Você está na página 1de 10

Modeling Reflective Higher-Order Constructs using Three Approaches with PLS Path

Modeling: A Monte Carlo Comparison.

Bradley Wilson, RMIT University, School of Applied Communication.


Jrg Henseler, Radboud University, Nijmegen School of Management, The Netherlands.

Abstract
Many studies in the social sciences are increasingly modeling higher-order constructs. PLS
can be used to investigate models at a higher level of abstraction (Lohmller, 1989). It is often
chosen due to its ability to estimate complex models (Chin, 1998). The primary goal of this
paper is to demonstrate the relative robustness of various item and construct characteristics on
the reproduction of parameter true scores when utilising the two-stage, hierarchical
components indicators (repeated indicators) and a newer hybrid technique (where indicators
are not repeated). Our Monte Carlo study mirrors a simple substantive branding example. We
vary pertinent dimensions such as: sample size, differing item reliabilities and inner weighting
schemes. Our contribution is twofold. Firstly, we provide an overview of the approaches to
model reflective second-order constructs with PLS. Secondly, based on our simulation, we
provide suggestions when to use each approach.
Three PLS-based Approaches to Estimating Path Models with Higher-Order Constructs
In the extant literature, two approaches have been suggested, the Two-Stage Approach, and
the Hierarchical Components Approach. Additionally, we propose a newer Hybrid Approach.
We acknowledge that is some definitional concern regarding the use of the terminology
component, factor and latent variables that affect the area of PLS, Principal Component
Analysis, Covariance-Based Structural Equation Modeling (CBSEM) and Exploratory Factor
Analysis domains (du Toit, du Toit, Joreskog and Sorbom 1999; Pedhazar and Pedhazur,
1991). We understand these issues but for the purposes of this paper we refer to constructs as
latent constructs or variables interchangeably. Others have also followed this convention in
PLS writing and reporting (Henseler, and Fassott, 2007; O'Cass, 2002). Chin, Marcolin and
Newsted, 2003, p. 197) refer to PLS as a components-based structural equation modeling
technique. CBSEM and PLS methods are seen as complementary methods (Barclay, Higgins
and Thompson, 1995). A description of the three approaches investigated is presented next.
The Two-Step Approach
The two-stage approach is when latent variable scores are initially estimated without the
second-order construct present, but with all of the first-order constructs only within the model
(Agarwal and Karahanna, 2000, Henseler, Wilson, Gtz and Hautvast, 2007). The latent
variable scores are subsequently used as indicants in a separate higher-order structural model
analysis. Hence, a two-stage approach. This is typical of how analysts previously used factor
scores prior to running further regression analyses. It may offer advantages when estimating
higher-order models with formative indicants (Diamantopoulos and Winklhofer, 2001;
Reinartz, Krafft, and Hoyer, 2004). A clear disadvantage of any two-stage approach is that
any construct that is investigated in stage two is not taken into account when estimating the
latent variable scores at stage one. This could encourage interpretation confounding (Burt,
1973). Similar arguments have followed the use of the two-step modeling approach advocated
791

by Anderson and Gerbing (1988) in the CBSEM literatures. The implementation is not one
simultaneous PLS run. Falk and Miller (1992) talk of the advantage of PLS estimation in that
it takes into account its nearest neighbor during iteration. To follow such an approach may
not fully capitalise on the consistency at large assumption that PLS is based around. PLS
can overcome some of the problems of indeterminancy experienced when using CBSEM
techniques (Falk and Miller, 1992). We assume that the reader is familiar with PLS estimation
procedures. i
The Hierarchical Components Approach
The hierarchical components model was suggested originally by Wold (1982) (see also
Lohmller, 1989, p. 130-133; Chin et al. 2003). Also known as the Repeated Indicators
Approach (Lohmller, 1989; Wold, 1982) or Superblock Approach (Tenenhaus, Esposito
Vinzi, Chatelin and Lauro, 2005) is the most popular approach when estimating higher order
constructs with PLS (Venaik, 1999; Wilson, 2007; Zhang, Li, and Sun, 2006). A second
order factor is directly measured by observed variables for all the first order factors. While
this approach repeats the number of manifest variables used, the model can be estimated by
the standard PLS algorithm (Reinartz, Krafft and Hoyer, 2003, p. 19). The manifest
indicators are repeated to also represent the higher order construct.
Ite m

Ite m

s t

o r d e r

o n s tr u c t

Ite m

o r d e r

Ite m

Ite m

Ite m

o n s tr u c t

3
1

Ite m

n d

Ite m

s t

o r d e r

o n s tr u c t

Figure. 1. Conceptual Representation of Hierarchical Components Model.


