Sie sind auf Seite 1von 29

JOURNAL OF BUSINESS LOGISTICS, Vol. 15, No.

2, 1994 145

LATENT VARIABLES IN BUSINESS LOGISTICS RESEARCH:


SCALE DEVELOPMENT AND VALIDATION

by

Steven C. Dunn
Idaho State University
Rohert F. Seaker
Penn State University
and
Matthew A. Waller
Western Michigan University

Logistics is confounded with an abundance of concepts that are not easily


operationalized for scientific analysis. Variables such as customer service,
partnerships, third parties, integration, and total quality management are unobservable
and termed "latent." A latent variable (a.k.a., construct) "is an unobserved entity
presumed to underlie observed variables. In science our real interest is more in
the relations among latent variables than it is in the relations among observed variables
because we seek to explain phenomena and their relations."'
Due to the difficulty of developing operational measures of latent variables,
business logistics research that tests hypotheses or propositions involving such
variables has been lacking. Conversely, research methods applied in logistics model
building generally tend to be more rigorous, following well developed methodological
standards common to other academic fields. The use of econometrics in transportation
research, simulation modeling for distribution network analysis, and analytical models
of inventory systems are all examples of logistics research that typically do not
involve the measurement of latent variables.^-^''*'^'^
The majority of logistics research can be categorized into three distinct areas:
(1) generalized descriptions of variables (e.g., case studies); (2) interpretation of
informant perceptions (e.g., surveys, interviews, expert panels); and (3) artificial
reconstruction of reality, such as model building.'
The strength of logistics research may lie in its diversity of approaches. What
is missing, however, is a stronger methodological approach within the perceptive
146 DUNN, SEAKER, AND WALLER

paradigm, especially when latent variables are involved. Surveys and structured
interviews account for almost half of the research published in this category. Survey
research in business logistics often does not use the data from the survey for
hypothesis testing. Instead the data is used to make inferences that are not statistically
verifiable. This is partially a result of the fact that latent variables have not been
scientifically measured, a necessary step for testing theory.
The purpose of this paper is to suggest a logistics research methodology for
the scientific analysis and testing of latent variables. The methodology is drawn
from well established techniques utilized in other areas of business research. This
paper is divided into three sections. The first provides an analysis of logistics research
methods, accomplished through an extension of the review of the articles in five
"top-tier" logistics research journals performed by Dunn et al.* The second section
assesses the need for, and describes the arguments in favor of, scientific research
within the framework of logistics. The concepts of theory-testing and the role of
science are discussed. In section three, a detailed technical description of the
methodology for construct measurement is presented as a guideline for interested
logistics researchers. The conclusion calls for the application of this methodology
in future logistics research.

LOGISTICS RESEARCH METHODS

According to La Londe and Dawson,' early references to logistics activity first


appeared in the early twentieth century with the works of Shaw,'° Cherington,''
Beckman,'^ and Borosodi.'-' Logistics rapidly gained attention as an integrated
concept during the early 1950s.'^ Since then, research methods used to study logistics
have taken fonnats ranging from systems simulations and econometric analysis to
case study development. Many of today's top logistics researchers started their careers
focused on transportation economics. Consequently, these researchers may not have
been well versed in construct measurement and scientific behavioral research.
These same researchers are not studying business logistics issues, which often
involve latent variables. While they may be applying econometric techniques, they
are still not adequately addressing the construct measurement issues. Econometric
techniques would produce more meaningful results if rigorous construct measurement
techniques were first applied.
JOURNALOF BUSINESS LOGISTICS, Vol. 15, No. 2, 1994 147

It is now appropriate to analyze the contemporary research in business logistics.


Using a framework for research paradigms developed by Meredith et al.,'^ Dunn
et al.'^ classified logistics research articles for the years 1986-1990. Five journals
were selected by the authors as a representative sample of logistics journals:
International Journal of Physical Distribution and Logistics Management, Journal
of Business Logistics, Journal of Purchasing and Materials Management, Logistics
and Transportation Review, and Transportation Journal. These journals were
included in Allen and Vellenga's*' assessment of top logistics journals. An explanation
of the Meredith et al. framework is presented in Figure 1. The Meredith model
has two continuums that enable categorization of research based upon the underlying
tenets of its methodology—the rational/existential (R/E) and the natural/artificial
(N/A).
The R/E continuum is based upon whether the research is independent of man
(deductive) or relative to one's individual experiences (inductive). This continuum
is more philosophical than the other. Inductive research is rooted in the researcher's
personal knowledge and experiences of the truth, while deductive research is based
upon logic and structure. This continuum is partitioned into four categories: axiomatic
or the "theorem-proof world," the most rational; logical positivist/empiricist, in
which "the facts are assumed to be independent of the laws and theories used
to explain them;" interpretive, a more inductive approach in which researchers study
people rather than objects, and where facts are not considered separate from the
theory or observer; and critical theory, the most subjective approach, which transcends
the inductive/deductive argument and attempts to generalize the phenomena without
scientific method.'*
The N/A continuum is concerned with the source of the infonnation utilized
in the research; in other words, its subjectivity or objectivity. This continuum is
partitioned into three categories, ranging from the empiricist extreme in which
explanation is derived from concrete, objective data, to subjectivism, in which
explanation is derived from interpretation and artificial reconstruction of reality.
The most objective, object reality, is based on direct observation of the phenomena
by the researcher. Peoples' perception of object reality, a less direct form of
observation, utilizes the observations and values of other people. The third and
most subjective category is artificial reconstruction of reality, in which the researcher
attempts to explain reality by using a model.
148 DUNN, SEAKER, AND WALLER

