This article was downloaded by: [University of Saskatchewan Library]

On: 03 January 2015, At: 10:27

Publisher: Routledge
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Total Quality Management & Business

Excellence
Publication details, including instructions for authors and
subscription information:
[Link]

Component-based Structural Equation

Modelling
a
Michel Tenenhaus
a
HEC School of Management (GREGHEC) , Jouy-en-Josas, France
Published online: 05 Sep 2008.

To cite this article: Michel Tenenhaus (2008) Component-based Structural Equation Modelling, Total
Quality Management & Business Excellence, 19:7-8, 871-886, DOI: 10.1080/14783360802159543

To link to this article: [Link]

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the
“Content”) contained in the publications on our platform. However, Taylor & Francis,
our agents, and our licensors make no representations or warranties whatsoever as to
the accuracy, completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content
should not be relied upon and should be independently verified with primary sources
of information. Taylor and Francis shall not be liable for any losses, actions, claims,
proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or
arising out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &
Conditions of access and use can be found at [Link]
and-conditions
 Total Quality Management
Vol. 19, Nos. 7 – 8, July –August 2008, 871– 886

Component-based Structural Equation

Modelling
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

Michel Tenenhaus

HEC School of Management (GREGHEC), Jouy-en-Josas, France

Two complementary schools have come to the fore in the field of Structural Equation Modelling
(SEM): covariance-based SEM and component-based SEM. The first approach has been
developed around Karl Jöreskog and the second one around Herman Wold under the name ‘PLS’
(Partial Least Squares). Hwang and Takane have proposed a new component-based SEM method
named Generalised Structured Component Analysis. Covariance-based SEM is usually used with
an objective of model validation and needs a large sample. Component-based SEM is mainly
used for score computation and can be carried out on very small samples. In this research, we
will explore the use of ULS-SEM, PLS, GSCA, path analysis on block principal components and
path analysis on block scales on customer satisfaction data. Our conclusion is that score
computation and bootstrap validation are very insensitive to the choice of the method when the
blocks are homogenous.

Keywords: component-based SEM; covariance-based SEM; GSCA; path analysis; PLS path
modelling; Structural Equation Modelling (SEM); Unweighted Least Squares (ULS)

Introduction
Two complementary schools have come to the fore in the field of Structural Equation Modelling
(SEM): covariance-based SEM and component-based SEM.
The first school developed around Karl Jöreskog. It can be considered as a generalisation of
path models, principal component analysis and factor analysis to the case of several data tables
connected by causal links. Covariance-based SEM is usually used with an objective of model
validation and needs a large sample (what is large varies from one author to another: more
than 100 subjects and preferably more than 200 subjects are often mentioned). The various
methods of estimation used for covariance-based SEM, like Maximum Likelihood (ML) or
Unweighted Least Squares (ULS), are full information methods.
The second school developed around Herman Wold under the name ‘PLS’ (Partial Least
Squares). It is a partial information method. It is a two-step method: (1) latent variables
scores are computed using the PLS algorithm and (2) OLS regressions are carried out on the

Email: tenenhaus@[Link]
1478-3363 print/1478-3371 online
# 2008 Taylor & Francis
DOI: 10.1080/14783360802159543
[Link]
 872 M. Tenenhaus

LV scores for estimating the structural equations. More recently, Hwang and Takane (2004)
have proposed a new full information method optimising a global criterion and named Gener-
alised Structured Component Analysis (GSCA). This second school can be considered as a gen-
eralisation of principal component analysis to the case of several data tables connected by causal
links. Component-based SEM is mainly used for score computation and can be carried out on
very small samples. A research based on six subjects has been published in Tenenhaus et al.
(2005b) and more recently, another one on 21 subjects, in Tenenhaus (2008).
Compared to covariance-based SEM, PLS suffers from several handicaps: (1) the diffusion of
path modelling software is much more confidential than that of covariance-based SEM software;
(2) the PLS algorithm is more a heuristic than an algorithm with well known properties; and
(3) the possibility of imposing value or equality constraints on path coefficients is easily
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

