Addressing Endogeneity in

International Marketing Applications of

Partial Least Squares Structural
Equation Modeling
G. Tomas M. Hult, Joseph F. Hair Jr., Dorian Proksch, Marko Sarstedt, Andreas Pinkwart, and
Christian M. Ringle

ABSTRACT
Partial least squares structural equation modeling (PLS-SEM) has become a key method in international marketing
research. Users of PLS-SEM have, however, largely overlooked the issue of endogeneity, which has become an integral
component of regression analysis applications. This lack of attention is surprising because the PLS-SEM method is
grounded in regression analysis, for which numerous approaches for handling endogeneity have been proposed. To
identify and treat endogeneity, and create awareness of how to deal with this issue, this study introduces a systematic
procedure that translates control variables, instrumental variables, and Gaussian copulas into a PLS-SEM framework.
We illustrate the procedure’s efficacy by means of empirical data and offer recommendations to guide international
marketing researchers on how to effectively address endogeneity concerns in their PLS-SEM analyses.

Keywords: control variable, endogeneity, Gaussian copula, instrumental variable, partial least squares structural
equation modeling (PLS-SEM)

he ambition to make causal claims is fundamental performance (Shu, Jin, and Zhou 2017)? How does strategic

T to international marketing research. Consider, for

example, a sampling of recent research questions
addressed in articles published in Journal of International
fit between host–home country similarity and exploration/
exploitation strategies influence firm performance (Cui,
Walsh, and Zou 2014)? Do macroenvironmental factors
Marketing: Is coopetition between foreign and local such as cultural dimensions and globalization forces mod-
partners good or bad for international joint venture erate the relationship between product diversification and
market value of large international firms (Qiu et al. 2015;
G. Tomas M. Hult is Professor and Byington Endowed Chair, Michigan
Tianjiao 2014)?
State University (email: hult@broad.msu.edu). Joseph F. Hair Jr. is
Professor of Marketing, Mitchell College of Business, University of Analyzing such research questions requires a research
South Alabama (email: jhair@southalabama.edu). Dorian Proksch is design that randomly assigns corporations or customers
Assistant Professor of Innovation Management and Entrepreneurship, to different experimental conditions and then tests for
HHL Leipzig Graduate School of Management (email: dorian.proksch@ differences among the groups (Antonakis et al. 2010).
hhl.de). Marko Sarstedt is Professor of Marketing, Otto von Guericke
However, as in many other fields of scientific inquiry
University Magdeburg and Adjunct Professor, Monash University Malaysia
(email: marko.sarstedt@ovgu.de). Andreas Pinkwart is Professor of
Innovation Management and Entrepreneurship, HHL Leipzig Graduate
School of Management (email: pinkwart@hhl.de). Christian M. Ringle Journal of International Marketing
is Professor of Management, Hamburg University of Technology (email: ©2018, American Marketing Association
c.ringle@tuhh.de). This article uses the statistical software SmartPLS 3 Vol. 26, No. 3, 2018, pp. 1–21
(http://www.smartpls.com). Christian M. Ringle acknowledges a financial DOI: 10.1509/jim.17.0151
interest in SmartPLS. Bulent Menguc served as associate editor for this article. ISSN 1069-031X (print) 1547-7215 (electronic)

Addressing Endogeneity in International Marketing Applications 1

(e.g., Clougherty, Duso, and Muck 2016; Sande and and van Heerde 2011; Park and Gupta 2012; Rossi 2014)
Ghosh 2018), researchers in international marketing as well as conjoint analysis (e.g., Liu, Otter, and Allenby
seldom undertake randomized controlled experiments. 2007). But little research addresses endogeneity in structural
This is not surprising, given that such experiments are equation modeling (SEM). Although several studies have
prohibitively expensive or are simply impossible to discussed endogeneity in the context of factor-based SEM
conduct because of managers and firms’ unwillingness to (e.g., Bollen, Kolenikov, and Bauldry 2014; Kirby and
be randomly assigned to the treatment and control groups Bollen 2009), there is a paucity of research on this topic
(Bascle 2008). As a consequence, researchers typically in partial least squares structural equation modeling
rely on survey-based primary data or secondary data (e.g., (PLS-SEM; Wold 1982), the use of which has recently gained
Nakata and Huang 2005) in which independent variables momentum in international marketing (Richter et al. 2016)
cannot be exogenously manipulated (e.g., Clougherty, and business research in general (e.g., Ali et al. 2018; Nitzl
Duso, and Muck 2016). 2016; Ringle et al. 2018). Some researchers even claim that
PLS-SEM does not allow endogeneity to be addressed at all
At the same time, in international marketing, opportu- (e.g., Antonakis et al. 2010; McIntosh, Edwards, and
nities for endogeneity to be present are widespread Antonakis 2014; Rönkkö and Evermann 2013). This as-
throughout empirical research designs and data contexts. sertion is astonishing and inaccurate given that PLS-SEM is
As Deng (2016, p. 1) notes, “endogeneity is pervasive and grounded in regression analysis (Lohmöller 1989, Chapter
sometimes inevitable in international business and market- 2), for which numerous approaches for handling endoge-
ing studies.” Numerous researchers have, therefore, called neity exist (e.g., Ebbes et al. 2005; Park and Gupta 2012;
for routine consideration of endogeneity in these and re- Staiger and Stock 1997). Indeed, Benitez, Henseler, and
lated fields (e.g., Hult et al. 2008; Sande and Ghosh 2018). Roldán (2016) recently made an advance in this direction by
combining the standard PLS-SEM algorithm with the two-
Endogeneity arises, for example, when a firm’s international stage least squares (2SLS) method, but they did not consider
marketing strategies (e.g., channel selection, entry modes, variables that control for endogeneity’s sources.
and advertising spending) are nonrandom and influenced
by firm- and industry-level factors. These factors may Overall, researchers have not yet developed a compre-
include a firm’s corporate culture and organizational ca- hensive framework that considers the multitude of ap-
pabilities or market forces such as industry growth and proaches applicable in PLS-SEM and offers guidelines for
competition intensity (Jin, Zhou, and Wang 2016). These assessing endogeneity in this context. It is therefore not
factors are typically difficult to measure and include in surprising that PLS-SEM applications do not use ap-
statistical models (e.g., Hult et al. 2008). Similarly, a proaches for assessing endogeneity. Specifically, our review
multitude of factors, which are difficult to consider in full, of articles published in the ten years between 2008 and 2017
potentially influence a firm’s international marketing per- in Journal of International Marketing, International Mar-
formance (i.e., the outcome of a firm’s decision making re- keting Review, and Journal of International Business
garding internationalization activities; Souchon et al. 2016). Studies—the three highest-ranked international marketing
In these situations, if an omitted construct or variable affects journals in Hult, Reimann, and Schilke’s (2009) journal
both the dependent variable and one or more independent ranking—shows that none of the 43 reviewed PLS-SEM
variable(s) in the regression model, the condition will induce studies tested for endogeneity (Appendix).
a correlation between the independent variable(s) and the
error term (Wooldridge 2013). Thus, variation in the omitted Addressing this gap in research, this study explains and
construct or variable will manifest in the error term, thereby illustrates how to deal with endogeneity in PLS-SEM-based
violating a key causal modeling assumption of regression international marketing studies. We discuss endogeneity’s
analysis: the error term is assumed to have an expected value roots and consequences, focusing on omitted variables as
of zero (Wooldridge 2013). As a consequence, the coefficient the principal, and most commonly examined, source of
estimates from standard regression analyses are biased and endogeneity (Bascle 2008; Clougherty, Duso, and Muck
inconsistent, thereby becoming causally uninterpretable and 2016).1 We then describe the different methods of (statis-
potentially triggering Type I and Type II errors (Papies, tically) responding to endogeneity in regression models and
Ebbes, and van Heerde 2017). translate these to a PLS-SEM context. Drawing on the
previously mentioned discussion, we offer a systematic
Dealing with endogeneity has been extensively discussed in 1Other sources of endogeneity include measurement error and
the marketing literature, especially with respect to different nonrecursive relationships between constructs (Wooldridge
forms of regression and panel models (e.g., Ebbes, Papies, 2010).

