Download as pdf or txt
Download as pdf or txt
You are on page 1of 21

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

net/publication/221942822

A paradigm for examining second-order factor models employing structural


equation modeling

Article  in  International Journal of Production Economics · August 2009


DOI: 10.1016/j.ijpe.2009.04.010

CITATIONS READS

203 4,266

3 authors, including:

Xenophon Koufteros Mehdi Kaighobadi


Texas A&M University Florida Atlantic University
67 PUBLICATIONS   5,471 CITATIONS    8 PUBLICATIONS   381 CITATIONS   

SEE PROFILE SEE PROFILE

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

Resilience View project

Buyer-Supplier Relationships View project

All content following this page was uploaded by Xenophon Koufteros on 14 October 2018.

The user has requested enhancement of the downloaded file.


ARTICLE IN PRESS
Int. J. Production Economics 120 (2009) 633–652

Contents lists available at ScienceDirect

Int. J. Production Economics


journal homepage: www.elsevier.com/locate/ijpe

A paradigm for examining second-order factor models employing


structural equation modeling
Xenophon Koufteros a,, Sunil Babbar b, Mehdi Kaighobadi c
a
Information & Operations Management, Mays Business School, Texas A&M University, 320 Wehner Building, College Station, TX 77843-4217, USA
b
Department of Information Technology & Operations Management, Barry Kaye College of Business, Florida Atlantic University, 777 Glades Road, Boca Raton,
FL 33431-0991, USA
c
Department of Information Technology & Operations Management, Barry Kaye College of Business, Florida Atlantic University, 2912 College Avenue,
LA Building 485, Davie, FL 33314, USA

a r t i c l e in fo abstract

Article history: We present here a paradigm for assessing second-order measurement models. Our
Received 20 June 2007 approach is hierarchical in nature. We discuss the need for higher-order models from a
Accepted 3 April 2009 conceptual perspective and illustrate how some common challenges in empirical research
Available online 24 April 2009
can be resolved through the deployment of higher-order modeling. Essentially, we argue
Keywords: that many constructs can be meaningfully described by a higher-order structure and
Research methodology testing for the existence of such structures requires a careful examination of alternative
Second-order factor models. There is a need for conceptual as well as empirical support. In order to
Measurement models demonstrate our paradigm, we use data that relate to airline service quality. Our sample
Structural equation modeling
includes two databases. Responses from 170 individuals are employed for exploratory
Service quality
purposes and responses from 437 individuals are used for subsequent data analyses.
Airlines
& 2009 Elsevier B.V. All rights reserved.

1. Introduction can be derived either from mathematical or analytical work


or from theory. The empirical testing of hypotheses in
Theory-driven and empirically based operations man- research is now more common place than ever before (Thun,
agement (OM) research has surfaced in the last 20–25 2008; Farris et al., 2009; Liljeblom and Vaihekoski, 2009).
years and its growth is remarkably strong. Over the past As the field of OM evolved and began to assemble an
four to five years alone the number of empirically based empirical tradition, new and more powerful methodolo-
studies submitted to academic journals more than gies have been adopted. While early empirical studies
doubled. Academics in the OM field came to the realiza- were relatively descriptive and inattentive to measure-
tion that mathematical modeling, which has been the ment issues, the field gradually moved towards more
primary research methodology in the field, is useful but sophisticated techniques such as exploratory factor ana-
empirical validation and assessment is also necessary. For lysis (Zhang and Chen, 2008) and path analysis. Even-
example, while a mathematical model may suggest that tually, the OM field caught up with other disciplines such
organizing work around manufacturing cells appears to be as Marketing and Psychology, which have a much longer
the most effective approach in a given environment, and richer tradition in empirical research, and started
empirical assessments may indicate otherwise. Empirical employing structural equation modeling (SEM) as a
research can render or deny support to hypotheses, which methodology of choice (Pal and Busing, 2008). Structural
equation modeling allows a researcher to test both a
 Corresponding author. Tel.: +1 979 845 2254; fax: +1 979 845 5653. measurement model and a structural model (substantive
E-mail addresses: Xkoufteros@tamu.edu (X. Koufteros), model) and affords an assessment of model fit and
Babbar@fau.edu (S. Babbar), Kaighoba@fau.edu (M. Kaighobadi). individual parameters through an array of fit-indices and

0925-5273/$ - see front matter & 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.ijpe.2009.04.010
ARTICLE IN PRESS
634 X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652

tools. Paradigms on the use of structural equation higher-order modeling from a conceptual as well as
modeling have appeared in the OM field (i.e., Koufteros, methodological perspective. We present some of the
1999) but several SEM techniques still remain unexplored. challenges that can afflict empirical research and demon-
Only recently, Koufteros and Marcoulides (2006) pre- strate some adverse affects that result from measurement
sented a paradigm on the use of multi-group analysis and model misspecifications. While we examine in depth only
factorial invariance tests. Studies that employ advanced a few principal challenges, such challenges can prove to be
SEM techniques such as higher-order modeling, latent consequential when testing substantive hypotheses.
growth modeling, and multilevel modeling remain scant. This exercise is predicated on presenting the constructs
The purpose of this manuscript is to present a of interest along with a description of the research design
paradigm on higher-order modeling and in particular, and the samples employed. We illustrate our paradigm for
second-order modeling. The OM literature has not testing second-order models and provide directions for
produced a sizable volume of manuscripts that employ future research along with a discussion. The survey data
higher-order modeling and those studies that have been we use is based on the US airline industry and includes
published or which we came across while serving as responses from 170 individuals for exploratory work and
editorial board members/reviewers of academic journals, from 437 for subsequent data analyses.
suffer from a variety of shortcomings. Often, there is no
explicit conceptual/theoretical advance that relates the
lower-level constructs to higher-level constructs. The 2. A need for higher-order modeling: conceptual and
assumption is that a higher-order structure does exist methodological issues
but no attempt is made to conceptually support why a
given construct can be specified at a higher-level of There are several constructs which can be mean-
abstraction. From a methodological perspective, some ingfully conceptualized at higher orders of abstraction.
researchers sum up the scores of manifest variables (the In such cases, a higher-order modeling approach would be
items) into an aggregate score (called ‘‘partial aggrega- the most suitable technique that can represent such
tion’’) for each sub-dimension (i.e., first-order construct). structures. Consider, for example, that an organization’s
They then treat these aggregates as manifest variables service quality (third-order level of abstraction) can be
anew, specifying them as reflective items of a higher-order conceptualized as a composite of two dimensions such as
construct (which in reality, however, is now specified at company deliverables (what the firm provides) and employ-
the first-order level). While this approach may be simple ee deliverables (what the firm’s employees furnish during
and swift, it fails to fully account for the variability of each the service encounter) which are at the second-order level
of the manifest variables, as their scores are summed up of abstraction (Fig. 1). Employee deliverables in turn can
into aggregates. This approach is also inconsistent with include several first-order latent variables such as in-
the conceptual specification of higher-order modeling as dividual attention, courtesy, promptness, and helpfulness
the higher-order construct is now specified at the first- that can be represented by observed or manifest indica-
order of abstraction. Yet other researchers specify a tors (the items on a survey for example). Likewise,
higher-order model correctly (relating manifest variables company deliverables may include fair policies, safe flights,
to first-order constructs and then first-order constructs to desirable assortment of flights, and reliable equipment.
second-order constructs etc) but never assess the efficacy Customers can aggregate their evaluations of the first-
of other measurement structures that can describe the order latent variables or facets to form their perceptions
data (e.g., Mentzer et al., 1999; Kettinger and Lee, 1994). It on the second-order dimensions and subsequently aggre-
is possible that other measurement models can represent gate evaluations of second-order dimensions to derive
the data equally well or even surpass the performance of perceptions of a higher-order overall construct.
second- or higher-order configurations. In the absence of A higher-order model can be posited and can relate the
testing for alternative measurement models, it is hazar- manifest variables to their respective first-order latent
dous to suggest that a higher-order specification is the variables which can then be related to their second-order
most suitable. latent variable(s). Second-order latent variables can be
To illustrate the efficacy and usefulness of higher-order specified as dimensions of third-order latent variables, if
modeling, we will make use of several first-order there is a conceptual and theoretical rationale for such
constructs from the service quality literature although advance. The contribution of each dimension to a higher-
the principles presented here are widely applicable. level construct can be assessed and delineated as
Specifically, we will employ a sample of facets of company compared to bundling all items together in a single
deliverables to be expected of any airline service offering. composite score. If all items are bundled together
These include fair policies, safe flights, desired assortment of (through just one first-order latent variable for example),
flights, and reliable equipment. These company deliverables the explication of the resultant construct is incomplete
do not include the deliverables expected of the employees (Gerbing et al., 1994) and the contribution of various
of airlines and can be captured by a different second-order content domains to the final scale score will not be
construct. known. In other words, if all items/manifest variables of
Each of these facets is operationalized through multi- individual attention, helpfulness, responsiveness etc are
ple manifest variables. In order to fully appreciate higher- posited as reflective items of a single first-order construct,
order modeling, customer satisfaction will assume the role then it would be difficult to ascertain the contribution of
of a dependent variable. We first elaborate on the need for each domain on the overall construct.
ARTICLE IN PRESS
X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652 635

Third-Order of
Abstraction

Airline
Service
Quality

Second-Order of
Abstraction

Employee First-Order of Company


Deliverables Deliverables
Abstraction

Desirable
Individual Fair Safe Reliable
Courtesy Promptness Helpfulness Assortment
Attention Policies Flights Equipment
of Flights

… CR1 CR2 CR3 CR4 ………………….. SF1 SF2 SF3 …………..

Manifest Variables
Fig. 1. An example of hierarchical structures.

