PLS-SEM: THE HOLY GRAIL FOR ADVANCED ANALYSIS - Marketing Management ...

Page created by Bernard Lang
 
CONTINUE READING
PLS-SEM: THE HOLY GRAIL FOR ADVANCED ANALYSIS - Marketing Management ...
PLS-SEM: The Holy Grail for Advanced Analysis                                     Matthews, Hair and Matthews

       PLS-SEM: THE HOLY GRAIL FOR ADVANCED ANALYSIS
                               LUCY MATTHEWS, Middle Tennessee State University
                                  JOE HAIR, The University of South Alabama
                                  RYAN MATTHEWS, RLM Enterprises, LLC

    Advanced analytical techniques are reviewed for researchers wanting to expand their knowledge of
    how partial least squares structural equation modeling (PLS-SEM) facilitates better understanding
    of complex data relationships. We provide a brief overview of the differences between covariance-
    based structural equation modeling (CB-SEM) and PLS-SEM along with guidelines for the
    appropriate application of each. The focus is on mediation, moderation, multi-group analysis, and
    hierarchical component models. We also summarize several emerging analytical tools available with
    PLS-SEM. The ease of applying these advanced analytical techniques in many different research
    contexts makes PLS-SEM the “holy grail” for advanced analysis.

                        INTRODUCTION                         several of the more advanced analytical tools
                                                             available when applying PLS-SEM.
    For many researchers, keeping up with
    advanced methods can seem daunting. Learning             PLS-SEM versus CB-SEM
    new software along with the application and
    interpretation guidelines can sometimes be               When it comes to structural equations modeling
    overwhelming. That is not the case, however,             (SEM), researchers have a choice of two
    with partial least squares structural equation           methods: covariance-based SEM (CB-SEM;
    modeling (PLS-SEM), particularly using                   Jöreskog, 1978, 1993) and variance-based
    SmartPLS 3.0 (Ringle, Wende, & Becker,                   partial least squares (PLS-SEM; Lohmöller,
    2015). The recent rise in popularity of PLS-             1989; Wold, 1982). A fundamental distinction
    SEM can be attributed, at least in part, to the          between the two approaches is that CB-SEM is
    ease of understanding and applying the basic             based on the common factor model, while PLS-
    analytical tools of the method (Hair, Ringle, &          SEM is based on the composite factor model
    Sarstedt, 2011). But learning to apply advanced          (Hair, Hult, Ringle, & Sarstedt, 2017). With
    methods such as mediation, moderation, multi-            common factor models, the analysis is based
    group analysis and more, is also relatively              only on the common variance in the data.
    straightforward.                                         Therefore, the solution begins by calculating
                                                             the covariances between the variables in the
    Most researchers are at least somewhat familiar          study and only that common variance is used in
    with covariance-based structural modeling (CB-           the analysis (Hair, Matthews, Matthews, &
    SEM), most often run with the AMOS or                    Sarstedt, 2018; Sarstedt, Hair, Ringle, Theile, &
    LISREL software. Few researchers are aware               Gudergan, 2016). With the composite factor
    of and understand the fundamentals of variance           model the constructs and their scores are
    -based structural modeling (PLS-SEM). The                represented by the total variance in the
    purpose of this paper is to introduce and                indicators, not just common variance that is the
    provide an overview of the rapidly emerging              case with CB-SEM (Hair, Hult, Ringle, &
    method of variance-based structural equation             Sarstedt, 2017). In addition, the statistical
    modeling. In this paper, we first explain the            objectives are substantially different between
    differences in variance-based structural                 the two methods.          Using CB-SEM, the
    equation modeling (PLS-SEM) and the                      statistical objective is to estimate model
    covariance-based CB-SEM method, and                      parameters that minimize the differences
    therefore the rationale for the selection of one         between the observed sample covariance
    approach over another. We then summarize                 matrix, which is calculated before the
    The Marketing Management Journal                         theoretical model solution is obtained, with the
    Volume 28, Issue 1, Pages 1-13
    Copyright © 2018, The Marketing Management Association
                                                             covariance matrix that is estimated after the
    All rights of reproduction in any form reserved          theoretical model solution is obtained (Hair,
                                                             Sarstedt, Ringle, & Mena, 2012). If goodness of
1                                                                  Marketing Management Journal, Spring 2018
PLS-SEM: THE HOLY GRAIL FOR ADVANCED ANALYSIS - Marketing Management ...
PLS-SEM: The Holy Grail for Advanced Analysis                              Matthews, Hair and Matthews