2 Journal of International Marketing

procedure for identifying and treating endogeneity in collect corresponding data. Furthermore, specifying a great
PLS-SEM and illustrate its use in an empirical example of number of control variables does not necessarily account
corporate reputation, a concept frequently examined in for all endogeneity in the model (Bascle 2008; Reeb,
international marketing and business research. Sakakibara, and Mahmood 2012) and contradicts the
basic philosophy of Occam’s razor (i.e., the principle of
TREATING ENDOGENEITY IN REGRESSION parsimony) on which the scientific discipline rests.
MODELS
Despite careful selection of control variables, researchers
Although endogeneity can have various roots, such as need to apply a statistical approach to address endogeneity
measurement errors, simultaneous causality, common if endogeneity could be a problem. Two broad statistical
method variance, and (un)observed heterogeneity (Bascle approaches have been developed to examine the presence of
2008; McIntosh, Edwards, and Antonakis 2014; Papies, endogeneity: the instrumental variable (IV) approach and
Ebbes, and van Heerde 2017; Sande and Ghosh 2018), IV-free approaches (Papies, Ebbes, and van Heerde 2017).
endogeneity problems most often arise from the omission
of variables that correlate with one or more independent Instrumental Variable Approach
variable(s) and the dependent variable(s) in the regression
model (e.g., Bijmolt, van Heerde, and Pieters 2005; Rossi In econometrics and other areas of applied research, the IV
2014). Omitting such variables induces a correlation be- approach is the technique of choice when dealing with
tween the corresponding independent variables and the endogeneity (e.g., Bound, Jaeger, and Baker 1995; Staiger
dependent variables’ error term (Wooldridge 2010). That is, and Stock 1997; Villas-Boas and Winer 1999). This ap-
the independent variables then explain not only the de- proach uses information in additionally specified IVs to
pendent variable but also the error in the model. decompose the endogenous independent variables’ variance
into two parts (Bascle 2008): the exogenous part, which
Consider the following regression model, where y repre- does not correlate with the error term, and the endogenous
sents the dependent variable, x1 and x2 are independent part, which correlates with the error term. The estimation of
variables, b0 is the intercept, b1 and b2 are the regression the regression model then uses only the exogenous part,
coefficients of x1 and x2, and e is the error term: which endogeneity does not affect, to provide unbiased and
consistent regression estimates. The IV approach is usually
(1) y = b0 + b1 x1 + b2 x2 + e.
implemented by means of the 2SLS algorithm. This algo-
Let us assume that the independent variable x2 is un- rithm’s first stage involves regressing an independent var-
correlated with e (i.e., x2 is exogenous), whereas x1 is en- iable on an IV and computing the predicted values from this
dogenous because it is correlated with the error term e equation. These values are then used as input to reestimate
(i.e., Cov(x1,e) „ 0). Thus, endogeneity occurs in this re- the original model in the second stage.
gression model example. Note that this terminology, which
is used in the endogeneity literature (Wooldridge 2010), To illustrate the approach, consider the regression model ex-
should not be confused with the terms used for the expla- ample given in Equation 1. Let us assume that the independent
nation of PLS path models with exogenous (i.e., independent) variable x1 is endogenous, whereas x2 is exogenous. The first
and endogenous (i.e., dependent) constructs (Chin 1998). stage of the IV approach involves regressing the endoge-
In the following, and in line with the literature, we use nous variable x1 on the exogenous variable x2 and the IV z:
endogenous and exogenous to identify variables that (2) x1 = g 0 + g 1 z + g 2 x2 + z.
endogeneity does or does not affect; we use dependent
and independent to identify constructs that explain other Here, g 0 represents the intercept, g 1 the regression co-
constructs in the PLS path model or are explained by them. efficient of z, g 2 the regression coefficient of x2, and z the
error term. Estimating Equation 2 yields predicted values of
A straightforward way of handling, or at least reducing, ^1 ), which express the exogenous part of the in-
x1 (i.e., x
endogeneity is to specify a set of control variables (Bernerth dependent variable x1. In the second stage, we replace x1 in
and Aguinis 2016) accounting for a part of the dependent the original regression model (Equation 1) with x ^1 to es-
variable’s variance (Ebbes, Papies, and van Heerde 2017). timate the following:
Although carefully chosen control variables often alleviate (3) y = b00 + b01 x
^1 + b02 x2 + e0 .
endogeneity problems substantially (Papies, Ebbes, and van
Heerde 2017), the success of this endeavor depends on the Endogeneity occurs when there is a significant difference
researcher’s ability to identify relevant control variables and between b1 (Equation 1) and b01 (Equation 3), with the latter

Addressing Endogeneity in International Marketing Applications 3

representing the influence of the independent variable’s Researchers have proposed the control function approach,
exogenous part on y. To test for this difference, researchers a modification of the 2SLS method, which also leverages
use the Hausman test (Hausman 1978). If the test indicates a IVs to control for endogeneity. Specifically, the control
significant difference between the b estimates, researchers function approach draws on the first-stage regression
should revert to the endogeneity-corrected estimates (Equation 2) to obtain the fitted residuals (^ x1 - x1 ), which
(Equation 3). If the Hausman test does not indicate a sig- are then included in the original model to obtain correct(ed)
nificant degree of endogeneity, researchers should use the estimates of the coefficients. Furthermore, interpreting the
noncorrected estimates from the original regression model significance of the fitted residuals requires running the
(Equation 1). Hausman test (Papies, Ebbes, and van Heerde 2017). That
is, if a coefficient is significant, researchers should revert to
To work effectively, the IVs must meet two criteria: the the endogeneity-corrected estimates.
instrument relevance criterion and the exclusion restriction.
First, according to the instrument relevance criterion, the The control function approach is particularly advanta-
correlation of each IV with the endogenous variable should geous when regression models with discrete dependent
be strong, which means that the IV’s effect g 1 on the de- variables are estimated (Petrin and Train 2010). How-
pendent variable x1 in Equation 2 must be pronounced and ever, such model types currently cannot be estimated in
significant (Papies, Ebbes, and van Heerde 2017). If it is not, PLS-SEM because the standard application of the al-
the corrected b coefficient estimates (Equation 3) may be gorithm requires the dependent constructs to be con-
considerably biased (Rossi 2014). But as Bound, Jaeger, and tinuous (Lohmöller 1989, Chapter 4). For linear models
Baker (1993) note, “the cure can be worse than the disease.” as typically estimated in PLS-SEM, the control function
To further substantiate an IV’s relevance and adequacy, approach yields exactly the same results as the 2SLS
researchers need to test whether its inclusion in the first- approach (Papies, Ebbes, and van Heerde 2017).
stage regression (Equation 2) improves the model’s R2 value
significantly, by referring to the F-statistic. As a rule of Instrumental Variable–Free Approaches
thumb, an F-statistic value of 10 or higher is generally
viewed as indicative that the IV is adequate (e.g., Staiger When the IV approach is used, identification of concep-
and Stock 1997). Alternatively, Stock, Wright, and Yogo tually fitting IVs is challenging (Rossi 2014), especially in
(2002) provide a table of critical F-statistic values for IV light of the data constraints characterizing a typical research
models based on the number of instruments used. Note project. To address this problem, researchers have sug-
that in situations with multiple endogenous variables and gested IV-free approaches that do not require IVs to be
multiple IVs, researchers must apply a multivariate F-test specified. For example, Park and Gupta (2012) introduced
(e.g., Angrist and Pischke 2009) to test for the IVs’ ad- the Gaussian copula approach, which controls for endo-
equacy. In addition, situations with multiple endogenous geneity by directly modeling the correlation between the
variables require researchers to consider at least the same endogenous variable and the error term by means of a
number of IVs to ensure that the approach draws on an copula. Another IV-free approach, proposed by Ebbes et al.
identified, or overidentified, model (Murray 2006b). (2005), is conceptually similar to the IV approach in that it
splits an endogenous variable into exogenous and endog-
Second, according to the exclusion restriction for evaluating enous parts. However, this latent instrumental variable
IVs when examining endogeneity, each IV should not be (LIV) approach does not draw on an observed variable
correlated with the error term in the original regression but generates an unobserved (i.e., latent) IV. Ebbes et al.
model (Murray 2006b). This requirement cannot be (2005) assume that the LIV is discrete and has at least two
assessed statistically unless the researcher specifies more IVs categories, separating the sample into L latent categories.
than endogenous variables. In this case, the Sargan test and Thus, with bivariate normally distributed error terms, the
the Hansen J test enable researchers to test for significant likelihood-based LIV approach belongs to the class of
correlations between the IVs and the error term (Bascle normal mixture models with L components (Papies, Ebbes,
2008; Wooldridge 2010). Rejecting the null hypothesis of and van Heerde 2017) whose model parameters can be
no correlation, however, does not help to verify which of the identified for L > 1 and nonnormality of the endogenous
IVs are endogenous (Papies, Ebbes, and van Heerde 2017). regressor (Ebbes et al. 2005).
Although this test provides some empirical substantiation
when the first-stage model is overidentified, researchers The LIV approach can be used in a PLS-SEM context, but a
should rely primarily on theoretical arguments when fundamental challenge in its application is the selection of
assessing the IVs’ adequacy (e.g., Murray 2006a). the number of categories, L. A priori, this number is

4 Journal of International Marketing

unknown, but it has a substantial effect on the results. with endogeneity in PLS-SEM. In the following sec-
Although Papies, Ebbes, and van Heerde (2017, p. 609) tions, we discuss each step in greater detail.
suggest that “the researcher fits at least L = 2, 3, 4 and 5,”
there is no clear guidance regarding the desired stability of Stage 1 of the procedure requires checking whether testing
the estimates to vary L and thus regarding the final choice of for endogeneity is necessary. Specifically, the meaning-
this parameter. At the same time, whereas the statistical fulness of considering endogeneity in PLS-SEM depends
software R facilitates the Gaussian copula approach, cur- on whether researchers use the method for predictive or
rent software support for the LIV approach is limited be- explanatory modeling purposes. In predictive modeling,
cause models can include only one independent and one the purpose of the analysis is to predict new or future
dependent variable (REndo version 1.3). Because these two observations (Shmueli 2010). Here, the main concern is
issues limit the LIV approach’s applicability, we focus on the predicting the case values of the dependent variables,
Gaussian copula approach, which enables correction for rather than (causally) interpreting the structural paths.
endogeneity in the absence of IVs. Controlling for endogeneity is not useful in this case, as it
may reduce the model’s predictive power (Ebbes et al.
Table 1 provides an overview of the advantages and dis- 2005). Conversely, in explanatory modeling, which in-
advantages of the control variable approach, the IV ap- volves “the use of statistical models for testing causal
proach, and two types of IV-free approaches (Gaussian explanations” (Shmueli 2010, p. 290), controlling for
copula and LIV) when dealing with endogeneity in endogeneity is crucial to adequately test hypotheses
PLS-SEM. (Papies, Ebbes, and van Heerde 2017). In practice, how-
ever, researchers using PLS-SEM generally balance pre-
diction and explanation in that they expect their model to
ADDRESSING ENDOGENEITY IN PLS-SEM have high predictive accuracy while being grounded in
Procedure well-developed causal explanations (Sarstedt, Ringle, and
Hair 2017a). Gregor (2006, p. 626) refers to this interplay
PLS-SEM is a commonly used method to estimate path as explanation and prediction theory, noting that this
models with latent variables (Lohmöller 1989, Chapter 2; approach “implies both understanding of underlying
Wold 1982). Unlike factor-based SEM, which considers causes and prediction, as well as description of theoretical
constructs as common factors that explain the covariation constructs and the relationships among them.” Such a
between the associated indicators (Jöreskog 1978), PLS- setting may require the consideration of endogeneity,
SEM is a composite-based approach to SEM that uses depending on whether explanation (i.e., testing a theory)
linear combinations of indicator variables to represent or prediction (e.g., deriving managerial recommendations)
conceptual variables in a statistical model (Henseler et al. is the prevailing intent of a certain research project.
2014; Sarstedt et al. 2016). PLS-SEM estimates the pa-
rameters of a set of equations in a structural equation model Stage 2 of the procedure (Figure 1) uses existing theory and
by combining an approach similar to principal components prior research findings to identify potential endogeneity
analysis with regression-based path analysis. Specifically, issues a priori. If prior research adequately theorized, de-
the PLS-SEM method uses partial regressions to estimate the tected, and treated endogeneity issues in a specific research
path coefficients between the latent variables and their in- context, researchers should apply this knowledge when
dicators in the measurement models, as well as between the proposing a PLS path model and collecting data. This
latent variables in the structural model. Lohmöller (1989, method typically involves including control variables in the
Chapter 2) provides an introduction to the PLS-SEM model, or gathering data on IVs for an ex post assessment,
algorithm. Approaches for dealing with endogeneity in using the IV approach (e.g., Ebbes, Papies, and Van Heerde
regression models can therefore be readily transferred to 2017; Reeb, Sakakibara, and Mahmood 2012). Thus, if
PLS-SEM. However, in contrast to regression analysis, in prior research suggests that endogeneity is an issue, or if
which the approaches use the (observed) indicator variables the results cast doubt on the assertion that all sources of
as input, in PLS-SEM the implementation of these variables endogeneity have been detected, researchers should test
draws on the composite scores obtained after the al- whether endogeneity substantially affects the model’s
gorithm’s convergence (Lohmöller 1989, Chapter 2).2 estimates.
Figure 1 illustrates our systematic procedure for dealing
2This To do so, in Stage 3 of the procedure (Figure 1), researchers
implementation is similar to other methodological
extensions, such as finite mixture PLS (Hair et al. 2018), which apply the Gaussian copula approach. If the results do not
uses composite scores as input for mixture regressions. indicate a problem, researchers can use the original model’s