Beyond higher-order modeling being a more elegant factor analysis. Though factor analysis applied to a group
representation of hierarchical constructs, it can serve to of items intended to be measures of a single facet can
alleviate numerous methodological problems that chal- indeed appear to produce a single meaningful factor,
lenge empirical studies. Nobody would deny for example when factor analysis is applied simultaneously across all
that promptness, courtesy, helpfulness, and individual items of all constructs there is a high likelihood that fewer
attention are important deliverables in a service encoun- meaningful factors will be extracted and manifest items
ter. Most would also agree that each of these latent might exhibit cross-loadings with non-intended factors.
variables is ‘‘different’’ or it has something idiosyncratic This is worrisome as composite scores of first-order
about it. However, it frequently happens that a person factors will be difficult to assemble and meaningfully
who is helpful is also courteous and pays individual describe and then employ in substantive hypotheses
attention to customers. Thus, these variables would be testing. Second-order models recognize the contribution
expected to be highly correlated, which then raises issues and retain the idiosyncratic nature of each first-order
of critical importance here. If all of these variables (e.g., construct, such as courtesy or promptness, and treat such
courtesy, individual attention, helpfulness, promptness) constructs as facets of the higher-order construct. Path
which are at the first-level of abstraction are entered coefficients relate each first-order construct to its second-
concurrently into a model that posits them as antecedents order construct and thus both the substantive and
to customer satisfaction, a researcher will soon find that statistical contribution of each first-order construct can
perhaps only one or none of these variables is statistically be estimated and evaluated. The second-order construct,
related to customer satisfaction. Or even, one may find which is now a composite of the first-order constructs, can
that the path coefficients may assume the opposite sign be posited as an explanatory variable to an endogenous
than what might be expected. This can be symptomatic of variable such as customer satisfaction.
the degree of multicollinearity present in the model. On Employing higher-order modeling can also resolve
an individual basis, each of these variables can have a problems with what Catell (1973) coined as ‘‘bloated
statistically significant relationship with an endogenous specifics’’ factors. These are factors/constructs/first-order
variable such as customer satisfaction but when tested as latent variables, which are characterized/anchored exclu-
a group, such effects might not be manifested. sively by a narrow content span of manifest/observed
Highly correlated constructs will also fail to perhaps variables. Gerbing et al. (1994) explain that as researchers
produce a ‘‘clean’’ factor structure when subjected to attempt to satisfy multiple-indicator unidimensionality
ARTICLE IN PRESS
636 X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652

concerns (especially when tested through sophisticated course, different determinants of aviation safety such as
structural equation modeling methodology), they often pilot attributes and aircraft equipment. Research has
concoct manifest variables, which are more or less accordingly examined some of these factors (McFadden,
‘‘restatements of each other.’’ While such items can satisfy 1996; Sanders and Hoffman, 1975; Sanders et al., 1976;
rigorous statistical criteria and can exhibit superb psycho- McFadden, 1997; Ford, 1997; Sheikh et al., 1996).
metric properties (e.g., composite reliability), such narrow Desirable assortment of flights: Selecting a particular
content domain curtails explanatory power and theore- airline to fly with is in part a function of how different
tical usefulness (Gerbing et al., 1994). Gerbing et al. state airlines position themselves and the routes they service.
that content tapped by bloated specific factors can be too Product positioning strategies include consideration of the
narrow to correspond to a substantive construct of breadth of offerings (Kalafatis et al., 2000; Vaagen and
interest. They consider instead, that such factors (i.e., Wallace, 2008) and research has recognized product
first-order factors) can be posited as the constituent facets selection as a predictor of retail choice and patronage
of constructs of interest. These facets are the ‘‘building (Arnold et al., 1983; Koelemeijer and Oppewal, 1999; Pan
blocks’’ of the second-order constructs. Each facet can be and Zinkhan, 2006) across varied service industries
defined by a unidimensional set of items and facets can be (Lindsley et al., 1991; Tamimi et al., 2005; Carpenter and
treated as the indicators of second-order factors. Moore, 2006). Availability of flight from the customer’s
Many of the aforementioned complications are fre- city of origin to the desired destination can be even more
quently encountered in empirical research studies but important than price in the air-travel industry just like
little effort has been exerted in evaluating higher-order assortment in a retail store can be as, if not more,
conceptualizations and how they can alleviate such important than a number of other key variables including
adverse effects. Before we illustrate in practice some of price (Stassen et al., 1999). It is important for a firm to link
these complications and a paradigm of examining higher- its product offerings to the needs of consumers (Heuvel et
order models, it is rather useful to discuss the constructs al., 2007) and to evaluate its offerings in an effort to
of interest, the instrument development process and position itself well (Penttinen and Palmer, 2007). Such
research design along with sample characteristics. consideration is essential for airlines as they attempt to
better serve their customers by partnering with other
airlines and servicing desired routes with desirable flight
3. Constructs of interest schedules.
Reliable equipment: Safety and reliability of equipment
There is indeed a wide array of elements shaping go hand in hand (Waikar and Nichols, 1997). Product
customer experience with an airline. We focus our reliability is critically important for both manufacturing
attention on a given set of company deliverables that are (Barber and Darrough, 1996) and service (Parasuraman et
to be expected of airlines. al., 1985, 1988) and helps create value for firms (Barber
Fair policies:Seiders and Berry (1998) provide an and Darrough, 1996) while shaping their performance
insightful treatise on the issue of fairness in service and (Reed et al., 1996). In this study, we limit the scope by
show how and why fairness matters to service customers. focusing attention on requisites of reliability as they relate
Fairness is closely related to justice and these terms are to the aircraft and equipment during transport. Unreliable
often used interchangeably (Sheppard et al., 1992; equipment may lead to flight delays, which subsequently
Colquitt et al., 2001) with fairness being defined as the can have an adverse effect on customer satisfaction.
customer’s perception of the degree of justice rendered by Collectively, customers expect an airline to offer fair
the service provider (Seiders and Berry, 1998). Social policies and a desirable selection of flights. Customers also
exchange theory (Adams, 1965) provides the foundation expect that the airline will assure the safety of flight
for assessments of justice and fairness, and perceptions of operations and the reliability of equipment so potential
justice are critical predictors of customer satisfaction flight delays and cancellations can be avoided. The four
(Martinez-Tur et al., 2006). Oliver and Swan (1989) helped constructs described above can be construed as facets of
introduce the justice framework into investigations of airline deliverables during the execution of the service
customer satisfaction with goods and services. encounter.
Safe flights: Considering the sensitivity of passengers to
safety matters, flight safety becomes paramount in
importance, and safety management an important aspect 4. Instrument development, research design and sample
of quality control in the aviation industry (Waikar and characteristics
Nichols, 1997). Safety and health issues are in general
receiving considerable attention in the management The service quality literature was extensively reviewed
literature (Law et al., 2006). In recent years, safety has in order to develop the measurement instrument. This
received particular attention in industries such as health literature spans numerous disciplines including Market-
care (Katz-Navon et al., 2005; McFadden et al., 2006; ing, Operations Management, Management Information
Walsh and Antony, 2007), food and agriculture (Sans et al., Systems, Hospitality Management, and Health Care Man-
2005; Griffith, 2006; Trienekens and Zuurbier, 2008), and agement and is influenced by the seminal work on
aviation (Ford, 1997; Waikar and Nichols, 1997; Rhoades SERVQUAL and its subsequent refinements. The SERVQ-
and Waguespack, 1999; McFadden and Hosmane, 2001; UAL instrument (Parasuraman et al., 1988, 1991) sepa-
Rhoades and Waguespack, 2004; Rose, 2006). There are, of rately measures the expected level of service and the
ARTICLE IN PRESS
X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652 637

experienced level of service. Service quality scores are (Table 1). Students also come from all walks of life with
then based on the difference between these two mea- substantial variation in their reported household incomes.
sures. The SERVQUAL instrument, despite being exten- A large number of these specific respondents are con-
sively used, has been controversial and has lead to sidered to be ‘‘non-traditional’’ students as they worked
modifications and the emergence of other competing and attended school at the same time.
instruments (e.g., SERVPERF—Cronin and Taylor, 1992;
TOPSIS—Mukherjee and Nath, 2005). Van Dyke et al.
(1997, 1999) provide an excellent summary of the 5. Complications/challenges with structural and
difficulties surrounding SERVQUAL. As such, we will not measurement models
engage in an exhaustive rendition of all of the conse-
quential issues pertaining to SERVQUAL’s performance. There are numerous complications/challenges that can
The difficulties with SERVQUAL are both conceptual and emerge without a careful examination of alternative
empirical. Conceptual difficulties include the operationa- measurement and structural model structures. One major
lization of perceived service quality as a difference or gap complication occurs when several exogenous variables
score, the ambiguity of the expectations construct, and the can be shown to each exhibit strong effects on a
unsuitability of using a single measure of service quality dependent variable, but when all are entered simulta-
across different industries. Empirical problems relate to neously in a single structural model, the effects may
reliability, poor convergent and discriminant validity, vanish and/or exhibit an opposite than expected sign.
unstable dimensionality, and poor predictive validity Consider for example the four first-order constructs of
(Van Dyke et al., 1997, 1999). Many of the empirical company deliverables (Fig. 2). Table 2 and Fig. 3 show the
problems stem from the use of gap scores and researchers results of four structural models where each of the
have suggested avoidance of the use of discrepancy scores constructs is posited as an exogenous variable and
for measuring service quality. satisfaction is hypothesized as the endogenous variable.
Despite its limitations, the SERVQUAL instrument does All the models exhibit adequate model fit and very strong
afford some of the most salient constructs of service effects are illustrated through t-values (minimum t-value
quality and it served as a basis to draw and adapt items is 5.73). The importance of each exogenous variable
from. We supplemented this with an examination of the cannot be denied. On the other hand, consider Table 3
marketing and service operations management literatures which shows the results of a model (Fig. 4) that examines
in which numerous other scales that measure service the simultaneous effects of all four exogenous variables on
quality have emerged since the influential publication of satisfaction. Though the model acknowledged the correla-
the SERVQUAL instrument. In this research, we focus on tions amongst the exogenous variables, it produced
the essential elements of airline service. Each respondent moderate fit indices and importantly, only fair policies
was first asked to identify the airline s/he uses most often exhibits a statistically significant relationship (t-value ¼
for domestic (US) flights and then rate that airline relative 6.46) with satisfaction. Moreover, the effects of desirable
to what the respondent’s expectations were for that assortment of flights (t-value ¼ 1.69) and reliable equip-
specific airline. After generating the items, we tested the ment (t-value ¼ 1.06) have a negative sign. We would
constructs and items through a formal pre-test study. We have expected that the effects would all be statistically
provided each construct, along with its definition and list significant and positive. Erroneously perhaps, one may
of items, to twelve faculty members from several large conclude that only fair policies have a meaningful relation-
universities in the United States representing multiple ship with satisfaction.
disciplines such as Marketing, Service Operations, Infor- Table 4 illustrates the results of yet one more model
mation Systems, and Strategy. These faculty members where all manifest variables are hypothesized to be
evaluated each item based on how each construct was related to just one first-order construct (Fig. 5). This
defined and provided additional comments as they relate construct exhibits a strong effect on satisfaction (t-value ¼
to the content domain of each construct. Feedback from 7.77) but model fit is lacking. In addition, ascertaining
them was useful and helped us reword and simplify some which particular aspects of company deliverables are
of the items. important for satisfaction becomes more problematic as
For exploratory purposes, the survey questionnaire was 14 manifest variables contribute directly to the composite.
administered to 170 respondents who were students Some are measures of fair policies, some are measures of
enrolled in a large university in the southeastern region safe flights, others are measures of desirable assortment of
of United States. The survey in its final form was flights and others are measures of reliable equipment.
administered to students (n ¼ 437) also enrolled at the When constructs are highly correlated (Table 5),
same university. Responses for this large-scale study were discriminant validity may also be difficult to support.
collected throughout the course of 2006 (the spring, Exploratory factor analysis (n ¼ 170) may fail to produce a
summer, and fall semesters). The respondents were all ‘‘clean’’ factor structure and thus the composition and
students majoring in Business. Although we acknowledge naming of factors can become difficult. While each
that our respondents may not be representative of the company deliverable exhibits a unidimensional factor
larger US population, students form an eligible respon- structure and a relatively high composite reliability
dent group as they also fly just like a great number of coefficient (Tables 5 and 6), a factor analysis of all 14
other residents of the US do. Respondents did have a fairly items fails to produce a ‘‘clean’’ factor structure (Table 7).
decent level of experience flying with various airlines Employing the sample of respondents derived for a pilot
ARTICLE IN PRESS
638 X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652

