollen, 2011; McDonald, 1996):
Roth (2008, p. 12111212) note, formative latent variables have a
I
number of proximal causes, which researchers try to identify when con- Y i1 wi  xi 3
ceptually specifying the construct. In many cases, however, researchers
will be unable to detect all possible causes, as there may be some causes Although researchers have often used composite models and causal
that have neither been discussed in prior literature nor revealed by ex- indicator models synonymously (e.g., Bollen & Lennox, 1991), more re-
ploratory research. The construct-level error term represents these cently they have started distinguishing composite from causal indica-
missing causes. Causal indicators themselves are, by denition, error tors (e.g., Bollen, 2011; Bollen & Diamantopoulos, 2016; Howell,
freethat is, they are not subject to any systematic or random error. Breivik, & Wilcox, 2013). Thus, there is still some ambiguity regarding
4002 M. Sarstedt et al. / Journal of Business Research 69 (2016) 39984010

their nature and areas of application. For example, Bollen (2011, p. 366) Alternatively, measurement models with composite indicators can
notes that it seems unlikely that there are many situations where an be interpreted as a prescription for dimension reduction, where the
error term would be absent (). This would mean that the latent vari- aim is to condense the measures so they adequately cover a conceptual
able that represents the unidimensional concept is an exact linear func- variable's salient features (Dijkstra & Henseler, 2011). For example, a re-
tion of its indicators, which would seem to be a rarity. Bollen (2011) searcher may be interested in measuring the salient aspects of per-
therefore treats the latent variables as if they were indeed the conceptu- ceived switching costs by means of three (composite) indicators,
al variables from a theoretical model (also see Bollen & Bauldry, 2011; which cover aspects particularly relevant to the study at hand
Bollen & Diamantopoulos, 2016). However, viewing latent variables as (e.g., evaluation costs, set-up costs, and monetary loss costs).
proxies for a conceptual variable seems more reasonable and realistic
(Rigdon, 2012), blurring the conceptual distinction between composite 3. Model estimation
and causal indicators.
Furthermore, Bollen (2011, p. 366) asserts that composite indica- 3.1. PLS and CBSEM
tors need not share unidimensional conceptual unity. That is, composite
indicators might be combined into a composite as a way to conveniently The previous sections described different routes to operationalize
summarize the effect of several variables that do not tap the same con- constructs as proxies for conceptual variables. This measurement per-
cept although they may share a similar theme. Following this logic, spective needs to be complemented with the model estimation per-
measurement models with composite indicators only offer a means to spective, which explains how the different SEM techniques arrive at a
model conceptual variables, for which elements are combined to form solution and which assumptions underlie them. Researchers typically
a new entity (Henseler, Hubona et al., 2016). This is particularly the use two approaches to estimate structural equation models. One is the
case when analyzing secondary data, which typically lack a comprehen- more widely applied CBSEM approach (Bollen, 1989; Diamantopoulos,
sive theoretical substantiation and are collected for a purpose other 1994; Jreskog, 1978); the other is PLS (Hair, Hult, Ringle, & Sarstedt,
than SEM (Rigdon, 2013). For example, a measurement model concep- 2017; Lohmller, 1989; Wold, 1982). While both complementary
tualization of information search activities could be based on capturing methods share the same basic aim, which is to estimate the relation-
the sum of the activities that customers engage in when seeking infor- ships among constructs and indicators, they differ fundamentally in
mation from dealers, promotional materials, the Internet and other their statistical conceptions and particularly in the way they treat mea-
sources. Another researcher might choose a different set of variables surement models of constructs (Jreskog & Wold, 1982).
to form a measure of information search activities. Thus, the items ulti- CBSEM initially divides the variance of each indicator into two parts:
mately determine the meaning of the construct, which implies that (1) the common variance, which is estimated from the variance shared
adding or omitting an indicator potentially alters the nature of the con- with other indicators in the measurement model of a construct, and
struct. While this interpretation of composite indicators may be conve- (2) the unique variance, which consists of both specic and error vari-
nient for communication, it remains largely unclear where to draw a ance (Bollen, 1989; Rigdon, 1998). The specic variance is assumed to
line between items having conceptual unity and sharing a similar be systematic and reliable while the error variance is assumed to be ran-
theme (Bollen, 2011, p. 366). dom and unreliable (i.e., measurement, sampling, and specication
In practice, researchers naturally choose items in operationalizing error). CBSEM initially calculates the covariances of a set of variables
measurement models that match their construct denition, regardless (common variance), and only that variance is included in any solutions
of whether the actual measurement conceptualization draws on reec- derived. CBSEM, therefore, follows a common factor model approach in
tive, causal or composite indicators. That is, they treat the constructs in the estimation of the construct measures, which assumes that the vari-
their studies as unitary entities just like Barroso and Picn (2012) do ance of a set of indicators can be perfectly explained by the existence of
when offering an in-depth literature review of the nature and dimen- one unobserved variable (the common factor) and individual random
sionality of the perceived switching costs concept prior to deriving indi- error (Spearman, 1927; Thurstone, 1947). The common factor model
cators in their operationalization of the construct's measurement estimation approach conforms to the measurement philosophy under-
model. As such, they fully comply with Rnkk et al. (2016) who note lying reective measurement models.
that only a guiding conceptual framework and careful development of In principle, CBSEM can also accommodate formative measurement
the indicator content imbues theoretical meaning upon factors. In fact, models even though the method follows a common factor model esti-
assuming that researchers use measures of composite indicators merely mation approach (e.g., Temme, Diamantopoulos, & Pfegfeidel, 2014).
as convenient summaries of the data (Bollen, 2011) implies that the Analogous to the scientic realist perspective assumed in the method's
common practice of aggregating items as composites to represent con- treatment of reective measurement models, formative measurement
structs, even though commonly done in practically all non-SEM studies models in CBSEM typically assume causal indicators (Diamantopoulos,
in every eld of research, is without any theoretical justication and un- 2011). To estimate models with causal indicators, researchers must fol-
dermines the fundamentals of appropriate measurement. However, the low rules that require specic constraints on the model to ensure model
very same measures in most instances have been carefully developed identication (Bollen & Davies, 2009; Diamantopoulos & Rieer, 2011).
and tested following conventional measurement model evaluation As Hair et al. (2012, p. 420) note, these constraints often contradict the-
guidelinesas extensively documented in standard measurement oretical considerations, and the question arises whether model design
scale handbooks (e.g., Bearden, Netemeyer, & Haws, 2011; Bruner, should guide theory or vice versa.
James, & Hensel, 2001). Thus, the very activity of forming composites As an alternative, CBSEM scholars have proposed the multiple indica-
from validated measurement scales interweaves composite and causal tors and multiple causes (MIMIC) model (e.g., Bollen, 1989; Jreskog &
indicators, casting doubt on the notion that the use of composites to Goldberger, 1975)that includes both formative and reective indicators
represent conceptual variables is an outright abandonment of measure- (e.g., Diamantopoulos & Rieer, 2011; Diamantopoulos et al., 2008).
ment theory as Rnkk et al. (2016) imply. While MIMIC models enable researchers to deal with the identication
Thus, composite indicators not only offer a way to conveniently sum- problem, they do not overcome the problem that formative measure-
marize the data but can be used to measure any type of property to which ment models with causal indicators invariably underrepresent the var-
the focal concept refers, including attitudes, perceptions, and behavioral iance in the construct, since correlated indicators are required by the
intentions (e.g., Rigdon, 2012). As with any type of measurement concep- CBSEM common factor model to produce a valid proxy and thereby ad-
tualization, however, researchers need to offer a clear construct deni- equately represent a conceptual variable. As Lee and Cadogan (2013,
tion and specify items that closely match this denitionthat is, they p. 243) note, researchers should not be misled into thinking that
must share conceptual unity. achieving statistical identication allows one to obtain information
M. Sarstedt et al. / Journal of Business Research 69 (2016) 39984010 4003