Latent scores are saved during analysis to be used within future analyses. At this point in
time, this approach appears to be the one most favoured by analysts when using PLS to model
higher constructs. We believe this is because it has been presented most clearly by key
prominent PLS methodologists (e.g., Wold and Lohmller). A disadvantage of this approach
is that there is a perceived effect of possibly biasing the estimates by relating variables of the
same type together via PLS estimation. That is, the exogeneous variables in effect becomes
the endogenous variables. We are yet to see any Monte Carlo study that compares the
performance of various modeling approaches.
The Hybrid Approach
The Hybrid Approach builds on an idea of Wold (1982) originally meant for modeling
nonlinear structural relationships. The hybrid approach was inspired by and adapted from the
work of Marsh, Wen, and Hau (2006) within the CBSEM literature when investigating
interactions and quadratic effects. They argue when creating product terms that each of the
multiple indicators should only be used once in the formation (cross-product terms).., to
avoid creating artificially correlated residuals when the same variable is used in the
construction of more than one product terms. p. 245. To implement this technique within
PLS would randomly split all variables so that half are represented on their respective first
order construct side and the other half of indicants are represented on the second order
construct side (E.g., items are not repeated ii ). So for instance, if we are to alter Figure 1

792

slightly, this would mean that item 1 and 3 may still measure the two first order constructs
and items 2 and 4 would be for the higher order construct. We believe this approach has not
been trialed in PLS and could overcome the criticism that is directed towards the hierarchical
components model iii in that the indicators are repeated and therefore via PLS iteration and
estimation the analyst could be in some way relating the same items together. Naturally, the
hybrid approach circumvents this criticism. During the runtime of the algorithm, the secondorder construct is generated by a proxy which is then assigned to the second-order construct
(to derive latent variable scores and path coefficients).
We believe, the need for research is clear given that researchers are using PLS with increasing
regularity (Dawes, Lee and Dowling, 1998; Fornell et al., 1996; OCass and Fenech, 2003).
Numerous approaches exist yet the analyst knows little about performance and robustness
properties of technique for these model types. Previous studies provide limited direction. This
study also makes a valuable contribution by investigating a newer hybrid approach.

An Example from Brand Management


Prior to this Monte Carlo study there has been no formal guidance for social researchers
modeling higher order constructs with PLS. The researcher often just selected what
procedures other researchers had followed in the past. Here is one such small example.
Fournier (1994) after extensive qualitative research developed an item battery to measure the
quality of the person-brand bond. She termed this Brand Relationship Quality. Brand
relationship quality (BRQ) is best thought of as a customer-based indicator of the strength and
depth of the person-brand relationship (Fournier, 1994, p. 124). We illustrate only the
repeated indicators approach here to reserve further presentation space. For a description of
the sample characteristics readers should consult Wilson (2007). There were two main item
batteries (BRQ Scale: 62 items, Future Purchase Intention: 10 items) utilised. Fournier
supplied an extended BRQ scale version iv . The items used a 7-point scale. All measures were
treated as being reflective in keeping with the initial mode they were specified. A final sample
size of 1290 was obtained (25.8% response rate). The example was analysed with SPAD 6.0.
Adequate unidimensionality and appropriate construct and discriminant validity was
established (Wilson, 2007). In this example, 27.69% of the variation in intention is explained
by BRQ. The path coefficient between BRQ and intention indicates a fairly strong positive
impact ( = 0.5254). Falk and Miller (1992) would indicate that this is a significant finding in
social research. Presentation space is reserved for the Monte Carlo design and selected results.