Table 1
Demographics.

Demographic Scale

I typically fly: Less than once a Once a year Twice a year Four times a year Once a month Once a More than once a
year week week

Frequency 78 96 123 107 28 2 1


Valid percent 17.9 22.1 28.3 24.6 6.4 0.5 0.2

I am: Male Female

Frequency 190 246


Valid percent 43.6 56.4

My age is: r20 years 21–30 years 31–40 years 41–50 years 51–60 years 460 years

Frequency 32 317 46 31 9 1
Valid percent 7.3 72.7 10.6 7.1 2.1 0.2

I fly most often for: Leisure Business

Frequency 379 57
Valid percent 86.9 13.1

My ethnic background is: Black Native Hispanic White Asian/Pacific Other


American Islander

Frequency 64 2 95 226 30 14
Valid percent 14.8 0.5 22.0 52.4 7.0 3.2

My annual household o$20,000 $20,001–40,000 $40,001–60,000 $60,001–100,000 4$100,000


income is:

Frequency 79 110 71 101 63


Valid percent 18.6 25.9 16.7 23.8 14.9

The class I most often fly in Economy Business First class


is:

Frequency 409 16 10
Valid percent 94.0 3.7 2.3

study (n ¼ 170) and Principal Axis Factoring as the with Xs, the measurement errors of the indicators with the
method of extraction with Oblimin rotations, Table 7 Greek letter d (delta), and the true correlation between
shows that only three factors are extracted, explaining the latent variables with the Greek letter f (phi). This
about 45% of the variance. Manifest variables from three specification is called X-model within LISREL terminology.
content domains load on the first factor while the second As can be seen in the path diagrams, LISREL performed the
factor includes items from two domains and some of its confirmatory factor analysis on each of the latent
items have cross-loadings above 0.20 with factor 3. Only variables.
two items (both from the fair policies domain) load on The posited measurement models seem to be sup-
factor 3 but they exhibit cross-loadings above 0.20 with ported by the various fit indices. There is evidence of very
factor 1. It is evident that discriminant validity cannot be strong relationships between observed and latent vari-
supported. A confirmatory analysis using maximum like- ables (t-values) and fit indices are suggestive of well-
lihood estimation and a covariance matrix as input (using fitting models (see Table 6). The descriptive statistics were
the large sample n ¼ 437) shows similar findings. presented in Table 5.
Confirmatory factor analysis for each exogenous vari- Relatively high levels of convergent validity (t-values
able (Table 6) has resulted in adequate model fit for each that describe the relationship between manifest variables
first-order factor (e.g., NNFI, CFI, and standardized RMR). and latent variable, desirable to be 4|2|), composite
Criteria for evaluation of model fit can be found in Byrne reliability (desirable to be 40.70), and average variance
(1998), Hu and Bentler (1999), and Koufteros and extracted (AVE) (measure of within variance, desirable to
Marcoulides (2006). Fig. 2 previously depicted the path be 40.50) (Koufteros, 1999) are also visible (Table 6).
diagrams comprising the measurement models of the five However, discriminant validity cannot be supported when
latent (i.e. unobservable) variables and their respective all four exogenous variables are considered. Discriminant
measurement indicators. We followed the LISREL termi- validity can be examined by comparing the squared
nology as outlined by Jöreskog and Sörbom (2001) and correlation between any two constructs against their
denoted the five latent variables with the Greek letter x individual AVE. Discriminant validity is exhibited when
(ksi), the (observed) measurement indicators or items each AVE is higher than the squared correlation between
ARTICLE IN PRESS
X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652 639

Fig. 2. Individual measurement models for exogenous and endogenous variables.

Table 2
Four structural models (one for each first-order factora).

Model Exogenous variable Completely standardized effect (t-value) Model fit indices

1 Fair policies 0.99 (5.73) NNFI ¼ 0.88, CFI ¼ 0.90, std. RMR ¼ 0.85
2 Safe flights 0.98 (7.57) NNFI ¼ 0.93, CFI ¼ 0.94, std. RMR ¼ 0.064
3 Desirable assortment of flights 0.99 (8.06) NNFI ¼ 0.89, CFI ¼ 0.91, std. RMR ¼ 0.078
4 Reliable equipment 0.98 (12.94) NNFI ¼ 0.91, CFI ¼ 0.93, std. RMR ¼ 0.066

a
Customer satisfaction is posited as the dependent variable in each model.

constructs (Fornell and Larcker, 1981). In almost all data. This exercise is necessary as there could be multiple
comparisons, there was no evidence of discriminant measurement model configurations that can characterize
validity and this can be attributed to the fact that the a given data set. The effort is hierarchical in nature.
factors are highly correlated (Table 5).
As we discussed and showed in Sections 2 and 5, there
are numerous complications/challenges that can potentially 6.1. Hierarchical approach
surface in empirical research. Many of these issues can be
effectively resolved with a conceptualization and specifica- A hierarchical approach is a systematic process for
tion of higher-order measurement and structural models. evaluating alternative models that have the potential to
Within the context of higher-order modeling, the idiosyn- describe relationships between observed and latent vari-
crasy of each facet can be retained while issues of ables. This process includes the construction of four
multicollinerarity, unidimensionality, discriminant validity, Models (as shown in Figs. 6–9). The first model (Fig. 6)
and ‘‘bloated’’ specific factors can be handled effectively. is hypothesized to include one first-order latent variable
with 14 observed indicators. Model 2 (Fig. 7) hypothesizes
four first-order uncorrelated (orthogonal) factors. Model 3
6. A paradigm on second-order modeling (Fig. 8) is similar to Model 2 except that factor correlations
are specified; and Model 4 (Fig. 9) includes one
Our paradigm begins here with the assessment of second-order factor and four first-order factors with
alternative measurement models that can describe the corresponding observed variables. This particular type of
ARTICLE IN PRESS
640 X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652

ST1 ε1
ST1 ε1
λi1 ST2 ε2
λi1 ST2 ε2
δ1 FR1
λi1 ST3 λi1
ε3
δ1 SL1 ST3
ε3
δ2 FR2
Fair Satisfaction
ST4 ε4
Policies
η1 ST5
δ2 SL2 Desirable
Satisfaction
ST4 ε4
δ3 FR3 ξ1 ε5 Assortment
δ3 SL3
of Flights η1 ST5
ε5
δ4 FR4
ST6 ε6 ξ1
ST7 δ4
ST6 ε6
ε7 SL4
ST7
ST8 ε7
ε8
ST8
ε8
ST9 ε9
ST9 ε9

ST1 ε1
ST1 ε1 λi1 ST2 ε2
λi1 ST2 ε2 λi1
ST3
δ1 ER1 ε3
λi1 ST3
ε3 Reliable
δ1 SF1
δ2 ER2 Equipment
Satisfaction
ST4 ε4
Safe Satisfaction
ST4 ε4 ξ1 η1 ST5
δ2 SF2 Flights ε5
ξ1 η1 ST5
ε5 δ3 ER3
δ3
ST6 ε6
SF3 ST6 ε6 ST7 ε7
ST7 ε7 ST8
ε8
ST8
ε8 ST9 ε9
ST9 ε9
Fig. 3. Individual structural models.