 fit is (GOF) demonstrated, the theoretical          order models, multi-group analysis, invariance
 structural model is confirmed. But if GOF is        and unobserved heterogeneity, making PLS-
 not possible the model is not confirmed. In         SEM a “go to” methodology for researchers.
 contrast, when using PLS-SEM, the statistical
 objective is to maximize the variance explained     Mediation
 in the dependent variable(s) (Hair, Sarstedt,
 Pieper, & Ringle, 2012a). Thus, the focus of        When a third variable, called a mediator,
 PLS is on optimizing prediction of the              intervenes between two other variables, the
 endogenous constructs and not on fit, which is      opportunity arises to examine mediation (Baron
 the focus of CB-SEM. Moreover, PLS-SEM is           & Kenny, 1986). Specifically, a change in the
 a variance-based approach and the analysis          exogenous construct produces a change in the
 does not start or end with a covariance matrix.     mediator variable, which then produces a
 Thus, a Chi-square type of GOF is not possible.     change in the endogenous construct, and the
                                                     mediator variable dictates the nature of the
 Determining when the application of each of         relationship between the exogenous and
 the methods is appropriate is straightforward. If   endogenous constructs (Hair, Hult, et al., 2017).
 the focus of the research is theory testing and     A crucial prerequisite for investigating
 confirmation (Sarstedt, Ringle, Henseler, &         mediating effects is strong apriori theoretical
 Hair, 2014), then CB-SEM is the appropriate         support (Preacher & Hayes, 2008).
 method. But if prediction, theory development
 and explanation are the focus of the research,      The foundation for mediation is a well-
 then PLS-SEM is the more appropriate method.        established theoretical relationship (c) between
 PLS-SEM is somewhat similar, both                   an exogenous (Y1) construct and an endogenous
 conceptually and practically, to using multiple     (Y3) construct (Figure 1) (Preacher & Hayes,
 regression analysis (Hair et al., 2011). But        2008). Testing for mediation in a model
 unlike multiple regression, PLS-SEM can be          requires a series of analyses beginning with
 applied to better understanding more complex        testing the significance of the indirect effect (a
 structural measurement and path models. At the      b) via the mediator variable (Y2) as seen in
 same time, PLS-SEM and CB-SEM are both              Figure 2. If the indirect effect is not significant,
 appropriate for metric data and reflective          then Y2 is not operating as a mediator in the
 measurement models. But PLS-SEM can easily          relationship. However, if (a b) is significant,
 be used with formative measurement models,          then the next test is to check the direct effect in
 non-metric data (e.g., ordinal & nominal),          the mediated model (c’). If c’ is not significant,
 continuous moderators, higher order models,         indirect-only (full) mediation has occurred.
 when latent variable scores are needed for          This occurs when the indirect effect is
 further analysis, and with small sample sizes (N    significant, but not the direct effect in the
 ≤ 100) as well as large samples (Hair,              mediated model.         Alternatively, if c’ is
 Hollingsworth, Randolph, & Chong, 2017).            significant, then partial mediation has occurred.
 Because it is nonparametric, PLS-SEM also has       When (a b c’) is positive, then complementary
 a wider application and greater flexibility in      mediation has occurred, but if (a b c’) is
 handling various modeling situations where it is    negative then competitive mediation has
 difficult to meet rigorous assumptions, such as     occurred (Hair, Hult, et al., 2017).
 a normal distribution and homoscedasticity, that
 are typically required with more traditional
                                                                     FIGURE 1:
 multivariate statistics (Vinzi, Chin, Henseler, &            Direct Effect of Exogenous
 Wang, 2010).          Therefore, PLS-SEM is               Variable on Endogenous Variable
 appropriate for exploratory research, theory
 development, and prediction.          It can be
                                                                           c
 executed on both small and large samples sizes,
 with reflective or formative measurement
 models, and does not assume the data has a               Y1                                 Y3
 normal distribution. Finally, the method can be
 used with metric and non-metric data,
 continuous moderators, secondary data, higher

Marketing Management Journal, Spring 2018                                                                   2
PLS-SEM: The Holy Grail for Advanced Analysis                                  Matthews, Hair and Matthews

                     FIGURE 2:                           complementary and competitive mediation
          Indirect Effect - Mediation Model              (Hair, Hult, et al., 2017). With complementary
                                                         mediation, the mediated effect (a b) and direct
                                                         effect (c’) both exist and point in the same
                                                         direction (i.e., the signs are either both positive
                           Y2
                                                         or both negative), while with competitive
               a                         b               mediation, the mediated effect (a b) and direct
                                                         effect (c’) are both present and point in opposite
                                                         directions (i.e., a sign for one relationship is
                                                         positive and the other relationship is negative)
                                c’                       (Zhao, Lynch, & Chen, 2010).
          Y1                                 Y3
                                                         The most common types of mediation are
                                                         simple mediation analysis and multiple
                                                         mediation analysis. Simple mediation analysis
                                                         is when one mediator variable is specified in
    Mediation has traditionally been executed using      the structural model. Often times, however,
    multiple regression. Baron and Kenny (1986)          exogenous constructs influence endogenous
    and more recently the PROCESS approach by            constructs through more than one mediating
    Preacher and Hayes (2008) both focus on              variable requiring multiple mediation analyses
    multiple regression and examine significance in      (Hair, Hult, et al., 2017). The mediators reveal
    mediation using the Sobel test, which assumes        the “real” relationship among the exogenous
    the data are normally distributed. The               and endogenous constructs. Figure 3 depicts a
    advantages of using PLS-SEM for mediation            model with two mediating variables, Y2 and Y4.
    are that bootstrapping makes no assumptions          The direct effect is measured by c’. But the
    about the shape of the variables’ distribution or    indirect effect of Y1 on Y3 now includes the Y4
    the sampling distribution of the statistics, and     mediator (d e) in addition to Y2, and the total
    all the mediated relationships are tested            indirect effect of Y1 on Y3 is measured by the
    simultaneously instead of separately, which          sum of the two indirect effects (i.e., a b + d e).
    reduces bias (Hair, Sarstedt, Hopkins, &             Therefore, the total effect of Y1 on Y3 is the
    Kuppelwieser, 2014).        Finally, mediation       sum of the direct effect and the total indirect
    testing using PLS-SEM can be applied with            effect (i.e., c’ + a b + d e).
    smaller sample sizes while yielding higher
    levels of statistical power compared to prior        Analyzing all mediators concurrently allows for
    testing methods like the parametric Sobel            a more thorough understanding of the overall
    (1982) test.                                         effect. If each mediator were analyzed in a
                                                         simple mediation analysis (i.e., with a
    Path models that include a mediator are still        regression model where all relationships are
    required to meet the quality criteria of the         tested separately), the indirect effect would
    measurement        models.      For    formative     likely be overstated due to the correlation of
    measurement models, convergent validity              one mediator to another (Hair, Hult, et al.,
    (redundancy), collinearity between indicators,       2017). When PLS-SEM is used, multiple
    and significance/relevance of outer weights are      mediation in which all relationships (direct and
    required (Hair, Hult, et al., 2017).           For   indirect) are tested simultaneously is possible.
    reflective measurement models, the quality           Thus, with multiple mediation the impact of
    criteria include internal consistency reliability,   one or multiple mediators can be tested
    convergent validity, and discriminant validity.      simultaneously, eliminating the overstatement
    It is important to also confirm that collinearity    of the correlation associated with each
    in the structural model is not at a critical level   mediator.
    since biased path coefficients may incur. When
    high collinearity exists, the direct effect may      The steps for processing multiple mediation
    suggest nonmediation via nonsignificance or an       analyses are the same as in simple mediation
    unexpected sign change may result in an              analysis. The testing process begins with
    erroneous         differentiation        between     examining the significance of each indirect