FIGURE 1

FRAMEWORK FOR RESEARCH METHODS


(MEREDITH ET AL.)*

Direct Peoples' Artificial


Observation of Perception of Reconstruction
Object Reality Object Reality of Object Reality
Rational Axiomatic •Reason/Logic/
Theorems
•Normative
Modeling
•Descriptive
Modeling
Logical •Field Studies •Structured •Prototyping
Positivist/ *Field Interviewing •Physical
Empiricist Experiments •Survey Modeling
Research •Laboratory
Experiments
•Simulation
Interpretive *Action •Historical •Conceptual
Research Analysis Modeling
*Case Studies •Delphi •Hermenuetics
•Intensive
Interviewing
•Expert Panels
•Futures
Scenarios
^ Cdtical •Introspective
Existential Theory Reflection

•J. Meredith, A. Raturi, K. Gyampah, and B. Kaplan, "Alternative


Research Paradigms in Operations," Journal of Operations
Management 8, no. 4 (October 1989): 297-326.
JOURNALOF BUSINESS LOGISTICS. Vol.15, No. 2, 1994 149

When the two axes of the Meredith et al. matrix Jire applied, the majority
of logistics research tends to fall into three areas. The interpretive/perceptive (27%),
artificial/axiomatic (24%), and logical-positivist/perceptive (23%) paradigms account
for three-quarters of the total research articles.
Figure 2 presents the findings of Dunn et a l . ' '

FIGURE 2

LOGISTICS JOURNAL PARADIGMS


1986-1990

Direct Perceptive Artifical


Axiomatic 24%
Log/Pos 1% 23% 3%
Interpretive 7% 27% 9%
Critical 6%

The perceptive/interpretive category (27%) consists of historical research,


intensive interviewing, Delphi panels, and futures scenarios. The hallmark of this
method is subjective interpretation by the researcher. The perceptive/logical positivist
category (23%) consists of surveys and structured interviews. The artificial/axiomatic
category (24%) involves rigorous, logical analyses that can be followed and replicated
by other researchers, such as analytical and mathematical models.
This paper focuses on the perceptive paradigm, specifically the surveys and
structured interviews of the logical positivist category. While the methods currently
employed in this area are appropriate for conceptualizing logistics theory, there
exists a great opportunity to apply additional methodologies that will contribute
to the building and testing of the theory. Section two addresses this statement in
detail. A review of the articles categorized as perceptive/interpretive and logical
positivist (e.g., survey and interview research) by the Dunn et al. article reveals
that the majority of this type of research did not incorporate a rigorous methodological
approach as will be advocated in this article.
150 DUNN, SEAKER. AND WALLER

THE SCIENTIFIC METHOD IN LOGISTICS RESEARCH

Theory is a precursor to generalized knowledge. Sutherland succinctly describes


theory as "an ordered set of assertions about a generic behavior or structure assumed
to hold throughout a significantly broad range of specific instances."^'^ Good theory
should be the intention of logistics theorists.
Building logistics theory through research that relies on ambiguous or general
descriptions of phenomena leaves logistics knowledge on treacherous footing.
Furthermore, much of the research tends to present such phenomena in qualitatively-
derived and prescriptive relationships. There needs to be more clarity in research
variables and more rigor in the methodologies. When theory is presented in testable
form and is eventually tested, knowledge becomes more objective, more dependable,
and less value-laden.^'
To achieve these objectives, increased scientific scrutiny must be applied to
logistics research, especially when latent variables are involved. This is important
since research involving latent variables is appearing more frequently in the logistics
literature. It is emphasized that this methodology is not intended to replace
the current approaches to logistics research, hut is intended to complement
them.
Some scientific research efforts, such as those in physics, are indebted to strong
inference. Strong inference is achieved through a consistent application of scientific
research methods, where they are used and taught on a systematic basis.^^ Such
methods provide an accumulation of inductive inference and are effective in
producing significant results. Logistics must begin to focus on a program of alternative
hypotheses and disproofs by invoking a strong inference approach as part of its
methodological regimen. The goal of logistics research, as with any research, should
be to ensure that theoretical systems and statements can be empirically tested. An
ability to consistently explain and predict phenomena must exist.^^
An explanation of the scientific method may best be approached by first
explaining the concept of science. Science is a structured approach toward developing
knowledge and requires observation, identification, description, experimental
investigation, and theoretical explanation to natural objects.^'* It is concerned with
quantitative formulation and general laws as a means of providing inductive under-
standing.^^ Knowledge derived from science provides a higher degree of
dependability.^^
JOURNAL OF BUSINESS LOGISTICS, Vol. 15, No. 2, 1994 151

Within science, there is the assumption that there is a real, tangible world
where phenomena are independent of human interpretation or conjecture. It contains
rules for analyzing natural phenomena. Science aims to observe and explain these
natural phenomena. It provides a dependability to knowing. Science is not concerned
with ideas lying within the metaphysical or theological realm. If there is to be
progress in the construction of knowledge, it is real world phenomena that will
provide observable, measurable structure to ongoing research processes. Therefore,
methodology used to test theories rooted in this objective world should also be
as free of human persuasion as possible.
Science can provide such tests. The scientific approach has a self-correction
characteristic so that scientific activities and conclusions are controlled and verified
so that the end result is dependable knowledge.
Scientific research is composed of objective observation and systematic
procedures. The purpose of the scientific method is to test theory. However, theory
can be built or adjusted based on unexpected test results. For example, a control
variable may unexpectedly provide a strong moderating effect to an independent
variable's explanation of the dependent variable. At this juncture, the original theory
may require some additional thought.
Science attempts to provide a consistent and unbiased view of observable
phenomena and how they relate to each other. It does so while maintaining objectivity
in both measurement and testing. Testing procedures are systematic and controlled
in that the investigation itself provides a strong assurance that the results are accurate
and consistent.
Science alone is not a panacea. There are weaknesses associated with the
scientific method. Howson and Urbach^^ have exposed some of the problems of
induction as they relate to knowledge. They probed for underlying assumptions
in scientific methodologies. For example, Hume's principle of the uniformity in
nature emphasizes that those ascribing to scientific research assume an a priori
"truth" that the future will resemble the past. It allows the scientist to assuredly
predict future occurrences based upon current experimental results.
Relationships between constructs built on perceptions (e.g., customer
satisfaction), prevalent in the behavioral sciences, have a tendency to be less
generalizable through time. In the natural sciences such as oceanography, tides can
be accurately and consistently predicted by the lunar positioning. In contrast, the
behavioral sciences (where, for example, transportation buying behavior might be
152 DUNN, SEAKER, AND WALLER

