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

net/publication/285089357

Path analysis vs. structural equation modeling

Article  in  Marketing Research · January 2000

CITATIONS READS

88 2,327

1 author:

Terry Grapentine
Simpson College
37 PUBLICATIONS   141 CITATIONS   

SEE PROFILE

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

Psychometrics View project

Theoretically Meaningful Concepts in Marketing View project

All content following this page was uploaded by Terry Grapentine on 03 April 2017.

The user has requested enhancement of the downloaded file.

Marketing research departments are often interested in quantifying the

relative importance of factors affecting issues such as brand equity, brand loyaity, and

customer satisfaction. Two widely used methods in this regard are path analysis (incorpo-

rating ordinary least-squares regression) and structural equation modeling (SEM) (typically

incorporating maximum likelihood estimation). The iatter is considered to be the more

"advanced" of the two, But is it? '

Miartccliiiiiresearcli H
 In most cases, the variables denoted by the rectangles are
E X E C U T I V E S U M M A R Y summated scales where two or more attributes measuring a com-
mon underlying construct are summed and divided by the num-
ber of items. For example, if four attribute statements measured
This article examines path analysis and structural equation the Durability construct—for each respondent—one simply adds
modeling when used to model marketing constructs such as the four items together and divides by four. (For more informa-
brand value and loyalty. The author examines whether these tion on how summated scales are constructed, see "Dimensions
two methods produce similar results when applied to an iden- of an Attribute," from the Summer 1995 issue of Marketing
tical data set and compares and contrasts the relative merits Research: A Magazine of Management and Applications, listed in the
of each. Because of both methods' potential shortcomings, Additional Readings section, pg. 19 |
the author says researchers may benefit from alternative Summated scales produce two benefits in models such as
research designs that more effectively control for mutticolloin- these. First, they help manage multicollinearity's effects on the
earity and other factors that can affect the reliability of path estimation of regression coefficients and second, they help focus
analysis and structural equation model findings. management's attention on more fundamental dimensions of
product/company performance, of which the individual attributes
This article addresses that question by pitting conventional are indicators.
path analysis against SEM and comparing and contrasting The straight arrows from one observed variable to another
the results from both when analyzing the same data set- denote the functional relationships between tfie variables, which
Additionally, 1 discuss the relative strengths and weaknesses of are specified by the marketing researcher prior to analysis, in
each approach and make recommendations to the practitioner. Exhibit 1, Durability 1 is functionally related to Value I. If the per-
To simplify the discussion, the term path anali^sis will always refer ceived performance of Durability I increases or decreases. Value 1
to regression-based path analysis, and SEM to LISREL-type increases or decreases as well The value construct, in turn, is
structural equation modeling incorporating maximum likelihood functionally related to brand loyalty (Loyaltyl).
estimation procedures. More formally, variables that only have straight arrows point-
ing from them to another variable—for example. Durability!.
OVERVIEW Ease of Use I, Sound I, and Price I—are called exogenous vari-
Both methods are forms of causal modeling that examine ables. This is because their values are not determined by any
relationships between and among one or more dependent vari- other variables in the model (i.e., no arrows from other constructs
ables and two or more predictor or independent variables. point to them). In contrast, Loyaltyl is an endogenous variable
Examples of dependent variables might be measures of customer because its value is determined by other constructs in the model
satisfaction, loyalty toward a brand or company, brand equity, or (e,g,, Durabilityl, Easeof Usel,Soundi, Pricel,and Valuel).The
perceived value. Independent variables frequently focus on issues various arrows pointing from these constructs to Loyalty 1 repre-
relating to brand or company performance such as perceived sent these functional relationships. Valuel is called a mediating
product quality, service quality, and price competitiveness. Both variable because it serves to mediate the effect selected exoge-
methods help decision makers understand the relationships nous variables have on the endogenous variable, (See the sidebar
between the independent and dependent variables. Testing for Mediation Effects on pg. 17.)
Neither path analysis nor SEM are methods for discovering The numbers near the straight arrows are the standardized
causative relationships. Rather, they are a means by which theo- regression coefficients obtained when an endogenous variable is
retical relationships can be tested, in applied research, one often regressed on the set of exogenous variables to which it is func-
sees path analysis employed to test relatively simple relation- tionally related. (The regression coefficients are standardized as
ships such as displayed in Exhibit 1. In contrast, SEM is some- a result of standardizing the independent and dependent vari-
times referred to as latent-variable analysis because these ables so each variable has a mean of zero and a standard devia-
models establish the relationships between "unobserved" vari- tion of one.) For example, in Exhibit I, regressing Valuel on the
ables—a topic that will be discussed in more detail shortly. In set of exogenous variables that point to it produces the following
the practical world, SEM is sometimes employed to examine standardized regression coefficients: Durabilityl=.26, Ease Of
more complicated relationships that cannot be handled by path Usel=,08, Soundl=.15,and.39=Pricel.
analysis (i.e., models incorporating reciprocal causation). (See The curved double-headed arrows linking the exogenous
the interview with Dr, R, Kenneth Teas on pg. 20.) variables to each other represent the fact that these exogenous
variables are correlated. (Unless one is conducting experimental
PATH ANALYSIS EXAMPLE design research, this situation is a fact of life in survey research.)
Exhibit 1 presents a path analysis example for a consumer The numbers by the curved arrows are the correlation coeffi-
electronics product. All variables are represented by rectangles, cients between each of the variables. For example, in Exhibit 1,
which denote "observed" variables. These variables are measured the correlation between Durabilityl and Ease of Usel is 66.
by asking respondents questions related to the construct repre-
sented by the rectangle. Such questions typically have respondents EFFECTS OF EXOGENOUS VARIABLES
rate a product on a series of attributes using some kind of scale In Exhibit I, each exogenous variable affects Loyaltyl
where higher scale numbers denote better-perceived performance. directly and indirectly. The arrows linking the exogenous vari-