managed in covariance-based SEM and does not exist in PLS. Of course, PLS has also some
advantages on covariance-based SEM (that’s why PLS exists) and we can list some of them: sys-
tematic convergence of the algorithm due to its simplicity, possibility of managing data with a
small number of individuals and a large number of variables, practical meaning of the latent vari-
able estimates, general framework for multi-block analysis.
It is often mentioned that PLS is to covariance-based SEM as PCA is to factor analysis. But
the situation seriously changed when Roderick McDonald showed, in his 1996 seminal paper,
that he could easily carry out a PCA with a covariance-based SEM software by using the
ULS criterion and constraining the measurement error variances to be equal to zero. Further-
more, the estimation of the latent variables proposed by McDonald is similar to using the
PLS mode A and the SEM scheme (i.e. using the ‘theoretical’ latent variables as the inner
LV estimates). Thus, it became possible to use a covariance-based SEM software to mimic
PLS. He concluded from this that he could in fact use the covariance-based SEM approach to
obtain results similar to those of the PLS approach, but with a precise optimisation criterion
in place of an algorithm with not well known properties.
When each block of variables is essentially uni-dimensional (the first eigenvalue of the cor-
relation matrix is much larger than one and the second one much smaller) and homogeneous (all
the variables have the same scale, all the correlations are positive and the Cronbach alpha is
large) it is a good procedure to summarise each block by the first principal component
or more simply by the sum of the block items (scale). For this kind of data, path analysis of
these summaries is a natural procedure and yields the same results as the previous methods.
First experiences have already shown that score computation and bootstrap validation are
very insensitive to the choice of the method.
In the first section of this paper, it is reminded how to use the ULS criterion for covariance-
based SEM and the PLS way of estimating latent variables for mimicking PLS path modelling.
This methodology is applied to customer satisfaction data (the ECSI example) in the next
section. In the third section ULS-SEM, PLS, path analysis of block first principal components
and path analysis of block scales are compared in this example.

Using the ULS estimation method for structural equation modelling and McDonald
approach for LV estimates
We describe in this section the use of the ULS estimation method applied to the SEM parameter
estimates and that of the McDonald estimation method for computing the LV values. In the first
part we consider the Structural Equation Model following Bollen (1989). A Structural Equation
Model consists of two models: the latent variable model and the measurement model.
 Total Quality Management 873

The latent variable model

Let h be a column vector consisting of m endogenous (dependent) centred latent variables, and j
a column vector consisting of k exogenous (independent) centred latent variables. The latent
variable model connecting the vector h to the vectors h and j is written as

h ¼ Bh þ Gj þ z (1)

where B is a zero-diagonal mm matrix of regression coefficients, G a mk matrix of regression

coefficients and z a centred random vector of dimension m.
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

The measurement model

Each latent (unobservable) variable is described by a set of manifest (observable) variables. The
column vector yj of the centred manifest variables linked to the dependent latent variable hj can
be written as a function of hj through a simple regression with usual hypotheses:

yj ¼ lyj hj þ 1j (2)

The column vector y, obtained by concatenation of the yj, is written as

y ¼ Ly h þ 1 (3)
m
where Ly ¼  lyj is the direct sum of ly1 ; . . . ; lym and 1 is a column vector obtained by
j¼1
concatenation of the vectors 1j. It may be remembered that the direct sum of a set of matrices
A1, A2, . . . , Am is a block diagonal matrix in which the blocks of the diagonal are formed by
matrices A1, A2, . . . , Am.
Similarly, the column vector x of the centred manifest variables linked to the independent
latent variables is written as a function of j:

x ¼ Lx j þ d (4)

Adding the usual hypothesis that the matrix I – B is non-singular, equation (1) can also be
written as:

h ¼ ðI  BÞ1 ðGj þ zÞ (5)

and consequently equation (3) becomes

y ¼ Ly ½ðI  BÞ1 ðGj þ zÞ þ 1 (6)

Factorisation of the manifest variable covariance matrix

Let F ¼ CovðjÞ ¼ Eðjj0 Þ, C ¼ CovðzÞ¼ Eðzz0 Þ, Q1 ¼ Covð1Þ ¼ Eð110 Þ and Qd ¼ CovðdÞ ¼ Eðdd0 Þ.
Suppose that the random vectors j, z, 1 and d are independent of each other and that the covariance
 874 M. Tenenhaus

matrices C, Q1, Qd, of the error terms are diagonal. Then, we get:

Sxx ¼ Lx FL0x þ Qd
Syy ¼ Ly ½ðI  BÞ1 ðGFG0 þ CÞ½ðI  BÞ0 1 L0y þ Q1

Sxy ¼ Lx FG0 ½ðI  BÞ0 1 L0y

From which we finally obtain:


Syy Syx
S¼
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

Sxy Sxx
" # (7)
Ly ½ðI  BÞ1 ðGFG0 þ CÞ½ðI  BÞ0 1 L0y þ Q1 Ly ½I  B1 GFL0x
¼
Lx FG0 ½ðI  BÞ0 1 L0y Lx FL0x þ Qd

Let u ¼ fLx, Ly, B, G, F, C, Q1, Qdg be the set of parameters of the model and S(u) the
matrix (7).