quantified) present an array of value-based variables in a multiple of possible


combinations. The predictive processes are less exact and, hence, more probabilistic.
As Nunnally states, "The intention of science is to try to find some underlying
order in all the particular events that occur in nature... Many generalizations in
all areas of science, and particularly in the behavioral sciences, must be stated
in statistical terms, with respect to the probability of an event occurring, rather
than specified with more exactness."^^
Although they will not be investigated at this time, other epistemological
questions can present challenges to the foundations of scientific research. Any and
all other research paradigms are also susceptible to attacks of this nature. Though
the underpinnings of science contain basic assumptions, it must be understood that
the importance of science is its ability to provide the following: (1) a consistency
in method, (2) exclusion of human values and emotions, and (3) external and internal
validity.
External validity is a direct function of the extent to which experimental results
can be applied to other scenarios, firms, industries, etc. Internal validity indicates
the degree of accuracy at which the test variables and the experiment measure
the conceptualized variables and their relationship(s), respectively.
Science does not rule out the importance of qualitative methods in research.
Eisenhardt^^ states that qualitative methods have the ability to indicate relationships
that would not otherwise be observable to the quantitative researcher. The case
study method, for example, can provide a grasp of the interactive dynamics within
a single setting. Such activities shape hypotheses and help build theory. However,
it may lend itself to a constant infusion of preconceived biases held but not consciously
recognized by the researcher.
Weiss supports qualitative methods, and states that "qualitative data are superior
to quantitative data in density of infomiation...not precision and reproducibility."^°
Flynn et al.'^' additionally point out that "without a conscious attempt at theory-
building (a focus of qualitative approaches), research can degenerate into simply
data-dredging or the ad hoc collection and analysis of data for whatever reason
can be found."^^ Consequently, divergent, creative approaches to research, as
provided by qualitative efforts, can help to ensure that a research programme (a
term used by Lakatos, Criticism and the Growth of Knowledge, New York: Cambridge
University Press, 1968, p. 155, to represent the collective research efforts within
a particular paradigm), will not become mired in one favorite tendency.
JOURNAL OF BUSINESS LOGISTICS, Vol. 15, No. 2, 1994 153

Research within the field of logistics may best be approached through the
application of multiple methods, both quantitative and qualitative. Eisenhardt,^-'
Jick,'''* and Campbell and Fisk^^ among others suggest that this triangulated approach
to research can create better assurances that variances are trait-related and not
method-related.
Logistics knowledge should be approached with an array of diverse
methodological tools to ensure the coexistence of both theory-testing and theory-
building. Weick^^ provides a model based on Thomgate's postulate of commensurate
complexity illustrating the strengths and weaknesses in the various approaches to
research in the social sciences. In this conception, the power of research outcomes
can be measured by its adeptness at providing the virtues of generalizability, internal
validity, or simplicity. There appears to be a form of mutual exclusivity regarding
each.
As any single research method attempts to satisfy any two of the three virtues,
the third virtue is increasingly alienated. For example, as empirical science attempts
to provide both external and internal validity in its outcomes, simplicity is sacrificed.
In contrast, as case study research attempts to capture both internal validity and
simplicity as its virtues, it loses generalizability.
Figure 3 incorporates Weick's model into an encompassing research framework.
By combining the large array of multiple logistics theories and multiple methods
into a collective research effort, a more powerful means of extracting new, meaningful,
and dependable knowledge is made available.
Research in business logistics tends to span most of the approaches in Figure
3. Although there is much survey research on business logistics topics, it frequently
lacks scientific rigor. This is especially true in terms of construct measurement.
The next section of this paper provides a framework for construct measurement
that should facilitate scientific survey research in business logistics.

MEASURING LATENT VARIABLES

Suppose a researcher has rigorously developed a theory that involves the


following proposition: Decentralized logistics functions lead to better customer
service. This proposition involves two constructs—decentralized logistics and
customer service—that must be measured in order to test the proposition.
154 DUNN, SEAKER, AND WALLER

FIGURE 3

METHODOLOGICAL TOOLS FOR


LOGISTICS RESEARCH

SNOWUDCS

Lab
Analytic
Models

Accuracy ' Scientific


Survey
Reseaichy

Case
Multiple
Case Studies
JOURNALOF BUSINESS LOGISTICS, Vol. 15, No. 2, 1994 155

Items or questions can then be listed that would measure, indirectly, the
constructs. For example, to measure decentralization several questions might be
asked. The following are examples of possible survey questions:
1. Routine logistics decisions are not formalized. (Circle one.)
A. Strongly Agree
B. Agree
C. Neutral
D. Disagree
E. Strongly Disagree

2. Logistics problems are resolved immediately without committees or


memoranda.
A. Strongly Agree
B. Agree
C. Neutral
D. Disagree
E. Strongly Disagree

These are only two of the possible items that might reflect a decentralized
logistics function. As mentioned above, suppose the researcher developed seven
such questions to measure, indirectly, the extent of decentralization in a logistics
function. Should all of these questions be included in the measurement of
decentralization? What statistical techniques are available to decide which questions
to use in measuring this construct? Do these seven questions adequately cover
the scope of what is meant by decentralization? Do these seven questions measure
decentralization per se or do they also measure another construct which is included
in the theory?
The balance of this section is aimed at answering these and other questions
related to latent variable measurement. Figure 4 is a flow chart of the process
of scale development and validation. It is important to note that this process is
iterative, as well as, sequential.
156 DUNN, SEAKER, AND WALLER