Addressing Endogeneity in International Marketing Applications 5

 Table 1. Comparison of Approaches to Deal with Endogeneity

Methods

Instrumental Variable–Free Approaches

Criteria Control Variable Approach Instrumental Variable Approach Gaussian Copula Latent Instrumental Variable

6 Journal of International Marketing

Number of variables Data on additional variables Instrumental variables have to be No additional variables No additional variables needed
must be collected identified, and data must be needed
collected
Distribution of variables No assumptions required No assumptions required Endogenous variables have to Endogenous variables have to
be nonnormally distributed be nonnormally distributed
Nature of dependent Discrete or continuous Continuous Discrete or continuous Continuous
variable
Statistical tests Not necessary Test for significance and Test for significance Test for significance
relevance
Acceptance in scientific Widely accepted and Widely accepted and commonly Relatively new and therefore Relatively new and therefore
community commonly used used rarely used rarely used
Implementation in software No additional implementation Supported by, for example, SPSS, Supported by the REndo Supported by the REndo (Gui,
necessary Stata, and R software packages (Gui, Meierer, and Meierer, and Algesheimer
Algesheimer 2017) package 2017) package with one
dependent variable and one
independent variable
 Figure 1. Procedure to Address Endogeneity in PLS-SEM

Stage 1: Decide whether you need to Stage 2: Use prior knowledge to Stage 3: Run the Gaussian copula Stage 4: Explain endogeneity.
address endogeneity issues. address and explain endogeneity issues. approach to detect endogeneity issues.

Does related prior research

Research goal? No
identify endogeneity issues? Assumptions of the
Gaussian copula No
approach fulfilled?

Yes
Primarily Primarily
Yes Include control
prediction explanation variables
Address and explain
endogeneity in line with prior Does the Gaussian copula Use the PLS -SEM results
Does endogeneity represent No Are suitable control
Yes research. Do you identify approach detect Yes Yes that include additional
a potential problem? critical endogeneity issues? endogeneity issues? variables available?
control variables.*

Include IVs
No Yes No No

Do you suspect additional

sources of endogeneity Yes Do the results of the original
problems? Is a valid and strong IV Use the results
Yes model and the IV model Yes
available? significantly differ? of the IV model.*

No
No
No
Use the PLS results and
address endogeneity line
with prior research.*
Are results of the Gaussian Use the results of the
Use the PLS results. No Yes
copula approach available? Gaussian copula approach.

*Consider continuing at Stage 3 and using the Gaussian copula approach to assess the extent to which the endogeneity problem has been successfully addressed by the model.

Addressing Endogeneity in International Marketing Applications 7

estimates. Otherwise, the sources of endogeneity must be
further explored and controlled for in the next step of the Figure 2. Simple PLS Path Model
procedure. Because endogeneity problems in international
marketing often arise from omitted variables (Jean et al.
2016), Stage 4 involves conceptually identifying the missing x11
information and adding control variables to the PLS path
model (Papies, Ebbes, and van Heerde 2017). The inclusion x12 X1
of suitable control variables deals with, or at least reduces, y11
the impact of endogeneity on the model estimates (Ebbes, x13
Papies, and van Heerde 2017). In situations in which no
Y1 y12
suitable control variables are available, researchers can
draw on the IV approach. If IVs are unavailable, or if re- x21
searchers cannot identify at least one suitable IV for each y13
endogenous variable in a partial regression model, they can x22 X2
use the Gaussian copula results, if available, and report that
they cannot identify omitted variables that adequately ex- x23
plain the endogeneity problem. If Gaussian copula results
are not available, because they do not meet the method’s
requirements, researchers should interpret the original PLS
path model estimates and specify the reason(s) for not
Ringle, and Hair 2017a). With reference to our example,
addressing the potential endogeneity issues as a limitation
the final stage of the PLS-SEM algorithm entails regressing
(Figure 1).
Y on X1 and X2 to estimate the structural model path
coefficients b1 and b2:
The Gaussian Copula Approach in PLS-SEM
(4) Y = b0 + b1 X1 + b2 X2 + e.
To illustrate the Gaussian copula approach in PLS-SEM,
Let us assume that the independent variable X1 is thought
consider a simple path model with one dependent latent
to be correlated with the error term (e) of the dependent
variable, Y1, explained by two independent latent vari-
variable Y and may therefore induce an endogeneity
ables, X1 and X2. Each of the latent variables is measured
problem in the estimation of Equation 4.
using three indicators. Figure 2 shows this simple path
model. The hexagons represent composites formed by linear
The Gaussian copula approach requires the composite
combinations of the indicator variables, as indicated by the
scores of the endogenous construct X1 to be nonnormally
arrows linking the indicators and the composites. These
distributed. If this requirement is met, the Gaussian copula
arrows do not necessarily correspond to the measurement
of X1 can be calculated as follows (Papies, Ebbes, and van
model specification, which can be reflective or formative
Heerde 2017):
(Sarstedt et al. 2016). However, we use this illustration
because PLS-SEM forms composites from the indicator (5) cp = F-1 ½HðX1 Þ,
variables to represent the conceptual variables in the model
estimation. where H(x) is the empirical cumulative density function and
F−1 is the inverse normal cumulative density function. Next,
Model estimation in PLS-SEM draws on a three-stage the copula is included as an additional independent variable
approach that belongs to the family of (alternating) least that controls for the correlation between the error term and the
squares algorithms (Mateos-Aparicio 2011). Stage 1 of the endogenous independent construct in the regression model:
0 0
PLS-SEM algorithm iteratively determines indicator (6) y = b0 + b1 x1 + b2 x2 + b3 cp1 + e0 .
weights, composite scores, and path coefficients. The al-
gorithm converges when the indicator weights change only To attend to the endogeneity problem, estimating Equation
0 0
marginally from one iteration to the next. Stages 2 and 3 use 6 yields the new path coefficients b1 and b2 .
the final composite scores from Stage 1 as input for a series
of ordinary least squares regressions. These regressions To determine whether endogeneity is at a critical level,
produce the final indicator loadings, indicator weights, and researchers need to assess the significance of the copula
path coefficients, as well as related elements, such as the coefficient (b3) using bootstrapping (Hausman 1978; Park
indirect and total effects, and R2 values (e.g., Sarstedt, and Gupta 2012). A significant copula coefficient indicates

8 Journal of International Marketing

a critical level of endogeneity. If the copula coefficient is
significant, researchers can assume that an endogeneity Figure 3. First Stage of the IV Approach
issue is present, even though it is treated by including the
copula in the regression model (Equation 6). If the copula
coefficient is not significant, no critical endogeneity issue z11
affects the regression results.
z12 Z1
Although the standard Gaussian copula approach assumes x11
a single endogenous variable, Park and Gupta (2012) dis- z13
cuss the case of modeling multiple endogenous variables
simultaneously. Doing so requires creating the copula for X1 x12