3                                                              Marketing Management Journal, Spring 2018
PLS-SEM: The Holy Grail for Advanced Analysis                               Matthews, Hair and Matthews

 effect, and then the direct effect between the       When including a moderator in the model, the
 exogenous and endogenous constructs. To              variable will appear twice, once as the variable
 determine the total indirect effect manual           itself (main effect) and again as the interaction
 calculations of the standard error for each          effect (a combination of the main effect and the
 specific indirect effect may be necessary.           moderator; see Figure 4). Unlike mediation
 Using SmartPLS 3.0 (Ringle et al., 2015),            where the exogenous construct acts as an
 however, this can be accomplished by simply          antecedent to the mediator, in moderation the
 obtaining the indirect effects results of the        moderator variable and exogenous construct
 bootstrapping routine and using spreadsheet          interact (Y1*M) at the same level to impact the
 software such as Microsoft Excel. Finally, to        endogenous variable. This is a multiplicative
 calculate the t value of the specific indirect       relationship.
 effect, divide the specific indirect effect by the
 standard error (Hair, Hult, et al., 2017).                           FIGURE 4:
                                                                    Moderation Model
                     FIGURE 3:
              Multiple Mediation Model
                                                                                         Y1 M
                         Y2
                                                                      M
                                                                                                 c
          a                          b                                            b

                                                                              a
                         c’
                                                           Y1                                 Y2
     Y1                                  Y3

          d                          e                While several analytical procedures exist for
                                                      estimating the measurement model with
                                                      moderation (e.g., product indicator approach,
                                                      orthogonalizing approach, and two-stage
                         Y4                           approach), the two-stage approach is typically
                                                      recommended. The two-stage approach is able
                                                      to handle both reflective and formative
 Moderation                                           moderators and additional exogenous constructs
                                                      in the path model. Moreover, compared to the
 Moderation (interaction effect) occurs when the      other approaches, the two-stage approach
 relationship between two constructs varies           exhibits higher levels of statistical power (Hair,
 depending on a third (moderator) variable            Hult, et al., 2017).
 (Henseler & Chin, 2010). The variation can
 influence the strength or direction of the           The two-stage approach begins by running the
 relationship (Baron & Kenny, 1986).                  main effects model without the moderation
 Moderator variables can be categorical (e.g.,        interaction term in the model to estimate the
 age, income, gender) and tested by means of          latent variable scores (Henseler & Chin, 2010).
 group comparisons using either PLS-SEM or            In the second stage, the latent variable scores
 CB-SEM.        Alternatively, with PLS-SEM           from stage one of the exogenous latent variable
 moderators can also be a continuous variable         and the moderator variables are multiplied to
 (e.g., attitudes about satisfaction, loyalty,        create a single-item measure to represent the
 commitment, brand passion) typically measured        interaction term (Hair, Hult, et al., 2017). At the
 using multi-item scales (note that continuous        same time, all the other latent variables are
 moderators cannot be used with CB-SEM).              represented by a single item measure (latent