FIGURE 4

THE PROCESS OF SCALE DEVELOPMENT AND VALIDATION


Define Constructs
\ yes
Develop Potential Items
^ yes
no Check Content Validity
^ yes
no
Confirm Substantive Validity
\ yes
Pilot Survey/Survey
yes
Item Purification
no Exploratory Factor Analysis
no
Item to Total Correlation 1
i yes
Confirmatory Factor Analysis
1 no
Dimensionality Purification | yes
1 \ yes
nn
Reliability
\ yes
no 1 Convergent Validity
\ yes
no Discriminant Validity
yes

no
1
1 Criterion Related Validity
Predictive Validity |
1
H Concurrent Validity
\ yes
h

no Nomologicai Validity
i yes
Test Theory
JOURNAL OF BUSINESS LOGISTICS, Vol. 15, No. 2, 1994 157

Content Validity
In order to measure latent variables, constructs should be carefully defined
from the literature and the author's understanding of the constructs. A set of tentative
items can then be produced to measure each construct. This approach helps ensure
content validity:
One can imagine a domain of meaning that a particular construct is intended
to measure. Content validity refers to the degree that one has
representatively sampled from that domain of meaning. ...[T]he researcher
should search the literature carefully to determine how various authors
have used the concept which is to be measured. ...[Also], researchers should
rely on their own observations and insights and ask whether they yield
additional facets to the construct under consideration.
Content validity exists when the scope of the construct is adequately reflected by
the items as a group. Unfortunately, there is no rigorous way to assess content
validity.^^ Multiple items are typically used to measure constructs so that construct
measurement will be thorough.^' Another term for measure is scale, that is, a
collection of items mapped into one variable. If content validity does not exist,
then there is no reason to proceed with the analysis because the desired construct
is not being properly represented by the group of items. This means that the
researchers will not be able to use the scale to test the hypothesis.
A logistics researcher may wish to measure customer service. If the researcher
uses only number of stockouts as a measure, the scale may lack content validity
because it does not sufficiently span the scope of the meaning of customer service.
The researcher may wish to change the name of the construct to "number of
stockouts." However, if the researcher is really interested in measuring customer
service, then other items should be included in the scale.

Substantive Validity
It is not possible for a scale to have content validity without having substantive
validity.'*^ If a measure has substantive validity, then its items are conceptually
or theoretically linked to the construct.'^' Content validity differs from substantive
validity in that content validity deals with a set of items (scale) whereas substantive
validity deals with each individual item of a construct. Anderson and Gerbing suggest
158 DUNN, SEAKER, AND WALLER

purifying the set of items, using a test for substantive validity."^^ Item purification
involves eliminating those items that do not agree with the set of items.
Many times researchers attempt to purify a set of items in a pretest by using
item-to-total correlations or contribution to Cronbach's coefficient alpha. The problem
with this approach is that the correlations are typically not statistically significant
because the sample size is often small in a pilot survey. As a result, the researcher
may be eliminating items that should not be eliminated or may be retaining items
that will weaken construct validity.
One solution would be to increase the sample size of the pretest. If the researcher
is to use exploratory factor analysis, the sample size must be fairly large. Exploratory
factor analysis is a statistical technique used to find which variables or items are
agreeing with one another. Those that do not agree with a given scale can be
eliminated as long as content validity is not jeopardized. This raises another problem,
namely, cost. To use exploratory factor analysis there should be at least ten
observations for each item'*-' and at least 200 observations'*'* even if there are fewer
than twenty items on the questionnaire. Therefore, use of exploratory factor analysis
in a pilot study is potentially very costly.
One solution to this dilemma is to assess substantive validity in a pretest setting
using the approach of Anderson and Gerbing.^^ They performed an empirical study
where their substantive validity assessment technique was able to predict which
items should have been retained and which should have been deleted. The results
of the substantive validity assessment were stable across multiple pretest samples.
The methodology for evaluating substantive validity is an item-sort task.'*^
Experts or a sample of the population to be surveyed are given a sheet of paper
that defines each of the constructs and another sheet of paper that contains all
of the items in random order. The participants are asked to match the items to
the constructs that they best represent. When this method is used as the pretest,
then the pretest population should be representative of the study population."*^
Anderson and Gerbing'** have developed two indices for evaluating the
substantive validity. One index is the proportion of substantive agreement (PSA),
which measures how well the items were assigned to their hypothesized constructs.
PSA is the number of participants assigning an item to its hypothesized construct
divided by the number of participants. The PSA index is an element of the unit
interval. A high value of the PSA index indicates agreement as to the construct
JOURNAL OF BUSINESS LOGISTICS. Vol. 15, No. 2, 1994 159

that a given item is assigned to. For example, if an item has been assigned to
a given construct by eight of the ten participants, the PSA value is 0.80.
The other index is the substantive validity coefficient (CSV). CSV is the
difference between the number of participants assigning an item to its hypothesized
construct and the highest number of assignments of an item to any other construct;
that quantity is then divided by the number of observations. For example, suppose
that six of ten participants assigned an item to the hypothesized construct, three
assigned it to another construct, and one assigned it to another construct. In this
case CSV = (6-3)/10 or 0.30.
The CSV index is a member of the interval from -1.0 to +1.0. Some items
are often eliminated in this stage in the research methodology. A critical value
of 0.5 was used to determine which items should be eliminated in the Gerbing
and Anderson study."*^ This stage helps assure substantive validity.
The substantive validity test together with the pilot study is intended to improve
the efficiency of the questionnaire as much as possible before data are collected.
This helps ensure high quality research. A rigorous statistical analysis of bogus
data is of little value. If substantive validity does not exist for an item and the
item is still included in the scale, then that item will probably have a low correlation
with the other items in the scale. Therefore, it will probably be eliminated during
the ensuing scale refinement stage of the analysis.
The research of Hunt, Sparkman, and Wilcox^° shows that the most common
type of error detected in a pretest is missing alternative errors, i.e., questionnaire
errors that result from not including the correct alternative in the set of alternatives.
The use of inappropriate vocabulary was the second most frequently detected error.
Telephone interviews were found superior to personal interviews for detecting errors.
After the pretest has been accomplished, the survey can be sent out. Several
books and articles exist on survey research^^ and related topics.^