about the variance of a formative latent variable. Clearly, CBSEM at best corresponds to correlation weights derived from bivariate correlations
only allows for approximating formative measurement models with between each indicator and the construct; Mode B corresponds to re-
causal indicators. gression weights, the standard in ordinary least squares regression anal-
Similarly, CBSEM can accommodate formative measurement models ysis. Regression weights not only take the correlation between each
with composite indicators (e.g., Diamantopoulos, 2011). Since con- indicator and the construct into account but also the correlations be-
structs measured with composite indicators are dened by having tween the indicators. No matter which mode for estimating the indica-
zero variances, the identication of the construct's error variance is tor weights is used, the resulting latent variable is always modeled as a
not an issue. Problems arise, however, with regard to the identication composite (Henseler, Ringle, & Sarstedt, 2016). That is, since all multi-
of all paths leading to as well as owing out from the construct. Grace item measures are converted into weighted componentseven in
and Bollen (2008) suggest solving this problem by specifying a single in- Mode APLS computes components by means of linear combinations
coming or outgoing path relationship to 1.0. While such specications of indicators.
overcome parameter identication issues, they severely limit the inter- PLS by default uses Mode A for reectively specied constructs and
pretability of the estimates of the magnitude and signicance of the Mode B for formatively specied constructs. Recent research, however,
xed paths in the structural model (Grace & Bollen, 2008). Because of suggests that selecting the appropriate weighting mode requires a more
these limitations, several researchers conclude that CBSEM is not well thoughtful approach. Specically, Becker, Rai, and Rigdon (2013) show
suited for estimating formative measurement models (Hair et al., that for formatively specied constructs, Mode A estimation yields bet-
2012; Peng & Lai, 2012; Reinartz, Haenlein, & Henseler, 2009). ter out-of-sample prediction for sample sizes larger than 100 and when
Different from CBSEM, PLS does not divide the variance into com- the R2 is moderate to large (i.e., R2 0.30). For large sample sizes and
mon and unique variance. More precisely, the objective of PLS is to ac- large R2 values, Mode A and Mode B perform equally well in terms of
count for the total variance in the observed indicators rather than to out-of-sample prediction. In terms of parameter accuracy in the struc-
explain only the correlations between the indicators (e.g., Tenenhaus, tural model, Mode A performs best when sample size or R2 values are
Esposito Vinzi, Chatelin, & Lauro, 2005). The logic of the PLS approach small to medium. For larger sample sizes or R2 values, Mode A and
is, therefore, that in estimating the model relationships, all of the vari- Mode B estimations do not differ in terms of parameter accuracy.
ance (common, unique and error) that the exogenous variables have From a measurement perspective, PLS and CBSEM both share an ap-
in common with the endogenous variables should be included proximation character as constructs do not necessarily fully correspond
(e.g., McDonald, 1996). The underlying notion is that the indicators to the conceptual variables they represent. As noted by Rigdon (2016,
can be (linearly) combined to form composite variables that are com- p. 19), common factor proxies cannot be generally assumed to carry
prehensive representations of the latent variables, and that these linear greater signicance than composite proxies in regard to the existence
combinations are valid proxies of the conceptual variables under inves- or nature of conceptual variables. A similar view is echoed in the in-
tigation (e.g., Henseler, Hubona et al., 2016). As such, PLS follows a com- tense debates on the relative advantages of component versus common
posite model approach in the estimation of the construct measures, factor analysis in the 90s, which witnessed a series of articles and com-
which generally conforms to the measurement philosophy underlying mentaries on the conceptual and philosophical underpinnings of the
formative measurement models. methods. Summarizing these debates, Bandalos and Boehm-Kaufman
PLS's designation as composite-based refers only to the method's (2009, p. 70) note that although methodologists still disagree about
way to represent constructs that approximate the conceptual variables which model is most appropriate, component analysis and common fac-
in a model. Although PLS draws on composites whose use has tradition- tor analysis have different goals and are based on different philoso-
ally been considered to be consistent with formative measurement phies. Rejecting the reex-like adherence to the common factor
models but not reective measurement models (e.g., Grace & Bollen, model, researchers have long warned that the common factor model
2008), the method readily accommodates both measurement model rarely holds in applied research (Schnemann & Wang, 1972). For ex-
types without identication issues (Hair et al., 2011). In estimating the ample, among 72 articles published during 2012 in what Atinc,
model parameters, however, PLS always follows a composite model ap- Simmering, and Kroll (2012) consider the four leading management
proach. That is, regardless of whether measurement models are reec- journals (Academy of Management Journal, Journal of Applied Psychology,
tive or formative, PLS always computes composite variables from sets Journal of Management, and Strategic Management Journal) that tested
of indicator variables as representations of the conceptual variables in one or more common factor model(s), fewer than 10% contained a com-
the model. Three aspects are important in this regard. mon factor model that did not have to be rejected. In light of these re-
First, in formative measurement models, PLS treats all indicators as sults, Henseler et al. (2014, p. 184) conclude from a philosophical
composite indicators. That is, the method does not allow for the explicit standpoint, there is no need for modeling constructs as common factors
modeling of a construct's error term measured with causal indicators (), and reducing SEM to common factor models is a very restrictive
(i.e., the error term z in Fig. 1 is constrained to zero). As a consequence (unnecessarily restrictive, we would argue) view about SEM.
and analogous to CBSEM, PLS only allows for approximating formative
measurement models with causal indicators. Note, however, that actu-
ally no method can estimate formative measurement models unless re- 4. Using PLS to estimate common factor models vs. using CBSEM to
ective measures are simultaneously available. estimate composite models
Second, researchers have long noted that since PLS is based on the
composite model logic, the method only approximates common 4.1. The parameter estimation bias
factor-based reective measurement models (Hui & Wold, 1982; also
see Rigdon, et al., 2014). That is, from a model estimation perspective, The previous discussions showed that PLS and CBSEM assume differ-
PLS will produce biased estimates if the common factor model ent ways of how the data represent measurement models that the
holdsjust like CBSEM will produce biased estimates when using researcherin line with a set of construct denitionsspecies in a re-
the method to estimate data generated from a composite model, as ective or formative way. CBSEM assumes the data follow a common
this study will show. However, the deviations in parameter estimates factor model in which the indicator covariances dene the nature of
should not be considered a bias as both methods estimate different the data, whereas PLS adheres to a composite model approach in
things and therefore may yield different values. which data are dened by means of linear combinations of indicators.
Third, to estimate the model parameters, PLS uses two modes, which So while the measurement models may follow a reective (or forma-
relate to the way the method estimates the indicators weights that rep- tive) specication, the underlying data model may be composite-
resent each indicator's contribution to the composite. Mode A based (or common factor-based).
4004 M. Sarstedt et al. / Journal of Business Research 69 (2016) 39984010