Table 3 values of the structural coefficients (i.e., gammas and


One structural model (four first-order factorsa). betas) can be used to test the research hypothesis. The
magnitudes of the structural coefficients and the statis-
Exogenous variable Completely Model fit indices
tical significance associated with such coefficients are
standardized
effect (t- useful in the assessment of potential relationships
value) between variables. The analysis can also be complemen-
ted by an evaluation of R-squared.
Fair policies 0.76 (6.46)
Safe flights 1.66 (1.63) NNFI ¼ 0.89, CFI ¼ 0.90,
std. RMR ¼ 0.065 6.2. Second-order specification and identification
Desirable assortment of 0.77 (1.69)
flights Jöreskog (1970) introduced confirmatory second-order
Reliable equipment 0.75 (1.06)
factor analysis (based on Thurstone’s (1947) treatment of
a second-order exploratory factor analysis) where each
Customer satisfaction is posited as the dependent variable in the
model. first-order factor is presumed to be affected by two
components. The second-order factor is hypothesized to
account for or explain all of the covariance among the
second-order model, which shows first- and second-order first-order factors. One component, the construct of
indicators being reflective, is frequently encountered in interest (x), is shared with the other first-order factors
the literature (Jarvis et al., 2003) and thus it is addressed (Z). The other component is unique to that factor (z) The
here. It is called a ‘‘total disaggregation second-order following equation relates a first-order factor to the two
factor model’’ (Bagozzi and Heatherton, 1994). Our components:
methods are reflective of the body of literature that posits
and tests higher-order models (e.g., Rindskopf and Rose,
Z ¼ Gx þ z
1988; Arnau and Thompson, 2000; Somers et al., 2003; where G is the pattern coefficient (regression paths)
Lai, 2006). relating the second-order factor to the first-order factor.
After selecting a measurement model, the next step is According to Gerbing et al. (1994), this conceptual
to test the substantive hypothesis. A structural model is advance provides ‘‘an explicit delineation of the domains
evaluated and, if the model fits the data adequately, the t- of content that operationalize the constructs of interest.
ARTICLE IN PRESS
X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652 641

δ1 FR1 λi1
δ2 FR2
Fair
Policies
δ3 FR3 ξ1
δ4 FR4

λi2 ST1 ε1
δ5 SF1
Safe ST2 ε2
δ6 SF2 Flights λi1
ξ2 ST3 ε3
δ7 SF3
Satisfaction ST4 ε4
η1
Φij ST5 ε5
λi3
δ8 SL1 ST6 ε6
δ9 SL2 Desirable ST7 ε7
Assortment
of Flights
δ10 SL3 ST8 ε8
ξ3
δ11 SL4 ST9 ε9

λi4
δ12 ER1
Reliable
δ13 ER2 Equipment
ξ4
δ14 ER3

Fig. 4. Structural model with four first-order correlated exogenous factors.

Table 4 order factors is given as:


One structural model (one first-order factora).
Y ¼ Ly Z þ 
Exogenous Completely standardized effect Model fit indices
variable (t-value) To compare the four posited models and assess whether
a second-order model is plausible, various fit indices can
Company 0.70 (7.77) NNFI ¼ 0.82, be compared. Such indices include chi-square per degrees
deliverables CFI ¼ 0.84, of freedom, the normed fit index (NFI), the comparative fit
std. RMR ¼ 0.099
index (CFI), the non-normed fit index (NNFI), and the
a
Customer satisfaction is posited as the dependent variable in the standardized root mean square residual. Of course, final
model. selection of a measurement model rests upon soundness
that goes beyond a mere comparison of fit indices.
That is, the failure of a unidimensional measurement Assuming that the models under comparison can be
model does not alone invalidate the operationalization of supported theoretically, one may consider the evaluation
the hypothesized construct. Rather, the problem may be through fit-indices as the ‘‘first-cut.’’ In other words,
operationalization of the construct as a first-order factor, models that exhibit poor fit indices ought not to advance
so that the construct may be represented as a second- to the next stage of scrutiny. On the other hand, a
order factor’’ (p. 863). Each facet is conceptualized as one measurement model that generates the best fit indices
of the building blocks of the higher-order construct. The cannot automatically secure its position as the leading
idiosyncrasy of each content domain is retained while model. Issues such as discriminant validity and the
unidimensionality and discriminant validity can be ad- degree of multicollinearity that can emerge have to be
dressed by the higher-order specification. ‘‘Bloated spe- considered.
cific’’ factors can be included as some or all of the An important issue in structural equation modeling is
constituent building blocks of higher-order structures and model identification and second-order modeling is not
highly-correlated factors can now be related through what immune to the requirements for identification. Byrne
they all share, a common higher-order factor. (1998) summarizes the issue of statistical identification.
The observed variables are represented by a vector of y. Byrne (1998) states that identification is concerned with
The matrix Ly contains the loadings of the observed ‘‘whether or not there is a unique set of parameters
variables on the first-order factors. The unique compo- consistent with the data’’ (p. 28). More specifically, if the
nents in the observed variables are represented by e. The model is identified, then there is a unique solution for the
equation for the observed variables in terms of the first- values of the parameters to be estimated and the model is
ARTICLE IN PRESS
642 X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652

Fig. 5. Structural model: one first-order factor.

Table 5
Correlations, reliability, average variance extracted, and descriptive statistics.

Latent variable (# of Fair policies Safe flights Desirable Reliable Latent variable Latent variable Mean per
manifest items) assortment of equipment mean std. deviation manifest
flights item

Fair policies (4) 1.00 (0.52a 0.81b) 15.8379 3.11157 3.96


Safe flights (3) 0.74c 0.55d 1.00 (0.47, 0.73) 13.2562 2.38327 4.42
Desirable assortment 0.74, 0.55 0.75, 0.56 1.00 (0.49, 0.79) 16.9294 3.53829 4.23
of flights (4)
Reliable 0.74, 0.55 0.85, 0.72 0.85, 0.72 1.00 (0.62, 0.83) 13.1609 2.58605 4.39
equipment (3)

a
Average variance extracted is on the diagonal in parentheses.
b
Composite reliability is on the diagonal in parentheses.
c
Correlation.
d
Squared correlation.

testable. However, if the model is not identified, this Some general rules for identification are applicable for
implies that ‘‘many sets of different parameter estimates second-order models (Rindskopf and Rose, 1988). These
could fit the data equally well, and, in this sense, any one rules can be conceptualized in terms of the two parts of
set of values would be arbitrary’’ (p. 29). There could be the structure. The first part relates first-order factors to
three kinds of models: just-identified, overidentified, and the second-order factor(s) while the second part relates
underidentified. Just-identified models occur when the the observed variables to the first-order factors. If each of
number of data variances and covariances is equal to the the two parts is identified, then the model is identified. If
number of parameters to be estimated. Such models will there is only one second-order factor then there must be
have zero degrees of freedom and can never be rejected at least three first-order factors. Even then, the model will
(Byrne, 1998). Overidentified models contain more data be just identified-which then implies that the overall test
points, such as variances and covariances of the observed for goodness of fit does not test the second-order
variables, than the number of parameters to be estimated. structure (Rindskopf and Rose, 1988). It is then suggested
This results in positive degrees of freedom which allows to have at least four first-order factors if there is only one
for a rejection of the model, rendering such a model second-order factor. If there are two or more oblique
scientifically useful and thus desirable. On the other hand, second-order factors, Rindskopf and Rose (1988) state that
underidentified models are manifested when the number ‘‘only two first-order factors are necessary per second
of parameters to be estimated is more than the number of order factor for the model to be identified.’’ (p. 54).
variances and covariances of the observed variables and However, more first-order factors per second-order factor
thus an infinite number of solutions are possible. are desirable. As far as the second part is concerned (that
ARTICLE IN PRESS
X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652 643

Table 6
Measurement models for the five constructs-four exogenous and one endogenous variable.

Variable Factor Completely std. coeff. Fit indicesconfirmatory results


loadingsexploratory (t-value)confirmatory
results (n ¼ 170) results (n ¼ 437)

FR1. Fairness of excess-baggage chargesa 0.54 0.61 Chi-square (df) ¼ 28.47 (5)
FR2. Fairness of Airline’s policies when flights 0.85 0.71 (12.30) Chi-square/df ¼ 5.69
are delayed
FR3. Fairness with which Airline takes care of 0.78 0.80 (13.36) NFI ¼ 0.98, NNFI ¼ 0.97, CFI ¼ 0.98
customers when flights are cancelled
Standardized RMR ¼ 0.029
FR4. Fairness of restrictions Airline places on 0.64 0.91 (14.50) Average variance extracted ¼ 0.63
tickets
Composite reliability ¼ 0.89

SF1. Reassuring passengers during bad weather 0.78 0.70 Chi-square (df) ¼ 28.91 (5)
SF2. Adhering to safety procedures with 0.79 0.71 (12.63) Chi-square/df ¼ 5.78
passengers while airborne
NFI ¼ 0.96, NNFI ¼ 0.93, CFI ¼ 0.97
SF3. Airline’s reputation for safety 0.73 0.72 (12.84) Standardized RMR ¼ 0.032
Average variance extracted ¼ 0.49
Composite reliability ¼ 0.83

SL1. Airline and its partners providing a 0.80 0.80 Chi-square (df) ¼ 43.37 (9)
network of flights covering desired destinations
Chi-square/df ¼ 4.82
SL2. Providing desired flight connections 0.76 0.89 (22.25) NFI ¼ 0.98, NNFI ¼ 0.98, CFI ¼ 0.99
SL3. Offering flights at desired times 0.81 0.95 (24.79) Standardized RMR ¼ 0.015
SL4. Offering non-stop flights 0.68 0.92 (23.66) Average variance extracted ¼ 0.80

ER1. Dependability of aircraft 0.83 0.79 Chi-square (df) ¼ 8.81 (5)


ER2. Working order of aircraft accessories such 0.77 0.84 (19.09) Chi-square/df ¼ 1.76
as air vents, lights, and call Buttons
NFI ¼ 0.99, NNFI ¼ 0.99, CFI ¼ 1.00
ER3. Airworthiness of aircraft 0.85 0.71 (15.56) Standardized RMR ¼ 0.014
Average variance extracted ¼ 0.65
Composite reliability ¼ 0.90

ST1. I like flying with this airlineb 0.93 Chi-square (df) ¼ 23.52 (27)
ST2. My expectations from this airline were met 0.88 (30.13) Chi-square/df ¼ 5.57
ST3. I get a good value for my money 0.74 (20.33) NFI ¼ 0.97, NNFI ¼ 0.97, CFI ¼ 0.98
ST4. Other things being the same, I would prefer Not applicable 0.88 (30.29) Standardized RMR ¼ 0.022
to fly this airline Average variance extracted ¼ 0.74
ST5. I would recommend this airline to others 0.93 (36.28) Composite reliability ¼ 0.96
ST6. I have chosen this airline over others in 0.80 (23.98)
recent years
ST7. I prefer flying this airline because of my 0.87 (28.99)
experience with it
ST8. My overall experience with this airline has 0.91 (33.28)
been great
ST9. I would rate the overall service quality of 0.80 (24.10)
this airline asc

a
Respondents were asked to rate the airline they have selected for evaluation relative to what their expectations were for that specific airline. This
was based on a scale where 1 ¼ exceptionally below my expectations, 2 ¼ much below my expectations, 3 ¼ below my expectations, 4 ¼ meets my
expectations, 5 ¼ above my expectations, 6 ¼ much above my expectations, and 7 ¼ exceptionally above my expectations.
b
Responses to items were based on a Likert type scale where 1 ¼ strongly disagree, 4 ¼ neutral, and 7 ¼ strongly agree.
c
Responses to items were based on a Likert type scale where 1 ¼ terrible, 2 ¼ poor, 3 ¼ somewhat poor, 4 ¼ neither good nor bad, 5 ¼ somewhat
good, 6 ¼ good, 7 ¼ excellent.

deals with observed variables), there should be at least 6.3. Convergent and discriminant validity issues
two measured variables per first-order factor while more
are desirable (if there is only one item for a given factor, On the basis that first-order factors are acting as
then an assumption about the reliability of that measure dependent variables within the confines of the type of
has to be made-and this assumption is sometimes second-order model (characterized by both reflective
untenable). manifest indicators and first-order constructs) we present
ARTICLE IN PRESS
644 X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652

Table 7
Simultaneous exploratory factor analysis of the four exogenous variables (n ¼ 170).