Model estimation using the ULS method

Let S be the empirical covariance matrix of the MVs. The object of the ULS method is to seek
the set of parameters u^ ¼ fL̂x ; L̂y ; B̂; Ĝ; F̂; Ĉ; Q̂1 ; Q̂d g minimising the criterion
 2
 
S  Sðu^ Þ (8)

The aim is therefore to seek a factorisation of the empirical covariance matrix S as a function of
the parameters of the structural model. In SEM softwares, the default is to compute the covari-
d
ance matrix estimates Q̂1 ¼ Covð1Þ d
and Q̂d ¼ CovðdÞ of the residual terms in such a way that
the diagonal of the reconstruction error matrix E¼S2S(û) is null, even when it yields negative
variance (the Heywood case).
Let us denote by ŝii the ith term L
of the diagonal of S ¼ ðL̂x ; L̂y ; B̂; Ĝ; F̂; Ĉ; 0; 0Þ and by ûii
the ith term of the diagonal of Q̂1 Q̂d . From the formula:

^ ii þ u^ ii
sii ¼ s (9)

we may conclude that ŝii is the part of the variance sii of the ith MV explained by its LV (except
in a Heywood case) and ûii is the estimate of the variance of the measurement error relative to
this MV. As all the error terms eii ¼ sii  ðs^ ii þ u^ ii Þ are null, this method is not oriented towards
the research of parameters explaining the MV variances. It is in fact oriented towards the recon-
struction of the covariances between the MVs, variances excluded.

The McDonald approach for parameter estimation

In his 1996 paper, McDonald proposes to estimate the model parameters subject to the con-
straints that all the ûii are null. The object is to seek the parameters L̂x ; L̂y ; B̂; Ĝ; F̂; Ĉ
 Total Quality Management 875

minimising the criterion

 2
 ^ y ; B;
^ x; L ^ F;
^ G; ^ 0; 0Þ
^ C;
S  SðL  (10)

The estimates of the variances of the residual terms 1 and d are integrated in the diagonal terms
of the reconstruction error matrix E ¼ S  SðL̂x ; L̂y ; B̂; Ĝ; F̂; Ĉ; 0; 0Þ. This method is there-
fore oriented towards the reconstruction of the full MV covariance matrix, variances included.
On a second step, final estimates Q̂1 and Q̂d of the variances of the residual terms 1 and d are
obtained by using again formula (9).
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

Goodness of Fit
The quality of the fit can be measured by the GFI (Goodness of Fit Index) criterion of Jöreskog
and Sorbum, defined by the formula
 2
 ^ x; L
^ y ; B; ^ F;
^ G; ^ d Þ
^ 1; Q
^ C; Q
S  SðL 
GFI ¼ 1  (11)
kSk2

i.e. the proportion of kSk2 explained by the model. By convention, the model under study is
acceptable when the GFI is greater than 0.90. 2
 ^ y ; B;
^ x; L ^ F;
^ G; ^ C; ^ d Þ
^ 1; Q
^ Q
The quantity S  SðL  can be deduced from the FMIN criterion
given in the covariance-based SEM software AMOS (Arbuckle, 2005):

1 ^ x; L
^ y ; B; ^ F;
^ G; ^ C; ^ 1; Q
^ Q
2
^ d Þ
FMIN ¼ S  SðL  (12)
2

Using the McDonald approach, the GFI given by AMOS is equal to

 2
 ^ x; L
^ y ; B; ^ F;
^ G; ^ 0; 0Þ
^ C;
S  SðL 
GFI ¼ 1  (13)
kSk2

The exact GFI computed with formula (11) is also equal to the following:
 2
 ^ x; L
^ y ; B; ^ F;
^ G; ^ C; ^ d Þ
^ 1; Q
^ Q
S  SðL 
GFI ¼ 1 
kSk2
 2 P (14)
 ^ x; L
^ y ; B; ^ F;
^ G; ^ 0; 0Þ
^ C;
S  SðL   u^ii2
i
¼1
kSk2

In practical applications of the McDonald approach, the difference between the GFI given by
AMOS
P 2 (formula (13)) and the exact GFI computed with formula (14) will be small as
^ 2
i uii =kSk is usually small. Furthermore, the exact GFI will always be larger than the GFI
given by AMOS.
 876 M. Tenenhaus

Evaluation of the latent variables

After having estimated the parameters of the model, we now present the problem of evaluating
the latent variables. Three approaches can be distinguished: the traditional SEM approach, the
‘McDonald’ approach, and the ‘Fornell’ approach. As usual in the PLS approach, we now des-
ignate one manifest variable with the letter x and one latent variable with the letter j, regardless
of whether they are of the dependent or independent type. The total number of latent variables is
n ¼ kþm and the number of manifest variables related to the latent variable jj is pj.

