Travel Reference
In-Depth Information
He points out that '[. . .] IRT is not a panacea for all measurement woes [. . .] rather, IRT is just
another approach that rests on different assumptions than CTT and, as a consequence, allows
researchers to tackle measurement issues differently' (2004: 185). Seeking a stronger link between
CTT and IRT would certainly improve the quality of measurement scales while taking into
account each paradigm's strengths and weaknesses.
Structural Equation Modelling
From the mid-1990s onwards, applications of SEM have experienced a tremendous upswing.
Implementations embrace a variety of themes; for a comprehensive overview and a discussion of
critical issues in the modelling process the reader is referred to Reisinger and Mavondo (2006).
The fi rst article published in the fi eld of tourism research is a JTTM paper by Bartkus (1995).
He assesses a mediation model where expertise acts as a mediator for the effect of work
experience on travel agents' sale performance.
In early occurrences of SEM in tourism marketing, authors praise SEM's methodological
usefulness, explain the basics and objectives of this method, give advice on what and how to
report and encourage applications in tourism research. This effort culminates in Reisinger and
Turner's (1999) paper providing a step-by-step guidebook to SEM based on the mainstream
software known as LISREL. They clearly point to the necessity of cross-validating one's results if
the model has been modifi ed and suggest using comparisons of competing or nested models in
order to strive for sound theory building and testing (Steenkamp and Baumgartner 2000). Even
though the paper is frequently cited (169 times, or 12.1 times per year according to Google
Scholar, July 2012) this advice does not seem to have become second nature to tourism
researchers.
Some general issues that have been critically noted by a number of authors deserve closer
attention.
Exploratory versus confi rmatory analysis
A perennial problem is the heavy reliance on a single cross-sectional data set for testing a
hypothesized model that is then repeatedly calibrated on the same data to arrive at a fi nal
model with good fi t. Clearly, this approach is prone to suffer from sample idiosyncrasies
(Baumgartner and Homburg 1996). Nevertheless, the vast majority of papers rely on a single
sample. Authors should at least admit that once the originally hypothesized model gets modifi ed
(to increase its fi t) SEM loses its confi rmatory character. While the most desirable approach to
deal with this problem is testing the fi tted model on a fresh sample, this is often not possible.
A random sample split is advisable. For a recent sample application, see Hallak, Brown and
Lindsay (2012).
Multinormality assumption
The most frequently applied estimation procedure is Maximum Likelihood (ML), due to its
default status in standard software packages. This method is based on multinormally distributed
data (Reinartz, Haenlein and Henseler 2009). Unsurprisingly, given the ordinal and categorical
data prevailing in most SEM applications in tourism research, this assumption is quite often not
met (Mazanec 2000). Even though a recent simulation study shows that ML estimation is fairly
robust against violations of the multinormality assumption (Reinartz
et al
. 2009), researchers
should be aware of this defi ciency and cross-validate with second-generation estimation software