each individual endogenous variable and adding all copulas x21

to the original model. The procedure separately computes x13
all the Gaussian copulas of the specified endogenous var- x22 X2
iables (Park and Gupta 2012).
x23
Control Variables in PLS-SEM
If the Gaussian copula approach indicates a significant
degree of endogeneity, researchers should use control
variables to identify its sources. The introduction of control Similar to the analysis of interaction effects in PLS-SEM (Hair
variables should always be grounded in strong theory et al. 2017; Chapter 7), the assessment of significant control
(Bernerth and Aguinis 2016). That is, researchers should be variables should include the use of the f2 effect size to assess
explicit (rather than implicit) regarding the role of control their relevance. According to Chin (1998), f 2 values of .02,
variables in the analysis and potentially even match the .15, and .35 suggest small, moderate, and substantial effects.
hypotheses precisely to their choice (Spector and Brannick
2010). Typical examples of control variables used in in- The Instrumental Variable Approach in PLS-SEM
ternational marketing models include a company’s location,
age, size, export experience, years of global experience, Implementing the IV approach in the context of PLS-SEM
percentage of products sold overseas, number of countries requires adjusting the final stage of the PLS-SEM algorithm.
entered, and cultural distance (e.g., Cui, Walsh, and Zou In its original form, this stage uses the composite scores
2014; Shu, Jin, and Zhou 2017). obtained after convergence as input for a series of ordinary
least squares regressions that estimate the structural model
The use of control variables in PLS-SEM is analogous to path coefficients. Instead of using ordinary least squares
their use in regression models. More precisely, control regressions, however, this stage now uses the 2SLS method
variables are included as single-item constructs in a PLS (Benitez, Henseler, and Roldán 2016). When implementing
path model. Each control variable is then linked to the the IV approach in PLS-SEM, researchers need to ensure
dependent latent variable whose predictor constructs have that they include at least one IV per endogenous variable.
been identified as (or are assumed to be) possibly exhibiting
endogeneity. Note that criticism of the use of single-item Analogous to the process of regression analysis, the IV
constructs (Diamantopoulos et al. 2012) does not apply to approach’s implementation requires researchers to find IVs
control variables, because they usually represent observable that are highly correlated with the independent variables
characteristics. However, if the control variables represent (instrument relevance) but are uncorrelated with the de-
unobservable concepts (e.g., cultural dimensions), they need pendent variable’s error term (exclusion restriction). In
to be measured by means of multiple items. Finally, the international marketing, previous regression-based models
interpretation of the control variables’ effect is also anal- used IVs such as organization quality, brand quality, en-
ogous to that in regression models and depends on each vironmental causes, brand strength (Swoboda and
control variable’s measurement scale, which can be cate- Hirschmann 2017), exposure to a multinational company’s
gorical or metric in PLS-SEM (Hair et al. 2017). marketing (Swoboda and Hirschmann 2016), and export
memory (Souchon et al. 2016). However, the identification
After the control variables are specified, researchers need to of suitable IVs is highly context-specific and is the most
use bootstrapping to assess these variables’ significance. challenging step in the analysis (e.g., Larcker and Rusticus

Addressing Endogeneity in International Marketing Applications 9

2010). Angrist and Pischke (2009, p. 17) argue that finding
good IVs requires “a combination of institutional knowl- Figure 4. The Simple Corporate Reputation Model
edge and ideas about processes determining the variable of
interest,” especially because no true test of their quality is comp1
available (Rossi 2014). comp2 COMP .016
comp3 .152
cusl1
To illustrate the implementation of the IV approach in .509
cusl2
PLS-SEM, consider the simple model in Figure 2 with the cusa CUSA CUSL

construct X1 possibly exhibiting endogeneity. We now like1 cusl3

.433
introduce an IV (i.e., Z1), measured by three indicators, that like2 LIKE .331
correlates strongly with the construct X1 but not with the like3
error term of Y1 in the original model. Thereafter, we create
and estimate the model shown in Figure 3.

Next, a 2SLS analysis uses the composite scores of X1 and X2, model originally presented by Eberl (2010). The goal of this
obtained from the first model (Figure 2), and the composite model is to explain the effects of competence (COMP) and
scores of Z1, derived from the second model (Figure 3). likability (LIKE), which represent corporate reputation’s
Specifically, we regress X1 on Z1 and X2 to obtain X ^ 1: two dimensions (Sarstedt, Wilczynski, and Melewar 2013;
Schwaiger 2004), on customer satisfaction (CUSA) and
(7) ^ 1 = g + g Z1 + g X2 .
X ultimately on customer loyalty (CUSL). This model has
0 1 2
been used to illustrate aspects of the PLS-SEM method (e.g.,
We first assess the IV’s strength by comparing the R2 values Hair et al. 2017; Hair et al. 2018) and has been frequently
of the first-stage regression (Equation 7) with and without applied in an international marketing context.3 For ex-
the IV. The results of an F-test indicate whether the R2 ample, Zhang and Schwaiger (2012) use this model to study
values of the regression model with and without the IVs the differentiating effects of corporate social responsibility
differ significantly. If significant differences are found, we and other antecedent dimensions on corporate reputation in
continue with the analysis by theoretically substantiating three Western countries (Germany, the United States, and
that Z1 is not correlated with the error term of Y1 (e.g., the United Kingdom) and in China. Similarly, Radomir and
Semadeni, Withers, and Certo 2014; Staiger and Stock Alan (2018) apply this model to contrast the effects of
1997). In the case of an overidentified model, statistical service quality and relationship investment on corporate
assessments by means of the Sargan test and the Hansen J reputation assessments of Romanian and U.K. consumers.
test may support this analysis. Furthermore, Walsh, Beatty, and Shiu (2009) developed
and validated a variant of this model using cross-country
Finally, we reestimate the original model by regressing Y on data; their model features prominently in the international
X2 and X ^ 1 from Equation 7 to obtain the corrected path
0 0 marketing literature. For example, several researchers have
coefficient estimates b1 and b2 : used this model to analyze the moderating effect of in-
(8)
0 0
^ 1 + b X2 + e 0 .
Y = b 0 + b1 X
0
stitutional country differences (Swoboda, Puchert, and
2
Morschett 2016) and cultural dimensions (Swoboda and
The Hausman test (Hausman 1978) enables us to test Hirschmann 2017) on the relationship between corporate
0
whether b1 (Equation 8) differs significantly from b1 reputation and customer loyalty.
(Equation 5). If so, researchers should revert to the IV
approach’s coefficients (Equation 8). If the difference is not Figure 4 shows the original model, which includes two
significant, researchers should continue using the results of partial regression models: (1) CUSL is regressed on COMP,
the partial regression model’s original PLS-SEM estimation. CUSA, and LIKE, while (2) CUSA is regressed on COMP
and LIKE. The following illustrations focus on the first,
more complex, regression model. The measurement
EMPIRICAL EXAMPLE TO ADDRESS models of COMP, LIKE, and CUSL draw on three re-
ENDOGENEITY IN PLS-SEM flective items each, whereas CUSA is measured with a
Model and Data single item. The model estimation draws on data from two

To illustrate our procedure for identifying and dealing with 3The model and data can be downloaded from https://www.
endogeneity (Figure 1), we draw on a corporate reputation pls-sem.com/.

10 Journal of International Marketing

major German mobile communications network providers We use the constructs’ standardized composite scores to
and two smaller competitors. A total of 344 respondents compute the Gaussian copula of the partial regressions in
rated each item on a seven-point Likert scale. Observations the structural model. To cross-validate the results, we used
with missing values were deleted, leaving a total sample two different statistical software packages, which returned
size of 336. We used the statistical software R (R Core the same results: Stata 14 (StataCorp 2015) and the REndo
Team 2018), Stata (StataCorp 2015), and SmartPLS 3 package of the R program (Gui, Meierer, and Algesheimer
(Ringle, Wende, and Becker 2015) for the computations. 2017). To implement Stata, we used specific commands (1)
For the PLS-SEM estimations, we use Mode A for all to compute the Gaussian cumulative distribution function,
composites (Sarstedt, Ringle, and Hair 2017a). The results (2) to create the copula as the inverse of the Gaussian
of the bootstrap routine with 5,000 samples, with the normal cumulative distribution function, and (3) to calcu-
no-sign-changes option applied, enable us to test for the late the model that includes the Gaussian copula. Because
coefficients’ significance. We find that the measurement the REndo package does not currently support the com-
models meet all the relevant evaluation criteria. In the putation of bootstrap standard errors, we used our own R
structural model assessment, all the path coefficient esti- code and the boot package (Canty and Ripley 2017) to
mates, except COMP to CUSL, are significant, and the run this part of the analysis.4
model’s R2 values are satisfactory (for this application,
Hair et al. [2017] provide further details of the PLS-SEM The results in Table 2 show that only one Gaussian copula
results and their assessment). (i.e., cCUSA) is significant (p < .1) when we consider one
endogenous variable; this finding points to a potential
Most corporate reputation model analyses take a predictive endogeneity issue. Including the significant Gaussian copula
perspective (e.g., Schwaiger 2004). However, in line with in the model changes the effect of CUSA on CUSL by .073
Eberl (2010), who elaborates on the causal effects of cor- units (from .509 to .582), which points to a potential
porate reputation’s two components (i.e., COMP and endogeneity problem for CUSA (model 3). Similarly, cCUSA
LIKE) on CUSA and CUSL, we consider our analysis pri- is also significant in the CUSA models in combination with
marily explanatory in nature. Thus, following Stage 1 of the LIKE (model 5, p < .1), COMP (model 6, p < .05), and LIKE
procedure (Figure 1), we continue by applying the Gaussian and COMP (model 7, p < .1). This finding confirms the
copula approach. possibility of CUSA being endogenous.5

Application of the Gaussian Copula Approach Application of the Control Variable Approach
in PLS-SEM in PLS-SEM
Before initiating the Gaussian copula approach, to meet its After completing Stage 3 of the procedure for identifying
assumptions, we first verify whether the variables that and treating endogeneity (Figure 1), we next include in the
potentially exhibit endogeneity are nonnormally distributed model control variables that, from a theoretical perspective,
(Figure 1). We do so by running the Kolmogorov–Smirnov are likely to influence customers’ loyalty (Eberl 2010). The
test with Lilliefors correction (Sarstedt and Mooi 2014) on first control variable (serviceprovider) distinguishes be-
the standardized composite scores of COMP, LIKE, and tween large, globally operating mobile phone service pro-
CUSA, which the PLS path model estimation provides. The viders (serviceprovider = 1; n(1) = 251) and smaller, locally
results show that none of the constructs have normally operating competitors (serviceprovider = 2; n(2) = 85). In
distributed scores; thus, we can consider them endogenous addition, we include a second control variable (servicetype),
in the Gaussian copula analysis. which is relevant from a service marketing perspective and
distinguishes between customers with a contract and those
Next, we create three regression models in which we with a prepaid plan.
consider the independent constructs COMP (model 1),
LIKE (model 2), and CUSA (model 3) as possibly exhibiting Analysis of the impact of the two control variables, both
endogeneity. In addition, to simultaneously account for separately and jointly, shows that neither has a significant
multiple endogenous variables when applying the Gaussian
copula approach, we create four regression models that 4The R code can be downloaded at https://www.pls-sem.com/.
5The parallel findings from the application of the Gaussian
include all possible combinations of multiple endogenous
copula approach to single and multiple endogenous variables is
variables in this PLS path model example: LIKE and COMP
not surprising, since the technique computes the copulas sep-
(model 4), LIKE and CUSA (model 5), COMP and CUSA arately when considering multiple endogenous variables in the
(model 6), and LIKE, COMP, and CUSA (model 7). PLS path model.

