Journal of Marketing Theory and Practice

ISSN: 1069-6679 (Print) 1944-7175 (Online) Journal homepage: https://www.tandfonline.com/loi/mmtp20

PLS-SEM: Indeed a Silver Bullet

Joe F. Hair , Christian M. Ringle & Marko Sarstedt

To cite this article: Joe F. Hair , Christian M. Ringle & Marko Sarstedt (2011) PLS-SEM:
Indeed a Silver Bullet, Journal of Marketing Theory and Practice, 19:2, 139-152, DOI: 10.2753/
MTP1069-6679190202

To link to this article: https://doi.org/10.2753/MTP1069-6679190202

Published online: 08 Dec 2014.

Submit your article to this journal

Article views: 6506

View related articles

Citing articles: 3603 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=mmtp20
 PLS-SEM: Indeed a Silver Bullet
Joe F. Hair, Christian M. Ringle, and Marko Sarstedt

Structural equation modeling (SEM) has become a quasi-standard in marketing and management research
when it comes to analyzing the cause–effect relations between latent constructs. For most researchers, SEM
is equivalent to carrying out covariance-based SEM (CB-SEM). While marketing researchers have a basic
understanding of CB-SEM, most of them are only barely familiar with the other useful approach to SEM—
partial least squares SEM (PLS-SEM). The current paper reviews PLS-SEM and its algorithm, and provides an
overview of when it can be most appropriately applied, indicating its potential and limitations for future
research. The authors conclude that PLS-SEM path modeling, if appropriately applied, is indeed a “silver
bullet” for estimating causal models in many theoretical models and empirical data situations.

Structural equation modeling (SEM) first appeared in the focusing on explained variance. While far less popular
marketing literature in the early 1980s (e.g., Bagozzi 1994; than CB‑SEM, PLS‑SEM has been increasingly applied in
Bagozzi and Yi 1988; Fornell and Larcker 1981a, 1981b), but marketing and other business disciplines (e.g., Henseler,
in recent years, its application has become quite widespread. Ringle, and Sinkovics 2009), with more than 100 published
Indeed, fewer than ten articles using SEM were published studies featuring PLS‑SEM in the top 20 marketing journals
in marketing journals prior to 1990, while more than two- (for the marketing journal ranking, see Hult, Reimann, and
thirds of all articles applying SEM appeared between 1995 Schilke 2009).
and 2007 (Babin, Hair, and Boles 2008). Despite the growing number of studies on PLS‑SEM,
The desire to test complete theories and concepts is one some scholars view the method as less rigorous and therefore
of the major reasons authors conducting business research— less suitable for examining relationships between latent
particularly marketing—have embraced SEM (Henseler, variables. A likely reason for these less favorable perceptions
Ringle, and Sinkovics 2009; Steenkamp and Baumgartner is individual author’s previous experience and familiarity
2000). For many researchers, SEM is equivalent to carrying with the CB‑SEM approach and software that match the
out covariance-based SEM (CB‑SEM) analyses using software needs of certain disciplines (e.g., psychology). Another is
such as Amos, EQS, LISREL, Mplus, and others. But SEM also that few researchers viewed PLS‑SEM as a “silver bullet”
needs to be thought of as including another unique and very or panacea for dealing with empirical research challenges
useful approach—partial least squares SEM (PLS‑SEM). such as smaller sample sizes (Marcoulides and Saunders
PLS-SEM is a causal modeling approach aimed at 2006; Sosik, Kahai, and Piovoso 2009). These negative
maximizing the explained variance of the dependent la- perceptions of PLS‑SEM are unfortunate and shortsighted.
tent constructs. This is contrary to CB‑SEM’s objective of When properly applied, the method has many benefits not
reproducing the theoretical covariance matrix, without offered by CB‑SEM.
CB‑SEM develops a theoretical covariance matrix based
on a specified set of structural equations. The technique
Joe F. Hair (Ph.D., University of Florida), Professor of Marketing, focuses on estimating a set of model parameters in such a
Coles College of Business, Kennesaw State University, Kennesaw, way that the difference between the theoretical covariance
GA, [email protected].
matrix and the estimated covariance matrix is minimized
Christian M. Ringle (Ph.D., University of Hamburg), Professor
of Management, Institute for Human Resource Management and (e.g., Rigdon 1998). The CB‑SEM model estimation requires
Organizations, Hamburg University of Technology (TUHH), a set of assumptions to be fulfilled, including the multivari-
Hamburg, Germany, [email protected], and Associate of ate normality of data, minimum sample size, and so forth
the Faculty, School of Marketing, University of Technology Sydney (e.g., Diamantopoulos and Siguaw 2000). But if CB‑SEM
(UTS), Australia, [email protected].
assumptions cannot be met, or the research objective is
Marko Sarstedt (Ph.D., Ludwig-Maximilians-University), Professor
prediction rather than confirmation of structural relation-
of Quantitative Methods in Marketing and Management, Institute
for Market-Based Management, Ludwig-Maximilians-University ships, then variance-based PLS‑SEM is the preferred method.
Munich, Germany, [email protected]. In comparison with CB‑SEM results, which can be highly

Journal of Marketing Theory and Practice, vol. 19, no. 2 (spring 2011), pp. 139–151.
© 2011 M.E. Sharpe, Inc. All rights reserved.
ISSN 1069-6679/2011 $9.50 + 0.00.
DOI 10.2753/MTP1069-6679190202
140 Journal of Marketing Theory and Practice