14 Fall 2000
EXHIBIT 1 Consumer Electronic Product Path Analysis

,71

ables to Valuel, and then linking Valuel to Loyaltyl, suggest the their courtesy, cleanliness, product knowledge, and friendliness.
indirect linkage. In this role. Valuel is a mediating variable. To If in creating this construct measure (through a summated scale,
calculate the total effect that an exogenous variable has on for example) we omit a measure of product knowledge, we inad-
Loyaltyl, one simply adds the direct and indirect effects together. vertently create a biased sampling of the domain of the concept,
In this example, the direct effect of Durability 1 on Loyalty 1 is which results in systematic error in our measure.
62. The indirect effect is calculated by multiplying the coefficient Random measurement error, by implication, is non-system-
represented by the arrow from Durabilityl to Valuel (.261 by the atic error. An example of random error might be the fact that
coefficient represented by the arrow from Valuel to Loyaltyl respondents are not always consistent in their answers to survey
(,I2l,or.26x.l2 = .03. The total effect of Durabilityl on Loyaltyl, questions from one point in time to another (assuming of course
therefore, is 62-1-,03 = 6 5 . there is no change in the attitude or perception being measured).
As far as the computer is concerned, the data that are input
MEASUREMENT WITHOUT ERROR into a regression analysis computer program are error free. This
A critical assumption underlying the use of regression analy- assumption is often false and, consequently, the measurement
sis in calculating these coefficients is the endogenous and error "effect" is contained in the error term of the equation:
exogenous variables are measured without random error. This
assumption usually isn't true. Measurement error creeps into Y = a + pX| + |3x2 + [3x3 +e
questionnaires from many sources, which generally can be clas-
sified into two categories—systematic and random error. Also contained in the error term are the effects of other vari-
Systematic error refers to measurement error that is directional ables important in predicting Y, but are missing in the equation.
or constant in nature Gilbert Churchill describes this type of
error, which results in biased measures, in his 1999 book STRUCTURAL EQUATION MODELING EXAMPLES
Marketing Research: Methodological foundations as: "Systematic error
Refer to Exhibit 2 on page 16, which gives the SEM counter-
is also known as constant error, because it affects the measure-
part for the path analysis model of Exhibit I. This SEM model is
ment in a constant way. An example would be the measurement
calculated using the Amos software package. Version 4.0. Clearly,
of a man's height with a poorly calibrated wooden yardstick."
there are many visual differences between path analyses and SEM.
A marketing research example could be a study that pur- First, and perhaps foremost. SEM attempts to account for
ports to measure the quality of waiters or waitresses in a restau- random measurement error. That is what all the small circles with
rant. The domain of this construct might include issues such as the letter "e" in the middle of them suggest. Consequently, the