Addressing Endogeneity in International Marketing Applications 11

 Table 2. Results of the Gaussian Copula Approach

12 Journal of International Marketing

Gaussian Gaussian Gaussian Gaussian Gaussian
Copula Model 1 Gaussian Gaussian Copula Model 4 Copula Model 5 Copula Model 6 Copula Model 7
(Endogenous Copula Model 2 Copula Model 3 (Endogenous (Endogenous (Endogenous (Endogenous
Variable: (Endogenous (Endogenous Variables: LIKE, Variables: LIKE, Variables: Variables: LIKE,
Original Model COMP) Variable: LIKE) Variable: CUSA) COMP) CUSA) COMP, CUSA) COMP, CUSA)

Variable Value p-Value Value p-Value Value p-Value Value p-Value Value p-Value Value p-Value Value p-Value Value p-Value

COMP .016 .746 .014 .848 .017 .763 .021 .707 −.006 .939 .021 .705 −.027 .717 −.037 .653
LIKE .331 <.01 .331 <.01 .370 <.01 .331 <.01 .381 <.01 .341 <.01 .333 <.01 .362 <.01
CUSA .509 <.01 .509 <.01 .511 <.01 .582 <.01 .509 <.01 .580 <.01 .592 <.01 .589 <.01
cCOMP .002 .974 .019 .737 .039 .432 .047 .405
cLIKE −.033 .245 −.041 .283 −.008 .790 −.024 .509
cCUSA −.041 .063 −.039 .084 −.049 .045 −.047 .061
Table 3. Results of the IV Approach

Coefficients R2 Values Weak ID Test Hausman Test

First Stage First Stage

Endogenous Variable COMP LIKE CUSA Without IV with IV F-Value Significant? p-Value

LIKE −.075 .506 .454 .496 .560 49.478 Yes .216

COMP .120 .278 .490 .433 .560 97.647 Yes .216
CUSA −.087 .081 1.113 .295 .300 2.575 No .216

influence on the CUSL construct. Specifically, the servicetype determinant of customer loyalty (e.g., Berné, Múgica, and
variable has a −.003 effect (p-value = .917; f 2 = .000) and Yagüe 2001). Because this tendency is a psychological trait
the service provider variable has a −.038 effect (p-value = that is independent of individuals’ preference structures and
.319; f 2 = .003) on CUSL when included separately. In- is stable over time (Menon and Kahn 1995), variety seeking
cluding both control variables simultaneously changes the is probably unrelated to customers’ perception of switching
results only marginally. Because none of the control variables costs. Network quality is another key determinant of cus-
identify a source of endogeneity, we continue with the tomer loyalty that is relevant in the context of our study
procedure (Figure 1) by running the IV approach on the data. (e.g., in terms of coverage, stability, and voice quality; Lim,
Widdows, and Park 2006). Network quality can differ
Application of the IV Approach in PLS-SEM substantially, especially in rural areas, and cannot be fully
observed a priori. Poor network coverage will have a strong
The application of the IV approach in the context of our negative effect on customer loyalty, independent of the
model requires selection of an IV that is correlated with customers’ perception of the switching costs.6
COMP, CUSA, and LIKE. At the same time, the IV needs to
be uncorrelated with the omitted determinants of customer We use an extended version of Jones, Mothersbaugh, and
loyalty and thus be uncorrelated with CUSL’s error term. Beatty’s (2000) scale and measure the switching costs by
Our implementation of the IV approach uses customers’ using four reflective indicators (switch1 to switch4), each
perceived switching costs as an IV. Switching costs comprise measured on a five-point Likert scale. For our analysis, we
the time, money, and effort that consumers expect to expend extract all (standardized) composite scores from the
when changing their service provider. In the mobile phone SmartPLS 3 software. We subsequently run the IV approach
market, switching costs come primarily in the form of by using Stata’s ivregress method (StataCorp 2017), as well
transfer costs (Klemperer 1987) incurred when a consumer as the statistical software R with its AER package (Kleiber
ends a relationship with one provider and starts a new one and Zeileis 2017) to validate the results. Both programs
with another. Specifically, the implementation and config- provide the same results. Table 3 shows the results of the
uration of the new relationship require the customer to analysis for the IVs.
invest time, effort, and other assets, while ending an old
service relationship may also require procedural tasks (Ray, To assess whether the IV is strong (Figure 1), we use the
Kim, and Morris 2012). Thus, in terms of the instrument weak ID test’s F-value, which compares the R2 values of the
relevance criterion, it is reasonable to assume that the first-stage regression with and without the IV. We consider
perceived switching costs correlate with the customers’ three models, each of which considers another predictor
satisfaction and reputation assessments, because significant construct of CUSL as likely to exhibit endogeneity
transfer costs limit their freedom to configure the business (i.e., LIKE, COMP, and CUSA). The results in Table 3 show
relationship. Given the manifold factors that influence that including the IV significantly increases the first-stage
loyalty, arguing that perceived switching costs are uncor-
related with omitted variables that affect customer loyalty 6While these arguments support the use of perceived

(exclusion restriction) is more difficult. However, research switching costs, we note that this is not an ideal IV, as one could
expect this variable to correlate with other service perceptions
has identified customers’ variety-seeking tendency—the
(e.g., perceived service quality) that are relevant for customer
“tendency of individuals to seek diversity in their choices loyalty and which the two reputation dimensions capture only
of services or goods” (Kahn 1995, p. 139)—as a major indirectly (e.g., Raithel and Schwaiger 2015).

Addressing Endogeneity in International Marketing Applications 13

R2 value when COMP and LIKE are considered as en- have overlooked this issue. A potential reason for this
dogenous, as evidenced by the weak ID test values well disregard is the predictive nature of PLS-SEM, which, if
above 10. Conversely, the model’s R2 value increases only taken into full consideration when the method is used,
marginally when CUSA is considered likely to exhibit renders endogeneity irrelevant (Ebbes, Papies, and van
endogeneity. Thus, in the following step, we use the Heerde 2011). However, researchers conventionally use
Hausman test to assess whether the coefficients of the IV PLS-SEM to test formal hypotheses, which corresponds
approach differ significantly from the original coefficients. to an explanatory perspective, while simultaneously de-
The results in Table 3 show that, in our analysis, the original riving managerial implications, which follows a predictive
and IV coefficients do not differ significantly. When LIKE perspective. Endogeneity becomes an important concern,
and COMP are considered endogenous, the Hausman test therefore, depending on whether researchers emphasize
specifically returns p-values of .216.7 explanation or prediction in their analyses.

Including switching costs as an IV does not explain the Similar to scholars who have applied PLS-SEM in inter-
endogeneity problem’s origin. More specifically, the IV national marketing and other disciplines, methodological
correction of CUSA increases the corresponding point es- researchers have failed to consider endogeneity in PLS-
timate to greater than one, further indicating that the SEM, with some authors suggesting that endogeneity
switching costs construct is not a suitable IV. cannot be accounted for in a PLS-SEM framework (e.g.,
Antonakis et al. 2010; Rönkkö and Evermann 2013). Such
Following the procedure in Figure 1, we examine the results notions are astonishing because PLS-SEM is a regression-
of the Gaussian copula approach. Specifically, because the based technique, which makes common approaches for
inclusion of CUSA’s significant Gaussian copula in the model identifying and treating endogeneity accessible in this spe-
changes the relationship of CUSA and CUSL from .509 to cific methodological context. Benitez, Henseler, and Roldán
.582, researchers should consider CUSA as endogenous (2016) took a first step in this direction by combining the
(Table 2). This finding is in line with prior research that has standard PLS-SEM algorithm with the 2SLS method;
theoretically and empirically substantiated endogeneity however, they did not consider actual IVs that allow the
problems in studies of customer satisfaction, especially its independent variables’ variance to be decomposed into
relationship to customer loyalty (Ivanov, Joseph, and endogenous and exogenous parts. Broadening the scope of
Wintoki 2013). Nevertheless, this analysis illustrates that, in their research, this study introduces a systematic procedure
the context of our empirical example, considering endoge- for identifying and treating endogeneity in PLS-SEM that
neity does not, from a theory-testing perspective, change any takes into account the most recent statistical developments
conclusions drawn from the model estimation. in the literature on endogeneity. Specifically, implementing
the Gaussian copula approach addresses the significant
To summarize, by carrying out the procedure shown in concerns associated with identifying strong and valid IVs.
Figure 1, we identified a potential endogeneity issue based Moreover, combining this novel approach with the in-
on the Gaussian copula results. However, we could not clusion of control variables and the classic IV approach
explain the problem in light of the available control and enables a comprehensive assessment of endogeneity when
instrumental variables. In this situation, the researcher PLS-SEM is applied for explanatory purposes.
should report the solution found with the Gaussian copula
approach, acknowledging that the findings do not reveal the Our application of the procedure to empirical data reveals
origins of the endogeneity. endogeneity in Eberl’s (2010) corporate reputation model,
which has been frequently used in different variations in
international marketing and international business research
SUMMARY AND CONCLUSIONS (e.g., Swoboda, Puchert, and Morschett 2016; Zhang and
Schwaiger 2012). Our analysis suggests that the CUSA
Although addressing endogeneity has become an integral construct is subject to endogeneity, but control variables or
component of regression-based analyses in international IVs were not sufficient to handle this problem. Correcting
marketing (Jean et al. 2016), users of the PLS-SEM method for endogeneity with the use of the Gaussian copula ap-
proach leads to a pronounced change in the model esti-
7We also ran the control function approach on the model
mates, while leaving the general conclusions intact.
using switching costs as the IV. We used the statistical software
R for the calculation and the boot package to bootstrap the
standard errors (Canty and Ripley 2017). As expected, the We hope that this article will be the starting point for future
results parallel those of the IV approach. PLS-SEM research on endogeneity. Researchers may