imprecise when the assumptions are violated, PLS‑SEM The philosophical distinction between CB‑SEM and
often provides more robust estimations of the structural PLS‑SEM is straightforward. If the research objective is
model (e.g., Lohmöller 1989; Reinartz, Haenlein, and theory testing and confirmation, then the appropriate
Henseler 2009; Ringle et al. 2009; Wold 1982). method is CB‑SEM. In contrast, if the research objective is
This paper provides insights into PLS‑SEM’s current state prediction and theory development, then the appropriate
of the art by merging contributions from the marketing, method is PLS‑SEM. Conceptually and practically, PLS‑SEM
statistics, and management disciplines. The overall objective is similar to using multiple regression analysis. The primary
is to encourage the appropriate use of PLS‑SEM in empiri- objective is to maximize explained variance in the depen-
cal research by describing the method’s strengths while at dent constructs but additionally to evaluate the data quality
the same time highlighting potentially problematic issues. on the basis of measurement model characteristics. Given
Thus, we call for a more balanced and informed perspec- PLS‑SEM’s ability to work efficiently with a much wider
tive of PLS‑SEM’s application in marketing and related range of sample sizes and increased model complexity, and
disciplines. its less restrictive assumptions about the data, it can ad-
dress a broader range of problems than CB‑SEM. Moreover,
STRUCTURAL EQUATION because the constructs’ measurement properties are less
MODELING OVERVIEW restrictive with PLS‑SEM, constructs with fewer items (e.g.,
one or two) can be used than those that CB‑SEM requires.
Path analysis models were first developed by Sewall Wright Overall, when measurement or model properties restrict the
(1921), a biostatistician, in the early 1920s. Structural mod- use of CB-SEM or when the emphasis is more on exploration
els as applied in the social sciences only began appearing than confirmation, PLS‑SEM is an attractive alternative to
in the 1970s (Bollen 1989; Jöreskog 1973) with their in- CB‑SEM and often more appropriate.
creasing application paralleling the availability of software While PLS‑SEM can be applied to a wider range of situ-
(Arbuckle 2010; Bentler 1995; Jöreskog and Sörbom 1996), ations, researchers must always be aware of the differences
all of which executed CB‑SEM. While Herman Wold—who in interpretation of the results, particularly as they relate to
was also the academic advisor of Karl Jöreskog, one of the the constructs’ measurement properties. For example, can
LISREL CB‑SEM software package developers—originated PLS‑SEM be appropriately applied when a measurement
variance-based SEM in the 1970s (Wold 1973, 1975), software theory fails to meet the confirmatory factor analysis’s crite-
packages executing PLS‑SEM were developed much later ria, including convergent validity and discriminant validity
(e.g., SmartPLS; Ringle, Wende, and Will 2005). tests? PLS‑SEM estimates loadings of the indicator variables
Jöreskog and Wold (1982) viewed CB‑SEM and PLS‑SEM for the exogenous constructs based on their prediction of
as complementary rather than competitive statistical meth- the endogenous constructs, not their shared variance among
ods. More specifically, Wold (1982) recognized CB‑SEM’s indicator variables on the same construct. Thus, the loadings
potential for the social sciences but was concerned about in PLS‑SEM are in a way their contribution to the path coef-
the informational and distributional requirements that ficients. Hair et al. (2010) give a numerical example with a
he regarded as unrealistic for empirical research. He also prespecified covariance matrix to highlight the differences
believed that estimation and description were emphasized between the two methods. While CB‑SEM provides poor
too much and prediction too little (Dijkstra 2010). It is im- results for the measurement model but a significant struc-
portant to note that next to PLS‑SEM, another PLS culture tural model relationship, PLS‑SEM offers acceptable results
has arisen from Wold’s original works—PLS regression. This for the measurement models whereas the structural model
approach generalizes and combines features from principal relationship is not significant. On the contrary, Tenenhaus
component analysis and multiple regression but generally (2008) offers an example in which PLS‑SEM and CB‑SEM
does not allow for the evaluation of complex cause–effect results are in close correspondence. Both examples show
relationships between latent constructs (for a notable excep- that the differences in results are primarily a matter of
tion, see Arteaga, Gallarza, and Gil 2010). Natural science measurement model quality. Using “good” measures and
disciplines, such as chemometrics, generally use PLS regres- data, both approaches practically yield the same results. As a
sion (e.g., Wold, Sjöström, and Eriksson 2001), but PLS‑SEM consequence, a case whether the one or the other methods is
is the approach that has become established in marketing better should not be made, especially because the majority
and business research (e.g., Fornell and Bookstein 1982; of prior studies comparing CB‑SEM and PLS‑SEM did not
Henseler, Ringle, and Sinkovics 2009). explicitly control for CB‑SEM specification biases, which
 Spring 2011 141

is an important issue in order to conduct a meaningful construct, and changes in the indicators determine changes
comparison of the two approaches. in the value of the latent construct (Diamantopoulos and
To correctly apply CB‑SEM and PLS‑SEM, researchers Winklhofer 2001; Diamantopoulos, Riefler, and Roth 2008).
must understand the purposes for which each approach was Formative indicators are represented by single-headed ar-
developed and apply them accordingly. Structural equation rows pointing toward the latent construct inward from the
models with good measurement properties generally achieve indicator variables; the associated coefficients for these
comparable results with either approach, especially when formative relationships are called outer weights in PLS‑SEM.
the CB‑SEM’s model specifications have been correctly set Researchers using PLS‑SEM often refer to reflective mea-
up (Reinartz, Haenlein, and Henseler 2009). Moreover, both surement models (i.e., scales) as Mode A, and formative
approaches should still consider issues such as the appro- measurement models (i.e., indices) are labeled Mode B (e.g.,
priate use and interpretation of formative versus reflective Rigdon, Ringle, and Sarstedt 2010).
measures (Diamantopoulos, Riefler, and Roth 2008). These Figure 1 shows an example of a simple path model. The
situations are often the ones in which the measurement model has a single endogenous (dependent) latent construct
properties are questionable and the results may diverge, (Y3) and two exogenous (independent) latent constructs (Y1
thus requiring the researcher to make a reasoned judgment and Y2), which are indicated as ovals. The differentiation
as to which approach is most appropriate. between endogenous and exogenous constructs is clear-cut in
this PLS‑SEM example. However, if a fourth latent construct
The PLS‑SEM Algorithm (Y4) were included in the SEM model with a relationship
from Y3 to Y4, both latent constructs would be described as
A structural equation model with latent constructs has two endogenous. The latent constructs Y1 and Y2 are measured
components. The first component is the structural model— with two formative indicator variables each, indicated as
typically referred to as the inner model in the PLS‑SEM boxes (X1 to X4, with arrows pointing toward the constructs).
context—which shows the relationships (paths) between the On the contrary, Y3 is measured with three reflective indica-
latent constructs. PLS‑SEM only permits recursive relation- tor variables (X5 to X7, with arrows pointing away from the
ships in the structural model (i.e., no causal loops). Therefore, construct). Most theoretical constructs (especially formative
the structural paths between the latent constructs can only ones) will be measured by six or seven indicator variables,
head in a single direction. In the structural model, we distin- but our example includes fewer indicators to facilitate the
guish between exogenous and endogenous constructs. The presentation.
term exogenous is used to describe latent constructs that do The basic PLS‑SEM algorithm (Lohmöller 1989) follows a
not have any structural path relationships pointing at them. two-stage approach. In the first stage, the latent constructs’
Thus, the term endogenous describes latent target constructs scores are estimated via a four-step process as shown in
in the structural model that are explained by other constructs Table 1. The second stage calculates the final estimates of the
via structural model relationships. outer weights and loadings as well as the structural model’s
The second component of the structural equation model path coefficients. The path modeling procedure is called
comprises the measurement models, also referred to as outer partial because the iterative PLS‑SEM algorithm estimates the
models in the PLS‑SEM context. The measurement models coefficients for the partial ordinary least squares regression
include the unidirectional predictive relationships between models in the measurement models and the structural model.
each latent construct and its associated observed indicators More specifically, when a formative measurement model is
(Figure 1). Multiple relations are not permitted, therefore assumed, a multiple regression model is estimated with the
indicator variables are associated with only a single latent latent construct as the dependent variable and the assigned
construct. PLS‑SEM can handle both formative and reflec- indicators as independent variables (computation of outer
tive measurement models. Reflective indicators are seen as weights). In contrast, when a reflective measurement model
functions of the latent construct, and changes in the latent is assumed, the regression model includes single regressions
construct are reflected in changes in the indicator (manifest) with each indicator individually being the dependent variable,
variables. Reflective indicators are represented as single- whereas the latent construct is always the independent vari-
headed arrows pointing from the latent construct outward able (computation of outer loadings). When the structural
to the indicator variables; the associated coefficients for model relationships are calculated, each endogenous latent
these relationships are called outer loadings in PLS‑SEM. In construct represents the dependent variable with its latent
contrast, formative indicators are assumed to cause a latent construct antecedents as independent variables in a partial
142 Journal of Marketing Theory and Practice