IS
standardized coefficients in a structural equation model possess ables are contained in rectangular boxes, and most of the struc-
more reliable estimates of how an exogenous variable affects an tural equation model variables are represented by ovals. Each
endogenous variable than what is produced with path analysis oval has one or more small boxes linked to it with arrows point-
ing from the ovals to the small boxes.
COEFFICIENT ESTIMATION IN PATH ANALYSIS AND SEM In Exhibit 2, qia, qls, qlu, and qlbb are four attribute state-
In calculating path analysis coefficients, ordinary least- ments that serve as measures or indicators of the unobserved
squares regression analysis minimizes the squared distance variable, Durability2. In this structural equation model,
between the estimated Y variable (the dependent or exogenous Durability2 is an unobserved or latent variable. (In Exhibit I,
variablel and the actual value of Y In contrast, SEM typically Durability! is an observed variable because it's a summated
incorporates the covariance matrix of the independent and scale comprised of these same four variables.) The arrows point-
dependent variables. It uses a maximum likelihood estimation ing from Durability2 to the q's denote the variance in the
procedure to derive the "most likely" coefficient values, given the observed variables (the q's) is caused by an underlying construct
actual covariance matrix. (The covariance between any two vari- that we cannot directly observe, but which the researcher
ables measures the extent that a change in one variable is associ- believes is important in predicting variables Value2 and Loyalty2,
ated with a change in another variable. Covariance is not the The "e's," on top of the q's, represent the random error associated
same as correlation, but they are related. The correlation between with Durability2's ability to completely explain the variance in
two variables is equal to the covariance of the two variables the q's. SEM takes this measurement error into account when
divided by the product of the variables' standard deviations.) calculating the model's coefficients—path analysis doesn't.
The numbers beside the arrows pointing from Durability2 to
OBSERVED VS. LATENT VARIABLES the q's can be considered as standardized regression coefficients.
One obvious difference between the path model and the The larger the number, the more a q variable can be considered a
structural equation model is that all path analysis model vari- good indicator of the latent variable. The numbers beside the

EXHIBIT 2 Consumer Electronic Product Maximum Likelihood Estimation

el 7

16 Fall 2000
arrows—From the e's to the q's—represent the amount of vari- how each goes about estimating the model's coefficients.
ance in the q's that can be explained by the unobserved variable, Moreover, SEM uses a confirmatory factor analysis approach to
Durability2. The larger the number, the more the unobserved measure unobserved variables, whereas the path model only
variable can explain the variance in the q's, uses observed variables.
As with path analysis, the structural equation model also
allows us to derive the total effect that each of the exogenous ARE MODEL RESULTS DIFFERENT?
variables has on Loyalty2 (the direct and indirect effects com- The rank order of importance of the total effects of the
bined) by using the same procedures discussed earlier for exogenous variables on Loyalty is comparable between the path
path analysis. In the path analysis model, the total effect of and SEM models but not identical (See Exhibit 3 on page 18).
Durability I on Loyalty I is .65. In the corresponding structural The most important variable in the path analysis model is also
equation model, the total effect of Durability2 on Loyalty2 is the most important variable in the structural equation model.
,65 + 1,22 X .07| = .67. These results are pretty close. And keep in The other variables' rank orders are similar but not identical.
mind, both models use the same data set. They differ primarily in Management decisions based on either model would likely be