Numerous studies have explored PLS's performance in terms of pa- applies to reective measurement models, while formative measure-
rameter accuracy when data are assumed to follow a common factor ment models remain unchanged.
model approach (e.g., Barroso, Cepeda Carrin, & Roldn, 2010; The path model and path coefcient specications used in the simu-
Hwang, Malhotra, Kim, Tomiuk, & Hong, 2010; Marcoulides, Chin, & lations (Fig. 2) are identical to Reinartz et al. (2009) with low (i.e., 0.15;
Saunders, 2012; Reinartz et al., 2009). Overall, these studies suggest p1, p2, p12), medium (i.e., 0.30; p5), and high (i.e., 0.50; p3, p4, p6, p9, p10,
that the bias that PLS produces when estimating common factor models p11) pre-specied path coefcients. Accounting for corresponding calls
is comparably small provided that the measurement models meet min- in the literature (Marcoulides et al., 2012), we extended the original
imum recommended standards in terms of the number of indicators model by adding a construct (Y5) with two null paths (p7 and p8). Also
and indicator loadings. Recent efforts to dramatize the differences be- analogous to Reinartz et al. (2009), all measurement models are reec-
tween CBSEM and PLS estimates (Rnkk et al., 2016) in, for example, tive. Table 1 illustrates the design factors and their levels manipulated in
Reinartz et al.'s (2009) study focused on descriptive differences be- the simulation study. The simulation study uses a factorial design. We
tween population values and parameter estimates only, disregarding conducted 300 replications of each factor-level combination to obtain
the concept of statistical inference. As Reinartz et al. (2009, p. 338; stable average outcomes for our analysis. In summary, the analysis in-
emphasis added by the authors) note in their results description of all cludes 4 4 3 5 300 = 72,000 datasets for Study I (i.e., the common
simulation conditions, parameter estimates do not differ signicantly factor-based simulation) and 8 3 5 300 = 36,000 datasets for
from their theoretical values for either ML-based CBSEM (p-values be- Study II (i.e., the composite-based simulation), which results in a total
tween 0.3963 and 0.5621) or PLS (p-values between 0.1906 and number of 324,000 computations for the three methods under research.
0.3449). Only when the model estimation draws on a very large sample In line with related research in the eld (e.g., Becker, Rai, Ringle, &
size (N = 10,000) and includes measurement models with many indi- Vlckner, 2013; Reinartz et al., 2009), common factor model-based data
cators with high loadings, did statistically signicant differences occur. generation was performed by means of Mattson's (1997) method (also
Correspondingly, empirical studies using both methods suggest that see Reinartz, Echambadi, & Chin, 2002), where univariate random vari-
the divergence between PLS and CBSEM results when estimating com- ables initially serve the generation of the latent variables in the structural
mon factor models is of little practical relevance for the results' implica- model, followed by the computation of the observed variables. The com-
tions (e.g., Astrachan, Patel, & Wanzenried, 2014). posite model-based data generation used in this study draws on a proce-
The question, however, is whether the bias identied in prior studies dure similar to the one that Schlittgen (2015) presents in his SEGIRLS
results from using composite-based PLS on common factor model data package for the statistical R software (R Core Team, 2014). We rst gener-
or if the method is inherently biased, including when estimating com- ate the model-implied covariance matrix of the indicators, followed by a
posite models. Similarly, while the (supposed) PLS bias has been exten- Cholesky decomposition, and nally extract the indicator data for a pre-
sively debated in the literature, the bias that CBSEM produces when specied number of observations and the sought data distribution. For
mistakenly estimating composite models has not yet been explored. model estimation based on PLS, PLSc, and CBSEM, we use the semPLS
For this reason, the following simulation studies focus on revealing the (Monecke & Leisch, 2012), matrixpls R (Rnkk, 2016), and sem (Fox
biases that occur when using (1) composite-based PLS to estimate com- et al., 2015) packages of the R software. As in Reinartz et al. (2009),
mon factor models, and (2) common factor-based CBSEM to estimate CBSEM estimation draws on the standard maximum likelihood approach;
composite models. Furthermore, both studies consider PLSc. PLS uses Mode A estimation while PLSc uses Mode A estimation followed
by the correction for attenuation in both studies.
4.2. Simulation studies
5. Results
Our studies replicate Reinartz et al.'s (2009) simulation study on the
comparative performance of PLS and CBSEM, which in its original form The assessment of the methods' parameter accuracy occurs on the
assumed a common factor model. We extended the original study, how- grounds of the mean absolute error MAE, which is dened as
ever, by additionally generating composite model-based data. Further-
more, our studies also consider PLSc, which follows a composite 1 t ^

MAE j1 j j ; 4
modeling logic but mimics a common factor model (Sarstedt, Ringle, & t
Hair, 2014). To do so, the method rst estimates the model parameters
using the standard PLS algorithm and corrects these estimates for atten- where t equals the number of parameters, j is the prespecied param-
uation using the consistent reliability coefcient A. This correction only eter and ^ j is the parameter estimate in any replication. Tables 2 and 3

Fig. 2. Simulation model.


M. Sarstedt et al. / Journal of Business Research 69 (2016) 39984010 4005

Table 1
Simulation design.

Design factors Study

Study I: Common factor-based simulation Study II: Composite-based simulation

Representation of the constructs and data generation Common factor model Composite model
Structural model Reinartz et al. (2009), extended by null paths Reinartz et al. (2009), extended by null paths
Loadings/weights and number of indicators All possible combinations of number of indicators (2, 4, 6, 8) Eight combinations of number of indicators and
and indicator loadings (equal: 0.50, 0.70 or 0.90; and unequal: indicator weights:
half of the indicators 0.50, the other half 0.90) Unequal weights
0.30/0.90;
0.10/0.30/0.50/0.70;
0.10/0.175/0.25/0.325/0.40/0.475; and
0.075/0.125/0.175/0.225/0.275/0.325/
0.375/0.425
Equal weights
0.625/0.625;
0.40/0.40/0.40/0.40;
0.30/0.30/0.30/0.30/0.30/0.30; and
0.25/0.25/0.25/0.25/0.25/0.25/0.25/0.25
Data distribution Three variations of skewness/kurtosis: none, moderate, and high Three variations of the normal distribution:
symmetric normal, log-normal, and diff-normal
Sample size 100, 250, 500, 1000, and 10,000 100, 250, 500, 1000, and 10,000

Notes: In composite models, the indicators fully explain the latent variable, which imposes some restrictions on the possibilities of cases that can be drawn; log-normal = random variable
which has a standard normally distributed logarithm; diff-normal = difference of two log-normal distributions.

