Está en la página 1de 6

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/316430135

Full collinearity as a new criterion to assess


discriminant validity of composite (formative)
and reflective measurement...

Conference Paper June 2017

CITATIONS READS

0 154

6 authors, including:

S. Mostafa Rasoolimanesh Mehran Nejati


Universiti Sains Malaysia Edith Cowan University
50 PUBLICATIONS 114 CITATIONS 70 PUBLICATIONS 354 CITATIONS

SEE PROFILE SEE PROFILE

T. Ramayah Nordin Abd Razak


Universiti Sains Malaysia Universiti Sains Malaysia
427 PUBLICATIONS 3,670 CITATIONS 41 PUBLICATIONS 171 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Chidren drawing a mirror to their perception of an ideal school View project

Developing A Predictive Model of Crowdsourcing Based on Consumer-Brand Engagement View


project

All content following this page was uploaded by S. Mostafa Rasoolimanesh on 24 April 2017.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
Full collinearity as a new criterion to assess discriminant validity
of composite (formative) and reflective measurement models

Keywords: Discriminant validity, Fornell-Larcker criterion, Heterotriat-monotriat (HTMT) ratio, Full


collinearity, Reflective, Composite (Formative)

Introduction

Partial least squares structural equation modeling (PLS-SEM) is a variance-based SEM that has witnessed
growing attention by researchers in recent years across marketing, strategic management, information systems,
and other social sciences disciplines. Assessment of a model using PLS-SEM technique generally follows a two-
step process including assessment of the measurement model and structural model [1, 2]. Assessment of the
measurement model entails the evaluation of validity and reliability of involved latent variables (LVs) in the
model. This refers to evaluating the relationships between the LV and the associated items; whereas the
assessment of the structural model focuses on the causal relationships between LVs [1, 2].
However, there are two types of measurement model known as reflective and composite (formative). The criteria
for assessing reflective and composite (formative) measurement models are very different when using PLS-SEM
[2]. For the reflective and composite (formative) measurement models, reliability and validity (i.e. convergent
and discriminant validity) should be assessed. However, the suggested criteria for assessment of reliability and
validity of reflective and composite (formative) constructs are totally different [1]. Reliability and convergent
validity refers to the assessment of strong correlation of indicators to their corresponding latent variables, and are
deemed acceptable when indicators load strongly on their associated constructs [2]. On the other hand,
discriminant validity refers to the distinction between LVs in the model. Discriminant validity is established
when the constructs are distinct from each other [3].
Discriminant validity is a crucial part of model assessment using PLS-SEM as it will gurantee the uniquness of
each construct in the model and avoide biased results arising from shared meaning by different LVs [4].
Neglecting discriminant validity in the model assessment may increase the posibility of statistical discrepancies
with real results [4]. Therefore, discriminant validity assessment is essential to the model assessment process.
However, the suggested approaches for assessing discriminant validity in the extant literature can only be
applied to reflective measurement model. Given that there are a number of studies including both reflective and
composite (formative) measurement models, the assessment of discriminant validity becomes paramount [14, 15,
16]. As such, the current study aims to introduce a new criterion to establish discriminant validity for both
reflective and composite (formative) measurement models, and examine the effectiveness of this new criterion
using a Monte Carlo simulation study.

1 Discriminant Validity

Previous studies suggested a few approaches to assess discriminant validity using PLS-SEM including the
Fornell and Larcker [5] criterion, cross-loadings [6], and more recently the conservative Heterotriat-monotriat
(HTMT) ratio [3]. Nonetheless, all of these suggested approaches can only be applied to the reflective constructs,
not the models involving composite (formative) constructs which is common in Social Sciences discipline. As
dealing with a model including both reflective and composite (formative) constructs is one of the key advantages
of PLS-SEM [7], lack of an approperiate approach to assess discriminant validity for models involving both
composite (formative) and reflective constructs is a very significant gap which requires urgent attention.
In order to establish discriminant validity, researchers are required to verify all the constructs in a model are
distinct from each other [4]. Lack of discriminant validity in a model leads to questionable conclusions, disputing
whether results can truely be supported by the data or they are obtained because of using a construct twice in the
model [4]. Assessing discriminant validity in a model can be at the construct level and at the observed items level
[4]. At the construct level, the disciminant validity assessment shows the degree to which the meaning and
attribute of constructs are distinct, whereas in the observed items level, the disciminant validity assessment
identifies to what extent items are the reflection of a unique construct in the model [4].
Among the three common methods to assess discriminant validity, the cross-loadings and HTMT approaches are
more item-based. However, the Fornell-Larcker criterion that compares the average variance extracted from the
associated items of each construct with the squared correlation between the construct and other constructs is
more construct-based [4]. Nonetheless, all of these approaches for assessing discriminant validity can only be
applied to reflective constructs. Therefore, when the model includes composite (formative) constructs, the
assessment of discriminant validity becomes problematic [4].