Variable Factor

1 2 3

ER3. Airworthiness of aircraft 0.938


ER1. Dependability of aircraft 0.674
SF3. Airline’s reputation for safety 0.547
SF2. Adhering to safety procedures with passengers while airborne 0.511
ER2. Working order of aircraft accessories such as air vents, lights, and call buttons 0.507
SF1. Reassuring passengers during bad weather 0.467
FR1. Fairness of excess-baggage charges 0.202
SL3. Offering flights at desired times 0.736
SL1. Airline and its partners providing a network of flights covering desired destinations 0.698
SL2. Providing desired flight connections 0.606
SL4. Offering non-stop flights 0.521 0.241
FR4. Fairness of restrictions Airline places on tickets 0.406 0.270
FR2. Fairness of Airline’s policies when flights are delayed 0.252 0.806
FR3. Fairness with which Airline takes care of customers when flights are cancelled 0.237 0.447

Percent of variance 34.165 6.226 4.821


Cumulative percent of variance 34.165 40.391 45.212

Note: Loadings below 0.20 are suppressed on this table to afford a better visualization of the factor structure.

δ1 FR1 λi1
δ1 FR1
δ2 FR2
δ2 FR2 Fair
δ3 FR3 Policies
δ3 ξ1
FR3
δ4 FR4

λij δ4 FR4
δ5 SF1
λi2
δ6 SF2 δ5 SF1
δ7 Company
Safe
SF3
Deliverables δ6 SF2 Flights
ξ1 ξ2
δ8 SL1
δ7 SF3
δ9 SL2
λi3
δ8 SL1
δ10 SL3
Desirable
δ9 SL2
δ11 SL4 Assortment
of Flights
δ10 SL3 ξ3
δ12 ER1

δ13 ER2
δ11 SL4

λi4
δ14 ER3 δ12 ER1

Fig. 6. Measurement Model 1: one first-order factor. Reliable


δ13 ER2 Equipment
ξ4
here, issues of discriminant validity are of less signifi- δ14 ER3
cance. The facets, posited as first-order constructs, are
treated as reflective indicators of the second-order factor Fig. 7. Measurement Model 2: four first-order orthogonal (uncorrelated)
and thus are expected to be highly correlated. It is in fact factors.
ARTICLE IN PRESS
X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652 645

δ1 FR1
λi1 ε1 FR1
ζ1
ε2 FR2
δ2 FR2 Fair Fair
Policies Policies
ε3 FR3
δ3 FR3
ξ1 η1
ε4 FR4
δ4 FR4 λi1

λi2 λi2 ζ2
δ5 SF1 γ11
ε5 SF1
Safe
δ6 SF2 Flights
ε6
Safe
Flights
ξ2 SF2

δ7 η2
SF3 ε7 SF3 γ12

Φij Company
γ13 Deliverables
λi3 λi3 ζ3 ξ1
δ8 SL1
ε8 SL1
Desirable
δ9 SL2 Assortment ε9 SL2
Desirable
Assortment
of Flights of Flights
δ10 SL3 ξ3 ε10 SL3 η3
δ11 SL4 ε11 SL4

γ14

ζ4
λi4 λi4
δ12 ER1 ε12 ER1
Reliable Reliable
δ13 ER2 Equipment ε13 ER2 Equipment
ξ4 η4
δ14 ER3 ε14 ER3

Fig. 8. Measurement Model 3: four first-order correlated factors. Fig. 9. Measurement Model 4: four first-order factors, one second-order
factor.

on the basis of what they share that we put them together 6.5. Results
under the umbrella of a second-order construct. They share
the more abstract construct, the second-order factor. More The first model (Model 1: Fig. 6, Table 8) specifies that
relevant here is the assessment of convergent validity at the all 14 items are reflective of one latent variable. The chi-
first-order latent construct to second-order latent construct square per degree of freedom was above 5 and other fit
level. At this level, one cannot demonstrate and support indices are indicative of a poor model fit. The poor model
convergent and discriminant validity at the same time. fit also illustrates, by the way, one of the adverse
Establishing convergent validity takes precedence. We consequences of combining manifest variables from
would like to note that if multiple second-order factors various content domains into one first-order latent
exist, then discriminant validity of the second-order factors variable. Although they can be correlated, the items of
can be examined through conventional means. each domain may be reflective of distinct facets which are
not acknowledged through a specification such as the one
6.4. Multicollinearity issues represented by Model 1. Combining items representing
distinct facets into a single first-order construct is
Because the second-order factor is posited to act as the conceptually and methodologically perilous.
independent variable to both the first-order latent factors The second model (Model 2: Fig. 7, Table 8) posits four
that represent it (facets) and the dependent variable, only uncorrelated latent variables that are related to their
one path coefficient (if there is only 1 second-order factor) respective observed variables. In view of the strong
is estimated from the independent variable (the second- correlations between the latent variables as noted pre-
order factor) to the dependent variable (Fig. 10). Multi- viously, an orthogonal specification for the relationships
collinearity is a moot issue here as only one independent between latent variables would be expected to produce
and one dependent variable are regressed. However, if poor model fit. Indeed, such was in fact the case where all
multiple second-order factors exist and are posited to fit indices here failed to pass muster. In situations where
impact the same dependent variable then the degree of first-order factors are poorly correlated, this specification
multicollinearity has to be examined through ordinary could prove to be appropriate. Here however, we are
means. presented with the reality of high correlations between
ARTICLE IN PRESS
646 X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652

Fig. 10. Structural model.

Table 8
Alternative measurement model structures (n ¼ 437).

Goodness of fit indices for alternative Measurement models


models of factor structure
Model 1 Model 2 Model 3 Model 4
One first-order factor Four first-order Four first-order Four first-order factors,
factorsuncorrelated factorscorrelated one second-order factor

Chi-square (df) 496.31 (77) 1207.83 (77) 216.22 (71) 221.67 (73)
Chi-square/df 6.45 15.69 3.04 3.04
Normed-fit index (NFI) 0.84 0.64 0.92 0.92
Non-normed fit index (NNFI) 0.84 0.59 0.93 0.93
Comparative fit index (CFI) 0.86 0.65 0.95 0.94
Standardized root mean square 0.06 0.34 0.04 0.04
residual (RMR)

first-order factors and thus a specification such as the one model specification such as the one represented by
represented by Model 2 would not be conducive. Model 3 would be advisable.
Model 3 (Fig. 8, Table 8) was similar to Model 2 except Model 4 (Fig. 9, Table 8) presents a second-order factor
for the fact that the latent variables were free to correlate. (i.e., company deliverables) that is related to four first-
The model fit was quite acceptable as all fit indices met order factors/facets that in turn are related to the
respective criteria. However, retention of such a model respective observed variables. At a first glance, it appears
would not resolve the issue of discriminant validity as that the first-order factors act as both independent and
argued earlier. In addition, when first-order factors are dependent variables (Byrne, 1998). In Structural Equation
highly correlated, multicollinearity emerges as a critical Modeling, if a variable has an arrow pointing to it then it is
concern as we have illustrated earlier in Section 5 (see targeted as a dependent variable and it will maintain such
also Table 3). In instances however where first-order status throughout. A variable can either operate as a
factors exhibit moderate correlations, a measurement dependent variable or an independent variable but not
ARTICLE IN PRESS
X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652 647

both (Byrne, 1998). Within the context of higher-order order model that rivals the performance of a first-order
modeling the first-order factors are conceptualized as correlated model can be an attractive alternative. Model 4
dependent variables. This implies that their variances and is well-fitting and only immaterially divergent in perfor-
covariance are not estimable within the model. Rather, mance from Model 3. Again, given the deficiency
‘‘such variation is presumed to be accounted for by the (e.g., discriminant validity, issues of multicollinearity) of
higher-order factor’’ (Byrne, 1998, p. 33). Model 3, our prior discussion for the need of second-order
In order to determine whether the second-order model models from a conceptual point of view, and the literature
is identified, we need to first calculate the number of support for second-order models in the context of service
parameters to be estimated. Here we have 10 first-order quality (e.g., Kettinger and Lee, 1994; Kettinger et al.,
regression coefficients (for identification purposes a 1995; Myerscough, 2002; Somers et al., 2003; Rodgers
constraint is placed on one coefficient of each first-order et al., 2005; Anitsal and Paige, 2006; Lai, 2006), Model 4
factor to be equal to one, thus if we have a total of 14 appears to be the most prudent choice. We examine
observed items and four first-order factors, then only 10 Model 4 in more detail.
parameters, ls, need to be estimated), four second-order Using the LISREL 8.51 software and applying MLE on
regression coefficients (gs)—one for each first-order the entire set of items in the model, the fit indices along
factor, 14 measurement error variances (es)—one for each with the t-values provide evidence of convergent validity
observed variable, and four residual error terms (zs)—one (Tables 8 and 9). The NFI was 0.92, NNFI was 0.93 and CFI
for each first-order factor, for a total of 32. Given that was 0.94, while the chi-square per degree of freedom was
there are 105 pieces of information (the total is 105 as we 3.04. The standardized RMR was 0.04. All of the items
have 14 items and it is computed by adding all the have statistically significant relationships with their
elements in the lower-left triangle of the variance- factors. All factor loadings were above 0.53 and a good
covariance matrix plus the principal diagonal elements majority of them above 0.70. The significance of the
in the sample variance—covariance matrix) the model is t-values (also known as ‘‘Wold statistics’’ or ‘‘critical
over-identified with 73 degrees of freedom. ratios’’, t-values are ratios of the absolute values of the
According to the literature (Marsh and Hocevar, 1985; parameters to their respective standard errors), as shown
Arnau and Thompson, 2000), a model that includes a in Table 9, associated with factor-to-item loadings exceeds
second-order model structure can never produce a model the critical value at the 0.001 significance level. Notice
fit that is ‘‘better’’ (i.e., better fitting in terms of the model that the loadings of the first-order factors to the second-
fit indices) than a model that specifies only first-order order factor are all above 0.81 and that the t-values are
correlated factors (such as Model 3). However, a second- quite substantive and statistically significant. This points

Table 9
Measurement model for four first-order latent factors and one second-order latent factor (n ¼ 437).