illustrate the results of the simulation studies. Our illustration focuses measurement models with 2 indicators and unequal weights, PLSc's
on the case of normally distributed data as the analysis of non-normal MAE values decrease from 100 to 500 observations, increase at 1000 ob-
data yields highly similar results. servations, and nally decrease at 10,000 observations. Overall our sim-
Our results conrm the well-known PLS bias when using the meth- ulation study results suggest that when the underlying model type is
od to estimate the path model with common factor model-based data. unknown, researchers are well advised to draw on PLS in order to
PLS produces biased estimates with average MAE values of around avoid substantial parameter biases that result from using PLSc or
0.10 when the measurement models only have two indicators or CBSEM in case the composite model holds.
when the loadings are low (i.e., 0.50). Conrming PLS's consistency at In summary, we nd that the methods' parameter bias depends on
large (Hui & Wold, 1982), PLS's MAE values decrease for greater num- the underlying model and data. If one assumes a common factor
bers of indicators per measurement model or higher sample sizes. Com- model and draws on data of such a nature, CBSEM generally
pared to PLS, CBSEM achieves lower MAE values across all conditions performsas expectedvery well. The same generally holds for PLSc,
except for small sample sizes of 100. In this condition, PLSc shows pro- except when the sample size is small. The PLS method offers a very
nounced MAE values of up to 0.34 and also performs weak at 250 obser- good approximation in this case. At the same time, PLS performsas
vations when measurement models have only two indicators or low expectedvery well for composite models, if one draws on data of
loadings. However, PLSc's performance increases considerably with such a nature. In this case, however, CBSEM and PLSc perform very
more indicators and higher sample sizes. On average across all simula- poorly.
tion conditions, PLS and PLSc have a higher MAE (0.07) compared to Table 4 summarizes the results of prior research on the methods'
CBSEM (0.05). Clearly, the differences between the three methods performance in terms of parameter bias when estimating common fac-
when used on common factor model-based data are overall only mar- tor models with effect, causal, and composite indicators. Furthermore,
ginal, however. the table summarizes the results of this paper's simulation studies re-
A different picture emerges when estimating data from a composite garding the methods' performances when estimating composite models
model population. Whereas PLS has an overall MAE value of 0.04, the with effect indicators. In line with the nature of each data generation ap-
parameter biases of CBSEM (0.76) and particularly PLSc (3.70) are proach and the methods' way of treating construct measures
much more pronounced. PLSc shows a bewildering performance across (e.g., Diamantopoulos, 2011; Henseler, Hubona et al., 2016), we differ-
the simulation conditions with MAE values ranging from 0.64 to 17.89. entiate between (1) effect indicator models and causal indicator models
Specically, in conditions with four indicators, equal weights and 500, when the underlying population is common factor-based, and (2) be-
and 10,000 observations, respectively, MAE values bounce up to values tween effect indicator models and composite indicator models when
higher than 10. To rule out potential problems resulting from the PLSc the underlying population is composite-based.
implementation of the matrixpls package, we re-ran the simulation
study using our PLSc extension of the semPLS package (Monecke & 6. Conclusion
Leisch, 2012). Results from this additional analysis provided support
for the extent and variation of PLSc's bias with MAE values well above Professional statisticians tend to know little about factor analysis
10 for several simulation conditions. Similar to PLSc, CBSEM shows pro- and seldom practice it. Indeed, statisticians mostly have a cool negative
nounced parameter estimation biases across all simulation conditions attitude towards the subject. They hardly ever write about it. [] I can
but at a much lower level with MAE values ranging from 0.51 to 1.72. see nothing advantageous in factor analytic methods. Factor analysis is
Nevertheless, CBSEM's bias when estimating composite models is on technically under-developed and at times appears almost cretinous. Its
average 11 times higher than PLS's bias when estimating common fac- practitioners seem to be largely unaware of the technical and methodo-
tor models. Clearly, the use of PLS to estimate common factor models logical problems, which they let themselves in for. This text, which is
is much less of an issue than using CBSEM on data consistent with the more than fty years old and taken from Ehrenberg's (1962, p. 191
composite model. Finally, while PLS's MAE values decrease when sam- and p. 206) article Some Questions About Factor Analysis, appears sur-
ple sizes increase, this is not the case with CBSEM and PLSc. For these prising considering that today factor analysis is one of the success
two methods, the MAE values show no clear pattern. For example, for stories of statistical analysis (Cudeck & MacCallum, 2007). This
4006 M. Sarstedt et al. / Journal of Business Research 69 (2016) 39984010

Table 2 Table 3
Coefcients' mean absolute error (MAE) in the common factor model situation. Coefcients' mean absolute error (MAE) in the composite model situation.

Design factor Mean absolute error Design factor Mean absolute error (MAE)
(MAE)
Observations Indicators Weights PLS PLSc CBSEM
Observations Group Loadings PLS PLSc CBSEM
100 2 Equal 0.07 2.90 0.84
100 2 indicators Mixed 0.11 0.30 0.13 4a 0.07 4.42 0.74
4 indicators 0.09 0.13 0.10 6a 0.07 3.57 0.63
6 indicators 0.08 0.09 0.08 8a 0.07 2.05 0.52
8 indicators 0.08 0.12 0.08 2 Unequal 0.07 5.05 0.61
Loadings: 0.5 Equala 0.13 0.34 0.16 4a 0.08 3.03 0.92
Loadings: 0.7 0.09 0.10 0.10 6a 0.07 5.83 0.54
Loadings: 0.9 0.06 0.07 0.07 8a 0.07 2.89 0.48
Loadings: 0.5/0.9 Unequala 0.08 0.10 0.08 250 2 Equal 0.05 5.33 0.81
250 2 indicators Mixed 0.10 0.13 0.09 4a 0.04 5.56 0.81
4 indicators 0.08 0.07 0.06 6a 0.05 3.22 0.73
6 indicators 0.06 0.05 0.05 8a 0.04 3.71 0.57
8 indicators 0.06 0.06 0.05 2 Unequal 0.05 2.94 0.54
Loadings: 0.5 Equala 0.12 0.14 0.10 4a 0.05 4.06 0.82
Loadings: 0.7 0.07 0.06 0.06 6a 0.05 2.76 0.58
Loadings: 0.9 0.04 0.04 0.04 8a 0.05 4.20 0.59
Loadings: 0.5/0.9 Unequala 0.06 0.05 0.05 500 2 Equal 0.03 5.21 0.90
500 2 indicators Mixed 0.10 0.07 0.06 4a 0.03 11.55 0.81
4 indicators 0.07 0.05 0.04 6a 0.03 2.87 0.82
6 indicators 0.06 0.04 0.04 8a 0.03 2.89 0.75
8 indicators 0.05 0.04 0.04 2a Unequal 0.03 2.29 0.55
Loadings: 0.5 Equala 0.12 0.08 0.07 4a 0.03 1.17 0.62
Loadings: 0.7 0.07 0.04 0.04 6a 0.03 2.52 0.60
Loadings: 0.9 0.03 0.03 0.03 8a 0.03 4.96 0.56
Loadings: 0.5/0.9 Unequala 0.05 0.04 0.03 1000 2 Equal 0.02 2.28 0.98
1000 2 indicators Mixed 0.09 0.05 0.04 4a 0.02 5.52 0.80
4 indicators 0.06 0.03 0.03 6a 0.02 1.38 0.95
6 indicators 0.05 0.03 0.03 8a 0.02 2.74 0.88
8 indicators 0.05 0.03 0.03 2a Unequal 0.03 4.39 0.55
Loadings: 0.5 Equala 0.12 0.05 0.05 4a 0.02 0.81 0.65
Loadings: 0.7 0.06 0.03 0.03 6a 0.02 2.29 0.65
Loadings: 0.9 0.03 0.02 0.02 8a 0.02 6.88 0.65
Loadings: 0.5/0.9 Unequala 0.05 0.03 0.02 10,000 2 Equal 0.01 1.00 1.34
10,000 2 indicators Mixed 0.09 0.01 0.01 4 0.01 17.89 0.74
4 indicators 0.06 0.01 0.01 6a 0.01 0.85 1.72
6 indicators 0.04 0.01 0.01 8a 0.01 1.06 1.22
8 indicators 0.04 0.01 0.01 2a Unequal 0.01 2.44 0.66
Loadings: 0.5 Equala 0.11 0.02 0.01 4a 0.01 0.64 0.51
Loadings: 0.7 0.06 0.01 0.01 6a 0.01 1.04 1.02
Loadings: 0.9 0.02 0.01 0.01 8a 0.01 1.59 0.58
Loadings: 0.5/0.9 Unequala 0.04 0.01 0.01 Total 0.04 3.70 0.76
Total 0.07 0.07 0.05 a
Instances in which CBSEM converged in b50% of the simulation runs.
a
Across all numbers of indicators.