14 Journal of International Marketing

 APPENDIX. International Marketing Studies That Apply PLS-SEM

Journal

Journal of International Business

Year International Marketing Review Studies Journal of International Marketing

2017 Gomes et al. (2017); Moon and Oh (2017); Ashraf et al. (2017); Obadia, Vida, and Pla-Barber
Pinho and Thompson (2017); Rahman et al. (2017)
(2017)
2016 Andéhn and L’Espoir Decosta (2016); Henseler Lew et al. (2016)
et al. (2016b); Jean et al. (2016); Richter et al.
(2016); Schlegelmilch et al. (2016)
2015 Rippé et al. (2015); Sinkovics et al. (2015) Khan et al. (2015); Obadia et al. (2015) Griffith and Yanhui (2015); Morgeson et al.
(2015); Sheng et al. (2015)
2014 Freeman and Styles (2014); Griffith et al. (2014) Ashraf et al. (2014); Diamantopoulos et al. (2014)
2013 Sinkovics et al. (2013) Schotter and Beamish (2013) Magnusson et al. (2013); Obadia (2013); Swoboda
and Elsner (2013)
2012 Lam et al. (2012) Hilmersson and Jansson (2012); Johnston et al.
(2012)
2011 Hortinha et al. (2011); Nijssen and Douglas
(2011); Sichtmann et al. (2011)
2010 Duque and Lado (2010) Shi et al. (2010) Navarro et al. (2010); Sichtmann and von
Selasinsky (2010)
2009 Fey et al. (2009) Lages et al. (2009); Nijssen and Herk (2009)
2008 Brettel et al. (2008); Nijssen and Douglas (2008)
Number of articles 15 7 21
(2008–2017)

Addressing Endogeneity in International Marketing Applications 15

endeavor to introduce bootstrapping-based approaches to from the data, and recover the segment-specific model es-
test the strength of an IV, its validity, and its relevance in timates. However, researchers need to assess the interplay
PLS-SEM. Further, the weak ID test and the Hausman test between the various sources of heterogeneity on the one
are parametric tests and therefore violate the nonparametric hand and endogeneity on the other and develop guidelines
nature of PLS-SEM, thus requiring other tests to be iden- to assess whether endogeneity is at a noncritical level in the
tified. Moreover, further research should advance the un- segment-specific solutions.
derstanding and applicability of IV-free approaches, such as
Gaussian copulas or Ebbes et al.’s (2005) LIV method. For Consideration of the fuzzy-set qualitative comparative
example, simulation study results could provide researchers analysis of regression models (Woodside 2013) would be
with thresholds for deciding whether a significant Gaussian particularly interesting. Joint consideration of these tech-
copula entails a negligible endogeneity problem or a sub- niques may enable researchers to better assess and control
stantial one. Future research should also identify ways to for endogeneity embedded in unobserved heterogeneity
parameterize the LIV approach, which we did not consider (Ivanov, Joseph, and Wintoki 2013; Schlittgen et al. 2016).
in our study because of its strong dependence on the number
of latent instrument categories (Papies, Ebbes, and Van Moreover, in this study we did not address how to deal with
Heerde 2017) and lack of software support. Such research endogeneity issues in PLS path models that include in-
would pave the way for a benchmarking study of the teraction terms, which are used, for example, in PLS-SEM-
Gaussian copula and LIV approaches in the context of based moderator analyses (Hair et al. 2017; Henseler and
PLS-SEM. In this regard, a Monte Carlo simulation study, Chin 2010). Because international marketing researchers
in which prespecified levels of endogeneity bias the pa- frequently apply such analyses (e.g., Jin, Zhou, and Wang
rameter estimates, would enable comparison of the IV 2016; Leonidou et al. 2017), future studies should use in-
and IV-free approaches’ efficacy in identifying and treating sights from research on regression analysis (e.g., by using
endogeneity. special 2SLS analyses to account for interaction terms and
IVs or the control function approach; see, for example,
Recent PLS-SEM developments, such as the consistent Sande and Ghosh 2018) to furnish recommendations on
PLS-SEM approach (Dijkstra and Henseler 2015), which al- how to adequately address this issue in PLS-SEM.
lows for mimicking factor-based SEM results, and the in-
troduction of model fit criteria into PLS-SEM (e.g., Henseler Finally, in its standard form PLS-SEM does not support the
et al. 2014), may help researchers achieve specific explanatory estimation of longitudinal or panel data without limitations
research goals. The next logical step is to extend PLS-SEM’s (Roemer 2016). After the PLS-SEM method advances in
capabilities to address endogeneity issues in empirical inter- these directions, assessing the efficacy of other approaches
national marketing research when the consistent PLS-SEM for handling endogeneity, such as the lagged variable ap-
approach is used to estimate models on the grounds of proach (Jean et al. 2016), the cross-lagged correlation
common factor model data (Rigdon, Sarstedt, and Ringle procedure (Tyagi and Wotruba 1993), the Arellano–Bond
2017; Sarstedt et al. 2016). This advancement should be (1991) estimator (Kumar and Pansari 2016), and the
straightforward, in terms of our research findings, because the Granger causality test (Filipescu et al. 2013), would be
consistent PLS-SEM method uses the composite scores ob- particularly interesting. Extending the PLS-SEM algorithm
tained from a standard PLS-SEM-based estimation as input. to accommodate binary dependent variables would provide
an opportunity to consider endogeneity methods specific to
Future research should also explore other approaches to such model settings, such as the control function approach
detect endogeneity problems in PLS-SEM. Research has and its extensions (e.g., Garen 1984; Wooldridge 2015).
brought forward a number of latent-class techniques, such These advances would be particularly beneficial in inter-
as finite mixture PLS (Hair et al. 2018), PLS prediction- national marketing research, which frequently relies on
oriented segmentation (Becker et al. 2013), and PLS iter- longitudinal and panel data (e.g., Kumar and Pansari 2016)
ative reweighted regression segmentation (Schlittgen et al. as well as binary outcome variables (e.g., Obadia, Vida, and
2016), to uncover unobserved heterogeneity. Sarstedt, Pla-Barber 2017; Westjohn, Roschk, and Magnusson 2017).
Ringle, and Hair (2017b) recently introduced a systematic
procedure for identifying and treating unobserved hetero- REFERENCES
geneity in PLS path models. Combining several latent-class Ali, Faizan, S. Mostafa Rasoolimanesh, Marko Sarstedt,
techniques, the procedure enables researchers to check Christian M. Ringle, and Kisang Ryu (2018), “An Assessment
whether unobserved heterogeneity has a critical impact on of the Use of Partial Least Squares Structural Equation
the results, decide on the number of latent segments to retain Modeling (PLS-SEM) in Hospitality Research,” International

16 Journal of International Marketing

 Journal of Contemporary Hospitality Management, 30 (1), Canty, Angelo, and Brian Ripley (2017), “R Package Boot:
514–38. Bootstrap Functions (Version 1.3-20)” (accessed April 5,
2018), https://cran.r-project.org/web/packages/boot/.
Angrist, Joshua D., and J örn-Steffen Pischke (2009), Mostly
Harmless Econometrics: An Empiricist’s Companion. Chin, Wynne W. (1998), “The Partial Least Squares Approach
Princeton, NJ: Princeton University Press. to Structural Equation Modeling,” in Modern Methods for
Business Research, G.A. Marcoulides, ed. Mahwah, NJ:
Antonakis, John, Samuel Bendahan, Philippe Jacquart, and Rafael Lawrence Erlbaum Associates.
Lalive (2010), “On Making Causal Claims: A Review and
Recommendations,” Leadership Quarterly, 21 (6), 1086–120. Clougherty, Joseph A., Tomaso Duso, and Johannes Muck
(2016), “Correcting for Self-Selection Based Endogeneity in
Arellano, Manuel, and Stephen Bond (1991), “Some Tests of Management Research: Review, Recommendations, and Sim-
Specification for Panel Data: Monte Carlo Evidence and an ulations,” Organizational Research Methods, 19 (2), 286–347.
Application to Employment Equation,” Review of Economic
Studies, 58 (2), 277–97. Cui, Annie Peng, Michael F. Walsh, and Shaoming Zou (2014),
“The Importance of Strategic Fit Between Host–Home
Bascle, Guilhem (2008), “Controlling for Endogeneity with Country Similarity and Exploration Exploitation Strategies on
Instrumental Variables in Strategic Management Research,” Small and Medium-Sized Enterprises’ Performance: A Con-
Strategic Organization, 6 (3), 285–327. tingency Perspective,” Journal of International Marketing,
22 (4), 67–85.
Becker, Jan-Michael, Arun Rai, Christian M. Ringle, and
Franziska Völckner (2013), “Discovering Unobserved Hetero- Deng, Ziliang (2016), “Endogeneity in International Marketing
geneity in Structural Equation Models to Avert Validity Threats,” Research, 1995–2014” (accessed March 20, 2018), http://en.
Management Information Systems Quarterly, 37 (3), 665–94. paper.edu.cn/en_releasepaper/content/4687265.