Marketing Management Journal, Spring 2018                                                                   4
PLS-SEM: The Holy Grail for Advanced Analysis                                   Matthews, Hair and Matthews

    variable score) that was calculated in stage one.     f2 effect size is examined. The f2 measures the
    The moderator hypothesis is supported if the          extent to which the endogenous latent variable
    interaction effect (c) is significant (Hair,          is explained by the moderation. The f2 effect
    Sarstedt, Ringle, & Gudergan, 2018).                  sizes of 0.02, 0.15, and 0.35 suggest small,
                                                          medium, and large effect sizes, respectively
    The results indicate that the value of c              (Cohen, 1988).
    (interaction effect) represents the strength of the
    relationship between Y1 and Y2 when the               Interpreting and drawing conclusions from the
    moderator variable M has a value of zero (Hair,       moderation results can be difficult. Slope plots
    Hult, et al., 2017). However, since many scales       are typically used as a visual illustration to gain
    either do not include a value of zero or a value      a better understanding of the moderation effect.
    of zero does not make sense, standardization is       Figure 5 displays a two-way interaction of the
    often necessary. Standardization facilitates          relationship between Y1 and Y2. The horizontal
    interpretation as well as reduces collinearity        x-axis represents the exogenous construct (Y1)
    among the exogenous construct, the moderator,         and the vertical y-axis represents the
    and the interaction term. To standardize, the         endogenous construct (Y2). The two lines
    variable’s mean is subtracted from each               illustrate the relationship between Y1 and Y2 for
    observation and divided by the variable’s             both low and high levels of the moderator
    standard error (Sarstedt & Mooi, 2014). When          construct (M). The low level of M (solid line)
    using SmartPLS 3, the software executes many          is one standard deviation unit below the
    types moderation, standardizes when necessary,        average, while the high level of M (dotted line)
    and produces a simple slope analysis for              is one standard deviation unit above the
    interpreting moderation results.                      average. There is a negative moderating effect
                                                          of -0.80 between the interaction term and the
    To assess the impact on the R2 value when the         endogenous construct. The high moderator’s
    interaction effect is omitted from the model, the     slope is relatively flat but decreases slightly as

                                                 FIGURE 5:
                                 Graphical Illustration of Moderation Effect

5                                                               Marketing Management Journal, Spring 2018
PLS-SEM: The Holy Grail for Advanced Analysis                              Matthews, Hair and Matthews

 the exogenous variable changes from low to           enough (See Hair, Hult, et al., 2014).
 high. Thus, the relationship between Y1 and Y2       Additionally, the subgroups should be similar
 becomes weaker with high levels of the               in size to avoid introducing error (Becker, Rai,
 moderator construct. But for low levels of the       Ringle, & Volckner, 2013).This involves
 moderator variable, the slope is quite steep and     coding the master data file into subgroups that
 the relationship between Y1 and Y2 becomes           can be executed with PLS-SEM (Matthews,
 much stronger with high levels of M. To              2017).
 facilitate interpretation of the interaction,
 SmartPLS 3 computes and the output displays a        While a number of approaches can be used to
 simple slope plot.                                   compare the path coefficients of the group
                                                      SEMs, most researchers recommend the
 Multi-Group                                          permutation test (Hair, Sarstedt, Ringle, &
                                                      Gudergan, 2018).        Permutation is non-
 Multi-group analysis (MGA) or between-group          parametric and more conservative than the
 analysis is a means of testing apriori defined       parametric test, and controls well for Type 1
 groups to determine if there are significant         error (Henseler, Ringle, & Sarstedt, 2016). To
 differences in group-specific parameter              execute the permutation test, the correlations
 estimates (e.g., outer weights, outer loadings       between the composite scores using the weights
 and path coefficients) obtained when using PLS       obtained from the first group are computed
 -SEM (Hair, Hult, Ringle, & Sarstedt, 2014;          against the composite scores using the weights
 Henseler & Chin, 2010). By applying MGA,             obtained from the second group during each
 researchers are able to test for differences         permutation run (Henseler et al., 2016).
 between two identical models for different
 apriori specified groups within the data set. In     The focus of multi-group analysis is to examine
 contrast to standard approaches to testing           the path coefficients of the theoretical models
 moderation, which examine a single structural        for the two groups to determine if they are
 relationship at a time, MGA via PLS-SEM is an        significantly different. This begins by running
 efficient way to assess moderation across            the model for each group separately, using the
 multiple relationships (Hair, Sarstedt, Ringle, et   guidelines set out for evaluation of a
 al., 2012).                                          measurement model (Hair, Hult, et al., 2014,
                                                      2017). This determines if there are group
 This type of analysis enables researchers to         specific differences. Then, it is necessary to
 identify differences between the structural paths    determine if the difference between the two
 of multiple groups. For example, Matthews            groups is significant, which can be
 (2017) illustrated how skill discrepancy             accomplished by running the Permutation Test.
 partially mediates the relationship between          A permutation p-value of less than or equal to
 autonomy and cognitive engagement for male           0.10 indicates a significant difference between
 salespeople, but not for female salespeople.         the two groups being compared (Matthews,
 PLS-MGA also facilitates a more accurate and         2017).
 comprehensive assessment of group differences
 and strategy implementation based on more            MGA allows researchers to determine
 specific outcomes for the heterogeneous groups       significant differences among observed
 in the data (Matthews, 2017). Finally, the           characteristics. These differences may not be
 differences identified can be used to highlight      evident in aggregate data since significant
 the potential error if these subpopulations are      positive and negative group-specific results
 considered a single homogeneous group                may offset one another resulting in non-
 (Schlagel & Sarstedt, 2016).                         significant relationships (Hair, Hult, et al.,
                                                      2017). Although the path coefficients for the
 The first step in conducting MGA involves            subdivided groups will often display numerical
 generating data groups that are based on the         differences, MGA assists in identifying when
 categorical variable of interest (e.g., gender,      those differences are statistically significant.
 country of origin, age, or income). Once the
 data is subdivided, it is necessary to ensure that
 the sample sizes of the new subgroups are large