Scale Refinement
Scale reliability, hereafter referred to as reliability, refers to the internal
consistency of the items that are used to measure a latent construct. Internally
consistent items form a homogeneous set in that they vary together statistically.
Therefore, in this context the term "reliability" refers to the accuracy or precision
160 DUNN, SEAKER, AND WALLER

of the scale.^-' Reliability is most commonly estimated using Cronbach's coefficient


alpha.^"^ Cronbach's coefficient alpha is given by the following expression:

a = Np/[1 + p(N-l)] (1)


In equation (1) N is the number of items and p is the mean inter-item correlation.
Cronbach's coefficient alpha is an element of the closed unit interval (i.e., 0 <
a < 1).
A higher level of Cronbach's coefficient alpha indicates a higher reliability
of the scale.^^ As can be seen from equation (1), Cronbach's coefficient alpha
is monotonically increasing in p and N. The reason for this is as follows: Higher
inter-item correlations indicate statistical agreement among the items; as N increases,
the probability of spurious correlation decreases. Nunnally^^ strongly encouraged
the use of Cronbach's coefficient alpha for assessing reliability. He also presented
a rule of thumb, stating that alpha levels higher than 0.70 indicate intemal consistency
among the items of a scale and that alpha levels as low as 0.60 are acceptable
for new scales.
There are other dimensions of reliability; however, the dimension described
above is the one most commonly referred to in the literature. The other types of
reliabilities and their measures will not be described here.
Item-to-scale (i.e., item-to-total correlation) correlation is the statistical
correlation between a given item and the scale to which it belongs. If a scale
contains an item with a low item-to-scale correlation, then the scale may be deemed
to be unreliable in that the items are not internally consistent.
Item-to-scale correlation is mainly used for scale refinement (Nunnally^^; Saraph,
Benson, and Schroeder^^). Scale refinement refers to the elimination of items from
a scale that do not statistically agree with the other items of the scale. Scales
are refined before reliability is assessed.
The two most common methods of scale refinement are the following: (1)
analysis of item-to-scale correlations, and (2) analysis of item-to-item correlation
within a given scale. Using item-to-scale correlations for scale refinement entails
elimination of those items with low (often < 0.3) item-to-scale correlations. This
clearly results in higher reliability estimates using Cronbach's coefficient alpha.
In general, the same procedure applies when item-to-item correlation within a given
scale is used.
JOURNAL OF BUSINESS LOGISTICS, Vol. 15, No. 2, 1994 161

If a scale has a low alpha value (below 0.60 for a new scale), then the researchers
can examine the item correlation matrix of that scale. Those items with low
item-to-item correlations can be deleted from the scale as long as the scale will
retain its content validity.
Exploratory factor analysis is a technique that is used to find latent variables
(factors) that are represented by a group of items or observed variables. It is often
referred to as a variable reduction technique,^^ and can be used to further refine
the scales by analyzing the observed variables. Items that are poorly related to
all factors or clearly represent more than one dimension are removed, resulting
in more reliable scales.
Gerbing and Anderson^^ encouraged the use of both item-to-scale correlation
(e.g., item-total correlations) and exploratory factor analysis to provide preliminary
scales. Exploratory factor analysis differs from confirmatory factor analysis, which
will be discussed in the section on construct validity. However, both types of factor
analysis have the following underlying assumption in common: Factor analysis is
based on the fundamental assumption that some underlying factors, which are smaller
in number than the number of observed variables, are responsible for the covariation
among the observed variables.^'
Scale refinement, within the context of this paper, is conducted on the same
data that is used to test the hypotheses. This is standard practice. However, it would
be better to refine the scales with one set of data and then test the hypotheses
with another set of data. The problem with this is that it is often too expensive
to acquire the extra data set.

Unidimensionality
Construct validity depends on how well the scale of a construct actually measures
that construct. A scale is construct valid ...
(1) to the degree that it assesses the magnitude and direction of a
representative sample of the characteristics of the construct; and (2) to
the degree that the measure is not contaminated with elements from the
domain of other constructs or error.
However, a scale cannot have construct validity unless it is unidimensional. It
is acceptable to have a multidimensional construct, but scales must be unidimensional.
^62 DUNN, SEAKER, AND WALLER

If a researcher is attempting to measure a multidimensional construct, several


unidimensional scales must be used. This is known as a second order factor model.^'^
Product quality may be a construct that a logistics researcher wishes to measure.
According to Garvin^^ there are eight dimensions of quality: reliability, durability,
conformance, etc. In this case, since product quality is a multidimensional construct
and a scale must be unidimensional, one scale should be used for each dimension.
If there is a sufficient level of correlation between each of the scales, then product
quality may be considered a second order factor model.
It is impossible to have construct validity without a unidimensional scale but
having a unidimensional scale is not sufficient for construct validity.^^ That is,
unidimensionality is a necessary but not sufficient condition for construct validity.
A scale is unidimensional when the items of a scale estimate one factor. Gerbing
and Anderson^' demonstrated how unidimensionality can be assessed using
confirmatory factor analysis.
Confirmatory factor analysis is similar to exploratory factor analysis except
that confirmatory factor analysis is more flexible. Exploratory factor analysis assumes
the following:
(1) all common factors are correlated ...; (2) all observed variables are
directly affected by all common factors; (3) unique factors are uncorrelated
with one another; (4) all observed variables are affected by a unique factor;
(5) all [constructs] are uncorrelated with all [error terms].^^
The assessment of the unidimensionality of a scale involves further scale
refinement. In fact, unidimensionality should be assessed before reliability is assessed.
Scales should not be developed based on selection of items that maximize reliability.^'
The dimensionality of a scale is different from the reliability of a scale:
The dimensionality of a scale can be evaluated by examining the patterning
of its component indicator correlations, whereas the reliability of a scale
is determined by the number of items that define the scale and the
reliabilities of those items.'°
To implement the test for unidimensionality, the researcher uses confirmatory
factor analysis. Those items that load weakly on the hypothesized factors are
eliminated from the scale, thus resulting in a unidimensional scale. A problem that
can arise at this point is that content validity can be lost. That is, after the
JOURNAL OF BUSINESS LOGISTICS, Vol. 15, No. 2, 1994 163