A Monte Carlo Simulation Study


In order to complete this project we created our own implementation of the PLS algorithm.
We used R 2.3.1 (R Development Core Team, 2006). Firstly, we define an underlying true
population model (see figure 2) and determine the factor attributes for the Monte Carlo
design. Secondly, we generated random data that emerges from the model parameters.
Thirdly, given the random data, we let each PLS approach estimate each model under each
factor combination. The simulation model mirrors our BRQ future intentions example that
is presented previously. In this study, 1000 Monte Carlo samples per condition were run
resulting in 216,000 observations.
Fixed Factor:
1. Approach (Two-Stage; Repeated Indicators; Hybrid)

793

We used the following random factors (with categories) in the Monte Carlo experiment:
2. effect size of the effect of the second order construct on first-order construct (f2: 0.02,
0.15, 0.35)
3. loadings () in the model, including loadings between first order constructs and second
order construct (Square root of 0.5 (ca. 0.7071); square root of 0.75 (ca. 0.8660))
4. number of observations or sample size (50, 200, 800)
5. number of indicators (k: 4, 8)
6. inner weighting scheme (Centroid Scheme, Factor Weighting Scheme)
The three approaches were compared on the basis of their capability to:
capture the true relationship parameter between second-order and
create reliable latent variable scores for both and
predict accurately the endogenous variable
x

11

...

x 1k
...

...

...

...

...

...

P o p u la tio n M o d e l

yk

y1
...

k: 4; 8
: 0 .1 4 0 0 ; 0 .3 6 1 2 ; 0 .5 0 9 2
: 0 .7 0 7 1 ; 0 .8 6 6 0

Figure. 2. True Population Simulation Model

Results and Discussion


The PLS estimation outcomes have been measured for each run. Extensive analysis and
reporting is currently in progress and full results will be presented at the conference. We are
aiming to increase the complexity of our design prior to the conference by adding cells such
as expanding the number of indicators, loadings and parameters (homogenous and
heterogeneous), great sample size and investigating maybe another model structure to allow
for greater contrasts. The results in short at this preliminary stage of analysis include:
The reliability of the second-order construct is almost completely independent of the
chosen modelling approach (significant at the 10%-level only), and depends only on
the item loadings (p<.01), which was to be expected.
We see that for small effects, the reliability-corrected hybrid approach and the
reliability-corrected two-stage approach deliver the most consistent estimates.
Particularly in the case of fewer indicators (4), these approaches highly outperform the
repeated indicators approach in this regard. The same results pattern holds true for
medium effects. The repeated indicators approach performs best when the number of
indicants is larger and at determining larger effects sizes.
In terms of consistency, we see that the centroid scheme is best chosen for small
effects and that the factor weighting scheme is best utilised for medium effects. For
large effects, the choice of the weighting scheme is of limited influence.
Results reveal that with small sample sizes (50), the repeated indicators approach is
not able to deliver consistent estimates for small effects.
The centroid scheme performs poorly for small sample sizes (50), but is not likely to
overestimate effects as does the factor weighting scheme.

794

The repeated indicators approach delivers heavily biased estimates under less reliable
measurement compared with the other approaches. With higher loadings this approach
starts to improve dramatically for medium and large effects and in some respects is
superior.
There has been adequate convergence for all methods using different inner weighting
schemes with the PLS R programme.