2012), research on the performance of PLS has repeatedly ignored the


implications of using a composite-based method for estimating com-
assessment sounds familiar to everyone who has been exposed to re- mon factor models (Becker, Rai, & Rigdon, 2013). Recent efforts to
cent papers critically referring to the PLS method. Authors have repeat- align reective measurement and composite-based modeling (Dijkstra
edly suggested that PLS has largely been ignored in research methods & Henseler, 2015; Henseler, Hubona et al., 2016)while commendable
journals (Rnkk & Evermann, 2013, p. 426), that its use is restricted from a methodological viewpointhave instead contributed to the
to few domains (Rnkk et al., 2016; Rnkk et al., 2015) and that confusion, leaving researchers with little guidance regarding when to
PLS is not useful for statistical estimation and testing (Rnkk et al., apply each method and how to align their use with measurement
2015, p. 76). While we do not suggest that PLS will undergo a similar de- considerations.
velopment as factor analysis, the statements about the limitations of The framework in Fig. 3 merges our theoretical discussions and sim-
factor analysis and PLS nicely show how unsubstantiated some method- ulation results. Whereas the theoretical layer serves to dene the
ological discussions can become. As noted elsewhere, any extreme po- conceptual variable, the conceptual layer delivers the operational de-
sition that (oftentimes systematically) neglects the benecial features of nition of the conceptual variables, which then serves as the basis for
the other technique, and may result in prejudiced boycott calls, is not the measurement operationalization using effect, causal, or composite
good research practice and does not help to truly advance our under- indicators on the operational layer. This conceptualization and
standing of methods and any other research subject (Sarstedt, Ringle, operationalization of construct measures represents the measurement
Henseler et al., 2014, p. 158). perspective. This perspective needs to be complemented with the
Our discussions show that researchers need to clearly distinguish model estimation perspective. The estimation layer intertwines with
between (conceptual) measurement approaches and the (statistical) the measurement model layer that expresses how the data represent
estimation perspectives when judging the appropriateness of or choos- reectively or formatively specied measurement models.
ing a specic SEM method. Model estimation does not occur in a meth- By exploring the performance of CBSEM, PLS, and PLSc when esti-
odological vacuum detached from measurement considerations but mating composite models, the simulation studies overcome a crucial
rests on specic assumptions, which need to be considered when con- limitation of prior studies, which univocally relied on data from com-
ceptualizing and operationalizing models and vice versa. Despite fre- mon factor model populations to judge their universal efcacy (Chin,
quent warnings (Chin, 2010; Henseler et al., 2014; Marcoulides et al., 2010; Marcoulides & Chin, 2013). Therefore, our studies address
M. Sarstedt et al. / Journal of Business Research 69 (2016) 39984010 4007

Table 4
Efcacy of PLS, PLSc, and CBSEM for estimating common factor and composite models.

Method Data type Measurement Performance Description Sample references


model
operationalization

PLS Common Effect indicators (+) Small bias Reinartz et al. (2009), this
factor study
Causal indicators PLS cannot model the construct-level error term Diamantopoulos (2011)
Composite Effect indicators + Very small bias This study
Composite + Small bias for small sample sizes, which approaches zero for Becker, Rai, Rigdon (2013)
indicators increasing sample sizes
PLSc Common Effect indicators (+) Small bias, which is however higher than that of CBSEM for Dijkstra and Henseler (2015),
factor small sample sizes this study
Causal indicators PLSc cannot model the construct-level error term Diamantopoulos (2011)
Composite Effect indicators Pronounced bias This study
Composite + Results parallel those from PLS as no correction for Becker, Rai, Rigdon (2013)
indicators attenuation occurs
CBSEM Common Effect indicators + Very small bias except for small sample sizes of 100 Reinartz et al. (2009), this
factor study
Causal indicators (+) Model identication via MIMIC models or by specication Diamantopoulos and Rieer
of select parameters (2011)
Composite Effect indicators Increased bias This study
Composite Model identication by specication of select parameters but strong limitations in Grace and Bollen (2008)
indicators terms of inference

Notes: not recommended/possible, (+) acceptable, + recommended.

corresponding calls for future research, such as expressed by Hwang generalized structured component analysis. We adopted the procedure
et al. (2010, p. 710) in their comparative study on parameter recovery because it was rather difcult to arrive at an impartial way of generating
of common factor-based SEM, PLS, and generalized structured compo- synthetic data for all three different approaches. Nevertheless, the same
nent analysis: We generated simulated data on the basis of covariance procedure has been used in other studies that compared the perfor-
structure analysis. This data generation procedure may have had an un- mance of covariance structure analysis with that of partial least squares
favorable effect on the performance of partial least squares and (). In any case, it appears necessary in future studies to investigate

Fig. 3. Measurement and model estimation framework.