The traditional SEM approach

To construct an estimate ĵj of jj one proceeds by multiple regression of jj on the whole set of the
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

P̂
centred manifest variables x11  x
11 ; . . . ; xnpn  x
npn . In other words, if one denotes as xx the
implied (i.e. predicted by the structural model) covariance matrix between the manifest vari-
ables, and as S^ xjj the vector of the implied covariances between the manifest variables x and
the latent variable jj, one obtains an expression of ĵj as a function of the whole set of manifest
variables:

j^ j ¼ X~ S^ 1 ^
xx Sxjj (15)

where X
¼ ½x11  x
11 ; . . . ; xnpn  x
npn . This method is not really usable, as it is more natural to
estimate a latent variable solely as a function of its own manifest variables.

The ‘McDonald’ approach for LV evaluation

Let xjl,. . ., xjpj be the manifest variables relative to the latent variable jj. McDonald (1996) pro-
poses evaluating the latent variable with the aid of the formula
X
j^ j / wjk ðxjk  x
jk Þ (16)
k

d jk ; jj Þ is the implied covariance between the MV xjk and the LV jj and where
where wjk ¼ Covðx
/ means that the left term is the standardised version of the right term.
The regression coefficient ljk of the latent variable jj in the regression of the manifest variable
xjk on the latent variable jj is estimated by

d jk ; jj Þ
Covðx
l^ jk ¼ (17)
d jj Þ
Varð

From this, we deduce that formula (16) can also be written as

X
j^ j / l^ jk ðxjk  x
jk Þ (18)
k

The McDonald approach thus amounts to estimating the latent variable jj with the aid of the first
PLS component computed in the PLS regression of the latent variable jj on the manifest vari-
ables xjl, . . . , xjpj. This approach could enter into the PLS framework. In the usual PLS approach
(Wold, 1985; Tenenhaus et al., 2005a), under mode A, the outer weights are obtained by simple
 Total Quality Management 877

regression of each variable xjk on the inner estimate zj of the latent variable jj. It is necessary to
calculate expressly the inner estimate zj of jj to obtain these weights. Three procedures are pro-
posed in PLS softwares: the centroid, factorial and structural schemes. The covariance-based
SEM software on the other hand, gives directly the weights (loadings) that for each xjk represent
an estimate of the regression coefficient of the ‘theoretical’ latent variable jj in the regression of
xjk on jj. Consequently, instead of the regression coefficient of the inner estimate zj, the estimated
regression coefficient of the ‘theoretical’ latent variable jj can be used. We have proposed this
procedure for calculating the weights based simply on the outputs of a covariance-based SEM
software in Tenenhaus et al. (2005a). We called it the ‘LISREL’ scheme and, without
knowing it, found the choice of weights proposed by McDonald.
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

The ‘Fornell’ approach

When all the coefficients l^ jk relative to a latent variable have the same sign and the manifest
variables are of a comparable order of magnitude, Fornell proposes building up a score taking
into account the level of the manifest variables xjk:
X X
j^ j ¼ l^ x =
k jk jk k
l^ jk (19)

This approach is standard in customer satisfaction studies.

Use of ULS-SEM on the ECSI data

We return to the ECSI (European Consumer Satisfaction Index) example presented in detail in
Tenenhaus et al. (2005a). The ECSI model is described in Figure 1. The data refers to a mobile
phone provider. Table 1 lists the manifest variables associated with each latent variable. Variable
h of block j is denoted by xjh. The manifest variables xjh are standardised to make the comparison
easier with other methods. It is a good practice to study, in a first step, the uni-dimensionality of
each block (the first eigenvalue of the correlation matrix is much larger than 1 and the second one
much smaller) and, when all the variables have the same scale and all the correlations are posi-
tive, its homogeneity (large Cronbach alpha). This information is given in Table 2. We detect
two facts: block ‘customer expectation’ is not very homogeneous and item 2 of block
‘loyalty’ is not a good representative of its block.
We have applied the usual constraint used in causality models to the weights of the manifest
variables: for each latent variable, one of the manifest variables has a weight equal to 1. The
estimation of the causality model of Figure 1 by minimisation of criterion (8) did not work
because a (small) negative error variance was obtained. Therefore, we have used the McDonald
approach for parameter estimation. Figure 2 shows the results of the estimation of the causality
model by minimisation of criterion (10). The outer model parameters normalized according to
the Fornell approach are given in Table 3. The confidence intervals of the MV weights were cal-
culated by bootstrapping (the percentile confidence interval option in AMOS 6 0). In Table 3, the
non-significant coefficients are shown in bold italic. LV scores are computed using these
weights. For example the ‘customer satisfaction score’ is computed as:

1  x51 þ 1:02  x52 þ 1:10  x53

Customer satisfaction score ¼
1 þ 1:02 þ 1:10
¼ 0:320  x51 þ 0:327  x52 þ 0:353  x53
 878 M. Tenenhaus
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

Figure 1. ECSI model.