Marketing Management Journal, Spring 2018                                                                6
PLS-SEM: The Holy Grail for Advanced Analysis                                Matthews, Hair and Matthews

    Hierarchical Component Models                        structural model, making the model more
                                                         parsimonious and easier to understand. In
    Hierarchical component models (HCM), or              addition, introducing a HCM into a structural
    higher order models, involve testing                 model can reduce multicollinearity among first-
    measurement structures that contain two layers       order constructs, or formative indicators that
    of constructs, the higher order component            exhibit high levels of collinearity. In either
    (HOC) of the model and the lower order               situation, the use of HCMs should be supported
    component (LOC). An example of a HCM is              by theory. Note that higher order models can
    when lower order components of job                   also be used with CB-SEM, but the
    satisfaction are specified as multi-faceted and      assumptions are much more restrictive and
    the higher order component is a single overall       therefore limit their application with that
    construct of job satisfaction. Figure 6 displays a   method.
    theoretical HCM for job satisfaction that
    includes the LOCs (lower order components)           There are four main types of HCMs (Jarvis,
    and the HOC (higher order component). In the         MacKenzie, & Podsakoff, 2003; Wetzels,
    figure, the LOCs represent the first-order multi-    Odekerken-Schroder, & van Oppen, 2009) used
    item constructs for job satisfaction, and the        in PLS-SEM models (Ringle, Sarstedt, &
    HOC is the second-order overall (combined)           Straub, 2012). The HCM model begins with the
    construct for job satisfaction. Researchers may      lower-order components (LOCs), which are
    find the use of a HCM helpful when trying to         used to make up the higher-order component
    reduce the number of relationships in the            (HOC) (Hair, Hult, et al., 2017). Each model is

                                            FIGURE 6:
                    Hierarchical Component Model for Multi-faceted Job Satisfaction

7                                                              Marketing Management Journal, Spring 2018
PLS-SEM: The Holy Grail for Advanced Analysis                             Matthews, Hair and Matthews

 characterized by the different relationships        the underlying components LOC1, LOC2, and
 between the LOCs and the HOC as well as the         LOC3 in the measurement model. However,
 indicators with each construct. The first type is   some issues arise using the repeated indicator
 the Reflective-Reflective, where the indicator      approach when the model is formative-
 measures for the first order components are         formative (type 4) or reflective-formative
 reflective and the measures from the LOC to         (type 2). In this situation, when the relationship
 the HOC are reflective (Figure 7). Type two is      from the LOC to the HOC is formative, almost
 Reflective-Formative. For this model, the LOC       all of the HOC variance is explained by the
 indicators are reflective but the LOC to the        LOC (R2 close to 1.0). This can be an issue if
 HOC is formative. Type three is Formative-          there are other relationships pointing to the
 Reflective, such that the first-order indicators    HOC, as they will have a very small and
 are measured formatively and the HOC from           insignificant impact. Therefore, for type 2 and
 the LOC is reflective. The final type (type         type 4 models, a two-stage HCM analysis
 four), is Formative-Formative where the             should be used (Hair, Hult, et al., 2017).
 indicators of the first-order are measured          Similar to the two-stage approach for
 formatively and the measures from the LOC to        moderation, in stage one the repeated indicator
 the HOC are formative.                              approach is used to obtain the latent variable
                                                     scores for the LOCs. Then in the second stage,
 When creating the HOC in PLS-SEM, all the           the LOC constructs and first-order indicators
 indicators from the LOC are assigned to the         used for the HOC are replaced with latent
 HOC using a repeated indicators approach            variable scores for the LOC from stage one
 (Hair, Hult, et al., 2017). Therefore, the          (Figure 8). The two-stage HCM analysis
 indicators for the HOC x 1 to x 9 are the same as

                                           FIGURE 7:
                          Four Types of Hierarchical Component Models

Marketing Management Journal, Spring 2018                                                                 8
PLS-SEM: The Holy Grail for Advanced Analysis                              Matthews, Hair and Matthews

    allows other latent variables outside of the      understanding and explaining the outcomes of
    HOC to explain some of the variance.              this technique (e.g., Becker, Klein, & Wetzels,
                                                      2012; Kuppelwieser & Sarstedt, 2014; Ringle et
    When using HCM it is important that a similar     al., 2012).
    number of indicators are used for all the LOCs.
    Otherwise, the relationship between the HOC       Other Advanced Topics
    and the LOC can be biased due to the
    disproportionate number of indicators (Hair,      In addition to the topics addressed above,
    Hult, et al., 2017). Note that the number of      researchers have further opportunities to
    indicators on the LOCs does not have to be        improve their analysis and understanding of
    equal (as shown in Figures 6 & 7), but should     theoretical relationships. Measurement model
    be comparable. Additionally, for the inner PLS    invariance, which tests datasets for differences
    path model, not all algorithmic weighting         in measurement model estimates, is a useful
    schemes apply when estimating HCMs in PLS-        tool that should be combined with multigroup
    SEM. In particular, the centroid weighting        analysis.    By employing the measurement
    scheme should not be used (Hair, Sarstedt,        invariance of composite models (MICOM)
    Ringle, et al., 2012). Prior research using and   procedure (Henseler et al., 2016), configural
    explaining HCM models can further assist in       and compositional invariance can be

