19th Australasian Conference on Information Systems Critique of PLS use in IS
3-5 Dec 2008, Christchurch Rouse & Corbitt
845
There’s SEM and “SEM”: A Critique of the Use of PLS Regression in
Information Systems Research
Anne C. Rouse
Deakin Business School, Deakin University
Burwood, Victoria, Australia
Email: anne.rouse@deakin.edu.au

Brian Corbitt
RMIT University
Melbourne, Australia
Email: brian.corbitt@rmit.edu.au
Abstract
In disciplines other than IS, the use of covariance-based structural equation modelling (SEM) is the mainstream
method for SEM analysis, and for confirmatory factor analysis (CFA). Yet a body of IS literature has developed
arguing that PLS regression is a superior tool for these analyses, and for establishing reliability and validity.
Despite these claims, the views underlying this PLS literature are not universally shared. In this paper the
authors review the PLS and mainstream SEM literatures, and describe the key differences between the two
classes of tools. The paper also canvasses why PLS regression is rarely used in management, marketing,
organizational behaviour, and that branch of psychology concerned with good measurement – psychometrics.
The paper offers some practical options to Australasian researchers seeking greater mastery of SEM, and also
acts as a roadmap for readers who want to check for themselves what the mainstream SEM literature has to say.
Keywords
Structural equation modelling, PLS, CFA, reliability
INTRODUCTION
The last twenty-five years have seen a growing use in business research of “structural equation modelling” or
SEM. These are a class of tools for analysing the structure of measures of theoretical (unobserved) constructs
(such as abilities, attitudes, and judgements) and for analysing the relationships between these constructs. Most
of the developments in SEM have come from studies in psychology, where the focus is on psychological
constructs. Since much of information systems (IS) research is also concerned with psychological constructs,
these tools offer the potential for development of better measures, and for identification of statistically-validated
theories based on complex path relationships. Drawing on mainstream texts on SEM and prior SEM training, the
authors set out to explain in layman’s language what SEM does and why it has been embraced so widely for
research involving psychological constructs. We canvass, in particular how it differs from a tool routinely
described, in Information Systems, as SEM, known as “partial least squares” (PLS) regression. Marcoulides and
Saunders (2006) have criticized the tendency by some IS researchers to treat PLS regression as a “silver bullet”
(p iii), and this paper responds to their editorial by alerting IS academics that alternative views exist of SEM,
and of PLS regression, outside of our discipline.
DEFINITIONS — BROAD AND PARTICULAR
Exactly what is “structural equation modelling” (SEM)? At its broadest, SEM involves a number of largely
linear modelling techniques that are based on solving a set of structural equations using matrix algebra. Using
this definition, SEM would include most of the multivariate statistical tools including multiple regression, path
analysis, factor analysis, and principal components analysis (see SEMNET, 2008b). It is in this sense that the
analytical tool used widely in information systems research – partial least squares regression or PLS – can be
said to be a type of “structural equation modelling”. However, outside of information systems, this broad use of
the label SEM is rare (a claim that can be confirmed by consulting the range of mainstream SEM texts listed in
the reference section).
The broad definition of SEM above would also embrace what is more typically described as “structural equation
modelling” in the disciplines of psychology, marketing, education, and management. This introduces an element
of confusion because the same term denotes quite different things. In those disciplines the term SEM is reserved
to mean that particular class of analytical tools that are based on iterative analysis of covariances and calculation

19th Australasian Conference on Information Systems Critique of PLS use in IS
3-5 Dec 2008, Christchurch Rouse & Corbitt
846
of a “discrepancy measure”. A covariance is the unstandardized (or “raw”) form of a correlation. SEM software
(including LISREL, AMOS, EQS, but not PLS) uses algorithms and analytical methods initially developed by
the psychologist Karl Joreskog in the 1970s.