For the inner model (structural equations) we do not use ULS-SEM results because the
obtained parameters are related to theoretical LVs and not to LV scores. Therefore, these par-
ameters are not comparable to the regression coefficients obtained with the other methods.
We prefer to use the path model built on the standardised LV scores shown in Figure 3. The par-
ameters of this model have been estimated using the maximum likelihood method available in
AMOS 6.0 and are given in Table 4. For each block, we compute the AVE (Average Variance
Extracted) given by the formula

p
1X j

AVEj ¼ Cor 2 ðxij ; j^j Þ

pj h¼1

The AVEs are given in Table 5. Then, for each endogenous LV jj, we compute the R-square
between ĵj and the other LV’s ĵk explaining ĵj. These R-squares are given in Table 6. Finally
the absolute goodness-of-fit (GoF) defined by the formula

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u 1 X 1 X
GoF ¼ u
t P pj AVEj  R2 ðj^j ; j^k explaining j^j Þ
pj j:pj .1 Nb of endogenous LV Endogenous LV
j:pj .1

is given in Table 7. Using the maximum likelihood method of estimation on the path model con-
structed on the ULS-SEM LV scores, we obtain the modification indices given in Tables 8 and 9.
These results suggest a strong new link from Image to Perceived quality and maybe another one
from Perceived value to Loyalty.
 Total Quality Management 879

Table 1. Manifest variables associated with each latent variable.

All the items are on a scale of 0 to 100. The value 0 corresponds to a very negative point of view of the
product, and a value of 100 to a very positive opinion.

Latent variables Manifest variables

Image (j1) (a) lt can be trusted in what it says and does

(b) It is stable and firmly established
(c) It has a social contribution for the society
(d) It is concerned about customers
(e) It is innovative and forward looking
Customer expectations of the (a) Expectations for the overall quality of ‘your mobile phone
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

overall quality (h1) provider’ at the moment you became a customer of this provider
(b) Expectations for ‘your mobile phone provider’ to provide
products and services to meet your personal need
(c) How often did you expect that things could go wrong with ‘your
mobile phone provider’?
Perceived quality (h2) (a) Overall perceived quality
(b) Technical quality of the network
(c) Customer service and personal advice offered
(d) Quality of the services you use
(e) Range of services and products offered
(f) Reliability and accuracy of the products and services provided
(g) Clarity and transparency of information provided
Perceived value (h3) (a) Given the quality of the products and services offered by ‘your
mobile phone provider’ how would you rate the fees and prices that
you pay for them?
(b) Given the fees and prices that you pay for ‘your mobile phone
provider’ how would you rate the quality of the products and services
offered by ‘your mobile phone provider’?
Customer satisfaction (h4) (a) Overall satisfiction
(b) Fulfilment of expectations
(c) How well do you think ‘your mobile phone provider’ compares
with your ideal mobile phone provider?
Customer complaints (h5) (a) You complained about ‘your mobile phone provider’ last year.
How well, or poorly, was your most recent complaint handled or
(b) You did not complain about ‘your mobile phone provider’ last
year. Imagine you have to complain to ‘your mobile phone provider’
because of a bad quality of service or product. To what extent do you
think that ‘your mobile phone provider’ will care about your
complaint?
Customer loyalty (h6) (a) If you would need to choose a new mobile phone provider how
likely is it that you would choose ‘your provider’ again?
(b) Let us now suppose that other mobile phone providers decide to
lower their fees and prices, but ‘your mobile phone provider’ stays at
the same level as today. At which level of difference (in %) would
you choose another mobile phone provider?
(c) If a friend or colleague asks you for advice, how likely is it that
you would recommend ‘your mobile phone provider’?
 880 M. Tenenhaus

Table 2. Block uni-dimensionality and Cronbach alpha on the standardised items.