                                            FIGURE 8:
                                Two-Stage Approach for HCM Analysis

9                                                           Marketing Management Journal, Spring 2018
PLS-SEM: The Holy Grail for Advanced Analysis                               Matthews, Hair and Matthews

 established. Doing so ensures that variations in      Summary
 the path relationships between latent variables
 is a result of the true differences in the            When analyzing research that requires
 structural relationships, and is not the result of    advanced analytical approaches, it is important
 different meanings in the groups’ responses           to understand the differences between CB-SEM
 attributed to the phenomena being measured            and PLS-SEM, as well as other multivariate
 (Hair, Hult, et al., 2017; Henseler et al., 2016).    analysis methods. Because PLS-SEM has a
 Failure to establish data equivalence using           much greater capacity for handling a variety of
 MICOM may potentially result in measurement           modeling issues and does not impose restrictive
 error and thus misleading results (Hult et al.,       assumptions (Vinzi et al., 2010), the use of PLS
 2008), reduce the overall power of the                -SEM is highlighted. For mediation, the
 statistical tests, and influence the precision of     advantage of using PLS-SEM is the lack of
 the estimators (Hair, Hult, et al., 2017).            restrictive distribution assumptions, the
                                                       flexibility to execute with both formative and
 The importance-performance map analysis               reflective measurement models, and the ability
 (IPMA), or importance-performance matrix              to yield higher levels of statistical power with
 analysis, displays the structural model total         smaller sample sizes, while overcoming the
 effects on a specific endogenous construct. The       limitations of multiple regression approaches
 total effects of the predecessor variables are        (Hair, Sarstedt, et al., 2014). The two-stage
 used to assess each exogenous construct’s             approach for moderation using PLS-SEM
 importance in shaping the endogenous                  exhibits high levels of statistical power and is
 construct. The average latent variable scores of      capable of handling both reflective and
 the exogenous constructs measure their                formative moderators when the structural
 performance (Hair, Hult, et al., 2017) using a        model includes other exogenous constructs.
 rescaling technique (Höck, Ringle, & Sarstedt,        Multi-group analysis via permutation in PLS-
 2010). Combined, researchers can identify             SEM enables researchers to easily identify
 constructs with relatively high importance            heterogeneous groups in the data to more
 (strong total effect) and low performance (low        accurately assess group differences (Matthews,
 average latent variable scores) as areas for          2017). Finally, higher order component models
 further research. IPMA can be conducted at the        (HCMs) can be applied with PLS-SEM to
 indicator level as well to identify and improve       obtain more accurate solutions for structural
 upon those indicators that are most relevant.         models exhibiting high multicollinearity.

 Finally, rather than using a priori characteristics   Beyond these advanced analysis approaches,
 to partition datasets into groups, as was             PLS-SEM can: (1) establish data equivalence
 described in multi-group analysis, tools like         via the three stage MICOM process (Henseler
 finite mixture PLS (FIMIX-PLS, Sarstedt,              et al., 2016) to minimize measurement error, (2)
 Becker, Ringle, & Schwaiger, 2011) or                 identify the importance and performance of
 prediction-oriented segmentation (FIMIX-POS,          antecedent constructs to target areas for further
 Sarstedt, Ringle, & Hair, 2017) can be used to        research (Hair, Hult, et al., 2017), and (3)
 uncover unobserved heterogeneity.            Since    uncover unobserved heterogeneity so structural
 sources of heterogeneity in the data aren’t           and measurement models can be examined
 always known a priori, identifying and treating       either at the individual group level or the
 unobserved heterogeneity allows researchers to        aggregate level (Matthews et al., 2016; Hair et
 feel confident about analyzing data at an             al. 2016). Therefore, when facing complex
 aggregate level (Hair, Sarstedt, Matthews, &          research models, even though there are a
 Ringle, 2016). Examples of FIMIX-PLS are              variety of multivariate methods available, the
 available to aid researchers in the application to    numerous flexible analysis options, limited
 their own dataset (Matthews, Sarstedt, Hair, &        assumptions, and user-friendliness of PLS-SEM
 Ringle, 2016; Sarstedt, Schwaiger, & Ringle,          make it the “holy grail” for advanced methods.
 2009). Failure to consider heterogeneity may
 also result in invalid outcomes (Becker et al.,
 2013).

Marketing Management Journal, Spring 2018                                                              10
PLS-SEM: The Holy Grail for Advanced Analysis                            Matthews, Hair and Matthews