Variable First-order factors Second-


order factor
Fair policies Safe flights Desirable Reliable
assortment of equipment
flights

FR1. Fairness of excess-baggage charges 0.53


FR2. Fairness of Airline’s policies when flights are delayed 0.81 (10.66)
FR3. Fairness with which Airline takes care of customers 0.78 (10.53)
when flights are cancelled
FR4. Fairness of restrictions Airline places on tickets 0.73 (10.17)

SF1. Reassuring passengers during bad weather 0.63


SF2. Adhering to safety procedures with passengers while 0.68 (11.09)
airborne
SF3. Airline’s reputation for safety 0.75 (11.80)

SL1. Airline and its partners providing a network of flights 0.64


covering desired destinations
SL2. Providing desired flight connections 0.71 (12.09)
SL3. Offering flights at desired times 0.80 (13.09)
SL4. Offering non-stop flights 0.65 (11.20)

ER1. Dependability of aircraft 0.79


ER2. Working order of aircraft accessories such as air vents, 0.78 (16.85)
lights, and call buttons.
ER3. Airworthiness of aircraft 0.77 (16.53)

Fair policies 0.81 (9.99)


Safe flights 0.88 (12.17)
Desirable Assortment of Flights 0.89 (11.65)
Reliable equipment 0.94 (17.21)

Chi-square (df) ¼ 221.67 (73), chi-square/df ¼ 3.04, normed-fit index (NFI) ¼ 0.92, non-normed fit index (NNFI) ¼ 0.93, comparative fit index
(CFI) ¼ 0.94, standardized root mean square residual (RMR) ¼ 0.04.
ARTICLE IN PRESS
648 X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652

to a very strong relationship between first-order factors nested models. A number of empirical manuscripts have
and the second-order factor, attesting to the convergent employed this hierarchical approach successfully. Exam-
validity of the posited second-order model. As such, a ples include Babbar and Koufteros (2008), Somers et al.
second-order model is quite effective in representing the (2003), and Myerscough (2002). The hierarchical process,
data. It is important to note that company deliverables (the however, should not be reminiscent of a ‘‘fishing expedi-
second-order factor) contributes the most to reliable tion.’’ The process should only involve the specification
equipment (completely std. coefficient ¼ 0.94, t-value ¼ and comparative evaluation of a series of measurement
17.21) and the least to fair policies (completely std. models that can be defended conceptually. Moreover, the
coefficient ¼ 0.81, t-value ¼ 9.99). selection of a given model goes beyond a prototypical
Once we selected Model 4 as the best representing assessment of model fit-indices. As we illustrated in
measurement model, we proceed with the testing of the Section 5, measurement models have a decisive impact on
structural model, which builds on Model 4. In Fig. 10 we the outcomes of substantive hypotheses testing. We thus
postulate a structural model where satisfaction is specified caution researchers that after a selection of a measure-
as the dependent variable and the second-order factor ment model based on fit indices, discriminant validity
(i.e., company deliverables) is hypothesized to affect it. The should be examined followed by a critical assessment of
model fit meets the generally acceptable guidelines substantive hypotheses-testing findings. As illustrated on
(Table 10). The chi-square per degree of freedom was Table 8, Model 3 (Fig. 8) produces the best measurement
2.65 and NFI was 0.92. Both NNFI (0.94) and CFI (0.95) are model fit. However, in Section 5 we demonstrated that
above 0.90 and standardized RMR was 0.039. The t-value this measurement model presents discriminant validity
associated with the relationship between company deli- shortcomings and generates undesirable challenges while
verables and satisfaction (g ¼ 0.74) was 15.49 and it is testing substantive hypotheses (Table 3 and Fig. 4). Thus,
rather potent. The model explained 51% of the variance in we opted instead to select Model 4 (a second-order
passenger satisfaction. Collectively, the four variables can specification). To summarize, measurement models that
have a significant impact on passenger satisfaction. Thus, produce poor fit indices can be excluded from further
we cannot discount their impact. There is a substantive consideration but a model that generates the best fit
size effect as well as statistical significance. indices may not always be the model of choice. Good
model fit is necessary but it should not be the only
determining factor for final model selection.
7. Discussion and directions for future research
The model we presented in Figs. 9 and 10 can be
enlarged to also include other second-order factors, such
Conducting empirical research is a reality for many
as employee deliverables (Fig. 1). The approach of testing
operations and supply chain management faculty around
the enlarged measurement model would be identical to
the world. The field has progressed through the last 20–25
the process we illustrated. However, we can also extend
years but much needs to be done just to catch up with the
the current model by adding exogenous variables, which
sophistication of methodologies employed in other rela-
may only stand at the first-order of abstraction. In other
tive or distant disciplines. With this manuscript, we wish
words, within the same model some of the exogenous
to present a paradigm on second-order modeling. Higher-
variables can be conceptualized at the second-order of
order modeling is an important advance as several
abstraction while others can be specified purely as first-
constructs that operations and supply chain management
order factors. For example, competitive pricing could be
researchers employ can be meaningfully conceptualized
specified as another antecedent to customer satisfaction.
through higher-order modeling. Beyond the conceptual
Fig. 11 shows a ‘‘mixed model’’ where competitive pricing
elegance that higher-order models present, there are also
(with its four manifest variables) is modeled as a first-
practical advantages that are gained through the use of
order construct which directly affects satisfaction. While
higher-order modeling.
the first-order constructs of company deliverables are part
We presented a hierarchical approach to evaluate
of a Y-model specification (they are after all dependent on
whether a higher-order conceptualization is prudent
the second-order construct), the specification of competi-
given that other structures could equally represent the
tive pricing has an X-model specification.
data. This approach involves the assessment of several
If a researcher is interested in testing for moderation
effects (within the context of second-order modeling), this
Table 10
Structural model.
can be accomplished through factorial invariance tests
(Koufteros and Marcoulides, 2006). For instance, a
Company deliverables (path researcher may postulate that class of service (economy
coefficient, t-value) vs business) may moderate the relationship between
company deliverables and satisfaction. To test for a
Fair policies 0.82 (10.04)
Safe flights 0.88 (12.12) moderation effect, measurement model invariance is
Desirable assortment of flights 0.88 (12.57) examined first followed by tests to see if the structural
Reliable equipment 0.94 (17.47) model is invariant across the two groups (i.e., economy vs
Satisfaction 0.71 (15.49) R2 ¼ 0.51 business class). The paradigm presented by Koufteros and
Chi-square (df) ¼ 596.33 (225), chi-square/df ¼ 2.65, normed-fit index
Marcoulides’s can be readily adopted except that the
(NFI) ¼ 0.92, non-normed fit index (NNFI) ¼ 0.94, comparative fit index invariance of the gamma coefficients that relate the
(CFI) ¼ 0.95, standardized root mean square residual (RMR) ¼ 0.039. second-order construct to the four first-order constructs
ARTICLE IN PRESS
X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652 649

Fig. 11. Expanded structural model: mixed model.

should be examined in light of the fact that these disaggregation approach. Moreover, given that indicators
coefficients are now part of the measurement model. can be either reflective or formative, researchers may also
Thus, the invariance of these specific gamma coefficients employ a formative indicator specification possibly both
should be assessed prior to testing for structural model at the first- and second-order factor levels.
invariance. With a ‘‘partial disaggregation’’ approach, two or more
Researchers may consider a second-order specification composite variables, the parcels, can be constructed for
as a form of aggregation. Aggregation is useful as it can each first-order construct which subsequently are treated
represent the relationships between variables more as indicators of a second-order factor model (Bagozzi and
parsimoniously. For example, several manifest/observed Heatherton, 1994). The parcels can be constructed from
variables can be represented by a single first-order latent like indicators (items that load on the same first-order
variable. Likewise, a second-order factor can embody the factor), and thus are called homogenous parcels, or they
meaning of many first-order latent variables. Thus, the can be ‘‘domain representative parcels’’ (Kishton and
number of variables that need to be estimated in Widaman, 1994) where ‘‘each parcel is made up of items
the context of a structural model can be reduced that load on different first-order factors’’ (Coffman and
significantly. MacCallum, 2005, p. 236). The ‘‘partial disaggregation’’
The second-order model we describe in this manu- approach represents the closest alternative to our con-
script is coined as ‘‘total disaggregation second-order ceptual approach so we discuss its pros and cons. The
factor model’’ (Bagozzi and Heatherton, 1994). It is merits and criticism of using parcels relates primarily to
described as a total disaggregated model as each and the philosophy and practicality of summing/averaging
every manifest/observed variable is employed as a indicators.
reflective indicator for a latent variable and each and There is plenty of criticism and praise of parceling as a
every first-order latent variable is employed as a reflective measurement practice. Little et al. (2002) provide an
indicator of the second-order model (Coffman and extensive body of work in regards to the question of
MacCallum, 2005). This approach however is not the only parceling and present the arguments for and against it.
means through which higher-order models and their They provide a multitude of arguments for each position.
motivation can be represented by. Researchers, for They state that ‘‘From an empiricist-conservative philoso-
example, can represent second-order models via a partial phy of science perspective, parceling is akin to ‘cheating’
ARTICLE IN PRESS
650 X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652