Block correlation 1st Block correlation 2nd Cronbach

eigenvalue eigenvalue alpha

Image 2.394 0.913 0.723

Customer expectation 1.444 0.903 0.452
Perceived quality 4.040 0.771 0.877
Perceived value 1.700 0.300 0.824
Customer satisfaction 2.082 0.518 0.779
Customer loyalty 1.561 0.983 0.472
Customer loyalty (without 1.542 0.458 0.703
item 2)
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

Use of PLS path modelling, GSCA, path analysis of block principal component and path
analysis of block scales on the ECSI data
PLS path modelling
The PLS estimation method for Structural Equation Modelling proposed by Wold (1982) and
Lohmöller (1989) and also fully described in Tenenhaus et al. (2005a) is now used on the
ECSI data. We have run XLSTAT-PLSPM (XLSTAT, 2008) on the standardised data, using
mode A and centroid scheme. Weights are standardised according to the Fornell approach.
Results are given in Tables 3 to 7. They are similar to ULS-SEM followed by path analysis
of ULS-SEM LV scores. The only difference between both analyses is that the weight related
to item 2 of the loyalty block is not significant in ULS-SEM and considered as significant in
PLS. The 95% confidence interval for this weight in PLS is equal to [0.012 – 0.245]. So,
PLS fails to detect that this weight is not significant. This result suggests that item 2 of block
‘Loyalty’ should have been deleted from its block after inspection of the Cronbach alpha (see
Table 2) and consequently not used in the analysis.

Generalised Structured Component Analysis

In 2004, Hwang and Takane proposed a new method, named GSCA (Generalised Structured
Component Analysis). This is a competing method with PLS path modelling. Its main advantage
is that it is based on a global criterion to be minimised. Let us describe the criterion. All the latent
variables are denoted as ĵ j. However, endogenous and exogenous latent variables as well as
reflective and formative ones are distinguished. A standardised estimate of the latent variable
jj is denoted by ĵ j ¼Xjwj. The objective of GSCA is to search vectors of weights wj, cj and
regression coefficients bjk minimising the criterion
 2
X 

2
 X  ^ X 

Xj  j^ j c0j  þ jj  bjk j^ k  (20)
jj reflective jj endogenous;
 k

jk explaining jj

where cj is a column vector of weights related to the manifest variables xjh belonging to block j.
The first term of criterion (20) corresponds exactly to PCA and the second term to OLS
regressions on variables similar to ‘principal components’. GSCA is here a compromise
between PCA and OLS regressions. We have used the software program VisualGSCA 1.0 of
 Total Quality Management 881
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

Figure 2. Estimation of the ECSI model using ULS-SEM.

Heungsun Hwang (2007). This software program is downloadable free of charge from the
following site [Link] Results are
given in Tables 3 to 7. MV weights have been standardised according to the Fornell approach.
All the results are similar to ULS-SEM results.

Path analysis of block principal components

If the blocks are good blocks – uni-dimensional and homogeneous – we can use path analysis on
the standardised first principal components of each block. We have used AMOS 6.0 with the ML
 882 M. Tenenhaus

Table 3. Estimation of the outer model parameters (Fornell normalization). Non-significant (5% level)
parameters in bold italic.

Outer model parameter

ULS-SEM PLS GSCA PCA SCALE

x11 Image 0.220 0.208 0.218 0.218 0.200

x12 Image 0.180 0.183 0.169 0.170 0.200
x13 Image 0.156 0.153 0.174 0.175 0.200
x14 Image 0.229 0.229 0.225 0.223 0.200
x15 Image 0.214 0.227 0.214 0.214 0.200
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

x21 Customer expectation 0.364 0.363 0.389 0.384 0.333

x22 Customer expectation 0.336 0.325 0.349 0.347 0.333
x23 Customer expectation 0.300 0.312 0.262 0.269 0.333
x31 Perceived quality 0.156 0.163 0.150 0.150 0.143
x32 Perceived quality 0.116 0.110 0.123 0.122 0.143
x33 Perceived quality 0.152 0.152 0.145 0.146 0.143
x34 Perceived quality 0.139 0.136 0.148 0.147 0.143
x35 Perceived quality 0.142 0.138 0.142 0.143 0.143
x36 Perceived quality 0.142 0.138 0.147 0.148 0.143
x37 Perceived quality 0.153 0.164 0.145 0.144 0.143
x41 Perceived value 0.459 0.448 0.500 0.500 0.500
x42 Perceived value 0.541 0.551 0.500 0.500 0.500
x51 Customer satisfaction 0.320 0.314 0.321 0.322 0.333
x52 Customer satisfaction 0.327 0.318 0.345 0.344 0.333
x53 Customer satisfaction 0.353 0.368 0.334 0.334 0.333
x71 Customer loyalty 0.401 0.363 0.446 0.439 0.500
x72 Customer loyalty 0.095 0.106 0.103 0.117 0
x73 Customer loyalty 0.505 0.531 0.451 0.445 0.500

method of estimation. MV weights have been standardised according to the Fornell approach.
All the results given in Tables 3 to 7 are similar to ULS-SEM results.