4008 M. Sarstedt et al. / Journal of Business Research 69 (2016) 39984010

whether a particular data generation procedure may inuence the rela- Our ndings suggest that composite-based methods are going to
tive performance of the different approaches. Our results show that play a greater role in future SEM applications. To date our understand-
Hwang et al.'s (2010) notion can be answered with a resounding "yes." ing of this strand of methods is incomplete, however, as prior assess-
The results outlined in this paper show that PLS entails practically no ments universally drew on common factor model-based data and
bias when estimating data from a composite model population, regard- thereby relied on misspecied populations (Rigdon, 2016). Therefore,
less of whether the measurement models are reective or formative future research should aim at broadening our knowledge of the relative
(Table 3). Biases are somewhat higher for common factor model popu- performance of the different approaches on the grounds of composite
lations (Table 2), but low in absolute terms. Clearly, PLS is optimal for model-based data. For example, studies should contrast PLS's perfor-
estimating composite models while simultaneously allowing the ap- mance with other composite-based SEM techniques such as generalized
proximation of common factor models involving effect indicators with structured components analysis (Hwang et al., 2010) or regularized
practically no limitations (see the solid lines between composite indica- generalized canonical correlation analysis (Tenenhaus & Tenenhaus,
tors/effect indicators and composite model and the dashed line between 2011).
effect indicators and common factor model in Fig. 3; also see Table 4). In In doing so, future research should consider a broader range of
contrast, CBSEM and PLSc estimation of reectively measured con- model constellations and more complex model structures such as hier-
structs when the data stem from a composite population entails severe archical component models, moderating effects, or nonlinear effects.
biases in parameter estimates, rendering their use inappropriate in Such assessments would help disclose the different methods' efcacy
these instances (no line between effect indicators and composite for different situations that researcher encounter in their studies. By ex-
model in Fig. 3; also see Table 4). Particularly PLSc shows a bewildering amining CBSEM's performance on composite model data, this study
behavior with strong biases across practically all conditions, which do complements prior research, which univocally examined PLS's perfor-
not diminish as sample size increases. When using PLSc to estimate mance on common factor model data. However, future research should
measurement models with composite indicators using data that stem compare PLS, PLSc, and CBSEM on data where both models t in the
from a composite model population, the PLSc results parallel those population. Such a design would provide supplementary insights, be-
from PLS as no correction for attenuation occurs (see the solid line be- cause CBSEM may work well on some composite measures but not
tween composite indicators and composite model in Fig. 3; also see others. In addition, with regards to PLS, future research should explore
Table 4). the interplay between measurement specications, population type,
When estimating data from common factor populations, CBSEM's and PLS's estimation modes (i.e., Mode A and B). These results would
parameter bias is small for a sample size of 250 and quickly diminishes help clarifying the estimation modes' efcacy for out-of-sample predic-
for higher sample sizes (see solid line between effect indicators and tion, in-sample-prediction, and parameter bias under different model
common factor model in Fig. 3; also see Table 4). PLSc shows a similar specication and data conditions.
pattern when estimating data from common factor populations but per-
forms less well for small sample sizes of 100, where MAE values peak at
0.34 (see dashed line between effect indicators and common factor References
model in Fig. 3; also see Table 4). In this situation, PLS outperforms Albers, S. (2010). PLS and success factor studies in marketing. In V. Esposito Vinzi, W. W.
the other methods but overall, the differences are marginal (Chin, Chin, J. Henseler, & H. Wang (Eds.), Handbook of partial least squares: Concepts,
1998; Fornell & Bookstein, 1982; also see Goodhue et al., 2012). Note methods and applications (Springer handbooks of computational statistics series, vol.
II) (pp. 409425). Heidelberg, Dordrecht, London, New York: Springer.
that other CBSEM estimators than ML (e.g., GLS, ULS, and ADF) entail Antonakis, J., Bendahan, S., Jacquart, P., & Lalive, R. (2010). On making causal claims: A re-
further biases when estimating common factor models; see for example view and recommendations. The Leadership Quarterly, 21(6), 10861120.
Boomsma and Hoogland (2001) and Dijkstra and Henseler (2015). Astrachan, C. B., Patel, V. K., & Wanzenried, G. (2014). A comparative study of CB-SEM and
PLS-SEM for theory development in family rm research. Journal of Family Business
The obvious problem with these observations is that researchers can Strategy, 5(1), 116128.
hardly know whether the data's nature is common factor- or Atinc, G., Simmering, M. J., & Kroll, M. J. (2012). Control variable use and reporting in
composite-based. Fit measures such as the standardized root mean macro and micro-management research. Organizational Research Methods, 15(1),
5774.
square residual (SRMR) may provide an indication of whether the
Bandalos, D. L., & Boehm-Kaufman, M. R. (2009). Four common misconceptions in explor-
data follow a common factor model. If the specic measurement atory factor analysis. In C. E. Lance, & R. J. Vandenberg (Eds.), Statistical and method-
model does not meet the required level (e.g., 0.08 and smaller for the ological myths and urban legends: Doctrine, verity and fable in the organizational and
social sciences (pp. 6187). New York, NY: Routledge.
SRMR; Hu & Bentler, 1998), this result suggests that the data follow a
Barroso, C., & Picn, A. (2012). Multi-dimensional analysis of perceived switching costs.
composite model. Alternatively, an improper CBSEM solution may Industrial Marketing Management, 41(3), 531543.
point to an underlying composite model population. Our results show Barroso, C., Cepeda Carrin, G., & Roldn, J. L. (2010). Applying maximum likelihood and
that CBSEM produces improper solutions in up to 99% of cases when PLS on different sample sizes: Studies on SERVQUAL model and employee behavior
model. In V. Esposito Vinzi, W. W. Chin, J. Henseler, & H. Wang (Eds.), Handbook of
the composite model holds. In consideration that in practical applica- partial least squares (pp. 427447). Springer Berlin Heidelberg.
tions improper solutions often occur in CBSEM use (Rigdon, 2012; Baxter, R. A. (2009). Reective and formative metrics of relationship value: A commen-
Sarstedt, Ringle, Henseler et al., 2014), these results offer a potential ex- tary essay. Journal of Business Research, 62(12), 13701377.
Bearden, W. O., Netemeyer, R. G., & Haws, K. L. (2011). Handbook of marketing scales.
planation why, more often than not, the common factor model cannot Multi-item measures of marketing and consumer behavior research (3rd ed.). Sage.
be supported in practice (Atinc et al., 2012). At the same time, howev- Becker, J. -M., Rai, A., & Rigdon, E. E. (2013). Predictive validity and formative measurement
er, reasons for improper solutions are manifold and not restricted to in structural equation modeling: Embracing practical relevance. Proceedings of the 34th
International Conference on Information Systems, Milan, Italy.
the misspecication of the model type. Therefore, interpreting im- Becker, J. -M., Rai, A., Ringle, C. M., & Vlckner, F. (2013). Discovering unobserved hetero-
proper solutions as clear evidence for an underlying composite geneity in structural equation models to avert validity threats. MIS Quarterly, 37(3),
model is not reasonable. As an alternative, researchers can follow a 665694.
Bentler, P. M., & Huang, W. (2014). On components, latent variables, PLS and simple
multi-methods approach, in which they combine CBSEM with PLS
methods: Reactions to Rigdon's rethinking of PLS. Long Range Planning, 47(3),
to see whether the results align in that specic research situation. 138145.
Substantial differences between the methods indicate that the un- Bollen, K. A. (1989). Structural equations with latent variables. New York: Wiley.
Bollen, K. A. (2002). Latent variables in psychology and the social sciences. Annual Review
derlying population is composite-based, supporting the use of
of Psychology, 53(1), 605634.
composite-based SEM methods. Nevertheless, in light of the biases Bollen, K. A. (2011). Evaluating effect, composite, and causal indicators in structural equa-
that come with a CBSEM and PLSc-based estimations of composite tion models. MIS Quarterly, 35(2), 359372.
model data, PLS is certainly the safer option when estimating data Bollen, K. A., & Bauldry, S. (2011). Three Cs in measurement models: Causal indicators,
composite indicators, and covariates. Psychological Methods, 16(3), 265284.
from an unknown population until research has proposed clear Bollen, K. A., & Davies, W. R. (2009). Causal indicator models: Identication, estimation,
guidelines on how to identify the population type. and testing. Structural Equation Modeling, 16(3), 498522.
M. Sarstedt et al. / Journal of Business Research 69 (2016) 39984010 4009