2 Full Collinearity

There is another concept in multivariate analysis literature called multi-collinearity which refers to measuring the
same attribute by two or more constructs in a model [8]. If the items of two constructs refer to the same object
attribute, it reveals the existence of multi-collinearity in the model [9]. However, the concept of collinearity is
different from correlation. Multi-collinearity is a multivariate concept, while correlation only refers to a pair of
variables [9]. Multi-collinearity can occur both laterally or vertically, and is called full collinearity when we
consider both lateral and vertical collinearity [8]. In variance-based SEM like PLS-SEM, full collinearity can
still remain if different constructs are redundant [8] even when constructs have passed the discriminant validity
assessment using the common Fornell-Larcker criterion, and cross-loading test. The redundancy among the
constructs may occur because of similarity of observed items, or the overall meaning of construct [4]. The recent
literature has proposed calculation of variance inflation factor (VIF) as a measure for collinearity for each
construct and then compare these VIFs with a threshold of 10, 5, or more conservative 3.3 [8]. Unfortunately,
most of researchers over the past three decades just focused on vertical collinearity [9] and simply ignored lateral
collinearity. However, lateral collinearity that refers to the collinearity among predictors and criterion can also
distort the results of multivariate analysis [8].
To this end, full collinearity can be considered as a significant criterion to identify redundancy or distinction
among the constructs in a model [10]. The distinction can be calculated based on multivariate approach and not
only correlation between pairs that is common among the current approaches to assess discriminant validity. In
addition, in PLS-SEM, full collinearity can be calculated for both reflective and composite (formative)
constructs, and even for single item constructs, based on the LV scores estimated at the first stage of PLS-SEM.
Therefore, full collinearity can serve as a new and improved criterion to assess discriminant validity in PLS-
SEM [10], particularly for situations where the model includes both reflective and composite (formative) or
single item constructs.
To assess the appropriateness of full collinearity to establish discriminant validity for a model including
reflective, single item, and composite (formative) constructs, and compare it with other common methods of
assessing discriminant validity, we performed a Monte Carlo simulation. Results of the simulation study will be
discussed in the following sections.

3 Research Method

To compare the traditional Fornell-Larcker criterion and more recent HTMT ratio with full collinearity to assess
discriminant validity, we conducted a simulation study. A model including two reflective constructs, one single
item construct, and one composite (formative) construct using PLS-SEM was adapted to undertake simulation
study. Two reflective constructs (X & Z) were generated randomly involving three indicators, and assuming the
loadings equal to 0.6, 0.7, and 0.8 respectively. Different levels of correlations (0.1, 0.3, 0.5, 0.7, and 0.9)
between X and Z were assumed to conduct simulations. Then, the composite construct (W) was generated
assuming five levels of correlations (0.1, 0.3, 0.5, 0.7, and 0.9) between Zs indicators and Ws indicators. The
composite construct was created with the constant weights of 0.3, 0.3, and 0.4 respectively (See 1-7 equations).
To conduct the simulation study, the correlation between W and Y as a single item construct was fixed to 0.9. To
generate 300 sample size data for simulation study, Excel software was employed. Moreover, SmartPLS 3.2.6
[11], and WarpPLS 5.0 software packages were employed to calculate Fornell-Larcker criterion, HTMT ratio
and full collinearity of constructs in the model.

x1= 0.6*X + e1 (1)


x2= 0.7*X + e2 (2)
x3= 0.8*X + e3 (3)

z1= 0.6*Z + e4 (4)


z2= 0.7*Z + e5 (5)
z3= 0.8*Z + e6 (6)
W=0.3*w1 + 0.3*w2 + 0.4*w3 (7)