Path analysis of block scales

On a good block, there is practically no difference between the standardised first principal com-
ponent and the ‘scale’ defined here by the standardised sum of the standardised items. Block
‘Loyalty’ has been made good by deleting item 2. MV weights have been standardised according
to the Fornell approach. All the results given in Tables 3 to 7 are similar to ULS-SEM results.

Comparison between PLS, GSCA, ULS-SEM path analysis, PCA path analysis and
Scale path analysis
We may compare the block components computed with the five methods. We give in Table 10
the correlation matrix for each block. The conclusion is clear: all methods yield to comparable
components. The results seem a little less comparable for block ‘loyalty’. But if we compute all
 Total Quality Management 883
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

Figure 3. Estimation of the ECSI model using Path analysis on ULS-SEM LV scores.

Table 4. Estimation of the inner model parameters (Standardised LV scores). Non-significant (5% level)
parameters in bold italic.

Inner model

OLS ML

Parameter PLS GSCA Path ULS-SEM Path PCA Path SCALE

Cust. expectation Image 0.505 0.487 0.505 0.504 0.508

Perc. quality Cust. expectation 0.557 0.533 0.555 0.551 0.553
Perc. value Perc. quality 0.557 0.545 0.553 0.545 0.538
Perc. value Cust. expectation 0.051 0.041 0.052 0.053 0.062
Cust. satisfaction Perc. value 0.192 0.173 0.190 0.186 0.187
Cust. satisfaction Cust. expectation 0.064 0.076 0.069 0.072 0.076
Cust. satisfaction Image 0.179 0.186 0.179 0.178 0.172
Cust. satisfaction Perc. quality 0.513 0.506 0.509 0.511 0.513
Complaints Cust. satisfaction 0.526 0.509 0.523 0.520 0.519
Cust. loyalty Cust. satisfaction 0.483 0.463 0.476 0.456 0.464
Cust. loyalty Image 0.195 0.212 0.198 0.210 0.206
Cust. loyalty Complaints 0.071 0.044 0.067 0.063 0.054
 884 M. Tenenhaus

Table 5. Average Variance Extracted computed on McDonald LV estimates.

AVE

ULS-SEM PLS GSCA PCA SCALE

Image 0.479 0.478 0.479 0.479 0.478

Customer expectation 0.481 0.480 0.481 0.481 0.479
Perceived quality 0.577 0.577 0.577 0.577 0.577
Perceived value 0.849 0.849 0.850 0.850 0.850
Customer satisfaction 0.694 0.693 0.694 0.694 0.694
Complaints 1.000 1.000 1.000 1.000 1.000
Customer loyalty 0.519 0.517 0.520 0.520 0.516
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

Weighted average AVE (Complaints not included) 0.575 0.574 0.575 0.575 0.574

Table 6. R-Square for structural equation computed on McDonald LV estimates.

R-Square

ULS-SEM PLS GSCA PCA SCALE

Customer expectation 0.255 0.255 0.253 0.254 0.258

Perceived quality 0.308 0.310 0.303 0.304 0.306
Perceived value 0.341 0.345 0.331 0.331 0.330
Customer satisfaction 0.677 0.680 0.673 0.672 0.671
Complaints 0.274 0.277 0.271 0.270 0.270
Customer loyalty 0.447 0.457 0.431 0.431 0.427
Average R-Square 0.384 0.387 0.377 0.377 0.377

Table 7. GoF for the various models.

ULS-SEM PLS GSCA PCA SCALE

GoF 0.470 0.471 0.465 0.465 0.465

Table 8. Path analysis on standardised ULS-SEM LV scores. Model FIT and

Modification Indices (.4) for LV structural model.

CMIN 172
df 9
CMIN/df 19.111
RMSEA .270
GFI .859
MI Perceived quality Image 78.5
Loyalty Perceived value 6.0
 Total Quality Management 885

Table 9. Path analysis on standardised ULS-SEM LV scores. Revised model

with an arrow added from Image to Perceived quality. Model FIT and
Modification Indices (.4) for LV structural model.

CMIN 34.8
df 8
CMIN/df 4.354
RMSEA 0.116
GFI 0.963
MI Loyalty Perceived value 5.96
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

Table 10. Correlation matrices between the LV scores computed with five methods.