Figure 1
Path Model Example

Table 1 as a proxy for the latent construct. PLS‑SEM software pro-

Stages and Steps in Calculating the grams, such as SmartPLS (Ringle, Wende, and Will 2005),
Basic PLS-SEM Algorithm use a uniform value of 1 as an initial value for each of the
Stage One: Iterative estimation of latent construct scores outer weights (W1 to W7).
Step 1: Outer approximation of latent construct scores (the In Step 2, the PLS‑SEM algorithm computes proxies for
scores of Y1, Y2, and Y3 are computed based on the manifest the structural model relationships (P1 and P2). Several differ-
variables’ scores and the outer coefficients from Step 4) ent weighting schemes are available to estimate these prox-
Step 2: Estimation of proxies for structural model relationships
between latent constructs (P1 and P2) ies. We recommend applying the path weighting scheme
Step 3: Inner approximation of latent construct scores (based on that uses combinations of regression analyses and bivariate
scores for Y1, Y2, and Y3 from Step 1 and proxies for structural correlations based on latent construct scores as proxies
model relationships, P1 and P2, from Step 2) for structural model relationships. This method develops
Step 4: Estimation of proxies for coefficients in the measurement
models (the relationships between indicator variables and
latent construct scores that maximize the final R² value es-
latent constructs with scores from Step 3; W1 to W7) timations of the endogenous latent constructs (Lohmöller
Stage Two: Final estimates of coefficients (outer weights and 1989). In Step 3, the inner proxies of the latent construct
loadings, structural model relationships) are determined using the scores (Y1, Y2, and Y3) are calculated as linear combinations
ordinary least squares method for each partial regression in the
of their respective adjacent latent construct outer proxies
PLS‑SEM model.
(from Step 1) using the previously determined (Step 2)
inner weights. Finally, in Step 4, the outer weights (W1 to
regression model. All partial regression models are estimated W7) are calculated in two different ways, depending on the
by the iterative procedures of the PLS‑SEM algorithm. type of measurement model represented by each construct.
Stage One has four steps. In Step 1 of Stage One, the la- If a construct is measured reflectively, then the correlations
tent construct scores’ outer proxies are computed as linear between the inner proxy of each latent construct and its
combinations of the values of all (standardized) indicators indicator variables are applied (outer loadings). If a construct
associated with a particular latent construct. For example, is measured formatively, then regression weights (i.e., outer
values of X1 and X2 are used to compute the proxy score for weights) are applied that are the result of the ordinary least
the latent construct Y1. Later iterations use the estimated squares regression of each latent construct’s inner proxy on
coefficients of the paths (e.g., W1 and W2 for Y1, whereby its indicator variables.
W represents the outer weight or loading coefficient; for The four steps in Stage One are repeated until the sum
the simplicity of representation in this example, we do of the outer weights’ changes between two iterations is suf-
not use different symbols for the Mode A and Mode B ficiently low. That is, the sum of the changes drops below
measurement models) between the latent constructs and a predetermined limit. The recommended predetermined
the indicator variables from Step 4 of Stage One. For the limit is a threshold value of 10–5 to ensure the PLS‑SEM
initial iteration, any combination of indicators can serve algorithm’s convergence while simultaneously minimizing
 Spring 2011 143

computational requirements. If the algorithm converges in (e.g., Reinartz, Haenlein, and Henseler 2009; Ringle et al.
Step 4 of Stage One, then the final outer weights are used 2009). This important feature holds especially for research
to compute the final latent construct scores in Stage Two. situations in which maximum likelihood–based CB‑SEM
The final latent construct scores are used to run the ordinary typically exhibits inflated standard errors (Sosik, Kahai,
least squares regressions for each construct to determine the and Piovoso 2009) and the method’s premises are violated
structural model relationships’ estimates (path coefficients). (e.g., small sample size, nonnormal data, high model com-
Most PLS‑SEM software packages (e.g., SmartPLS) provide a plexity). This increased efficiency in parameter estimation
graphical user interface to create the model and implement is manifested in PLS‑SEM’s greater statistical power than
the basic PLS algorithm for model estimation. that of CB‑SEM.
The choice of either PLS‑SEM or CB‑SEM is determined
Choosing PLS‑SEM versus CB‑SEM by the relevant research’s objective. In causal modeling situ-
ations where prior theory is strong and further testing and
PLS‑SEM is a promising method that offers vast potential confirmation are the goals, CB‑SEM is the more appropriate
for SEM researchers especially in the marketing and man- statistical methodology. Global goodness-of-fit criteria that
agement information systems disciplines. PLS‑SEM is, as also emphasize theory testing rather than theory building
the name implies, a more “regression-based” approach can be used to assess CB‑SEM results (Anderson and Gerbing
that minimizes the residual variances of the endogenous 1988). However, like all statistical methods, CB‑SEM has
constructs. Compared to CB‑SEM, it is more robust with shortcomings. For example, CB‑SEM approaches largely ig-
fewer identification issues, works with much smaller as well nore the prediction objective. If the structural relationships
as much larger samples, and readily incorporates formative between the latent constructs are the primary concern in
as well as reflective constructs. theory testing, researchers are generally less concerned with
These advantages are constrained by some disadvantages. predictive accuracy. In situations where theory is less devel-
One disadvantage is PLS-SEM’s focus is on maximizing par- oped, however, researchers need an alternative approach to
tial model structures. Specifically, the PLS-SEM algorithm examine structural models if the primary objective is not
first optimizes measurement model parameters and then, in theory confirmation. Thus, because of its prediction orien-
a second step, estimates the path coefficients in the struc- tation, PLS‑SEM is the preferred method when the research
tural model. Thus, researchers applying PLS-SEM first have objective is theory development and prediction.
to examine the measurement models’ characteristics and Moreover, PLS‑SEM avoids the indeterminacy problem
deal with those that are unacceptable. If this approach is (difficulty with estimating stable factor scores) and develops
followed, PLS‑SEM findings can be interpreted after assess- more precise estimates of factor scores (Fornell 1982) as the
ment of the model’s measurement characteristics—similar algorithm calculates latent variable scores as exact linear
to CB‑SEM methods. Another issue that restricts the use of combinations of the observed indicator variables. As a con-
PLS‑SEM for theory testing and confirmation is that there is sequence, PLS‑SEM is particularly useful when subsequent
no adequate global measure of goodness of model fit. In addi- analyses employ these scores. For example, PLS‑SEM impact-
tion, it should be noted that, in general, PLS‑SEM parameter performance analyses (e.g., Höck, Ringle, and Sarstedt 2010;
estimates are not optimal regarding bias and consistency—a Völckner et al. 2010), which enable researchers to easily
property frequently referred to as PLS‑SEM bias. Although prioritize activities for potential improvements and render
the CB‑SEM versus PLS‑SEM discussion strongly emphasizes the required managerial decisions, make use of latent vari-
this characteristic, simulation studies show that the differ- able scores. Likewise, researchers have used differences in
ences between CB‑SEM and PLS‑SEM estimates are at very latent construct scores to evaluate experimental setups (e.g.,
low levels (e.g., Reinartz, Haenlein, and Henseler 2009). Schwaiger, Sarstedt, and Taylor 2010).
Thus, the extensively discussed PLS‑SEM bias is often of CB‑SEM and PLS‑SEM methodologies differ from
minor relevance for practical applications because estimates a statistical point of view, but PLS‑SEM estimates can
will be asymptotically correct under consistency at large con- be good proxies of CB‑SEM results. For example, when
ditions (i.e., both a large sample size and large numbers of CB‑SEM assumptions are violated with regard to normal-
indicators per latent variable) (Jöreskog and Wold 1982). ity of distributions, minimum sample size, and maximum
Furthermore, although the estimates that result from model complexity, or when related methodological matters
PLS‑SEM are, on average, biased, they also exhibit a lower emerge, such as Heywood cases and inflated parameter esti-
degree of variability than those produced by CB‑SEM mates’ standard errors (Rindskopf 1984), PLS‑SEM is a good
144 Journal of Marketing Theory and Practice