because modeled data should be as close to the response factor while retaining all possible sources of variance. If
of the individual as possible in order to avoid the potential the researcher has an adequate sample size at his or her
imposition, or arbitrary manufacturing, of a false struc- disposal and reasonable model fit indices can be obtained,
ture. Any potential source of subjective bias on the part of then a total disaggregated second-order factor model
the data analyst is to be avoided at all costs, a simple a would be advisable. On the other hand, when the
priori principle which most would agree. In this sense, researcher encounters relatively few observations and
allowing the researcher to create parcels from items there are model fit and reliability concerns, then a partial
fundamentally undermines the objective empirical pur- disaggregation approach would be most appealing.
pose of the techniques that have been developed to model Here in this manuscript we describe a process through
multivariate data’’ (p. 152). Countering this philosophical which the viability of a particular type of second-order
argument against parceling, Little et al. state, ‘‘From a models is addressed. Although the type of a second-order
more pragmatic-liberal philosophical perspective, parcels model we assess is the most frequently encountered in the
have potential merits as the lowest level of data to be literature, three other types exist along the reflective
modeled. Given that measurement is a strict, rule-bound versus formative divide and further research can aid
system that is defined, followed, and reported by the towards their proper construction and evaluation. Jarvis
investigator, the level of aggregation used to represent the et al. (2003) describe four main types of second-order
measurement process is a matter of choice and justifica- models. The first-order constructs can have either for-
tion on the part of the investigator. With a compelling mative or reflective indicators while those first-order
justification, using parcels would not be seen as a constructs can themselves be either formative or reflec-
transgression against truth because the operational tive indicators of an underlying second-order construct.
aspects of an investigation are, fundamentally, a public The four combinations that result appear as Fig. 2 in Jarvis
process’’ (pp. 152–153). Another argument against parcel- et al. (2003). The model we selected to address (Type I) is
ing is articulated by Little et al. as they further state, ‘‘The the aforementioned ‘‘total disaggregation second-order
empiricist-conservative position appears rooted in the factor model’’ (Bagozzi and Heatherton, 1994) and is
stance that all sources of variance in each item must be characterized by first- and second-order reflective indica-
represented in any multivariate statistical models invol- tors. Models that include a formative specification at
ving a given scale. Failing to represent one or more sources either the first-order and/or second-order level are more
of variance, however minor, may lead to bias in estimates rare but they can be meaningful representations. The work
of other key parameters throughout the model’’ (p. 154). of Diamantopoulos et al. (2008) as it pertains to formative
Offering an argument in contrast to the statement above, measurement models would be quite useful in gaining a
Little et al. posit ‘‘In contrast, the pragmatic-liberal better understanding of the three types of second-order
position holds that representing each and every source models we are not examining here. Future research can
of variance in each item, particularly on an a priori basis, is perhaps modify our process to accommodate second-
impossible. Under this position, researchers cannot know order models that include a formative conceptualization.
every single source of variance in every single item; one Second-order models are not particularly difficult to
can only hope that one’s models will represent the construct or evaluate but the OM literature has yet to
important common sources of variance across samples make full use of the capabilities and insights they afford.
of items. When minor influences, which are substantively Other techniques that could prove catalytic for the
trivial yet empirically significant, cannot be predicted on advancement of empirical tradition in operations manage-
an a priori basis, they will be difficult or impossible to ment also beg for attention. Two such approaches that we
distinguish from chance findings’’ (p. 154). deem to be quite pertinent are latent growth modeling
One can argue perhaps that the motivation for and multilevel modeling. Latent growth modeling is not
parceling is empirically driven. Compared to aggregated/ only useful in discerning ‘‘causal’’ effects, but it is quite
parceled data, ‘‘item-level data contain one or more of the useful in research studies that examine changing pro-
following disadvantages: lower reliability, lower com- cesses in individuals or individual firms or industries. One
munality, a smaller ratio of common-to-unique factor reason for their popularity in educational, marketing,
variance, and a greater likelihood of distributional viola- organizational behavior and psychological research is that
tions’’ (p. 154). A second empirical advantage pertains to repeated measurements allow researchers to investigate
the number of parameters required to model items versus individual (intra-individual) change across time as well as
parcels and the overall fit of structural models. Bagozzi between individual differences (inter-individual) and
and Edwards (1998) and Bagozzi and Heatherton (1994) similarities in change patterns (Raykov and Marcoulides,
argue that ‘‘because fewer parameters are needed to 2000).
define a construct when parcels are used, parcels are Organizational research also presents an opportunity
preferred, particularly when sample sizes are relatively to study phenomena that are multilevel, or hierarchical, in
small’’ (p. 155). nature. For example, individual strategic business units
The motivation for second-order modeling is primarily (SBUs) are embedded within a firm which subsequently is
a conceptual one where ‘‘the central scientific interest lies embedded within an industry which is embedded within
in the general construct’’ (Chen et al., 2006, p. 90). This a country. A hierarchy consists of units grouped at
higher-order factor is hypothesized to account for the different levels. SBUs may be the level 1 units in a
relationship among the lower-order factors. It acknowl- 2-level structure where the level 2 units are the firms:
edges the multidimensional nature of the higher-order SBUs may be the level 1 units clustered within firms that
ARTICLE IN PRESS
X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652 651

are the level 2 units. Multi-level methodology (also Fornell, C., Larcker, D.F., 1981. Evaluating structural equations models
known as hierarchical linear models, generalized linear with unobservable variables and measurement error. Journal of
Marketing Research 18 (1), 39–50.
mixed models, nested models, and random-effects mod- Gerbing, D.W., Hamilton, J.G., Freeman, E.B., 1994. A large-scale second-
els) recognizes that the group and its members both order structural equation model of the influence of management
influence and are influenced by the group membership participation on organizational planning benefits. Journal of Man-
agement 20 (4), 859–885.
(Heck and Thomas, 2000). Multilevel modeling offers a
Griffith, C.J., 2006. Food safety: where from and where to? British Food
richer and probably more accurate examination of Journal 108 (1), 6–15.
organizational life. Heck, R.H., Thomas, S.L., 2000. An Introduction to Multilevel Modeling
The operations and supply chain management litera- Techniques. Lawrence Earlbaum Associates, Publishers, Mahwah, NJ.
Heuvel, T.V.D., Trijp, H.V., Woerkum, C.V., Renes, R.J., Gremmen, B., 2007.
ture has no evidence of publications using either the Linking product offering to consumer needs: inclusion of credence
latent growth or the multilevel methodology and we call attributes and the influences of product features. Food Quality and
for researchers to consider employing them in future Preference 18 (2), 296–304.
Hu, L.T., Bentler, P.M., 1999. Cutoff criteria for fit indices in covariance
research. With this manuscript we want to encourage OM structure analysis: conventional criteria versus new alternatives.
researchers to step out of the confines of existing research Structural Equation Modeling 6 (1), 1–55.
methodology and consider techniques which may shed Jarvis, C.B., MacKenzie, S.B., Podsakoff, P.M., 2003. A critical review of
construct indicators and measurement model misspecification in
additional light into organizational life.
marketing and consumer research. Journal of Consumer Research 30,
199–218.
References Jöreskog, K.G., 1970. A general method for analysis of covariance
structures. Biometrika 57 (2), 239–251.
Jöreskog, K.G., Sörbom, D., 2001. LISREL 8: User’s Reference Guide.
Adams, J.S., 1965. Inequity in social exchange. In: Berkowitz, L. (Ed.), Scientific Software International, Lincolnwood, IL.
Advances in Experimental Social Psychology, vol. 2. Academic Press, Kalafatis, S.P., Tsogas, M.H., Blankson, C., 2000. Positioning strategies in
New York, NY, pp. 267–299. business markets. Journal of Business and Industrial Marketing 15
Anitsal, I., Paige, R.C., 2006. An exploratory study on consumer (6), 416–437.
perceptions of service quality in technology-based self-service. Katz-Navon, T., Naveh, E., Stern, Z., 2005. Safety climate in health care
Services Marketing Quarterly 27 (3), 53–67. organizations: a multidimensional approach. Academy of Manage-
Arnau, R.C., Thompson, B., 2000. Second-order confirmatory factor ment Journal 48 (6), 1075–1089.
analysis of the WAIS-III. Assessment 7 (3), 237–246. Kettinger, W.J., Lee, C.C., 1994. Perceived service quality and user
Arnold, S.J., Oum, T.H., Tigert, D.J., 1983. Determining attributes in retail satisfaction with the information systems function. Decision
patronage: seasonal, temporal, regional, and international compar- Sciences 25 (5/6), 737–766.
isons. Journal of Marketing Research 20 (2), 149–157. Kettinger, W.J., Lee, C.C., Lee, S., 1995. Global measures of information
Babbar, S., Koufteros, X.A., 2008. The human element in airline service service quality: a cross-national study. Decision Sciences 26 (5),
quality: contact personnel and the customer. International Journal of 569–588.
Operations and Production Management 28 (9), 804–830. Kishton, J.M., Widaman, K.F., 1994. Unidimensional versus domain
Bagozzi, R.P., Edwards, J.R., 1998. A general approach to representing representative parceling of questionnaire items: an empirical
constructs in organizational research. Organizational Research example. Educational and Psychological Measurement 54, 757–765.
Methods 1, 45–87. Koelemeijer, K., Oppewal, H., 1999. Assessing the effects of assortment
Bagozzi, R.P., Heatherton, T.F., 1994. A general approach to representing and ambience: a choice experimental approach. Journal of Retailing
multifaceted personality constructs: application to state self-esteem 75 (3), 319–345.
1 (1), 35–67. Koufteros, X.A., 1999. Testing a model of pull production: a paradigm for
Barber, B.M., Darrough, M.N., 1996. Product reliability and firm value: the manufacturing research using structural equation modeling. Journal
experience of American and Japanese automakers, 1973–1992. of Operations Management 17 (4), 467–488.
Journal of Political Economy 104 (5), 1084–1099. Koufteros, X.A., Marcoulides, G.A., 2006. Product development practices
Byrne, B.M., 1998. Structural Equation Modeling: Basic Concepts, and performance: a structural equation modeling-based multi-group
Application, and Programming. Lawrence Earlbaum Associated, Inc, analysis. International Journal of Production Economics 103 (1),
Mahwah, NJ. 286–307.
Carpenter, J.M., Moore, M., 2006. Consumer demographics, store Lai, J.Y., 2006. Assessment of employee’s perceptions of service quality
attributes, and retail format choice in the US grocery market. and satisfaction with ebusiness. International Journal of Human–
International Journal of Retail and Distribution Management 34 (6), Computer Studies 64 (9), 926–938.
434–452. Law, W.K., Chan, A.H.S., Pun, K.F., 2006. Prioritising the safety manage-
Catell, R.B., 1973. Questionnaire psychometric properties: test evaluation ment elements: a hierarchical analysis for manufacturing enter-
and improvement. Personality and Mood by Questionnaire. Jossey- prises. Industrial Management and Data Systems 106 (6), 778–792.
Bass, San Fransisco. Liljeblom, E., Vaihekoski, M., 2009. Corporate ownership and managerial
Chen, F.F., West, S.G., Sousa, K.H., 2006. A comparison of bifactor and short-termism: results from a Finnish study of management
second-order models of quality of life. Multivariate Behavioral perceptions. International Journal of Production Economics 117 (2),
Research 41 (2), 189–225. 427–438.
Coffman, D.L., MacCallum, R.C., 2005. Using parcels to convert path Lindsley, W.B., Blackburn, J.D., Elrod, T., 1991. Time and product variety
analysis models into latent variable models. Multivariate Behavioral competition in the book distribution industry. Journal of Operations
Research 40 (2), 235–259. Management 10 (3), 344–362.
Colquitt, J.A., Conlon, D.E., Wesson, M.J., Porter, C.O.L.H., Ng, K.Y., 2001. Little, T.D., Cunningham, W.A., Shahar, G., Widaman, K.F., 2002. To parcel
Justice at the millennium: a meta-analytic review of 25 years of or not to parcel: exploring the question, weighing the merits.
organizational justice research. Journal of Applied Psychology 86 (3), Structural Equation Modeling 9 (2), 151–173.
425–445. Marsh, H.W., Hocevar, D., 1985. Application of confirmatory factor
Cronin, J.J., Taylor, S.A., 1992. Measuring service quality: a reexamination analysis of the study of selfconcept: first and higher order factor
and extension. Journal of Marketing 56 (3), 55–68. models and their invariance across groups. Psychological Bulletin 97
Diamantopoulos, A., Riefler, P., Roth, K.P., 2008. Advancing for- (3), 562–582.
mative measurement models. Journal of Business Research 61, Martinez-Tur, V., Peiro, J.M., Ramos, J., Moliner, C., 2006. Justice
1203–1218. perceptions as predictors of customer satisfaction: the impact of
Farris, J.A., Aken, E.M.V., Doolen, T.L., Worley, J., 2009. Critical success distributive, procedural, and interactional justice. Journal of Applied
factors for human resource outcomes in Kaizen events: an empirical Psychology 36 (1), 100–119.
study. International Journal of Production Economics 117 (1), McFadden, K.L., 1996. Comparing pilot-error accident rates of male and
42–65. female airline pilots. Omega 24 (4), 443–450.
Ford, T., 1997. Three aspects of aerospace safety. Aircraft Engineering and McFadden, K.L., 1997. Policy improvements for prevention of alcohol
Aerospace Technology 69 (3), 254–264. misuse by airline pilots. Human Factors 39 (1), 1–8.
ARTICLE IN PRESS
652 X. Koufteros et al. / Int. J. Production Economics 120 (2009) 633–652