Bollen, K. A., & Diamantopoulos, A. (2016). In defense of causal-formative indicators: A Henseler, J., Ringle, C. M., & Sarstedt, M. (2016). Testing measurement invariance of com-
minority report. Psychological Methods (forthcoming). posites using partial least squares. International Marketing Review, 33(3), 405431.
Bollen, K. A., & Lennox, R. (1991). Conventional wisdom on measurement: A structural Howell, R. D., Breivik, E., & Wilcox, J. B. (2013). Formative measurement: A critical per-
equation perspective. Psychological Bulletin, 110(2), 305314. spective. The DATA BASE for Advances in Information Systems, 44(4), 4455.
Boomsma, A., & Hoogland, J. J. (2001). The robustness of LISREL modeling revisited. In R. Hu, L. -t., & Bentler, P. M. (1998). Fit indices in covariance structure modeling: Sensitivity
Cudeck, S. du Toit, & D. Srbom (Eds.), Structural equation modeling: Present and future to underparameterized model misspecication. Psychological Methods, 3(4),
(pp. 139168). Chicago: Scientic Software International. 424453.
Borsboom, D., Mellenbergh, G. J., & van Heerden, J. (2003). The theoretical status of latent Hui, B. S., & Wold, H. O. A. (1982). Consistency and consistency at large of partial least
variables. Psychological Review, 110(2), 203219. squares estimates. In K. G. Jreskog, & H. O. A. Wold (Eds.), Systems under indirect ob-
Bruner, G. C., James, K. E., & Hensel, P. J. (2001). Marketing scales handbook: A compilation servation, part II (pp. 119130). Amsterdam: North Holland.
of multi-item measures: Butterworth-Heinemann. Hwang, H., Malhotra, N. K., Kim, Y., Tomiuk, M. A., & Hong, S. (2010). A comparative study
Chang, W., Franke, G. R., & Lee, N. (2016). Comparing reective and formative measures: New on parameter recovery of three approaches to structural equation modeling. Journal
insights from relevant simulations. Journal of Business Research, 69(10), 31773185. of Marketing Research, 47(4), 699712.
Chin, W. W. (1998). Commentary: Issues and opinion on structural equation modeling. Jarvis, C. B., MacKenzie, S. B., & Podsakoff, P. M. (2003). A critical review of construct indi-
MIS Quarterly, 22(1), xiixvi. cators and measurement model misspecication in marketing and consumer re-
Chin, W. W. (2010). How to write up and report PLS analyses. In V. Esposito Vinzi, W. W. search. Journal of Consumer Research, 30(2), 199218.
Chin, J. Henseler, & H. Wang (Eds.), Handbook of partial least squares: Concepts, Jones, M. A., Mothersbaugh, D. L., & Beatty, S. E. (2000). Switching barriers and repurchase
methods and applications (Springer handbooks of computational statistics series, vol. intentions in services. Journal of Retailing, 76(2), 259274.
II) (pp. 655690). Heidelberg, Dordrecht, London, New York: Springer. Jreskog, K. G. (1978). Structural analysis of covariance and correlation matrices.
Churchill, G. A. (1979). A paradigm for developing better measures of marketing con- Psychometrika, 43(4), 443477.
structs. Journal of Marketing Research, 16(1), 6473. Jreskog, K. G., & Goldberger, A. S. (1975). Estimation of a model with multiple indicators
Cliff, N. (1983). Some cautions concerning the application of causal modeling methods. and multiple causes of a single latent variable. Journal of the American Statistical Asso-
Multivariate Behavioral Research, 18(1), 115126. ciation, 70(351), 631639.
Coltman, T., Devinney, T. M., Midgley, D. F., & Venaik, S. (2008). Formative versus reec- Jreskog, K. G., & Wold, H. O. A. (1982). The ML and PLS techniques for modeling with la-
tive measurement models: Two applications of formative measurement. Journal of tent variables: Historical and comparative aspects. In H. O. A. Wold, & K. G. Jreskog
Business Research, 61(12), 12501262. (Eds.), Systems under indirect observation, part I (pp. 263270). Amsterdam: North-
Cudeck, R., & MacCallum, R. C. (Eds.). (2007). Factor analysis at 100: Historical Holland.
developments and future directions. Mahwah, NJ: Lawrence Erlbaum Associates. Kaufmann, L., & Gaeckler, J. (2015). A Structured Review of Partial Least Squares in Supply
Diamantopoulos, A. (1994). Modelling with LISREL: A guide for the uninitiated. Journal of Chain Management Research. Journal of Purchasing and Supply Management
Marketing Management, 10(13), 105136. (forthcoming).
Diamantopoulos, A. (2005). The C-OAR-SE procedure for scale development in market- Law, K. S., & Wong, C. -S. (1999). Multidimensional constructs in structural equation anal-
ing: A comment. International Journal of Research in Marketing, 22(1), 19. ysis: An illustration using the job perception and job satisfaction constructs. Journal of
Diamantopoulos, A. (2006). The error term in formative measurement models: Management, 25(2), 143160.
Interpretation and modeling implications. Journal of Modelling in Management, 1(1), Law, K. S., Wong, C. -S., & Mobley, W. H. (1998). Toward a taxonomy of multidimensional
717. constructs. The Academy of Management Review, 23(4), 741755.
Diamantopoulos, A. (2011). Incorporating formative measures into covariance-based Lee, N., & Cadogan, J. W. (2013). Problems with formative and higher-order reective var-
structural equation models. MIS Quarterly, 35(2), 335358. iables. Journal of Business Research, 66(2), 242247.
Diamantopoulos, A., & Rieer, P. (2011). Using formative measures in international mar- Lohmller, J. -B. (1989). Latent variable path modeling with partial least squares. Heidel-
keting models: A cautionary tale using consumer animosity as an example. Advances berg: Physica.
in International Marketing, 10(22), 1130. MacCallum, R. C., & Browne, M. W. (1993). The use of causal indicators in
Diamantopoulos, A., & Winklhofer, H. M. (2001). Index construction with formative indi- covariance structure models: Some practical issues. Psychological Bulletin, 114(3),
cators: An alternative to scale development. Journal of Marketing Research, 38(2), 533541.
269277. MacCallum, R. C., Browne, M. W., & Cai, L. (2007). Factor analysis models as approxima-
Diamantopoulos, A., Rieer, P., & Roth, K. P. (2008). Advancing formative measurement tions: Some history and some implications. In R. Cudeck, & R. C. MacCallum (Eds.),
models. Journal of Business Research, 61(12), 12031218. Factor analysis at 100: Historical developments and future directions (pp. 153175).
Dijkstra, T. K., & Henseler, J. (2011). Linear indices in nonlinear structural equation Mahwah, NJ: Laurence Erlbaum Associates.
models: Best tting proper indices and other composites. Quality & Quantity, 45(6), MacKenzie, S. B. (2003). The dangers of poor construct conceptualization. Journal of the
15051518. Academy of Marketing Science, 31(3), 323326.
Dijkstra, T. K., & Henseler, J. (2015). Consistent partial least squares path modeling. MIS MacKenzie, S. B., Podsakoff, P. M., & Podsakoff, N. P. (2011). Construct measurement and
Quarterly, 39(2), 297316. validation procedures in MIS and behavioral research: Integrating new and existing
Edwards, J. R., & Bagozzi, R. P. (2000). On the nature and direction of relationships be- techniques. MIS Quarterly, 35(2), 293A295.
tween constructs and measures. Psychological Methods, 5(2), 155174. Marcoulides, G. A., & Chin, W. W. (2013). You write, but others read: Common
Ehrenberg, A. S. C. (1962). Some questions about factor analysis. Journal of the Royal methodological misunderstandings in PLS and related methods. In H. Abdi, W.
Statistical Society Series D (The Statistician), 12(3), 191208. W. Chin, V. Esposito Vinzi, G. Russolillo, & L. Trinchera (Eds.), New Perspectives
Fornell, C. G., & Bookstein, F. L. (1982). Two structural equation models: LISREL and in Partial Least Squares and Related Methods. Vol. 56. (pp. 3164). New York:
PLS applied to consumer exit-voice theory. Journal of Marketing Research, Springer.
19(4), 440452. Marcoulides, G. A., & Saunders, C. (2006). PLS: A silver bullet? MIS Quarterly, 30(2), IIIIIX.
Fox, J., Nie, Z., Byrnes, J., Culbertson, M., DebRoy, S., Friendly, M., ... Monette, G. (2015). R Marcoulides, G. A., Chin, W. W., & Saunders, C. (2012). When imprecise statistical state-
package sem: Structural equation models (version 3.1-6). http://cran.r-project.org/ ments become problematic: A response to Goodhue, Lewis, and Thompson. MIS
web/packages/sem/ Quarterly, 36(3), 717728.
Gelhard, C., & von Delft, S. (2016). The role of organizational capabilities in achieving su- Mattson, S. (1997). How to generate non-normal data for simulation of structural equa-
perior sustainability performance. Journal of Business Research. http://dx.doi.org/10. tion models. Multivariate Behavioral Research, 32(4), 355373.
1016/j.jbusres.2016.03.053 (forthcoming). McDonald, R. P. (1996). Path analysis with composite variables. Multivariate Behavioral
Gilliam, D. A., & Voss, K. (2013). A proposed procedure for construct denition in market- Research, 31(2), 239270.
ing. European Journal of Marketing, 47(1/2), 526. Michell, J. (2013). Constructs, inferences and mental measurement. New Ideas in
Goodhue, D. L., Lewis, W., & Thompson, R. (2012). Does PLS have advantages for small Psychology, 31(1), 1321.
sample size or non-normal data? MIS Quarterly, 36(3), 9811001. Monecke, A., & Leisch, F. (2012). semPLS: Structural equation modeling using partial least
Grace, J. B., & Bollen, K. A. (2008). Representing general theoretical concepts in structural squares. Journal of Statistical Software, 48(3), 132.
equation models: The role of composite variables. Environmental and Ecological Nunnally, J. C., & Bernstein, I. (1994). Psychometric theory (3 ed.). New York: McGraw Hill.
Statistics, 15(2), 191213. Pedhazur, E. J., & Pedhazur Schmelkin, L. (1991). Measurement, design, and analysis: An in-
Guide, V. D. R., & Ketokivi, M. (2015). Notes from the editors: Redening some methodo- tegrated approach. Mahwah, NJ: Lawrence Erlbaum.
logical criteria for the journal. Journal of Operations Management, 37, vviii (July 2015). Peng, D. X., & Lai, F. (2012). Using partial least squares in operations management re-
Hair, J. F., Ringle, C. M., & Sarstedt, M. (2011). PLS-SEM: Indeed a silver bullet. Journal of search: A practical guideline and summary of past research. Journal of Operations
Marketing Theory and Practice, 19(2), 139151. Management, 30(6), 467480.
Hair, J. F., Sarstedt, M., Ringle, C. M., & Mena, J. A. (2012). An assessment of the use of par- R Core Team (2014). R: A language and environment for statistical computing. Vienna,
tial least squares structural equation modeling in marketing research. Journal of the Austria: R Foundation for Statistical Computing.
Academy of Marketing Science, 40(3), 414433. Reinartz, W. J., Echambadi, R., & Chin, W. W. (2002). Generating non-normal data for sim-
Hair, J. F., Hult, G. T. M., Ringle, C. M., & Sarstedt, M. (2017). A primer on partial least squares ulation of structural equation models using Mattson's method. Multivariate
structural equation modeling (PLS-SEM) (2nd ed.). Thousand Oaks, CA: Sage. Behavioral Research, 37(2), 227244.
Henseler, J., Dijkstra, T. K., Sarstedt, M., Ringle, C. M., Diamantopoulos, A., Straub, D. W., ... Reinartz, W. J., Haenlein, M., & Henseler, J. (2009). An empirical comparison of the efcacy
Calantone, R. J. (2014). Common beliefs and reality about partial least squares: Com- of covariance-based and variance-based SEM. International Journal of Research in
ments on Rnkk & Evermann (2013). Organizational Research Methods, 17(2), Marketing, 26(4), 332344.
182209. Rigdon, E. E. (1998). Structural equation modeling. In G. A. Marcoulides (Ed.), Modern
Henseler, J., Hubona, G. S., & Ray, P. A. (2016). Using PLS path modeling in new methods for business research (pp. 251294). Mahwah: Erlbaum.
technology research: Updated guidelines. Industrial Management & Data Systems, Rigdon, E. E. (2012). Rethinking partial least squares path modeling: In praise of simple
116(1), 119. methods. Long Range Planning, 45(56), 341358.
4010 M. Sarstedt et al. / Journal of Business Research 69 (2016) 39984010