4 Results of Simulations
Table 1 shows the results of simulation study for the model including two reflective constructs, one composite
(formative) construct, and one single item construct, assuming different levels of correlations between the
constructs. Among 25 rounds of simulations, the results demonstrated just one detection (4%) of discriminant
validity problem using Fornell-Larcker criterion. However, Fornell-Larcker criterion is only applicable to
reflective constructs to detect the discriminant validity problem in the model. The second criterion to assess the
discriminant validity in this simulation study was the HTMT ratio. HTMT can be applied to assess the
discriminant validity of reflective and single item constructs in a model. According to the results shown in Table
1, HTMT detected the problem of discriminant validity between constructs in 36 percent of simulations. This
finding is consistent with previous studies, which revealed that HTMT criterion for assessing discriminant
validity is more conservative than Fornell-Larcker criterion. However, if we calculate the score of composite
construct using equation 7, the model will include two reflective and two single item constructs. Therefore, in
this situation, the HTMT criterion is applicable for the current model. In this case, the results show discriminant
validity problem in 52 percent of simulations. Therefore, to apply HTMT criterion for assessing discriminant
validity, we need to use a two-stage approach. In the first stage, the score of composite (formative) construct
should be calculated, and the score is used in second stage. In the second stage, we can apply HTMT to assess
discriminant validity. However, the results of simulation study was surprising when full collinearity approach
was applied. With threshold of 3.3 for full collinearity that was suggested in some studies [10, 12, 13], in 60
percent of simulations, the problem of discriminant validity was detected in the model. Hence, full collinearity
could detect discriminant validity in more cases than even two-stage approach using HTMT. However, if we
apply threshold of 5 for multi-collinearity, which is more liberal, the discriminant validity problem is detected in
36 percent of the simulations.
The results showed that, in high correlation of reflective constructs (0.9), HTMT works better than full
collinearity, and can detect the discriminant validity problem in more cases. However, between reflective and
composite (formative) constructs, full collinearity could detect discriminant validity problem in more simulations
than HTMT, even when calculating the HTMT ratio using the score of composite (formative) construct.

5 Conclusion

The current study attempted to propose and examine a new criterion to assess discriminant validity, particularly
when the model includes both reflective and composite (formative) constructs using a simulation study. The
results of simulation study demonstrated the strong capability of full collinearity in detecting discriminant
validity problem in the model with reflective, and composite (formative) constructs compared to the traditional
Fornell-Larcker criterion, and even more recent HTMT approach. The findings revealed that HTMT is more
conservative to assess discriminant validity between reflective constructs, but for the model of combination of
reflective, single item, and composite (formative) constructs, full collinearity criterion can be more conservative
and applicable. The traditional Fornell-Larcker criterion, only can be applied to the models including reflective
constructs, but HTMT can be employed for reflective and single item constructs. It is impossible to apply HTMT
for composite (formative) constructs, unless, we use the score of construct and change the composite (formative)
to single item construct using a two-stage approach. However, full collinearity can be used for all types of
constructs.
This study compared the three different approaches for assessing discriminant validity using 25 rounds of
simulations assuming various correlations between constructs. Conducting a small number of simulations can be
a limitation of the current study. Thus, future research can consider large number of simulations to confirm the
results of the current study and compare full collinearity with other approaches of assessing discriminant
validity.
Table 1. The results of simulation study