Table 2
Rules of Thumb for Selecting CB-SEM or PLS-SEM

Research Goals
• If the goal is predicting key target constructs or identifying key “driver” constructs, select PLS-SEM.
• If the goal is theory testing, theory confirmation, or comparison of alternative theories, select CB-SEM.
• If the research is exploratory or an extension of an existing structural theory, select PLS-SEM.
Measurement Model Specification
• If formative constructs are part of the structural model, select PLS-SEM.
Note that formative measures can also be used with CB-SEM but to do so requires accounting for relatively complex and limiting
specification rules.
• If error terms require additional specification, such as covariation, select CB-SEM.
Structural Model
• If the structural model is complex (many constructs and many indicators), select PLS-SEM.
• If the model is nonrecursive, select CB-SEM.
Data Characteristics and Algorithm
• If your data meet the CB-SEM assumptions exactly, for example, with respect to the minimum sample size and the distributional
assumptions, select CB-SEM; otherwise, PLS-SEM is a good approximation of CB-SEM results.
• Sample size considerations:
– If the sample size is relatively low, select PLS-SEM. With large data sets, CB-SEM and PLS-SEM results are similar, provided that a large
number of indicator variables are used to measure the latent constructs (consistency at large).
– PLS-SEM minimum sample size should be equal to the larger of the following: (1) ten times the largest number of formative indicators
used to measure one construct or (2) ten times the largest number of structural paths directed at a particular latent construct in the
structural model.
• If the data are to some extent nonnormal, use PLS-SEM; otherwise, under normal data conditions, CB-SEM and PLS-SEM results are
highly similar, with CB-SEM providing slightly more precise model estimates.
• If CB-SEM requirements cannot be met (e.g., model specification, identification, nonconvergence, data distributional assumptions), use
PLS-SEM as a good approximation of CB-SEM results.
• CB-SEM and PLS-SEM results should be similar. If not, check the model specification to ensure that CB-SEM was appropriately applied. If
not, PLS-SEM results are a good approximation of CB-SEM results.
Model Evaluation
• If you need to use latent variable scores in subsequent analyses, PLS-SEM is the best approach.
• If your research requires a global goodness-of-fit criterion, then CB-SEM is the preferred approach.
• If you need to test for measurement model invariance, use CB-SEM.

methodological alternative for theory testing. Another at- PLS‑SEM assessment typically follows a two-step pro-
tractive feature for theory testing is PLS-SEM’s higher levels cess that involves separate assessments of the measure-
of statistical power compared to CB-SEM. Thus, researchers ment models and the structural model. The first step is to
should consider both CB‑SEM and PLS‑SEM when selecting examine the measures’ reliability and validity according
the appropriate statistical method. Table 2 displays rules of to certain criteria associated with formative and reflective
thumb that can be applied when deciding whether to use measurement model specification. This first step is based
CB‑SEM or PLS‑SEM. The rules of thumb are listed accord- on the logic that if you are not confident that the measures
ing to five types of decision considerations. represent the constructs of interest, there is little reason
to use them to examine the structural relationships. If the
What to Look for after measures are shown to be adequate, however, the second
running PLS‑SEM step involves an assessment of the structural model esti-
mates. As in any statistical analysis, the evaluation of the
Just like any multivariate analysis technique (Hair et al. measurement and structural models should include examin-
2010), applying PLS‑SEM requires a thorough assessment ing the parameter estimates’ stability by means of a holdout
of the final results (model evaluation). In many respects, sample. Generally, we recommend using 30 percent of the
model evaluation in PLS‑SEM is still heavily influenced by original sample as the holdout sample. However, researchers
CB‑SEM reporting—especially in terms of the evaluation of must consider the minimum sample size requirements as
measurement models—because it follows the notion of an described in Table 2.
underlying covariance-based latent construct–generating In the following, we provide an overview of aspects
mechanism. related to the evaluation of the measurement models and
 Spring 2011 145

Table 3
Rules of Thumb for Model Evaluation

Reflective Measurement Models

• Internal consistency reliability: Composite reliability should be higher than 0.70 (in exploratory research, 0.60 to 0.70 is considered ac‑
ceptable).
• Indicator reliability: Indicator loadings should be higher than 0.70.
• Convergent validity: The average variance extracted (AVE) should be higher than 0.50.
• Discriminant validity:
– The AVE of each latent construct should higher than the construct’s highest squared correlation with any other latent construct
(Fornell–Larcker criterion).
– An indicator’s loadings should be higher than all of its cross loadings.
Formative Measurement Models
• Examine each indicator’s weight (relative importance) and loading (absolute importance) and use bootstrapping to assess their signifi‑
cance. The minimum number of bootstrap samples is 5,000, and the number of cases should be equal to the number of observations in
the original sample. Critical t-values for a two-tailed test are 1.65 (significance level = 10 percent), 1.96 (significance level = 5 percent),
and 2.58 (significance level = 1 percent).
– When all the indicator weights are significant, there is empirical support to keep all the indicators.
– If both the weight and loading are nonsignificant, there is no empirical support to retain the indicator and its theoretical relevance
should be questioned.
• Multicollinearity: Each indicator’s variance inflation factor (VIF) value should be less than 5.
• Indicator weights should be examined to determine if they are affected by (observed or unobserved) heterogeneity, which results in
significantly different group-specific coefficients. If theory supports the existence of alternative groups of data, carry out PLS-SEM multi‑
group or moderator analyses. If no theory or information is available about the underlying groups of data, an assessment of unobserved
heterogeneity’s existence must be conducted by means of the finite mixture PLS (FIMIX-PLS) method.
• When many indicators are used to measure a formative construct, with some being nonsignificant, establish two or more distinct con‑
structs, provided there is theoretical support for this step.
Structural Model
• R ² values of 0.75, 0.50, or 0.25 for endogenous latent variables in the structural model can be described as substantial, moderate, or
weak, respectively.
• Use bootstrapping to assess the path coefficients’ significance. The minimum number of bootstrap samples is 5,000, and the number of
cases should be equal to the number of observations in the original sample. Critical t-values for a two-tailed test are 1.65 (significance
level = 10 percent), 1.96 (significance level = 5 percent), and 2.58 (significance level = 1 percent).
• Predictive relevance: Use blindfolding to obtain cross-validated redundancy measures for each construct. Make sure the number of valid
observations is not a multiple integer number of the omission distance d. Choose values of d between 5 and 10.
Resulting Q ² values of larger than zero indicate that the exogenous constructs have predictive relevance for the endogenous construct
under consideration.
• Heterogeneity: If theory supports the existence of alternative groups of data, carry out PLS-SEM multigroup or moderator analyses. If no
theory or information about the underlying groups of data is available, an assessment of unobserved heterogeneity’s existence must be
conducted by means of the FIMIX-PLS method, which is available in the SmartPLS software package.

the structural model. In addition, we address the frequently tency. Unlike Cronbach’s alpha, composite reliability does
asked question: Why does PLS‑SEM rely on bootstrapping? not assume that all indicators are equally reliable, making
We summarize the rules of thumb for model evaluation in it more suitable for PLS‑SEM, which prioritizes indicators
Table 3. Researchers and practitioners can conduct all the according to their reliability during model estimation.
analyses proposed in this section by using the SmartPLS Composite reliability values of 0.60 to 0.70 in exploratory
(Ringle, Wende, and Will 2005) software package. research and values from 0.70 to 0.90 in more advanced
stages of research are regarded as satisfactory (Nunnally and
Evaluation of Measurement Models Bernstein 1994), whereas values below 0.60 indicate a lack
of reliability. Likewise, each indicator’s reliability needs to
We must distinguish between reflective and formative mea- be taken into account, whereby each indicator’s absolute
surement models to evaluate them (Henseler, Ringle, and standardized loading should be higher than 0.70. Gener-
Sinkovics 2009). Reflective measurement models should be ally, indicators with loadings between 0.40 and 0.70 should
assessed with regard to their reliability and validity. Con- only be considered for removal from the scale if deleting
struct reliability assessment routinely focuses on composite this indicator leads to an increase in composite reliability
reliability as an estimate of a construct’s internal consis- above the suggested threshold value. Another consideration
146 Journal of Marketing Theory and Practice