The above preliminary results should give the reader an idea on what will be fully reported at
the conference. Even these preliminary results provide considerable guidance on what
approach to follow when faced with different study factors. Structural equation analysts are
often trying to find modeling solutions to these problems with literature that provides limited
consolidated guidance. v This is definitely a first step in what will be a more concerted
research agenda with a more complex research design.
This work is not without limitations. A Monte Carlo design can provide guidance for the
chosen population model under study. Often social researchers are selecting structural
equation models of greater complexity to investigate so the chosen population model may not
mirror our one structural path design. Also, Chin and Newsted (1999, p. 325) state that the
PLS estimates are "inconsistent" relative to the CBSEM model due to the fact that they are
aggregates of the observed variables, which in part includes measurement error. The estimates
will approach the "true" latent variable scores as both the number of indicators per block and
the sample size increases. Our preliminary work confirms these findings. The point at which
this happens is of interest within our Monte Carlo work to provide pragmatic guidelines for
appropriate use of PLS for different model types. Chin and Newsted (1999) go on to
emphasize that more indicants and also more cases are needed to improve PLS estimation
robustness. We also can confidently state that our results also indicate this is the case. The
hybrid approach although operating against these principles compared with the hierarchical
components approach showed relatively robust performance. This may offer the modeller a
new choice that has not been previously considered. Our design needs to be improved to have
smaller sample size graduations (at the lower range), smaller number of indicator graduations,
testing different distributional patterns and non-homogeneous outside loading patterns. Such
work is now in progress.
As we limited our study to PLS-based approaches, other structural equation modeling
techniques CBSEM [summated scales and congeneric models (Joreskog, 1971)] were not
considered. Some researchers have used CBSEM in estimating congeneric models before
using PLS for the structural modeling stages (Grace and OCass, 2005). This approach is also
worthy of future exploration within a Monte Carlo design to act as another basis for
comparison vi .
As we studied only reflective measurement models, it remains unclear whether our results
will be generalised to PLS path models with formative measures where PLS has significant
demonstrated advantages (Ringle, Wilson and Gtz, 2007; Vilares, Almeida and Coelho,
2007). All higher order model structure combinations represented within the Jarvis et al.
(2003) study need to be investigated. That is, all model types need to have separate PLS
Monte Carlo studies completed.
We believe that researchers need to approach tasks with more methodological certainty given
the increasing popularity for investigating complex higher order structural models. It is our
contention that such research will increasingly be undertaken even if it is in a model
generation exploratory capacity. Practically, market researchers are looking for such
flexibility. We believe that Monte Carlo work within PLS is some way behind what has been
established in the CBSEM domain (Boomsma, 1983; Gerbing and Anderson, 1993; Hu and
Bentler, 1999; Tanaka, 1987). Some recent studies in PLS are aiming to alter this dearth in

795

knowledge to provide researchers with greater guidance (Cassel, Hackl and Westlund, 1999;
Henseler, Wilson and Dijkstra, 2007; Tenenhaus et al., 2005). We encourage others to follow
a similar research agenda to expand the body of PLS knowledge.

796

References
Agarwal, R., Karahanna, E., 2000. Time Flies When You're Having Fun: Cognitive
Absorption and Beliefs About Information Technology Usage. MIS Quarterly 24 (4), 665694.
Anderson, J.C., Gerbing, D.W., 1988. Structural equation modeling in practice: A review and
recommended two-step approach. Psychological Bulletin 103, 411-423.
Barclay, D., Higgins, C., Thompson, R., 1995. The Partial Least Squares (PLS) Approach to
Causal Modeling: Personal Computer Adoption and Use as an Illustration. Technology
Studies 2 (2), 285-309.
Boomsma, A., 1983. On the robustness of LISREL (maximum likelihood estimation) against
small sample size and nonnormality. Unpublished doctoral dissertation. University of
Groningen, Groningen.
Burt, R.S., 1973. Confimatory Factor-Analytic Structures and the Theory Construction
Process. Sociological Methods and Research 2 (November), 131-190.
Chen, F., Bollen, K.A., Paxton, P., Curran, P.J., Kirby, J.B., 2001. Improper Solutions in
Structural Equation Models: Causes, Consequences, and Strategies. Sociological Methods and
Research 29 (4), 468-508.
Chin, W.W., 1998. The Partial Least Squares Approach to Structural Equation Modeling. In
Marcoulides, G.A. (Ed.), Modern Methods for Business Research. Lawrence Erlbaum,
Mahwah, NJ, pp. 295-336.
Chin, W.W., Newsted, P.R., 1999. Structural Equation Modeling Analysis with Small
Samples using Partial Least Squares. In Hoyle, R.H. (Ed.), Statistical Strategies for Small
Sample Research. Sage Publications, Thousand Oaks: CA, pp. 307-341.
Chin, W.W., Marcolin, B.L., Newsted, P.N., 2003. A partial least squares latent variable
modeling approach for measuring interaction effects: Results from a monte carlo simulation
study and an electronic-mail emotion/adoption study. Information Systems Research 14 (2),
189217.
Dawes, P.L., Lee, D.Y., Dowling, G.R., 1998. Information Control and Influence in Emergent
Buying Centers. Journal of Marketing 62 (3), 55.
Diamantopoulos, A., Winklhofer, H.M., 2001. Index construction with formative indicators:
An alternative to scale development. Journal of Marketing Research 38 (2), 269277.
Dillon, W.R., Mulani, N., Kumar, A., 1987. Offending Estimates in Covariance Structure
Analysis: Comments on the Causes and Solutions to Heywood Cases. Psychological Bulletin
101 (1), 126-135.
du Toit, S., du Toit M., Joreskog K.G., Sorbom D., 1999. Interactive LISREL Users Guide.
Scientific Software International., Chicago, IL.