Correlation Correlation Fornell-Larcker HTMT HTMT Full Collinearity


X &Z W&Z (model with composite construct) (Change composite to single item construct)
Reflective Single Composite Reflective Single Composite Reflective Single item Composite Reflective Single Composite
item (Formative) item (Formative) (Formative) item (Formative)
0.1 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 <0.85 <2 <3 <3
0.3 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 <0.85 <2 <3 <3
0.1 0.5 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 <0.85 <2 <3 <3
0.7 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 >0.9 <3.3 <3 5.77
0.9 Yes N/A N/A <0.85 >0.9 N/A <0.85 >0.9 >0.9 <2 3.7 18.3
Z=14.1
0.3 0.1 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 <0.85 <2 <3 <3
0.3 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 <0.85 <2 <3 <3
0.5 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 <0.85 <2 <3 4.0
0.7 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 >0.9 <3.3 <3.3 3.3
0.9 No N/A N/A <0.85 >0.9 N/A <0.85 >0.9 >0.9 <2 3.4 13.7
Z=11.05
0.5 0.1 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 <0.85 <2 <3 <3
0.3 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 <0.85 <2 <3 <3
0.5 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 <0.85 <2 <3 4.2
0.7 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 >0.9 <3.3 <3.3 5.4
0.9 No N/A N/A <0.85 >0.9 N/A <0.85 >0.9 >0.9 <2 3.4 17.0
Z=11.43
0.7 0.1 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 <0.85 <2 <3 <3
0.3 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 <0.85 <2 <3 <3
0.5 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 <0.85 <2 <3 4.46
0.7 No N/A N/A <0.85 <0.85 N/A <0.85 <0.85 >0.9 <3 3.47 7.3
0.9 No N/A N/A <0.85 >0.9 N/A <0.85 >0.9 >0.9 <3 3.74 19.3
Z=11.6
0.9 0.1 No N/A N/A >0.9 <0.85 N/A >0.9 <0.85 <0.85 <3 <3 <3
0.3 No N/A N/A >0.9 <0.85 N/A >0.9 <0.85 <0.85 3.27 <3 3.5
Z<2
0.5 No N/A N/A >0.85 <0.85 N/A >0.85 <0.85 >0.9 3.38 <3 4.4
Z<2
0.7 No N/A N/A >0.85 <0.85 N/A >0.85 <0.85 >0.9 3.16 3.2 7.11
Z<3
0.9 No N/A N/A >0.85 >0.9 N/A >0.85 >0.9 >0.9 3.3 3.5 19.0
Z=11.28
References
[1] J.F. Hair, C.M. Ringle, and M. Sarstedt, PLS-SEM: Indeed a silver bullet, Journal of Marketing
Theory and Practice, 19(2), pp.139-152, 2011.
[2] W. W. Chin, How to write up and report PLS analyses. In Handbook of partial least squares (pp. 655-
690). Springer Berlin Heidelberg, 2010.
[3] J. Henseler, C.M. Ringle, and M. Sarstedt, A new criterion for assessing discriminant validity in
variance-based structural equation modeling, Journal of the Academy of Marketing Science, 43(1)
pp.115-135, 2015.
[4] C.M. Voorhees, M.K. Brady, R. Calantone, and E. Ramirez, Discriminant validity testing in
marketing: an analysis, causes for concern, and proposed remedies, Journal of the Academy of
Marketing Science, 44(1), pp.119-134, 2016.
[5] C. Fornell, and D.F. Larcker, Evaluating structural equation models with unobservable variables and
measurement error, Journal of marketing research, 18(1), pp.39-50, 1981.
[6] W.W. Chin, The partial least squares approach to structural equation modeling, Modern methods for
business research, 295(2), pp.295-336, 1998.
[7] J.F. Hair, G.T.M. Hult, C.M. Ringle, and M. Sarstedt, A primer on partial least squares structural
equation modeling (PLS-SEM). Sage Publications, 2014.
[8] N. Kock, and G. Lynn, Lateral collinearity and misleading results in variance-based SEM: An
illustration and recommendations, Journal of the Association for Information Systems, 13(7), pp.546-
580, 2012.
[9] J.F. Hair, W.C. Black, B.J. Babin, and R.E. Anderson, Multivariate Data Analysis, NJ, Prentice Hall,
2010.
[10] N. Kock, Common Method Bias in PLS-SEM: A Full Collinearity Assessment
Approach, International Journal of e-Collaboration (IJeC), 11(4), pp.1-10, 2015.
[11] C.M. Ringle, S. Wende, and J.-M. Becker, SmartPLS 3. Boenningstedt: SmartPLS GmbH,
http://www.smartpls.com, 2015.
[12] D. Rezania, and R. Gurney, The Effect of Coaching Practices on Psychological Contract Fulfillment of
Student-Athletes, Physical Culture and Sport. Studies and Research, 71(1), pp.21-29, 2016.
[13] G. Urbonas, L. Kubilien, and A. Urbonien, Pharmacists Job Satisfaction and Its Effect on
Dispensing Precaution Taken at Community Pharmacies, Sveikatos mokslai/Health Sciences, 25(3),
pp.17-21, 2015.
[14] R. Baxter, Reflective and Formative Metrics of Relationship Value: A Commentary Essay, Journal of
Business Research, 62, pp. 1370-1377, 2009.
[15] A. Dickinger, and B. Stangl, Website Performance and Behavioral Consequences: A Formative
measurement Approach, Journal of Business Research, 66, pp. 771-777, 2013.
[16] S. M. Rassolimanesh, N. Dahlan, and M. Jaafar, Tourist Perceived Value and Satisfaction in a
Community-based Homestay in the Lenggong Valley World Heritage Site, Journal of Hospitality and
Tourism Management, 26, pp. 72-81, 2016.

View publication stats

También podría gustarte