unidimensionality refinement of the scale, the scale may lose its content validity
because it no longer reflects the scope of the construct. Content validity must always
be considered when eliminating items from a scale. Sometimes the researcher is
faced with a dilemma—namely, eliminate an item to achieve unidimensionality,
resulting in a loss of content validity or keep the item, maintaining content validity
while not achieving unidimensionality.
If the researcher uses a scale that is not unidimensional, then it will reduce
the scale's variance. In other words, items that are reflecting one construct in a
scale will, to some extent, off-set changes in items in the same scale that reflect
another construct. This may result in a reduction in the explanatory power of the
model as a whole.

Reliability
Once unidimensionality has been established, reliability can be assessed using
Cronbach's coefficient alpha. Cronbach's coefficient alpha was discussed prior to
this section because the concept of reliability is useful in the discussion of
unidimensionality.
Reliability is of little significance until unidimensionality has been established.
It is possible to have a reliable scale that is measuring more than one construct,
fulfilling a sufficient condition for lack of construct validity. This is why
unidimensionality should be established before reliability is established.

Construct Validity
Once unidimensionality has been established, construct validity can be
investigated. Construct validity is the extent to which a scale measures the construct
it was intended to measure. The concept of construct validity has been difficult
for scientists to operationalize.^' Convergent validity and discriminant validity are
the criteria most frequently used to support construct validity.'^
Convergent validity is the degree to which there is agreement between two
or more attempts to measure the same construct through dissimilar methods.'^
Discriminant validity depends on the degree to which scales measure distinct
constructs.^** When convergent validity and discriminant validity are found, construct
validity is supported.
For a thorough discussion of convergent validity and various ways of checking
for it see Bagozzi, Yi, and Phillips.^^ In this discussion, assessment of convergent
DUNN, SEAKER, AND WALLER

validity will be explained through a confirmatory factor analysis of the data. If


the factor loadings are all statistically significant, then convergent validity exists.
This means that the factor loadings were significantly different from zero according
to the t-values.
Convergent validity can easily be checked after unidimensionality has been
achieved using the confirmatory factor analysis approach. The process of achieving
unidimensionality results in the information needed to check for convergent validity.
The test for discriminant validity involves comparing the correlations between
the constructs of a model to a hypothetical model. In the hypothetical model all
of the correlations between the constructs are set to be one. Then the chi-squared
difference value can be calculated. If the value is associated with a p value of
less than 0.05, then discriminant validity is supported.'''^ Discriminant validity
indicates that the constructs are distinct.
If a particular pair of constructs has an unusually strong correlation, then it
is appropriate to perform a separate chi-square difference test on that individual
correlation. Even when the correlation is very high, the hypothesis that the population
correlation is one can be rejected.^'
If support is found for convergent and discriminant validities, then construct
validity is supported. However, one study cannot establish construct validity'^
because it is an on-going process of investigation (Cronbach, 1971). Construct validity
is a function of previous studies, current research, and future research.

Criterion-Related Validity
Criterion-related validity refers to how well a scale correlates with the criterion
it is trying to predict. If the criterion exists in the present, then it is called concurrent
validity; if the criterion exists in the future, then it is called predictive validity.
The following is an example of predictive validity: GMAT scores are used
by business schools to predict how well a given applicant for an MBA program
will perform if allowed to enter the program. Therefore, to measure the predictive
validity of GMAT scores, the researcher might find the correlation between GMAT
scores and grade-point-averages of graduating MBAs. The higher the correlation,
the greater the support for the test's predictive validity.
Suppose a researcher was using subjective items to measure a construct called
manufacturing quality. To assess concurrent validity, the researcher could conduct
JOURNALOF BUSINESS LOGISTICS, Vol. 15, No. 2, 1994 165

correlation analysis on an objective measure of manufacturing quality, such as scrap


rate, and the scale used to measure manufacturing quality.
Unfortunately, it is often not possible to find a reasonable criterion for which
to judge a scale's predictive validity in the social sciences:
[An] important limitation is that, for many if not most measures in the
social sciences, there simply do not exist any relevant criterion variables.
For example, what would be an appropriate criterion for measures of
a personality trait such as self-esteem? We know of no specific type of
behavior that people with high or low self-esteem exhibit such that it
could be used to validate a measure of this personality trait.
If criterion-related validity is found not to exist, then the scale used to measure
the construct of interest is of little value. However, the researcher should be confident
that the appropriate criterion is being used before reaching such a conclusion.

Nomologicai Validity
Nomologicai validity of a construct exists when the construct relates to the
other research constructs in a way that is consistent with the underlying theory.^
Therefore, at this stage of the analysis the line between construct validation and
theory testing becomes blurred. This obfuscation occurs because the approach to
investigation of nomologicai validity and the testing of theory are identical.
If a construct, in the nomologicai network, does not behave in the way theory
predicts, does that show that the construct is not measuring the appropriate latent
variable? Or is the theory wrong? Therefore, if nomologicai validity does not hold,
the situation becomes equivocal. However, if the nomologicai validity of a construct
is supported, then the part of the theory which involves that construct is also supported.