In the class of modelling tools defined by SEM (the term used in the particular), a theoretical structural equation
model is developed. This model predicts, or “implies” a particular structure for the data-based covariance
matrix. Once the theoretical model's parameters have been estimated using iterative matrix algebra, the software
attempts to produce an implied matrix based on the theoretical model that as closely as possible matches the
data. The implied (or theoretical) covariance matrix can then be compared to the actual covariance matrix for
the data that has been gathered. If the discrepancy between these two matrices is statistically small, then the
structural equation model can be considered a statistically plausible explanation for relations between the
measures. SEM is thus designed to maximize, then test, the degree of consistency between the theoretical model,
and the actual data (Byrne 2001; Kline 2005; Rigdon 1996).
The development of this class of statistical tools was a major breakthrough in social and business research,
because till they became available, the only way of establishing whether the gathered data matched theory was
to test whether it was possible to predict one or more outcomes well in the particular sample (with the
assumption that this prediction would be valid in other samples). However being able to predict a variable (say
through regression) is less powerful than being able to account for (explain) the whole pattern of co-variation
within the gathered data. In addition, by partitioning out the effects of measurement error (discussed below),
SEM reduces the attenuation in regression paths that results from this error. In other words, with SEM, it is
possible to obtain more accurate measures of the path coefficients, so revealing more of the underlying
relationship between latent variables.
To avoid confusion, in this paper, the authors will use the more generally-accepted particular definition of SEM
─ a statistical technique based on analysis of covariance structures that explicitly models measurement errors
and seeks to derive unbiased estimates for the relationships between latent constructs (SEMNET, 2008b). One
developer of PLS software – Wynne Chin – is critical of this particularizing of the term SEM by social scientists
(Chin 1998). However, since almost no business researchers typically describe other forms of regression or
factor analysis as “structural equation modelling” it is Chin’s use of the term SEM to signify the broader
meaning that is atypical. Furthermore, this atypical use can lead to confusion, and to beliefs that IS researchers
have mastered “SEM” when in fact they have little knowledge of SEM developments in reference disciplines.
SEM IN INFORMATION SYSTEMS VS. OTHER BUSINESS DISCIPLINES
In psychology, and in the range of business disciplines (including IS) that gather psychological data, there has
been a steady increase in the use of SEM to understand the underlying structure of unobservable psychological
constructs that are reflected in quantitative measures (such as survey responses). A psychological construct is a
theoretical, unobserved mental concept that is reflected in statements expressed by interviewees, or by responses
to survey measures. In psychology, and those disciplines that call on psychology, such as management and
marketing, SEM is increasingly replacing traditional tools like exploratory factor analysis for establishing the
unidimensionality, reliability and construct validity — including discriminant and convergent validity — of
quantitative measures (Ketchen and Bergh, 2006).
In the quantitative IS literature, too, there has been a growing reliance on what is termed “structural equation
modelling” for these purposes. However, rather puzzlingly, in our discipline the tool of choice is increasingly no
longer one of the commonly used SEM software tools (which tend to all output similar results if the same
estimation method is used). Instead it is increasingly PLS regression — although this PLS analysis is invariably
described as “structural equation modelling”. A perhaps-surprising observation is that outside of Information
Systems, the use of PLS regression is rare in mainstream “Tier 1” research journals. This is illustrated in Table 1
below. This tabulates the numbers of PLS regression studies reported in the business literature between 2004
and 2006. Table 1 is based on searches within Proquest/ABI Inform Global and Business Source
Premier/Academic (electronic databases) for articles reporting the use of PLS, or CFA based on PLS. The
journals in which these appear were classified according to the recent ERA (Excellence in Research for
Australia) listings using John Lamp’s web-based search engine (Lamp 2008). Not all Tier 1 IS journals are
captured by these databases, though Tier 1 journals in marketing and management have good coverage. This
means that the method employed will tend to understate the proportion of PLS articles that appears in the IS
literature.
Table 1 illustrates that in disciplines other than Information Systems, PLS is rarely used in Tier 1 (A* and A
level) journals. Instead, most PLS studies are reported in less well-established journals (Classified as B, C, or
unclassified by the ERA). Table 1 suggests that, for journals where audiences (and reviewers) are typically
required to undergo substantial quantitative methods training, PLS is generally not seen as an appropriate
research tool for complex multivariate research. Goodhue et al (2006) reported a similar finding based on their