Benitez, Jose, Jörg Henseler, and José L. Roldán (2016), “How Diamantopoulos, Adamantios, Marko Sarstedt, Christoph Fuchs,
to Address Endogeneity in Partial Least Squares Path Mod- Petra Wilczynski, and Sebastian Kaiser (2012), “Guidelines
eling,” paper presented at the 22nd Americas Conference on for Choosing Between Multi-Item and Single-Item Scales
Information Systems (ACIS), San Diego, CA (August 11–14). for Construct Measurement: A Predictive Validity Per-
spective,” Journal of the Academy of Marketing Science,
Berné, Carmen, José M. Múgica, and M. Jesús Yagüe (2001), 40 (3), 434–49.
“The Effect of Variety-Seeking on Customer Retention in
Services,” Journal of Retailing and Consumer Services, 8 (6), Dijkstra, Theo K., and Jörg Henseler (2015), “Consistent Partial
335–45. Least Squares Path Modeling,” Management Information
Systems Quarterly, 39 (2), 297–316.
Bernerth, Jeremy B., and Herman Aguinis (2016), “A Critical
Review and Best-Practice Recommendations for Control Ebbes, Peter, Dominik Papies, and Harald J. van Heerde (2017),
Variable Usage,” Personnel Psychology, 69 (1), 229–83. “Dealing with Endogeneity: A Nontechnical Guide for
Marketing Researchers,” in Handbook of Market Research,
Bijmolt, Tammo H.A., Harald J. van Heerde, and Rik G.M. Pieters C. Homburg, M. Klarmann, and A. Vomberg, eds. Cham, CH:
(2005), “New Empirical Generalizations on the Determinants of Springer International Publishing, 1–14.
Price Elasticity,” Journal of Marketing Research, 42 (2), 141–56.
Ebbes, Peter, Dominik Papies, and Harald J. van Heerde (2011),
Bollen, Kenneth A., Stanislav Kolenikov, and Shawn Bauldry “The Sense and Non-Sense of Holdout Sample Validation in
(2014), “Model-Implied Instrumental Variable—Generalized the Presence of Endogeneity,” Marketing Science, 30 (6),
Method of Moments (MIIV-GMM) Estimators for Latent 1115–22.
Variable Models,” Psychometrika, 79 (1), 20–50.
Ebbes, Peter, Michel Wedel, Ulf Böckenholt, and Ton Steerneman
Bound, John, David A. Jaeger, and Regina M. Baker (1993), (2005), “Solving and Testing for Regressor-Error (In)dependence
“The Cure Can Be Worse Than the Disease: A Cautionary When No Instrumental Variables Are Available: With New
Tale Regarding Instrumental Variables,” NBER Technical Evidence for the Effect of Education on Income,” Quantitative
Working Paper No. 137. Cambridge, MA: National Bureau of Marketing and Economics, 3 (4), 365–92.
Economic Research.
Eberl, Markus (2010), “An Application of PLS in Multi-Group
Bound, John, David A. Jaeger, and Regina M. Baker (1995), Analysis: The Need for Differentiated Corporate-Level
“Problems with Instrumental Variables Estimation: When the Marketing in the Mobile Communications Industry,” in
Correlation Between the Instruments and the Endogeneous Handbook of Partial Least Squares: Concepts, Methods and
Explanatory Variable Is Weak,” Journal of the American Applications, Vincenzo Esposito Vinzi, Wynne W. Chin, Jörg
Statistical Association, 90 (430), 443–50. Henseler, and Huiwen Wang, eds. Berlin: Springer, 487–514.

Addressing Endogeneity in International Marketing Applications 17

Filipescu, Diana A., Shameen Prashantham, Alex Rialp, and Jean, Ruey-Jer “Bryan,” Ziliang Deng, Daekwan Kim, and
Josep Rialp (2013), “Technological Innovation and Exports: Xiaohui Yuan (2016), “Assessing Endogeneity Issues in In-
Unpacking Their Reciprocal Causality,” Journal of Interna- ternational Marketing Research,” International Marketing
tional Marketing, 21 (1), 23–38. Review, 33 (3), 483–512.

Garen, John (1984), “The Returns to Schooling: A Selectivity Jin, Jason Lu, Kevin Zheng Zhou, and Yonggui Wang (2016),
Bias Approach with a Continuous Choice Variable,” Econ- “Exploitation and Exploration in International Joint Ven-
ometrica, 52 (5), 1199–218. tures: Moderating Effects of Partner Control Imbalance and
Product Similarity,” Journal of International Marketing,
Gregor, Shirley (2006), “The Nature of Theory in Information 24 (4), 20–38.
Systems,” Management Information Systems Quarterly,
30 (3), 611–42. Jones, Michael A., David L. Mothersbaugh, and Sharon E. Beatty
(2000), “Switching Barriers and Repurchase Intentions in
Gui, Raluca, Markus Meierer, and Rene Algesheimer (2017), Services,” Journal of Retailing, 76 (2), 259–74.
“R Package REndo: Fitting Linear Models with Endogenous
Regressors Using Latent Instrumental Variables (Version Jöreskog, Karl G. (1978), “Structural Analysis of Covariance
1.2),” (accessed April 5, 2018), https://cran.r-project.org/web/ and Correlation Matrices,” Psychometrika, 43 (4), 443–77.
packages/REndo/.
Kahn, Barbara E. (1995), “Consumer Variety-Seeking Among
Hair, Joseph F., G. Tomas M. Hult, Christian M. Ringle, and Goods and Services: An Integrative Review,” Journal of Re-
Marko Sarstedt (2017), A Primer on Partial Least Squares tailing and Consumer Services, 2 (3), 139–48.
Structural Equation Modeling (PLS-SEM), 2nd ed. Thousand
Oaks, CA: Sage Publications. Kirby, James B., and Kenneth A. Bollen (2009), “Using In-
strumental Variable Tests to Evaluate Model Specification in
Hair, Joseph F., Marko Sarstedt, Christian M. Ringle, and Latent Variable Structural Equation Models,” Sociological
Siegfried P. Gudergan (2018), Advanced Issues in Partial Methodology, 39 (1), 327–55.
Least Squares Structural Equation Modeling (PLS-SEM).
Thousand Oaks, CA: Sage Publications. Kleiber, Christian, and Achim Zeileis (2017), “R Package AER:
Applied Econometrics with R (Version 1.2-5)” (accessed April
Hausman, Jerry A. (1978), “Specification Tests in Econometrics,” 5, 2018), https://cran.r-project.org/web/packages/AER/.
Econometrica, 46 (6), 1251–71.
Klemperer, Paul (1987), “Markets with Consumer Switching
Henseler, Jörg, and Wynne W. Chin (2010), “A Comparison of Costs,” Quarterly Journal of Economics, 102 (2), 375–94.
Approaches for the Analysis of Interaction Effects Between
Latent Variables Using Partial Least Squares Path Modeling,” Kumar, V., and Anita Pansari (2016), “National Culture,
Structural Equation Modeling, 17 (1), 82–109. Economy, and Customer Lifetime Value: Assessing the Rel-
ative Impact of the Drivers of Customer Lifetime Value for a
Henseler, Jörg, Theo K. Dijkstra, Marko Sarstedt, Christian Global Retailer,” Journal of International Marketing, 24 (1),
M. Ringle, Adamantios Diamantopoulos, Detmar W. Straub, 1–21.
et al. (2014), “Common Beliefs and Reality About Partial Least
Squares: Comments on R önkkö & Evermann (2013),” Larcker, David F., and Tjomme O. Rusticus (2010), “On the Use
Organizational Research Methods, 17 (2), 182–209. of Instrumental Variables in Accounting Research,” Journal of
Accounting and Economics, 49 (3), 186–205.
Hult, G. Tomas M., David J. Ketchen, David A. Griffith,
Brian R. Chabowski, Mary K. Hamman, Bernadine Johnson Leonidou, Leonidas C., Bilge Aykol, Thomas A. Fotiadis, and
Dykes, et al. (2008), “An Assessment of the Measurement of Paul Christodoulides (2017), “Antecedents and Consequences
Performance in International Business Research,” Journal of of Infidelity in Cross-Border Business Relationships,” Journal
International Business Studies, 39 (6), 1064–80. of International Marketing, 25 (1), 46–71.

Hult, G. Tomas M., Martin Reimann, and Oliver Schilke Lim, Heejin, Richard Widdows, and Jungkun Park (2006),
(2009), “Worldwide Faculty Perceptions of Marketing Jour- “M-Loyalty: Winning Strategies for Mobile Carriers,” Journal
nals: Rankings, Trends, Comparisons, and Segmentations,” of Consumer Marketing, 23 (4), 208–18.
globalEdge Business Review, 3 (3), 1–23.
Liu, Qing, Thomas Otter, and Greg M. Allenby (2007), “In-
Ivanov, Vladimir, Kissan Joseph, and M. Babajide Wintoki vestigating Endogeneity Bias in Marketing,” Marketing Sci-
(2013), “Disentangling the Market Value of Customer Sat- ence, 26 (5), 642–50.
isfaction: Evidence from Market Reaction to the Unantici-
pated Component of ACSI Announcements,” International Lohmöller, Jan-Bernd (1989), Latent Variable Path Modeling
Journal of Research in Marketing, 30 (2), 168–78. with Partial Least Squares. Heidelberg: Physica.

18 Journal of International Marketing

Mateos-Aparicio, Gregoria (2011), “Partial Least Squares (PLS) Radomir, Lacramioara, and Wilson Alan (2018), “Corporate
Methods: Origins, Evolution, and Application to Social Sci- Reputation: The Importance of Service Quality and Relation-
ences,” Communications in Statistics: Theory and Methods, ship Investment,” in Partial Least Squares Structural Equation
40 (13), 2305–17. Modeling: Recent Advances in Banking and Finance, N.K.
Avkiran and C.M. Ringle, eds. Heidelberg: Springer, 77–123.
McIntosh, Cameron N., Jeffery R. Edwards, and John Antonakis
(2014), “Reflections on Partial Least Squares Path Modeling,” Raithel, Sascha, and Manfred Schwaiger (2015), “The Effects of
Organizational Research Methods, 17 (2), 210–51. Corporate Reputation Perceptions of the General Public on
Shareholder Value,” Strategic Management Journal, 36 (6),
Menon, Satya, and Barbara E. Kahn (1995), “The Impact of 945–56.
Context on Variety Seeking in Product Choices,” Journal of
Consumer Research, 22 (3), 285–95. Ray, Soumya, Sung S. Kim, and James G. Morris (2012),
“Research Note: Online Users’ Switching Costs: Their Nature
Murray, Michael P. (2006a), “Avoiding Invalid Instruments and Formation,” Information Systems Research, 23 (1), 197–213.
and Coping with Weak Instruments,” Journal of Economic
Perspectives, 20 (4), 111–32. Reeb, David, Mariko Sakakibara, and Ishtiaq P. Mahmood
(2012), “From the Editors: Endogeneity in International
Murray, Michael P. (2006b), “The Bad, the Weak, and the Ugly: Business Research,” Journal of International Business Studies,
Avoiding the Pitfalls of Instrumental Variables Estimation,” 43 (3), 211–18.
(accessed March 20, 2018), http://dx.doi.org/10.2139/
ssrn.843185. Richter, Nicole F., Rudolf R. Sinkovics, Christian M. Ringle,
and Christopher Schlägel (2016), “A Critical Look at the Use
Nakata, Cheryl, and Yili Huang (2005), “Progress and Promise: of SEM in International Business Research,” International
The Last Decade of International Marketing Research,” Marketing Review, 33 (3), 376–404.
Journal of Business Research, 58 (5), 611–18.
Rigdon, Edward E., Marko Sarstedt, and Christian M. Ringle
Nitzl, Christian (2016), “The Use of Partial Least Squares (2017), “On Comparing Results from CB-SEM and PLS-SEM:
Structural Equation Modelling (PLS-SEM) in Management Five Perspectives and Five Recommendations,” Marketing
Accounting Research: Directions for Future Theory Devel- ZFP—Journal of Research and Management, 39 (3), 4–16.
opment,” Journal of Accounting Literature, 37 (December),
19–35. Ringle, Christian M., Marko Sarstedt, Rebecca Mitchell,
and Siegfried P. Gudergan (2018), “Partial Least Squares
Obadia, Claude, Irena Vida, and José Pla-Barber (2017), Structural Equation Modeling in HRM Research,” Inter-
“Differential Effects of Bilateral Norms on SMEs’ Export national Journal of Human Resource Management (pub-
Relationships: A Dynamic Perspective,” Journal of Interna- lished electronically January 7), DOI:10.1080/09585192.
tional Marketing, 25 (3), 21–41. 2017.1416655.