797

Falk, R.F., Miller, N.B., 1992. A Primer for Soft Modeling. The University of Akron Press,
Akron, Ohio.
Fornell, C., Cha, J., 1994. Partial least squares. In Bagozzi, R. P. (Ed.), Advanced Methods of
Marketing Research. Basil Blackwell Ltd. Oxford, pp. 52-78.
Fornell, C., Johnson, M. D., Anderson, E. W., Cha, J., Bryant, B.E., 1996. The American
Customer Satisfaction Index: Nature, purpose, and findings. Journal of Marketing 60 (4), 719.
Fournier, S., 1994. A Consumer-Brand Relationship Framework for Strategic Brand
Management. Unpublished Doctoral Dissertation. University of Florida.
Gerbing, D. W., Anderson, J. C., 1993. Monte Carlo Evaluations of Goodness-of-Fit Indices
for Structural Equation Models. In .Bollen, K.A., Long, S.J. (Eds.), Testing Structural
Equation Models. Sage Publications, Newbury Park, US. pp. 40-65.
Grace, D., OCass, A., 2005. An examination of the antecedents of repatronage intentions
across different retail store formats. Journal of Retailing and Consumer Services 12, 227-243.
Henseler, J., Fassott, G., 2007. Testing Moderating Effects in PLS Path Models: An
Illustration of Available Procedures. In .Esposito Vinzi, V., Chin, W.W., Henseler, J., Wang,
H. (Eds.), Handbook of Partial Least Squares: Concepts, Methods and Applications in
Marketing and Related Fields (Springer Handbooks of Computational Statistics Series).
Springer Publishing. Berlin/Heidelberg.
Henseler, J., Wilson, B., Dijkstra, T., 2007. Testing Nonlinear Effects in PLS Path Models.
PLS07 - 5th International symposium on PLS and related methods: Causalities explored by
indirect observation. Matforsk, Aas, Norway. September 5th - 7th. forthcoming.
Henseler, J., Wilson, B., Gtz, O., Hautvast, C. 2007. Investigating the Moderating Role of
Fit on Sports Sponsoring and Brand Equity: A Structural Model. International Journal of
Sports Marketing and Sponsorship 9 (1), forthcoming.
Hu, L., Bentler, P.M., 1999. Cutoff Criteria for Fit Indexes in Covariance Structure Analysis:
Conventional Criteria Versus New Alternatives. Structural Equation Modeling 6 (1), 1-55.
Jarvis, C.B., MacKenzie, S.B., Podsakoff, P.M., 2003. A Critical Review of Construct
Indicators and Measurement Model Misspecification in Marketing and Consumer Research.
Journal of Consumer Research 30 (September), 199-218.
Joreskog, K.G., 1971. Statistical Analysis of Congeneric Tests. Psychometrika. 36, 109-133.
Landis, R.S., Beal, D.J., Tesluk, P.E., 2000. A Comparison of Approaches to Forming
Composite Measures in Structural Equation Models. Organizational Research Methods 3 (2),
186-207.
Lohmller, J.-B., 1989. Latent Variable Path Modeling with Partial Least Squares.
Heidelberg: Physica-Verlag.