in the decision to delete indicators is the extent to which weight and loading are nonsignificant, there is no em-
their removal affects validity. Weaker indicators are some- pirical support for the indicator’s relevance in providing
times retained on the basis of their contribution to content content to the formative index (Cenfetelli and Bassellier
validity. Indicators that exhibit very low loadings of 0.40 2009). Researchers must then decide whether to retain or
and lower should, however, always be eliminated from delete a nonsignificant indicator, which is always crucial
reflective scales. for the theoretical underpinnings and interpretation of
Reflective measurement models’ validity assessment empirical results. If the theory-driven conceptualization
focuses on convergent validity and discriminant validity. of the measure strongly supports the indicator’s inclusion
For convergent validity, researchers need to examine the (e.g., by means of face, expert, and content validity), it
average variance extracted (AVE). An AVE value of 0.50 and should be kept in the formative measurement model and
higher indicates a sufficient degree of convergent validity, the researcher should focus on explaining the empirical
meaning that the latent variable explains more than half of outcome. However, the researcher can also interpret the
its indicators’ variance. For the assessment of discriminant empirical finding as countering the conceptual foundations
validity, two measures have been put forward—the Fornell– that support the indicator’s inclusion and thus decide to
Larcker criterion and cross loadings. The Fornell–Larcker exclude the nonsignificant indicator from further analysis.
criterion (Fornell and Larcker 1981a) postulates that a latent Even though this step usually has almost no effect on the
construct shares more variance with its assigned indicators parameter estimates when reestimating the model, one
than with another latent variable in the structural model. has to bear in mind that eliminating formative indicators
In statistical terms, the AVE of each latent construct should can have adverse consequences for the derived measure’s
be greater than the latent construct’s highest squared cor- content validity (Diamantopoulos and Siguaw 2006).
relation with any other latent construct. The second crite- While nonsignificant indicators may simply imply a lack
rion of discriminant validity is usually a bit more liberal: of theoretical relevance, another potential reason for their
an indicator’s loading with its associated latent construct lack of significance may be the existence of heterogeneous
should be higher than its loadings with all the remaining data structures. If parameter estimates are affected by het-
constructs (i.e., the cross loadings). erogeneity, a formative indicator may not be significant
Traditional statistical evaluation criteria for reflective when solely evaluated on the aggregate data level. But there
scales cannot be directly transferred to formative indices. may be one or more subpopulation(s) in which an indica-
In a formative measurement model, indicators represent the tor is significantly related to the construct while this is not
latent construct’s (potentially) independent causes and thus the case in another subpopulation. Therefore, researchers
do not necessarily correlate highly. Furthermore, formative should examine whether heterogeneity affects the coef-
indicators are assumed to be error free (Edwards and Bagozzi ficients in formative measurement models significantly
2000). Consequently, the concepts of internal consistency before eliminating nonsignificant indictors. This can be
reliability and convergent validity are not meaningful when done by either partitioning the data on the basis of a priori
formative indicators are involved. Instead, theoretical ra- information (e.g., demographic variables; Rigdon, Ringle,
tionale and expert opinion play a more important role in and Sarstedt 2010) and estimating distinct models or by
the evaluation of formative indexes. Nevertheless, PLS‑SEM using the finite mixture PLS (FIMIX-PLS) method (Ringle,
also offers some statistical criteria for assessing formative Wende, an Will 2010), which is part of the SmartPLS soft-
measurement models’ quality. ware package.
At the indicator level, the question arises whether each Second, an indicator’s information can become redun-
indicator indeed contributes to forming the index in ac- dant due to high levels of multicollinearity in the forma-
cordance with its intended contents. Two specific issues tive measurement model which can cause indicators to
require a critical analysis and careful decisions with respect be nonsignificant. To determine redundancy, researchers
to the index’s contents. First, an indicator that was theo- should examine the degree of multicollinearity in the
rized to contribute to a formative index can be irrelevant. formative indicators (Cassel, Hackl, and Westlund 1999;
The bootstrapping procedure allows the significance of Diamantopoulos and Winklhofer 2001; Grewal, Cote, and
formative indicators’ coefficients to be tested. In addition Baumgartner 2004), for instance, by calculating the vari-
to considering the significance of the indicator’s weight, ance inflation factor (VIF). In the context of PLS‑SEM, a VIF
researchers should also evaluate an indicator’s absolute value of 5—which implies that 80 percent of an indicator’s
importance for its construct (i.e., the loading). When both variance is accounted for by the remaining formative in-
 Spring 2011 147

dicators related to the same construct—indicates potential ing procedure, a sample reuse technique that omits every
multicollinearity problems. This necessitates the recon- dth data point part and uses the resulting estimates to
sideration of the formative measurement model setup. In predict the omitted part. It is important to note that the
light of the discussion above, we recommend eliminating omission distance d must be chosen so that the number of
an indicator to relax multicollinearity problems (1) if the valid observations divided by d is not an integer. Experience
level of multicollinearity is very high (as indicated by shows that d values between 5 and 10 are advantageous.
a VIF value of 5 or higher), (2) if this indicator’s forma- The blindfolding procedure is only applied to endogenous
tive measurement model coefficient (outer weight) is not latent constructs that have a reflective measurement model
significantly different from zero, and (3) the remaining specification. Q ² comes in two forms—the cross-validated
indicators sufficiently capture the domain of the construct redundancy and communality. We recommend using
under consideration. the cross-validated redundancy, which—unlike the cross-
It is important to note that the higher number of indica- validated communality—uses the PLS‑SEM estimates of both
tors used to measure a formative latent construct, the more the structural model and the measurement models for data
likely it is that one or more indicators will have low or even prediction and, thereby, perfectly fits the PLS‑SEM approach.
nonsignificant weights. If this is the case, indicators should If an endogenous construct’s cross-validated redundancy
be grouped into two or more distinct constructs, provided measure value (i.e., Q ²) for a certain endogenous latent
there is theoretical support for this step (Cenfetelli and variable is larger than zero, its explanatory latent constructs
Bassellier 2009). exhibit predictive relevance.
Another important aspect of structural model evaluation
Evaluation of the Structural Model is heterogeneity of observations, which can be a threat to
the PLS‑SEM results’ validity, that is, different population
The primary evaluation criteria for the structural model are parameters are likely to occur for different subpopulations
the R² measures and the level and significance of the path such as segments of consumers, firms, or countries. Con-
coefficients. Because the goal of the prediction-oriented sequently, researchers should consider potential sources of
PLS‑SEM approach is to explain the endogenous latent vari- heterogeneity and test these by means of multigroup com-
ables’ variance, the key target constructs’ level of R² should parison or moderator analyses (Rigdon, Ringle, and Sarstedt
be high. The judgment of what R² level is high depends, 2010). However, the true sources of heterogeneity can never
however, on the specific research discipline. Whereas R² be fully known a priori. As a result, situations arise in which
results of 0.20 are considered high in disciplines such as differences are related to unobserved heterogeneity and can-
consumer behavior, R² values of 0.75 would be perceived as not be attributed to any predetermined variable(s). Several
high in success driver studies. In marketing research stud- tools for dealing with unobserved heterogeneity within
ies, R² values of 0.75, 0.50, or 0.25 for endogenous latent PLS‑SEM have recently been developed, with many new
variables in the structural model can, as a rule of thumb, be approaches appearing only in specialized literature devoted
described as substantial, moderate, or weak, respectively. to PLS‑SEM methods. Among these, FIMIX‑PLS (Hahn et al.
The individual path coefficients of the PLS structural 2002; Sarstedt and Ringle 2010) is to date regarded as the
model can be interpreted as standardized beta coefficients of primary approach and has become mandatory for evaluat-
ordinary least squares regressions. Just as with the indicators’ ing PLS path modeling results (Sarstedt 2008). Using this
weights and loadings, each path coefficient’s significance technique, researchers can either confirm that their results
can be assessed by means of a bootstrapping procedure. are not distorted by unobserved heterogeneity or they can
Paths that are nonsignificant or show signs contrary to the identify thus far neglected variables that describe the uncov-
hypothesized direction do not support a prior hypothesis, ered data segments. Sarstedt et al. (2011) provide a further
whereas significant paths showing the hypothesized direc- discussion of the FIMIX‑PLS approach, and Ringle, Sarstedt,
tion empirically support the proposed causal relationship. and Mooi (2010) illustrate its application via the American
Another assessment of the structural model involves the customer satisfaction index model.
model’s capability to predict. The predominant measure of
predictive relevance is the Stone–Geisser’s Q ² (Geisser 1974; Why PLS‑SEM Uses Bootstrapping
Stone 1974), which postulates that the model must be able
to adequately predict each endogenous latent construct’s PLS‑SEM does not presume that the data are normally
indicators. The Q ² value is obtained by using a blindfold- distributed. Consequently, PLS applies nonparametric
148 Journal of Marketing Theory and Practice