                REFERENCES                           Hair, J. F., Sarstedt, M., Matthews, L. M., &
                                                       Ringle, C. M. (2016). Identifying and
 Baron, R. M., & Kenny, D. A. (1986). The              Treating Unobserved Heterogeneity with
  moderator-mediator variable distinction in           FIMIX-PLS: Part I - Method. European
  social psychological research: Conceptual,           Business Review, 28(1) 63-76.
  strategic, and statistical considerations.         Hair, J. F., Sarstedt, M., Pieper, T. M., &
  Journal of Personality and Social                    Ringle, C. M. (2012). The Use of Partial
  Psychology, 51(6), 1173-1182.                        Least Squares Structural Equation Modeling
 Becker, J.-M., Klein, K., & Wetzels, M. (2012).       in Strategic Management Research: A
  Formative Hierarchical Latent Variable               Review       of     Past    Practices     and
  Models in PLS-SEM: Recommendations and               Recommendations for Future Applications.
  Guidelines. Long Range Planning, 45(5-6),            Long Range Planning, 45(5-6), 320-340.
  359-394.                                           Hair, J. F., Sarstedt, M., Ringle, C. M., &
 Becker, J.-M., Rai, A., Ringle, C. M., &              Gudergan, S. P. (2018). Advanced Issues in
  Volckner, F. (2013). Discovering unobserved          Partial Least Squares Structural Equations
  heterogeneity in structural equation models to       Modeling (PLS-SEM). Thousand Oaks, CA,
  avert validity threats. MIS Quarterly, 37(3),        Sage.
  665-694.                                           Hair, J. F., Sarstedt, M., Ringle, C. M., &
 Cohen, J. (1988). Statistical Power A nalysis for     Mena, J. A. (2012). An Assessment of the use
  the Behavioral Sciences. Mahwah, NJ:                 of Partial Least Squares Structural Equation
  Lawrence Erlbaum.                                    Modeling in Marketing Research. Journal of
 Hair, J. F., Hollingsworth, C. L., Randolph, A.       the Academy of Marketing Science, 40(3),
  B., & Chong, A. (2017). An Updated and               414-433.
  Expanded Assessment of PLS-SEM in                  Henseler, J., & Chin, W. W. (2010). A
  Information Systems Research. Industrial             Comparison of Approaches for the Analysis
  Management & Data Systems (3), 442-458.              of Interaction Effects Between Latent
 Hair, J. F., Hult, G. T. M., Ringle, C. M., &         Variables Using Partial Least Squares Path
  Sarstedt, M. (2014). A Primer on Partial             Modeling. Structural Equation Modeling: A
  Least Squares Structural Equation Modeling           Multidisciplinary Journal, 17(1), 82-109.
  (PLS-SEM). Los Angeles, London, New                Henseler, J., Ringle, C. M., & Sarstedt, M.
  Delhi, Singapore, Washington DC: SAGE.               (2016). Testing Measurement Invariance of
 Hair, J. F., Hult, G. T. M., Ringle, C. M., &         Composites Using Partial Least Squares.
  Sarstedt, M. (2017). A Primer on Partial             International Marketing Review, 33(3),
  Least Squares Structural Equation Modeling           405-431.
  (PLS-SEM) (2nd ed.). Los Angeles, London,          Höck, C., Ringle, C. M., & Sarstedt, M. (2010).
  New Delhi, Singapore, Washington DC:                 Management of Multi-purpose Stadiums:
  SAGE.                                                Importance and Performance Measurement of
 Hair, J. F., Matthews, L. M., Matthews, R. L.,        Service Interfaces. International Journal of
  & Sarstedt, M. (2017). PLS-SEM or CB-                Services Technology and Management, 14(2-
  SEM: Updated Guidelines on Which Method              3), 188-207.
  to Use. International Journal of Multivariate      Hult, G. T. M., Ketchen Jr, D. J., Griffith, D.
  Data Analysis,1(2), 107-123.                         A., Finnegan, C. A., Gonzalez-Padron, T.,
 Hair, J. F., Ringle, C. M., & Sarstedt, M.            Harmancioglu, N., & et al. (2008). Data
  (2011). PLS-SEM: Indeed a Silver Bullet.             Equivalence in Cross-Cultural International
  The Journal of Marketing Theory and                  Business     Research:     Assessment     and
  Practice, 19(2), 139-152.                            Guidelines. Journal of International Business
 Hair, J. F., Sarstedt, M., Hopkins, L., &             Studies, 39(6), 1027-1044.
  Kuppelwieser, V. G. (2014). Partial least          Jarvis, C. B., MacKenzie, S. B., & Podsakoff,
  squares structural equation modeling (PLS-           P. M. (2003). A Critical Review of Construct
  SEM): An emerging tool in business                   Indicators    and     Measurement      Model
  research. European Business Review, 26(2),           Misspecification in Marketing and Consumer
  106-121.                                             Research. Journal of Consumer Research, 30
                                                       (2), 199-218.

11                                                         Marketing Management Journal, Spring 2018
PLS-SEM: The Holy Grail for Advanced Analysis                              Matthews, Hair and Matthews