ULS-SEM PLS GSCA PCA SCALE

Image
ULS-SEM 1.000
PLS 1.000 1.000
GSCA 1.000 0.999 1.000
PCA 1.000 0.999 1.000 1.000
SCALE 0.998 0.997 0.998 0.998 1.000
Customer expectation
ULS-SEM 1.000
PLS 1.000 1.000
GSCA 0.998 0.997 1.000
PCA 0.999 0.997 1.000 1.000
SCALE 0.998 0.999 0.993 0.994 1.000
Perceived quality
ULS-SEM 1.000
PLS 1.000 1.000
GSCA 1.000 0.999 1.000
PCA 1.000 0.999 1.000 1.000
SCALE 0.999 0.999 1.000 1.000 1.000
Perceived value
ULS-SEM 1.000
PLS 1.000 1.000
GSCA 0.999 0.999 1.000
PCA 0.999 0.999 1.000 1.000
SCALE 0.999 0.999 1.000 1.000 1.000
Customer satisfaction
ULS-SEM 1.000
PLS 1.000 1.000
GSCA 1.000 0.999 1.000
PCA 1.000 0.999 1.000 1.000
SCALE 1.000 0.999 1.000 1.000 1.000
Loyalty
ULS-SEM 1.000
PLS 0.999 1.000
GSCA 0.998 0.995 1.000
PCA 0.998 0.995 1.000 1.000
SCALE 0.991 0.986 0.992 0.989 1.000
 886 M. Tenenhaus

the loyalty components without using item 2 from this block we obtain the same kind of corre-
lation matrix as the other ones. We may conclude two points: (1) when the blocks are good, the
computation of the components does not depend upon the method used, and (2) path analysis of
the components also seems a very simple and promising approach.

Conclusion
Roderick McDonald has thrown a bridge between the SEM and PLS approaches by making use
of three ideas: (1) using the ULS method, (2) setting the variances of the residual terms of the
measurement model to 0, and (3) estimating the latent variables by using the loadings of the MVs
on their LVs. The McDonald approach has some very promising implications. Using a SEM soft-
Downloaded by [University of Saskatchewan Library] at 10:27 03 January 2015

ware such as AMOS 6.0 makes it possible to get back to the analysis of multi-block data and to a
‘data analysis’ approach for SEM completely similar to the PLS approach. However, this
approach is limited to reflective blocks. Heungsun Hwang and Yoshio Takane have proposed
a full information method named GSCA and based on a global criterion to be minimised.
This method can be used for reflective and formative blocks. We have also considered PCA
path analysis and Scale path analysis. We have illustrated these five methods with one classical
example on customer satisfaction: the ECSI data. On these ‘good’ data, all the methods give
practically the same results. On more general data, ULS-SEM and PLS will probably still
give close results. GSCA and PCA path analysis will probably give similar results as the pre-
vious methods, except when a first principal component is not related to the other blocks. We
end this paper with a wish: that all these methods are included in a component-based SEM soft-
ware. The user would then have access to a very comprehensive toolbox for a ‘data analysis’
approach to Structural Equation Modelling.

References
Arbuckle, J.L. (2005). AMOS 6.0. Spring House, PA: AMOS Development Corporation.
Bollen, K.A. (1989). Structural equations with latent variable. New York: John Wiley.
Hwang, H. (2007). VisualGSCA 1.0, A Graphical User Interface Software Program for Generalized Structured Com-
ponent Analysis. Department of Psychology, McGill University.
Hwang, H., & Takane, Y. (2004). Generalized structured component analysis. Psychometrika, 69(1), 81–99.
Jöreskog, K.G., & Sorbum, D. (1989). LISREL 7: A guide to the program and application. (2nd ed.) Chicago, IL: Scient-
ific Software, Inc.
Lohmöller, J.-B. (1989). Latent variable path modeling with Partial Least Squares. Heildelberg: Physica-Verlg.
McDonald, R.P. (1996). Path analysis with composite variables. Multivariate Behavioral Research, 31(2), 239– 270.
Tenenhaus, M. (2008). Structural Equation Modelling for small samples. HEC Paris: Jouy-en-Josas, Working paper no. 885.
Tenenhaus, M., Esposito Vinzi, V., Chatelin, Y.-M., & Lauro, C. (2005a). PLS path modelling. Computational Statistics
& Data Analysis, 48, 159 –205.
Tenenhaus, M., Pagès, J., Ambroisine, L., & Guinot, C. (2005b). PLS methodology to study relationships between
hedonic judgements and product characteristics. Food Quality and Preference, 16(4), 3l5 –325.
XLSTAT. (2008). XLSTAT-PLSPM module, XLSTAT software. Paris: Addinsoft.
Wold, H. (l982). Soft modeling: the basic design and some extensions. In K.G. Jöreskog & H. Wold (Eds.), System under
indirect observation, Part 2 (pp. 1 –54). Amsterdam: North-Holland.
Wold, H. (1985). Partial Least Squares. In S. Kotz & N.L. Johnson (Eds.), Encyclopedia of Statistical Sciences (pp. 581–
91). New York: John Wiley & Sons.