CONCLUSION

This paper calls for a more scientific approach to empirical research in business
logistics, especially when latent variables are constituted in the theory. A
methodological road map is provided to guide researchers through the often enigmatic
maze of empirical science as it attempts to operationalize construct. Figure 4 is
a flowchart of this process.
166 DUNN, SEAKER, AND WALLER

The process recommended in this paper has important ramifications for logistics
research. The paper is intended as a recommendation to logistics researchers as
an important step forward conceming survey research, specifically, the methodologies
used in that research.
Adoption of the procedures presented in this paper will enable a common set
of standards to be set for measuring latent variables in logistics. It will allow the
logistics research community to find common ground in its understanding of often
elusive concepts. This will provide for a stronger inference within the paradigm,
resulting in improved logistics practice.
Application of standard methodological procedures will bring a common
discipline and order, enhancing both academic and professional acceptance of
research. Additionally, it will enable logistics researchers to contribute to and draw
from the scholarly work in other disciplines. This is important because logistics
is a cross-functional discipline.
The methodology utilized for survey research in the perceptive paradigm can
have ramifications for other forms of logistics research. Although the deductive
forms of logistics research such as model building are very sophisticated, they
are hindered by a lack of generalizability. The more inductive case study approach
suffers from the same problem. Empirical science will facilitate generalizability,
hence improve theory building.
Logistics researchers have successfully generated a plethora of hypotheses and
propositions. Now as logisticians enter the realm of theory testing, articles addressing
methodology can be expected to follow. These contributions, collectively, will bring
about a more vigorous and pervasive theoretical base to the discipline. It is hoped
that this paper will provide encouragement for researchers who are interested in
this genre of research by reducing some of the barriers that may have inhibited
them.
As the logistics discipline accumulates sound theory it will become more
applicable for practice, thus reducing lag between research and application. The
dependability and generalizability of the theory will facilitate such an outcome.
Logistics researchers will not need to worry about compromising relevance for
theory when this approach is taken.
JOURNAL OF BUSINESS LOGISTICS, Vol. 15, No. 2, 1994 167

NOTES

' F . Kerlinger, Foundations of Behavioral Research (Fort Worth, Tex.: Harcourt


Brace Jovanovich, 1986), p. 37.
^W. Kendal and C. Jordan, "Fare and Income Elasticities in the North Atlantic
Air Travel Market: Economic and Policy Implications," Transportation Journal 28,
no. 4 (1989): 35-41.
3j. I. Hsu and M. K. El-Najdawi, "Integrating Safety Stock and Lot-Sizing
Policies for Multi-Stage Inventory Systems Under Uncertainty," Journal of Business
Logistics 12, no. 2 (1991): 221-238.
'*P. D. Larson, "Business Logistics and the Quality Loss Function," Journal
of Business Logistics 13, no. 1 (1992): 125-148.
^M. Mahmoud, "Optimal Inventory Consolidation Schemes: A Portfolio Effect
Analysis," Journal of Business Logistics 13, no. 1 (1992): 193-214.
•^H. Min, J. Current, and D. Schilling, "The Multiple Depot Vehicle Routing
Problem withBackh&uVmg" Journal of Business Logistics 13, no. 1 (1992): 259-288.
''S. Dunn, R. Young, A. Stenger, and R. Seaker, "An Assessment of Logistics
Research Methods," Working Paper, Center for Logistics Research, The Pennsylvania
State University, no. 10 (1992).
^Same reference as Note 7.
' B . La Londe and Dawson in D. Bowersox, B. La Londe, and E. Smykay,
eds.. Physical Distribution Management (London: Macmillan, 1969), pp. 9-18.
'°A. Shaw, An Approach to Business Problems (Cambridge, Mass.: Harvard
University Press, 1916).
I'P. Cherington, The Elements of Marketing (New York: Macmillan, 1920).
l^T. Beckman, Wholesaling (New York: Ronald Press, 1926).
'3R. Borosodi, The Distribution Age (New York: D. Appleton, 1929).
''*D. Bowersox, Logistical Management (New York: Macmillan, 1974).
'5j. Meredith, A. Raturi, K. Gyampah, and B. Kaplan, "Alternative Research
Paradigms in Operations," Journal of Operations Management 8, no. 4 (1989):
297-326.
reference as Note 7.
DUNN, SEAKER, AND WALLER

'^B. Allen and D. Vellenga, "Affiliation of Authors in Transportation and


Logistics Academic Journals—An Update," Transportation Journal 27, no. 3 (1987)-
39-47.
reference as Note 15, at pp. 305-307.
''Same reference as Note 7.
^°J. Sutherland, Systems: Analysis, Administration, and Architecture (New York:
Van Nostrand, 1975), p. 9.
reference as Note I.
^^J. Platt, "Strong Inference," Science 146 (1964): 347-353.
^^S. Bacharach, "Organizational Theories: Some Criteria for Evaluation,"
Academy of Management Review 14 (1989): 496-515.
r's 3rd New International Dictionary (Springfield, Mass.: G. & C.
Merriam, 1983).
reference as Note 24.
reference as Note 1.
. Howson and P. Urbach, Scientific Reasoning: The Bayesian Approach
(LaSalle, III.: Open Court, 1989).
J. Nunnally, Psychometric Theory (New York: McGraw-Hill, 1978), p. 8.
. Eisenhardt, "Building Theories From Case Study Research," Academy of
Management Review 14, no. 4 (1989): 532-550.
3°R. Weiss, "Issues in Holistic Research," in Institutions and the Person, H.
Becker, B. Greer, and R. Weiss, eds. (Chicago: Aldine, 1968), pp. 344-345.
3 ' B . Flynn, S. Sakakibara, R. Schroeder, K. Bates and E. Flynn, "Empirical
Research Methods in Operations Management,"/owrna/ of Operations Management
9, no. 2 (1990): 250-284.
^^ reference as Note 31, at p. 253.
reference as Note 29.
T. Jick, "Mixing Qualitative and Quantitative Methods: Triangulation in
Action," Administrative Science Quarterly 24 (1979): 602-611.
^^D. Campbell and D. Fiske, "Convergent and Discriminant Validation by the
Multitrait—Multimethod Matrix," Psychological Bulletin 56 (1959): 81-105.
JOURNAL OF BUSINESS LOGISTICS, Vol. 15, No. 2, 1994 169