Rigdon, E. E. (2013). Partial least squares path modeling. In G. R. Hancock, & R. O. Mueller Sarstedt, M., Wilczynski, P., & Melewar, T. C. (2013). Measuring reputation in global mar-
(Eds.), Structural equation modeling: A second course, seconded. Vol. 1. (pp. 81116). kets A comparison of reputation measures' convergent and criterion validities.
Charlotte: Information Age. Journal of World Business, 48(3), 329339.
Rigdon, E. E. (2014). Rethinking partial least squares path modeling: Breaking chains and Sarstedt, M., Ringle, C. M., & Hair, J. F. (2014). PLS-SEM: Looking back and moving for-
forging ahead. Long Range Planning, 47(3), 161167. ward. Long Range Planning, 47(3), 132137.
Rigdon, E. E. (2016). Choosing PLS path modeling as analytical method in European Sarstedt, M., Ringle, C. M., Henseler, J., & Hair, J. F. (2014). On the emancipation of PLS-
management research: A realist perspective. European Management Journal SEM: A commentary on Rigdon (2012). Long Range Planning, 47(3), 154160.
(forthcoming). Schlittgen, R. (2015). R package SEGIRLS (version 0.5). http://www.wiso.uni-hamburg.
Rigdon, E. E., Preacher, K. J., Lee, N., Howell, R. D., Franke, G. R., & Borsboom, D. (2011). de/leadmin/bwl/statistikundoekonometrie/Schlittgen/SoftwareUndDaten/SEGIRLS_
Overcoming measurement dogma: A response to Rossiter. European Journal of Mar- 0.5.tar.gz
keting, 45(11/12), 15891600. Schneewei, H. (1991). Models with latent variables: LISREL versus PLS. Statistica
Rigdon, E. E., Becker, J. M., Rai, A., Ringle, C. M., Diamantopoulos, A., Karahanna, E., & Neerlandica, 45(2), 145157.
Dijkstra, T. K. (2014). Conating antecedents and formative indicators: A comment Schnemann, P. H., & Wang, M. -M. (1972). Some new results on factor indeterminacy.
on Aguirre-Urreta and Marakas, 25(4), 780784. Psychometrika, 37(1), 6191.
Ringle, C. M., Sarstedt, M., & Straub, D. W. (2012). A critical look at the use of PLS-SEM in Spearman, C. (1927). The abilities of man. London: MacMillan.
MIS quarterly. MIS Quarterly, 36(1), iiixiv. Temme, D., Diamantopoulos, A., & Pfegfeidel, V. (2014). Specifying formatively-measured
Rnkk, M. (2016). R package matrixpls: Matrix-based partial least squares estimation constructs in endogenous positions in structural equation models: Caveats and
(version 0.7.0). http://cran.r-project.org/web/packages/matrixpls/ guidelines for researchers. International Journal of Research in Marketing, 31(3),
Rnkk, M., & Evermann, J. (2013). A critical examination of common beliefs about partial 309316.
least squares path modeling. Organizational Research Methods, 16(3), 425448. Tenenhaus, A., & Tenenhaus, M. (2011). Regularized generalized canonical correlation
Rnkk, M., McIntosh, C. N., & Antonakis, J. (2015). On the adoption of partial least squares in analysis. Psychometrika, 76(2), 257284.
psychological research: Caveat emptor. Personality and Individual Differences, 87, 7684. Tenenhaus, M., Esposito Vinzi, V., Chatelin, Y. -M., & Lauro, C. (2005). PLS path modeling.
Rnkk, M., McIntosh, C. N., Antonakis, J., & Edwards, J. R. (2016). Partial least squares Computational Statistics & Data Analysis, 48(1), 159205.
path modeling: Time for some serious second thoughts. Journal of Operations Thurstone, L. L. (1947). Multiple factor analysis. Chicago, IL: The University of Chicago
Management (forthcoming). Press.
Rossiter, J. R. (2011). Marketing measurement revolution: The C-OAR-SE method and why Wickens, M. R. (1972). A note on the use of proxy variables. Econometrica, 40(4),
it must replace psychometrics. European Journal of Marketing, 45(11/12), 15611588. 759761.
Salzberger, T., Sarstedt, M., & Diamantopoulos, A. (2016). Measurement in the social sci- Wold, H. O. A. (1982). Soft modeling: The basic design and some extensions. In K. G.
ences: Where C-OAR-SE delivers and where it does not. European Journal of Marketing Jreskog, & H. O. A. Wold (Eds.), Systems under indirect observations: Part II
(forthcoming). (pp. 154). Amsterdam: North-Holland.