McFadden, K.L., Hosmane, B.S., 2001. Operations safety: an assessment of Rose, A., 2006. Measuring operational safety in aviation. Aircraft
a commercial aviation safety program. Journal of Operations Engineering and Aerospace Technology: An International Journal
Management 19 (4), 579–591. 78 (1), 26–31.
McFadden, K.L., Stock, G.N., Gowen III, C.R., 2006. Implementation of Sanders, M.G., Hoffman, M.A., 1975. Personality aspects of involvement in
patient safety initiatives in US hospitals. International Journal of pilot-error accidents. Aviation, Space, and Environmental Medicine
Operations and Production Management 26 (3), 326–347. 46, 186–190.
Mentzer, J.T., Flint, D.J., Kent, J.L., 1999. Developing a logistics service Sanders, M.G., Hoffman, M.A., Neese, T.A., 1976. Cross-validation study of
quality scale. Journal of Business Logistics 20 (1), 9–32. the personality aspects of involvement in pilot-error accidents.
Mukherjee, A., Nath, P., 2005. An empirical assessment of comparative Aviation, Space, and Environmental Medicine 47, 1977–1979.
approaches to service quality measurement. Journal of Services Sans, P., Fontguyon, G.D., Briz, J., 2005. Meat safety as a tool of
Marketing 19 (3), 174–184. differentiation for retailers. International Journal of Retail and
Myerscough, M.A., 2002. Information systems quality assessment: replicat- Distribution Management 33 (8), 618–635.
ing Kettinger and Lee’s USIF/SERVQUAL combination. In: 2002—Eighth Seiders, K., Berry, L.L., 1998. Service fairness: what it is and why it
Americas Conference on Information Systems, pp. 1104–1115. matters. Academy of Management Executive 12 (2), 8–20.
Oliver, R.L., Swan, J.E., 1989. Consumer perceptions of interpersonal Sheikh, A.K., Al-Garni, A.Z., Badar, M.A., 1996. Reliability analysis of
equity and satisfaction in transactions: a field survey approach. aeroplane tyres. International Journal of Quality and Reliability
Journal of Marketing 53 (2), 21–35. Management 13 (8), 28–38.
Pal, R., Busing, M.E., 2008. Teaching operations management Sheppard, B.H., Lewicki, R.J., Minton, J.W., 1992. Organizational Justice: The
in an integrated format: student perception and faculty Search for Fairness in the Workplace. Lexington Books, New York, NY.
experience. International Journal of Production Economics 115 (2), Somers, T.M., Nelson, K., Karimi, J., 2003. Confirmatory factor analysis of
594–610. the end-user computing satisfaction instrument: replication within
Pan, Y., Zinkhan, G.M., 2006. Determinants of retail patronage: a meta- an ERP domain. Decision Sciences 34 (3), 595–621.
analytical perspective. Journal of Retailing 82 (3), 229–243. Stassen, R., Mittelstaedt, J., Mittelstaedt, R., 1999. Assortment overlap: its
Parasuraman, A., Zeithaml, V.A., Berry, L.L., 1985. A conceptual model of effect on shopping patterns in a retail market when the distribu-
service quality and its implications for future research. Journal of tions of prices and goods are known. Journal of Retailing 75 (3), 371–386.
Marketing 49 (4), 41–50. Tamimi, N., Sebastianelli, R., Rajan, M., 2005. What do online customers
Parasuraman, A., Zeithaml, V.A., Berry, L.L., 1988. SERVQUAL: a multi- value? Quality Progress 38 (7), 35–39.
item scale for measuring perceptions of service quality. Journal of Thun, J., 2008. Empirical analysis of manufacturing strategy implementa-
Retailing 64 (1), 12–40. tion. International Journal of Production Economics 113 (1), 370–382.
Parasuraman, A., Zeithaml, V.A., Berry, L.L., 1991. Refinement and reassess Thurstone, L.L., 1947. Multi-factor Analysis. University of Chicago Press,
of the SERVQUALscale. Journal of Retailing 67 (4), 420–450. Chicago.
Penttinen, E., Palmer, J., 2007. Improving firm positioning through Trienekens, J., Zuurbier, P., 2008. Quality and Safety standards in the food
enhanced offerings and buyer–seller relationships. Industrial Mar- industry, developments and challenges. International Journal of
keting Management 36 (4), 552–564. Production Economics 113 (1), 107–122.
Raykov, T., Marcoulides, G.A., 2000. A First Course in Structural Equation Vaagen, H., Wallace, S.W., 2008. Product variety arising from hedging in
Modeling. Lawrence Earlbaum Associates, Publishers, Mahwah, NJ. the fashion supply chains. International Journal of Production
Reed, R., Lemak, D.J., Montgomery, J.C., 1996. Beyond process: TQM Economics 114 (2), 431–455.
content and firm performance. Academy of Management Review 21 Van Dyke, T.P., Kappelman, L.A., Prybutok, V.R., 1997. Measuring
(1), 173–202. information systems service quality: concerns on the use of the
Rhoades, D.L., Waguespack Jr., B.P., 1999. Better safe than service? The SERVQUAL questionnaire. MIS Quarterly 21 (2), 195–208.
relationship between service and safety quality in the US airline Van Dyke, T.P., Prybutok, V.R., Kappelman, L.A., 1999. Cautions on the use
industry. Managing Service Quality 9 (6), 396–400. of the SERVQUAL measure to assess the quality of information
Rhoades, D.L., Waguespack Jr., B.P., 2004. Service and safety quality in US systems services. Decision Sciences 30 (3), 877–891.
airlines: pre- and post-September 11th. Managing Service Quality 14 Waikar, A., Nichols, P., 1997. Aviation safety: a quality perspective.
(4), 307–316. Disaster Prevention and Management 6 (2), 87–93.
Rindskopf, D., Rose, T., 1988. Some theory and applications of Walsh, K., Antony, J., 2007. Improving patient safety and quality.
confirmatory second-order factor analysis. Multivariate Behavioral International Journal of Health Care Quality Assurance 20 (2),
Research 23 (1), 51–67. 107–115.
Rodgers, W., Negash, S., Suk, K., 2005. The moderating effect of on-line Zhang, X., Chen, R., 2008. Examining the mechanism of the value co-
experience on the antecedents and consequences of on-line creation with customers. International Journal of Production Eco-
satisfaction. Psychology and Marketing 22 (4), 313–331. nomics 116 (2), 242–250.

View publication stats

You might also like