bootstrapping (Davison and Hinkley 1997; Efron and specifications and data) and its comparatively high statisti-
Tibshirani 1993), which involves repeated random sam- cal power make the PLS method particularly adequate for
pling with replacement from the original sample to create SEM applications that aim at prediction or theory building
a bootstrap sample, to obtain standard errors for hypothesis such as in studies that focus on identifying critical success
testing. The process assumes that the sample distribution drivers (e.g., Höck and Ringle 2010; Sarstedt and Schloderer
is a reasonable representation of the intended population 2010; Sattler et al. 2010). Finally, PLS‑SEM can also be used
distribution. The bootstrap sample enables the estimated for confirmatory theory testing.
coefficients in PLS‑SEM to be tested for their significance Recent PLS‑SEM research reveals that the method is
(Henseler, Ringle, and Sinkovics 2009). developing rapidly and has demonstrated numerous tech-
The procedure creates a large, prespecified number of nical advancements. These include different approaches to
bootstrap samples (e.g., 5,000) by randomly drawing cases response-based clustering such as FIMIX‑PLS (Hahn et al.
with replacement from the original sample. Each bootstrap 2002; Ringle, Sarstedt, and Mooi 2010; Sarstedt and Ringle
sample should have the same number of cases as the original 2010), PLS genetic algorithm segmentation (Ringle and
sample. The PLS algorithm estimates the SEM results from Schlittgen 2007), and response-based unit segmentation
each bootstrap sample (e.g., 5,000 PLS‑SEM estimations). in PLS (REBUS‑PLS) (Esposito Vinzi et al. 2008). These
The repeated bootstrap parameter estimates are then used to methods have been constantly advanced, for example, by
create an empirical sampling distribution for each model pa- addressing model selection issues in FIMIX‑PLS (Sarstedt et
rameter, and the empirical sampling distribution’s standard al. 2011) and comparing several of the approaches by means
deviation is used as proxy for the empirical standard error of computational experiments (Esposito Vinzi et al. 2007;
for the parameter. The obtained path model coefficients Ringle, Sarstedt, and Schlittgen 2010). The confirmatory
form a bootstrap distribution, which can be viewed as an tetrad analysis in PLS path modeling (CTA‑PLS), as proposed
approximation of the sampling distribution. The PLS‑SEM by Gudergan et al. (2008), allows the empirical testing
results of all the bootstrap samples provide the standard of the measurement models’ mode ex post. More specifi-
error for each path model coefficient. With this informa- cally, the method tests the null hypothesis of a reflective
tion, a student’s t‑test can be performed to measure the measurement model. Furthermore, in case of rejection, the
significance of path model relationships. method provides support for using the formative measure-
The bootstrapping analysis allows for the statistical test- ment model alternative. Other studies broach the issue of
ing of the hypothesis that a coefficient equals zero (null moderating effects (Henseler and Chin 2010; Henseler and
hypothesis) as opposed to the alternative hypothesis that Fassott 2010), hierarchical component models (Wetzels,
the coefficient does not equal zero (two-tailed test). The Odekerken-Schroder, and van Oppen 2009), and higher-
effectiveness of the bootstrap depends on the sample’s order interaction effects in hierarchical component models
representativeness in terms of the targeted population. (Wilson 2010). All these research efforts expand PLS‑SEM’s
applicability in marketing and management research.
Summary and conclusions While the recent methodological developments in
PLS‑SEM are radical, there are several promising avenues
PLS‑SEM path modeling can indeed be a “silver bullet” for for further research. From a technical point of view, future
estimating causal models in many theoretical model and research should aim at developing a global goodness-of-fit
empirical data situations. Other researchers’ criticisms measure for PLS‑SEM. This kind of model evaluation crite-
of PLS‑SEM often focus on past abuses instead of on the rion is a key requirement for testing and comparing alterna-
method’s true potential. Researchers must understand the tive theories with their associated models. Future research
goal of the analysis in order to select an appropriate statisti- should also aim at providing corrections for the PLS‑SEM
cal method for SEM applications. bias regarding estimated relationships in the model. Based
Instead of using the model to explain the covariation on its recent as well as promised methodological advances,
between all the indicators, PLS‑SEM provides parameter PLS‑SEM continues to become an ever-more commanding
estimates that maximize the explained variance (R² values) multivariate analysis tool for marketing research, with
of the dependent constructs. The method therefore supports improved general usefulness in various research settings.
prediction-oriented goals (i.e., explaining/predicting the In particular, the method supports the theoretical devel-
target constructs in the structural model). Its flexibility opment of standard path models for assessing the success
(i.e., almost no limiting assumptions regarding the model drivers of certain target constructs with key relevance for
 Spring 2011 149