Papies, Dominik, Peter Ebbes, and Harald J. van Heerde (2017), Ringle, Christian M., Sven Wende, and Jan-Michael Becker
“Addressing Endogeneity in Marketing Models,” in Ad- (2015), SmartPLS 3. Bönningstedt: SmartPLS.
vanced Methods for Modeling Markets, P.S.H. Leeflang, J.E.
Wieringa, T.H.A. Bijmolt, and K.H. Pauwels, eds. New York: Roemer, Ellen (2016), “A Tutorial on the Use of PLS Path
Springer, 581–627. Modeling in Longitudinal Studies,” Industrial Management &
Data Systems, 116 (9), 1901–21.
Park, Sungho, and Sachin Gupta (2012), “Handling Endoge-
nous Regressors by Joint Estimation Using Copulas,” Mar- Rönkkö, Mikko, and Joerg Evermann (2013), “A Critical Ex-
keting Science, 31 (4), 567–86. amination of Common Beliefs About Partial Least Squares Path
Modeling,” Organizational Research Methods, 16 (3), 425–48.
Petrin, Amil, and Kenneth Train (2010), “A Control Function
Approach to Endogeneity in Consumer Choice Models,” Rossi, Peter E. (2014), “Even the Rich Can Make Themselves
Journal of Marketing Research, 47 (1), 3–13. Poor: A Critical Examination of IV Methods in Marketing
Applications,” Marketing Science, 33 (5), 655–72.
Qiu, Hanqin, Ben Haobin Ye, Billy Bai, and Heng Wang Wei
(2015), “Do the Roles of Switching Barriers on Customer Sande, Jon Bingen, and Mrinal Ghosh (2018), “Endogeneity in
Loyalty Vary for Different Types of Hotels?” International Survey Research,” International Journal of Research in
Journal of Hospitality Management, 46, 89–98. Marketing (published electronically February 3), DOI:
10.1016/j.ijresmar.2018.01.005.
R Core Team (2018), “R: A Language and Environment for
Statistical Computing,” (accessed April 5, 2018), https:// Sarstedt, Marko, Joseph F. Hair, Christian M. Ringle,
www.R-project.org. Kai O. Thiele, and Siegfried P. Gudergan (2016), “Estimation

Addressing Endogeneity in International Marketing Applications 19

 Issues with PLS and CBSEM: Where the Bias Lies!” Journal of StataCorp (2015), Stata Statistical Software: Release 14. College
Business Research, 69 (10), 3998–4010. Station, TX: StataCorp LP.

Sarstedt, Marko, and Erik A. Mooi (2014), A Concise Guide StataCorp (2017), “ivregress: Single-Equation Instrumental-
to Market Research: The Process, Data, and Methods Using Variables Regression,” Technical Report. College Station,
IBM SPSS Statistics. Heidelberg, DE: Springer. TX: StataCorp LP.

Sarstedt, Marko, Christian M. Ringle, and Joseph F. Hair Stock, James H., Jonathan H. Wright, and Motohiro Yogo
(2017a), “Partial Least Squares Structural Equation (2002), “A Survey of Weak Instruments and Weak Identifi-
Modeling,” in Handbook of Market Research, C. Homburg, cation in Generalized Method of Moments,” Journal of
M. Klarmann, and A. Vomberg, eds. Heidelberg, DE: Business & Economic Statistics, 20 (4), 518–29.
Springer.
Swoboda, Bernhard, and Johannes Hirschmann (2016), “Does
Sarstedt, Marko, Christian M. Ringle, and Joseph F. Hair Being Perceived as Global Pay Off? An Analysis of Leading
(2017b), “Treating Unobserved Heterogeneity in PLS-SEM: Foreign and Domestic Multinational Corporations in India, Ja-
A Multi-Method Approach,” in Partial Least Squares Struc- pan, and the United States,” Journal of International Marketing,
tural Equation Modeling: Basic Concepts, Methodological 24 (3), 1–30.
Issues, and Applications, R. Noonan and H. Latan, eds. Hei-
delberg: Springer, 197–217. Swoboda, Bernhard, and Johannes Hirschmann (2017), “Per-
ceptions and Effects of Cross-National Corporate Reputation:
Sarstedt, Marko, Petra Wilczynski, and T.C. Melewar (2013), The Role of Hofstede’s Cultural Value Approach,” Interna-
“Measuring Reputation in Global Markets: A Comparison of tional Marketing Review, 34 (6), 909–44.
Reputation Measures’ Convergent and Criterion Validities,”
Journal of World Business, 48 (3), 329–39. Swoboda, Bernhard, Cathrin Puchert, and Dirk Morschett
(2016), “Explaining the Differing Effects of Corporate Rep-
Schlittgen, Rainer, Christian M. Ringle, Marko Sarstedt, and utation Across Nations: A Multilevel Analysis,” Journal of the
Jan-Michael Becker (2016), “Segmentation of PLS Path Academy of Marketing Science, 44 (4), 454–73.
Models by Iterative Reweighted Regressions,” Journal of
Business Research, 69 (10), 4583–92. Tianjiao, Qiu (2014), “Product Diversification and Market Value
of Large International Firms: A Macroenvironmental Per-
Schwaiger, Manfred (2004), “Components and Parameters of spective,” Journal of International Marketing, 22 (4), 86–107.
Corporate Reputation: An Empirical Study,” Schmalenbach
Business Review, 56 (1), 46–71. Tyagi, Pradeep K., and Thomas R. Wotruba (1993), “An Ex-
ploratory Study of Reverse Causality Relationships Among
Semadeni, Matthew, Michael C. Withers, and S. Trevis Certo Sales Force Turnover Variables,” Journal of the Academy of
(2014), “The Perils of Endogeneity and Instrumental Vari- Marketing Science, 21 (2), 143–53.
ables in Strategy Research: Understanding Through Simula-
tions,” Strategic Management Journal, 35 (7), 1070–79. Villas-Boas, J. Miguel, and Russell S. Winer (1999), “Endo-
geneity in Brand Choice Models,” Management Science,
Shmueli, Galit (2010), “To Explain or to Predict?” Statistical 45 (10), 1324–38.
Science, 25 (3), 289–310.
Walsh, Gianfranco, Sharon E. Beatty, and Edward M.K. Shiu
Shu, Chengli, Jason Lu Jin, and Kevin Zheng Zhou (2017), “A (2009), “The Customer-Based Corporate Reputation Scale:
Contingent View of Partner Coopetition in International Joint Replication and Short Form,” Journal of Business Research,
Ventures,” Journal of International Marketing, 25 (3), 42–60. 62 (10), 924–30.

Souchon, Anne L., Paul Hughes, Andrew M. Farrell, Ekaterina Westjohn, Stanford A., Holger Roschk, and Peter Magnusson
Nemkova, and Jo~ ao S. Oliveira (2016), “Spontaneity and (2017), “Eastern Versus Western Culture Pricing Strategy:
International Marketing Performance,” International Mar- Superstition, Lucky Numbers, and Localization,” Journal of
keting Review, 33 (5), 671–90. International Marketing, 25 (1), 72–90.

Spector, Paul E., and Michael T. Brannick (2010), “Common Wold, Herman O.A. (1982), “Soft Modeling: The Basic Design
Method Issues: An Introduction to the Feature Topic in Or- and Some Extensions,” in Systems Under Indirect Observations:
ganizational Research Methods,” Organizational Research Part II, K.G. Jöreskog and H.O.A. Wold, eds. Amsterdam:
Methods, 13 (3), 403–06. North-Holland.

Staiger, Douglas, and James H. Stock (1997), “Instrumental Woodside, Arch G. (2013), “Moving Beyond Multiple Re-
Variables Regression with Weak Instruments,” Econometrica, gression Analysis to Algorithms: Calling for Adoption of a
65 (3), 557–86. Paradigm Shift from Symmetric to Asymmetric Thinking in

20 Journal of International Marketing

 Data Analysis and Crafting Theory,” Journal of Business Wooldridge, Jeffrey M. (2015), “Control Function Methods in
Research, 66 (4), 463–72. Applied Econometrics,” Journal of Human Resources, 50 (2),
420–45.
Wooldridge, Jeffrey M. (2010), Econometric Analysis of Cross
Section and Panel Data, 2nd ed. Cambridge, MA: MIT Press. Zhang, Yang, and Manfred Schwaiger (2012), “A Comparative
Study of Corporate Reputation between China and Developed
Wooldridge, Jeffrey M. (2013), Introductory Econometrics. A Western Countries,” in Handbook of Research in Interna-
Modern Approach, 5th ed. Mason, OH: South Western, tional Advertising, S. Okazaki, ed. Cheltenham: Edward
Cengage Learning. Elgar Publishing.

Addressing Endogeneity in International Marketing Applications 21