798

Marsh, H.W., Wen, Z., Hau, K-T., 2006. Structural Equation Models of Latent Interaction
and Quadratic Effects. In Hancock, G.R., Mueller, R.O. (Eds.), Structural Equation Modeling:
A Second Course. Information Age Publishing, Greenwich, pp. 225-268.
O'Cass, A., 2002. Political Advertising Believability and Information Source Value During
Elections. Journal of Advertising 31 (1), 63-74.
OCass, A., Fenech, T., 2003. Webretailing adoption: exploring the nature of internet users
Webretailing behaviour. Journal of Retailing and Consumer Services 10, 81-94.
Pedhazar, E.J., Pedhazur Schmelkin, L., 1991. Measurement, Design, and Analysis: An
Integrated Approach. Lawrence Erlbaum Associates., Hillsdale, New Jersey.
Reinartz, W., Krafft, M., Hoyer, W.D., 2003. Measuring the Customer Relationship
Management Construct and Linking it to Performance Outcomes. INSEAD Working Paper
Series. 2003/02/MKT, 1-51.
Reinartz, W., Krafft, M., Hoyer, W.D., 2004. The Customer Relationship Management
Process: Its Measurement and Impact on Performance. Journal of Marketing Research 41(3),
293305.
Rindskopf, D., 1984. Structural Equation Models. Sociological Methods and Research 13 (1),
109-119.
Ringle, C.M., Wende, S., Will, A., 2006. SmartPLS 2.0. Hamburg: University of Hamburg.
Ringle, C., Wilson, B., Gtz, O., 2007. A Monte Carlo Robustness Study on Formative
Measurement Model Specification in CBSEM and PLS. PLS07 - 5th International symposium
on PLS and related methods: Causalities explored by indirect observation. Matforsk, Aas,
Norway. September 5th - 7th.
Tanaka, J. S., 1987. "How big is big enough?": Sample size and goodness of fit in structural
equation models with latent variables. Child Development 58, 134-146.
Tenenhaus, M., Esposito Vinzi, V., Chatelin, Y-M., Lauro, C., 2005. PLS Path Modeling.
Computational Statistics & Data Analysis. 48 (1), 159-205.
Vilares, M.J., Almeida, M.H., Coelho, P.S., 2007. Comparison of likelihood and PLS
estimators for structural equation modeling: A simulation with customer satisfaction data. In
Vinzi, V.E., Chin, W.W., Henseler, J., Wang, H., (Eds.), Handbook of partial least squares:
concepts, methods and applications in marketing and related fields. Springer. Berlin,
forthcoming
Venaik, S., 1999. A Model of Global Marketing in Multinational Firms: An Empirical
Investigation. Unpublished Doctoral Dissertation. The Australian Graduate School of
Management, Sydney.
Wilson, B., 2007. Using PLS to Investigate Interaction Effects Between Higher Order Brand
Constructs. In Esposito Vinzi, V., Chin, W.W., Henseler, J., Wang, H. (Eds.), Handbook of

799

Partial Least Squares: Concepts, Methods and Applications in Marketing and Related Fields
(Springer Handbooks of Computational Statistics Series). Berlin/Heidelberg: Springer.
Wold, H., 1982. Soft modeling. The basic design and some extensions. In Jreskog, K.G.,
Wold, H., (Eds.), Systems under indirect observation. Causality, structure, prediction. Part I,
154. North-Holland: Amsterdam.
Zhang, P., Li, N., Sun, H., 2006. Affective Quality and Cognitive Absorption: Extending
Technology Acceptance Research. Proceedings of the 39th Hawaii International Conference
on Systems Sciences. Hawaii. IEEE Computer Society, 11p. vii
i

For an excellent explanation of the PLS algorithm its respective estimation in establishing outside and inside
parameter estimates see (Chin, 1998; Chin and Newsted, 1999; Fornell and Cha, 1994; Tenenhaus et al., 2005).
ii
We realise that this separation of items could occur in many ways. This is worthy of future investigation. This
in analogous to all the approaches that have been tried and tested when using parcelling or testlets within the
CBSEM literature (Landis, Beal and Tesluk, 2000).
iii
This criticism cannot be referenced but has come about from discussions with the PLS academic community.
Some researchers currently sparingly use the hierarchical components approach as they believe the jury is out
on its overall performance. Others avoid investigating models with higher-order relations altogether at this time
with PLS. However, some researchers are starting to use PLS when investigating higher order conceptual
models.
iv
The first author would like to acknowledge and thank Professor Fournier for her initial support and inspiration.
v
We acknowledge that this is our view from modeling experience with higher-order constructs.
vi
We acknowledge it is difficult to extend such research to incorporate LISREL approaches for higher order
constructs at the item level due to the inability to use the repeated indicators approach. There are also many
complex identification and estimation problems (Chen et al., 2001; Dillon, Mulani and Kumar, 1987; Rindskopf,
1984) that can become more prevalent with the use of CBSEM with higher order constructs.

800