marketing management (e.g., in specific industry and com- Cenfetelli, Ronald T., and Geneviève Bassellier (2009), “Interpre-
pany settings). Such models can become a valuable tool to tation of Formative Measurement in Information Systems
Research,” MIS Quarterly, 33 (4), 689–708.
frequently estimate, assess, monitor, and benchmark the Davison, Anthony C., and David V. Hinkley (1997), Bootstrap Meth‑
key drivers. ods and Their Application, Cambridge: Cambridge University
Finally, even though we believe that PLS‑SEM path mod- Press.
eling can indeed be a “silver bullet” if correctly applied, Diamantopoulos, Adamantios, and Judy A. Siguaw (2000), Introduc‑
ing LISREL, Thousand Oaks, CA: Sage.
it is important to stress that we do not argue against the
———, and ——— (2006), “Formative vs. Reflective Indicators in Mea-
statistically more demanding CB‑SEM method. CB‑SEM sure Development: Does the Choice of Indicators Matter?”
and PLS‑SEM are complementary in that the nonparamet- British Journal of Management, 13 (4), 263–282.
ric and variance-based PLS‑SEM approach’s advantages ———, and Heidi M. Winklhofer (2001), “Index Construction with
Formative Indicators: An Alternative to Scale Development,“
are the parametric and covariance-based SEM approach’s
Journal of Marketing Research, 38 (2), 269–277.
disadvantages and vice versa (Fornell and Bookstein 1982; ———, Petra Riefler, and Katharina P. Roth (2008), “Advancing For-
Henseler, Ringle, and Sinkovics 2009; Jöreskog and Wold mative Measurement Models,” Journal of Business Research,
1982; Lohmöller 1989; Schneeweiß 1991). Thus, the one 61 (12), 1203–1218.
method is not superior to the other in general—these kinds Dijkstra, Theo (2010), “Latent Variables and Indices: Herman
Wold’s Basic Design and Partial Least Squares,” in Handbook
of statements are usually biased by researchers’ method- of Partial Least Squares: Concepts, Methods and Applications
ological background and discipline. Depending on the in Marketing and Related Fields, Vincenzo Esposito Vinzi,
specific empirical context and objectives of an SEM study, Wynne W. Chin, Jörg Henseler, and Huiwen Wang, eds.,
PLS‑SEM’s distinctive methodological features make it a Berlin: Springer, 23–46.
Edwards, Jeffrey R., and Richard P. Bagozzi (2000), “On the Nature
valuable and potentially better-suited alternative to the and Direction of Relationships Between Constructs and Mea-
more popular CB‑SEM approaches. sures,” Psychological Methods, 5 (2), 155–174.
Efron, Bradley, and Robert J. Tibshirani (1993), An Introduction to
the Bootstrap, New York: Chapman Hall.
References Esposito Vinzi, Vincenzo, Christian M. Ringle, Silvia Squillac-
ciotti, and Laura Trinchera (2007), “Capturing and Treating
Anderson, James C., and David W. Gerbing (1988), “Structural Unobserved Heterogeneity by Response Based Segmentation
Equation Modeling in Practice: A Review and Recom- in PLS Path Modeling: A Comparison of Alternative Methods
mended Two-Step Approach,” Psychological Bulletin, 103 by Computational Experiments,” Working Paper no. 07019,
(3), 411–423. Cergy-Pontoise, France: ESSEC Research Center.
Arbuckle, James L. (2010), Amos 18.0 User’s Guide, Mount Pleasant, ———, Laura Trinchera, Silvia Squillacciotti, and Michel Tenenhaus
SC: Amos Development Corporation. (2008), “REBUS‑PLS: A Response-Based Procedure for Detect-
Arteaga, Francisco, Martin G. Gallarza, and Irene Gil (2010), “A ing Unit Segments in PLS Path Modelling,” Applied Stochastic
New Multiblock PLS Based Method to Estimate Causal Mod- Models in Business and Industry, 24 (5), 439–458.
els: Application to the Post-Consumption Behavior in Tour- Fornell, Claes (1982), “A Second Generation of Multivariate
ism,” in Handbook of Partial Least Squares: Concepts, Methods Analysis: An Overview,” in A Second Generation of Multivariate
and Applications in Marketing and Related Fields, Vincenzo Analysis, Claes Fornell, ed., New York: Praeger, 1–21.
Esposito Vinzi, Wynne W. Chin, Jörg Henseler, and Huiwen ———, and Fred L. Bookstein (1982), “Two Structural Equation Mod-
Wang, eds., Berlin: Springer, 141–169. els: LISREL and PLS Applied to Consumer Exit-Voice Theory,”
Babin, Barry J., Joseph F. Hair, and James S. Boles (2008), “Publish- Journal of Marketing Research, 19 (4), 440–452.
ing Research in Marketing Journals Using Structural Equa- ———, and David F. Larcker (1981a), “Evaluating Structural Equation
tions Modeling,” Journal of Marketing Theory and Practice, Models with Unobservable Variables and Measurement Error,”
16 (4), 279–285. Journal of Marketing Research, 18 (1), 39–50.
Bagozzi, Richard P. (1994), “Structural Equation Models in Market- ———, and ——— (1981b), “Structural Equation Models with Un-
ing Research: Basic Principles,” in Principles of Marketing Re‑ observable Variables and Measurement Error: Algebra and
search, Richard P. Bagozzi, ed., Oxford: Blackwell, 317–385. Statistics,” Journal of Marketing Research, 18 (3), 328–388.
———, and Youjae Yi (1988), “On the Evaluation of Structural Equa- Geisser, Seymour (1974), “A Predictive Approach to the Random
tion Models,” Journal of the Academy of Marketing Science, Effects Model,” Biometrika, 61 (1), 101–107.
16 (1), 74–94. Grewal, Rajdeep, Joseph A. Cote, and Hans Baumgartner (2004),
Bentler, Peter M. (1995), EQS Structural Equations Program Manual, “Multicollinearity and Measurement Error in Structural
Encino, CA: Multivariate Software. Equation Models: Implications for Theory Testing,” Marketing
Bollen, Kenneth A. (1989), Structural Equations with Latent Vari‑ Science, 23 (4), 519–529.
ables, New York: Wiley. Gudergan, Siegfried P., Christian M. Ringle, Sven Wende, and Alex-
Cassel, Claes, Peter Hackl, and Anders H. Westlund (1999), “Robust- ander Will (2008), “Confirmatory Tetrad Analysis in PLS Path
ness of Partial Least-Squares Method for Estimating Latent Modeling,” Journal of Business Research, 61 (12), 1238–1249.
Variable Quality Structures,” Journal of Applied Statistics, 26 Hahn, Carsten, Michael D. Johnson, Andreas Herrmann, and Frank
(4), 435–446. Huber (2002), “Capturing Customer Heterogeneity Using a
150 Journal of Marketing Theory and Practice