 Jöreskog, K. G. (1978). Structural Analysis of      Sarstedt, M., Ringle, C. M., & Hair, J. F. (2017).
   Covariance and Correlation Matrices.                Treating Unobserved Heterogeneity in PLS-
   Psychometrika, 43(4), 443-477.                      SEM: A Multi-method Approach. In H. Latan
 Jöreskog, K. G. (1993). Testing Structural            & R. Noonan (Eds.), Partial Least Squares
   Equation Models. In K. A. Bollen & J. S.            Structural Equation Modeling - Basic
   Long (Eds.), Testing Structural Equation            Concepts,     Methodological     Issues     and
   Models (pp. 294-316). Newbury Park: Sage.           Applications. Berlin: Springer International
 Kuppelwieser, V., & Sarstedt, M. (2014).              Publishing AG, 197-213.
   Applying the Future Time Perspective Scale        Sarstedt, M., Ringle, C. M., Henseler, J., & Hair,
   to Advertising Research. International              J. F. (2014). On the Emancipation of PLS-
   Journal of Advertising, 33(1), 113-136.             SEM: A Commentary on Rigdon (2012). Long
 Lohmöller, J.-B. (1989). Latent V ariable Path        Range Planning, 47(3), 154-160.
   Modeling with Partial Least Squares.              Sarstedt, M., Schwaiger, M., & Ringle, C. M.
   Heidelberg: Physica.                                (2009). Do We Fully Understand the Critical
 Matthews, L. M. (2017). Applying Multi-group          Success Factors of Customer Satisfaction with
   Analysis in PLS-SEM: A Step-by-Step                 Industrial Goods? Extending Festge and
   Process. In H. Latan & R. Noonan (Eds.),            Schwaiger's     Model      to  Account       for
   Partial Least Squares Structural Equation           Unobserved Heterogeneity. Journal of
   Modeling - Basic Concepts, Methodological           Business Market Management, 3(3), 185-206.
   Issues and Applications. Berlin: Springer         Schlagel, C., & Sarstedt, M. (2016). Assessing
   International Publishing AG.                        the Measurement Invariance of the Four-
 Matthews, L. M., Sarstedt, M., Hair, J. F., &         Dimensional Cultural Intelligence Scale
   Ringle, C. M. (2016). Identifying and               Across Countries: A Composite Model
   Treating Unobserved Heterogeneity with              Approach. European Management Journal, 34
   FIMIX-PLS: Part II - A Case Study.                  (6), 633-649.
   European Business Review, 28(2) 208-224.          Sobel, M. E. (1982). Asymptotic Confidence
 Preacher, K. J., & Hayes, A. F. (2008).               Intervals for Indirect Effects in Structural
   Asymptotic and Resampling Strategies for            Equations Models. Sociological Methodology,
   Assessing and Comparing Indirect Effects in         13, 290-312.
   Multiple Mediator Models. Behavior                Vinzi, V. E., Chin, W. W., Henseler, J., &
   Research Methods, 40(3), 879-891.                   Wang, H. (2010). Editorial: Perspectives on
 Ringle, C. M., Sarstedt, M., & Straub, D. W.          Partial Least Squares. In V. E. Vinzi, W. W.
   (2012). A Critical Look at the use of PLS-          Chin, J. Henseler, & H. Wang (Eds.),
   SEM in MIS Quarterly. MIS Quarterly, 36             Handbook of Partial Least Squares: Concepts,
   (1), iii-xiv.                                       Methods and Applications. Heidelberg
 Ringle, C. M., Wende, S., & Becker, J.-M.             Dordrecht London New York: Springer.
   (2015). Smartpls 3. Bonningstedt: SmartPLS.       Wetzels, M., Odekerken-Schroder, G., & van
   Retrieved from http://smartpls.com.                 Oppen, C. (2009). Using PLS Path Modeling
 Sarstedt, M., Becker, J.-M., Ringle, C. M., &         for Assessing Hierarchical Construct Models:
   Schwaiger, M. (2011). Uncovering and                Guidelines and Empirical Illustration. MIS
   treating unobserved heterogeneity with              Quarterly, 33, 177-195.
   FIMIX-PLS: which model selection criterion        Wold, H. (1982). Soft Modeling: The Basic
   provides an appropriate number of segments?         Design and Some Extentions. In K. G.
   Schmalenbach Business Review, 63, 34-62.            Jöreskog & H. Wold (Eds.), Systems Under
 Sarstedt, M., Hair, J. F., Ringle, C. M., Theile,     Indirect Observations: Part II (pp. 1-54).
   K. O., & Gudergan, S. P. (2016). Estimation         Amsterdam: North-Holland.
   Issues with PLS and CBSEM: Where the
   Bias Lies! Journal of Business Research, 69,                       GLOSSARY
   3998-4010.
 Sarstedt, M., & Mooi, E. A. (2014). A Concise       Bootstrapping: a resampling technique that
   Guide to Market Research: The Process,            draws a specified (large) number of subsamples
   Data, and Methods using IBM SPSS                  from the original data and using replacement,
   Statistics (2nd ed.). Berlin: Springer.           estimates models for each subsample. It is used
                                                     to assess statistical significance without relying
Marketing Management Journal, Spring 2018                                                            12
PLS-SEM: The Holy Grail for Advanced Analysis                      Matthews, Hair and Matthews

 on distributional assumptions to determine
 standard errors of coefficients.

 Collinearity: when two variables are highly
 correlated.

 Formative measurement model: a type of
 measurement model setup in which the
 direction of the arrows is from the indicator
 variables to the construct, thus indicating an
 assumption that the indicator variables cause
 the measurement of the construct.

 Orthogonalizing approach: an approach to
 model the interaction term when including a
 moderator variable in the model. This creates
 an interaction term with orthogonals indicators.
 In the moderator model, these orthogonal
 indictors are not correlated with independent
 variable indicators and the moderator variable
 indicators.

 Product indicator approach: an approach to
 model the interaction term when including a
 moderator variable in the model. This approach
 involves multiplying the indicators of the
 moderator with the indicators of the exogenous
 latent variable to establish a measurement
 model of the interaction term. The approach is
 only applicable when both moderator and
 exogenous latent variables are reflective.

 Reflective measurement: a type of measurement
 model setup in which measures the direction of
 the arrow is from the construct to the indicator
 variables, thus the measures represent the
 effects (or manifestations) of an underlying
 construct. Causality is from the construct to its
 measures (indicators).

13                                                   Marketing Management Journal, Spring 2018
You can also read