. Weick, The Social Psychology of Organizing (Reading, Mass.:


Addison-Wesley, 1979).
^^G. Bohmstedt, "Measurement," in Handbook of Survey Research, P. Rossi,
J. Wright and A. Anderson, eds. (San Diego: Academic Press, 1983), pp. 98-99.
reference as Note 37, at p. 100.
. Churchill, "A Paradigm for Developing Better Measures of Marketing
Constructs," Journal of Marketing Research 16 (1979): 64-73.
^^J. Anderson and D. Gerbing, "Predicting the Performance of Measures in
a Confirmatory Factor Analysis With a Pretest Assessment of Their Substantive
Validities," Journal of Applied Psychology 76, no. 5 (1991): 732-740.
^^ reference as Note 40.
reference as Note 40.
reference as Note 28.
. Loo, "Caveat on Sample Sizes in Factor Analysis," Perceptual and Motor
Skills 56, no. 2 (1983): 371-374.
reference as Note 40.
reference as Note 40.
reference as Note 40.
reference as Note 40.
reference as Note 40.
5°S. Hunt, R. Sparkman, and J. Wilcox, "The Pretest in Survey Research: Issues
and Preliminary Findings," Journal of Marketing Research 19 (1982): 269-273.
5 ' D . Dillman, Mail and Telephone Surveys: The Total Design Method (New
York: Wiley, 1978).
^^D. Lambert and T. Harrington, "Measuring Nonresponse Bias in Customer
Service Mail Surveys," Journal of Business Logistics 13, no. 1 (1990): 5-25.
reference as Note 1.
. Cronbach, "Coefficient Alpha and the Internal Structure of Tests,"
Psychometrika 16 (1951): 297-334.
170 DUNN, SEAKER, AND WALLER

^E. Carmines and R. Zeller, Reliability and Validity Assessment, Sage University
Paper Series on Quantitative Applications in the Social Sciences, 07-017 (Beverly
Hills, Calif.: Sage Publishing, 1979).
reference as Note 28.
reference as Note 28.
^^J. Saraph, P. Benson, and R. Schroeder, "An Instrument for Measuring the
Critical Factors of Quality Management," Decision Sciences 20, no. 4 (1989):
810-829.
J. Kim and C. Mueller, Introduction to Factor Analysis, Sage University
Paper Series on Quantitative Applications in the Social Sciences, 07-013 (Beverly
Hills, Calif.: Sage Publications, 1978).
J. Anderson and D. Gerbing, "Structural Equation Modeling in Practice: A
Revie and Recommended Two-Step Approach," Psychological Bulletin 103, no.
3 (1988): 411-423.
^ ' j . Kim and C. Mueller, Factor Analysis: Statistical Methods and Practical
Issues, Sage University Paper Series on Quantitative Applications in the Social
Sciences, 07-014 (Beverly Hills, Calif.: Sage Publications, 1978), p. 12.
^^J. Peter, "Construct Validity: A Review of Basic Issues and Marketing
Practices," Journal of Marketing Research 18 (1981): 133-145.
^^D. Gerbing and J. Anderson, "An Updated Paradigm for Scale Development
Incorporating Unidimensionaiity and Its Assessment," 7oMrna/ of Marketing Research
25 (1988): 186-192.
^"^K. Joreskog and P. Sorban, LISREL 7: A Guide to the Program and Application,
2nd ed. (Chicago, SPSS, 1989).
^^D. Garvin, "Competing on the Eight Dimensions of Qaaiiiy," Harvard Business
Review 65, no. 6 (1987): 101-109.
reference as Note 63.
reference as Note 63.
"^^J. Long, Confirmatory Factor Analysis, Sage University Paper Series on
Quantitative Applications in the Social Sciences, 07-034 (Beverly Hills, Calif.: Sage
Publications, 1983), p. 12.
reference as Note 63.
JOURNAL OF BUSINESS LOGISTICS, Vol. 15, No. 2, 1994 171

reference as Note 70, at p. 190.


' ' R . Bagozzi and L. Phillips, "Representing and Testing Organizational Theories:
A Holistic Construal," Administrative Science Quarterly 17 (1982): 459-489.
'^Same reference as Note 62.
'^R. Bagozzi, Y. Yi, and L. Phillips, "Assessing Construct Validity in
Organizational Research," Administrative Science Quarterly 36 (1991): 421-458.
'^ reference as Note 71.
reference as Note 73.
. Joreskog, "Statistical Analysis of Congeneric Tests," Psychometrika 36
(1971): 109-133.
reference as Note 71.
reference as Note 62.
reference as Note 28, at p. 19.
reference as Note 62.

ABOUT THE AUTHORS

Steven C. Dunn is assistant professor of management at Idaho State University.


He received his MBA from Boise State University, and his Ph.D. from The
Pennsylvania State University. He has more than eight years of industry experience
in logistics and operations management. His research interests include the integration
of logistics within the organization, service industry logistics, and diffusion of
advanced manufacturing technology, and he is a co-author of Logistics in Service
Industries.
Robert F. Seaker has more than ten years of operations and marketing experience
with various carriers, both truck and rail. He is currently a doctoral candidate in
business logistics at The Pennsylvania State University. He holds a B.S. from West
Chester University and an M.S. from Penn State. His research interests include
organizational structure and culture as they affect the functional integration of
logistics.
Matthew A. Waller is assistant professor of management at Western Michigan
University. He has a Ph.D. in business logistics and an M.S. in management science
from The Pennsylvania State University, and a B.S. in economics (summa cum
172 DUNN, SEAKER, AND WALLER

laude) from the University of Missouri. His research interests focus on quality
management, logistics strategy, and supply chain management.

Das könnte Ihnen auch gefallen