Finite Mixture PLS Approach,” Schmalenbach Business Review, Rindskopf, David (1984), “Structural Equation Models,” Sociologi‑
54 (3), 243–269. cal Methods and Research, 13 (1), 109–119.
Hair, Joseph F., William C. Black, Barry J. Babin, and Rolph E. Ringle, Christian M., and Rainer Schlittgen (2007), “A Genetic
Anderson (2010), Multivariate Data Analysis, Englewood Segmentation Approach for Uncovering and Separating
Cliffs, NJ: Prentice Hall. Groups of Data in PLS Path Modeling,” in Causalities Explored
Henseler, Jörg, and Wynne W. Chin (2010), “A Comparison of by Indirect Observation: Proceedings of the 5th International
Approaches for the Analysis of Interaction Effects Between Symposium on PLS and Related Methods (PLS’07), Harald Mar-
Latent Variables Using Partial Least Squares Path Modeling,” tens, Tormod Næs, and Magni Martens, eds., Åas, Norway:
Structural Equation Modeling, 17 (1), 82–109. MATFORSK, 75–78.
———, and Georg Fassott (2010), “Testing Moderating Effects in ———, Marko Sarstedt, and Erik A. Mooi (2010), “Response-Based
PLS Path Models: An Illustration of Available Procedures,” in Segmentation Using Finite Mixture Partial Least Squares:
Handbook of Partial Least Squares: Concepts, Methods and Ap‑ Theoretical Foundations and an Application to American
plications in Marketing and Related Fields, Vincenzo Esposito Customer Satisfaction Index Data,” in Data Mining: Annals of
Vinzi, Wynne W. Chin, Jörg Henseler, and Huiwen Wang, Information Systems, vol. 8, Robert Stahlbock, Sven F. Crone,
eds., Berlin: Springer, 713–735. and Stefan Lessmann, eds., New York: Springer, 19–49.
———, Christian M. Ringle, and Rudolf R. Sinkovics (2009), “The ———, ———, and Rainer Schlittgen (2010), “Finite Mixture and Ge-
Use of Partial Least Squares Path Modeling in International netic Algorithm Segmentation in Partial Least Squares Path
Marketing,” Advances in International Marketing, vol. 20, Modeling: Identification of Multiple Segments in a Complex
Rudolf R. Sinkovics and Pervez N. Ghauri, eds., Bingley, UK: Path Model,” in Advances in Data Analysis, Data Handling and
Emerald Group, 277–320. Business Intelligence, Andreas Fink, Berthold Lausen, Wilfried
Höck, Claudia, Christian M. Ringle, and Marko Sarstedt (2010), Seidel, and Alfred Ultsch, eds., Berlin: Springer, 167–176.
“Management of Multi-Purpose Stadiums: Importance and ———, Sven Wende, and Alexander Will (2005), “SmartPLS 2.0
Performance Measurement of Service Interfaces,” Interna‑ (Beta),” SmartPLS, Hamburg (available at www.smartpls
tional Journal of Ser vices Technology and Management, 14 .de).
(2–3), 188–207. ———, ———, and ——— (2010), “Finite Mixture Partial Least Squares
Höck, Michael, and Christian M. Ringle (2010), “Local Strategic Analysis: Methodology and Numerical Examples,” in Hand‑
Networks in the Software Industry: An Empirical Analysis book of Partial Least Squares: Concepts, Methods and Applica‑
of the Value Continuum,” International Journal of Knowledge tions in Marketing and Related Fields, Vincenzo Esposito Vinzi,
Management Studies, 4 (2), 132–151. Wynne W. Chin, Jörg Henseler, and Huiwen Wang, eds.,
Hult, G. Tomas M., Martin Reimann, and Oliver Schilke (2009), Berlin: Springer, 195–218.
“Worldwide Faculty Perceptions of Marketing Journals: Rank- ———, Oliver Götz, Martin Wetzels, and Bradley Wilson (2009), “On
ings, Trends, Comparisons, and Segmentations,” globalEDGE the Use of Formative Measurement Specifications in Struc-
Business Review, 3 (3), 1–10. tural Equation Modeling: A Monte Carlo Simulation Study to
Jöreskog, Karl G. (1973), “A General Method for Estimating a Linear Compare Covariance-Based and Partial Least Squares Model
Structural Equation System,” in Structural Equation Models Estimation Methodologies,” METEOR Research Memoranda
in the Social Sciences, Arthur S. Goldberger and Otis Dudley RM/09/014, Maastricht University.
Duncan, eds., New York: Seminar Press, 255–284. Sarstedt, Marko (2008), “A Review of Recent Approaches for Cap-
———, and Dag Sörbom (1996), LISREL 8 User’s Reference Guide, turing Heterogeneity in Partial Least Squares Path Model-
Chicago: Scientific Software. ling,” Journal of Modelling in Management, 3 (2), 140–161.
———, and Herman Wold (1982), “The ML and PLS Techniques for ———, and Christian M. Ringle (2010), “Treating Unobserved
Modeling with Latent Variables: Historical and Comparative Heterogeneity in PLS Path Modelling: A Comparison of
Aspects,” in Systems Under Indirect Observation, Part I, Her- FIMIX‑PLS with Different Data Analysis Strategies,” Journal
man Wold and Karl G. Jöreskog, eds., Amsterdam: North- of Applied Statistics, 37 (8), 1299–1318.
Holland, 263–270. ———, and Matthias Schloderer (2010), “Developing a Measurement
Lohmöller, Jan-Bernd (1989), Latent Variable Path Modeling with Approach for Reputation of Non-Profit Organizations,” Inter‑
Partial Least Squares, Heidelberg: Physica. national Journal of Nonprofit and Voluntary Sector Marketing,
Marcoulides, George A., and Carol Saunders (2006), “PLS: A Silver 15 (3), 276–299.
Bullet?” MIS Quarterly, 30 (2), iii–ix. ———, Jan-Michael Becker, Christian M. Ringle, and Manfred
Nunnally, Jum C., and Ira Bernstein (1994), Psychometric Theory, Schwaiger (2011), “Uncovering and Treating Unobserved
3d ed., New York: McGraw-Hill. Heterogeneity with FIMIX‑PLS: Which Model Selection
Reinartz, Werner J., Michael Haenlein, and Jörg Henseler (2009), Criterion Provides an Appropriate Number of Segments?”
“An Empirical Comparison of the Efficacy of Covariance- Schmalenbach Business Review, 63 (1), 34–62.
Based and Variance-Based SEM,” International Journal of Sattler, Henrik, Franziska Völckner, Claudia Riediger, and Christian
Market Research, 26 (4), 332–344. M. Ringle (2010), “The Impact of Brand Extension Success
Rigdon, Edward E. (1998), “Structural Equation Modeling,” in Factors on Brand Extension Price Premium,” International
Modern Methods for Business Research, G.A. Marcoulides, ed., Journal of Research in Marketing, 27 (4), 319–328.
Mahwah, NJ: Lawrence Erlbaum, 251–294. Schneeweiß, Hans (1991), “Models with Latent Variables: LISREL
———, Christian M. Ringle, and Marko Sarstedt (2010), “Structural Versus PLS,” Statistica Neerlandica, 45 (2), 145–157.
Modeling of Heterogeneous Data with Partial Least Squares,” Schwaiger, Manfred, Marko Sarstedt, and Charles R. Taylor (2010),
in Review of Marketing Research, vol. 7, Naresh K. Malhotra, “Art for the Sake of the Corporation Audi, BMW Group,
ed., Armonk, NY: M.E. Sharpe, 255–296. DaimlerChrysler, Montblanc, Siemens, and Volkswagen Help
 Spring 2011 151

Explore the Effect of Sponsorship on Corporate Reputations,” Wilson, Bradley (2010), “Using PLS to Investigate Interaction
Journal of Advertising Research, 50 (1), 77–90. Effects Between Higher Order Branding Constructs,” in
Sosik, John J., Surinder S. Kahai, and Michael J. Piovoso (2009), Handbook of Partial Least Squares: Concepts, Methods and Ap‑
“Silver Bullet or Voodoo Statistics? A Primer for Using Par- plications in Marketing and Related Fields, Vincenzo Esposito
tial Least Squares Data Analytic Technique in Group and Vinzi, Wynne W. Chin, Jörg Henseler, and Huiwen Wang,
Organization Research,” Group & Organization Management, eds., Berlin: Springer, 621–652.
34 (1), 5–36. Wold, Herman (1973), “Nonlinear Iterative Partial Least Squares
Steenkamp, Jan-Benedict E.M., and Hans Baumgartner (2000), (NIPALS) Modeling: Some Current Developments,” in Mul‑
“On the Use of Structural Equation Models for Marketing tivariate Analysis, vol. 3, Paruchuri R. Krishnaiah, ed., New
Modeling,” International Journal of Research in Marketing, York: Academic Press, 383–407.
17 (2–3), 195–202. ——— (1975), “Path Models with Latent Variables: The NIPALS Ap-
Stone, Mervyn (1974), “Cross-Validatory Choice and Assessment proach,” in Quantitative Sociology: International Perspectives
of Statistical Predictions,” Journal of the Royal Statistical on Mathematical and Statistical Modeling, H.M. Blalock,
Society, 36 (2), 111–147. A. Aganbegian, F.M. Borodkin, R. Boudon, and V. Capecchi,
Tenenhaus, Michel (2008), “Component-Based Structural Equa- eds., New York: Academic Press, 307–357.
tion Modelling,” Total Quality Management & Business Excel‑ ——— (1982), “Soft Modeling: The Basic Design and Some Exten-
lence, 19 (7–8), 871–886. sions,” in Systems Under Indirect Observations: Part I, Karl
Völckner, Franziska, Henrik Sattler, Thorsten Hennig-Thurau, and G. Jöreskog and Herman Wold, eds., Amsterdam: North-
Christian M. Ringle (2010), “The Role of Parent Brand Qual- Holland, 1–54.
ity for Service Brand Extension Success,” Journal of Service Wold, Svante, Michael Sjöström, and Lennart Eriksson (2001),
Research, 13 (4), 379–396. “PLS-Regression: A Basic Tool of Chemometrics,” Chemomet‑
Wetzels, Martin, Gaby Odekerken-Schroder, and Claudia van rics and Intelligent Laboratory Systems, 58 (2), 109–130.
Oppen (2009), “Using PLS Path Modeling for Assessing Wright, Sewall (1921), “Correlation and Causation,” Journal of
Hierarchical Construct Models: Guidelines and Empirical Agricultural Research, 20 (7), 557–585.
Illustration,” MIS Quarterly, 33 (1), 177–195.