T E S T I N G F O R E D i A T i O N E F F E C T S
The mediator variable in Exhibits 1 and 2 is Value because the exoge- According to Sanjeev Agarwal and R, Kenneth Teas, from their 1997 Iowa
nous variables can affect brand loyalty both directly and indirectly through State University working paper, "the procedure produces evidence ot media-
the value-loyalty linkage. tion when (a) explanatory variables are statisticaliy significant in the esti-
Examination of the mecfiafion issue is important for two reasons. First, mates of Step 1; (b) the explanatory variables are statistically significant in
from a theoretical perspective, if a construct sucti as Value mediates the the estimates of Step 2; and (c) the mediator variable is statistically signifi-
effect that exogenous variables have on a dependent variable such as Loyalty, cant in the estimates of Step 3."
fhe model should incorporate these functional relationships. Second, in creat- The figure below shows the mediator variable, Valuel, is statistically
ing models with mediator variables, the practitioner must answer the follow- significant, Additionally, as discussed below, there are three mediation out-
mg question: "Is the mediator variable in my model statistically significant?"' comes—no mediation, partial mediation, and full mediation.
Authors Ruben M. Barron and David Kenny developed a procedure for test- No mediation exists if the regression coefficient for an exogenous vari-
ing the statistical significance of mediator variables. {For more explanation, see able is insignificant in Step 1 or if the regression coefficient tor Valuel
their 1986 Journal of Marketing Research article listed in the Additional in Step 3 is insignificanf (which it's not), Partial mediation exists if the
Readings section on pg. 19.) Their approach utilizes ordinary least-squares regression coefficient for an exogenous variable is significant in Steps 1 and
multiple regression, which is available in most statistical software packages. 3 (as is the case with Durabiiityl. Ease of Usel. and Sound I).' Full mediation
The procedure involves the following three steps, using Exhibit 1 as an example: exists if the regression coefficient for an exogenous variable such as Pricel
is significant in Step 1 but not in Step 3. (Although Step 2 is not referenced
• Step 1: Regress the mediator variable, Valuel. on the explanatory in the Figure, conducting Step 2 provides a more comprehensive test for
variables. Durabiiityl, Ease Ot Usel, Soundl, and Pricel, mediation effects. For example, evidence for mediation is increased if the
standardized beta coefficients of the independent variables in Step 2 are
• Step 2; Regress the dependent variable. Loyaltyl, on the same set of significantly greater than comparable values in Step 3, which is not the case
explanatoryvariablesinStep I. in this particular example.)
In summary, there are three possible mediation outcomes for each
• Step 3: Regress Loyaltyl on the explanatory variables of Step 1 and exogenous variable—no mediation, partial mediation, and full mediation.
the mediator variable, Valuel. Exogenous variables that have partial mediation (i.e., a direct and indirect
linkage to a dependent variable) are generally more important predictors of a
Testing for Mediation Effects (Data drawn from Exhibit 1) dependent variable such as Loyalty than similar variables that have full
mediation. Consequently, if moderator effects are not examined appropriately,
management can be misled as to the relative importance of different factors
riandardized regression coefficient''
affecting a construct such as Loyalty.
I" (given it signiticant at .05 or less)

Dependent variable ' In reality, the mental processes by which consumers make decisions are too complex tor any
Step 1: Step 2: Step 3: model to specify all mediator variables—or all exogenous variables, for that matter. No matter how
VariabtB Value 1 Loyalty 1 Loyalty 1 Interpretation good a model appears to be it will always omit constructs, which represent attitudes or perceptions
of consumers that affect consumer behavior. We can only hope that we include In our models those
Durability 1 .26 ,65 .62 . Partial mediation
Ease of Use 1 ,08 .07 .06 Partial mediation mediator variables that theory suggests are important in our attempt to understand in greater
Sound 1 .15 ,14 .12 Partial mediation [fetaii bow consumers make decisions.
Price 1 ,39 ,08 ns Full mediation * A situation where an exogenous variable is significant in Step 3 and insignificant In Step 2 is
Value 1 .12 Significance is required tiighfy unlikely. The converse is not true and would represent a case of full mediation it the exoge-
1
lor any kind of mediation nous variable were significant in Step 1, which is the case with Price 1.

17
comparable as well. Yet, the academic community says that SEM Finally, random error in data sets attenuates the strength
is the better of the two methods for the following reasons: ofthe relationship among functionally related variables.
By controlling for random error, structural equation mod-
• The coefficient estimates are more valid because the els will generally produce higher r's than path models that
estimation procedures take random measurement error don't take this source of error into account, everything else
into account. held constant.

• r- is higher in the structural equation models. Examine the correlations among the independent or exoge-
nous variables between the path model and the structural equa-
• Structural equation models with latent variables use a tion model, In nearly all instances, the correlations are higher in
confirmatory factor analysis technique that provides infor- the structural equation model than the path mode!, indicating a
mation on which observed variables are the best indicators higher level of multicollinearity. For example, the correlation
ofthe unobserved variables, between Ease of Use and Durability in the path model is ,66, The
correlation between the same two constructs in the correspond-
(For additional insights into the relative merits of SEM, see ing structural equation model is ,95, which is nearly perfectly
the interview with Dr, R- Kenneth Teas on page 20). collinear! The effects of multicollinearity ate as follows:
And yet, there may be a problem. When SEM takes random
measurement error into account, multicollinearity can increase, • Model coefficients may be far from the true, unknown
and the parameter estimates become less stable. (This is a matter parameters of which they are estimates. In extreme cases,
of degree, however, depending upon how reliable the summated individual coefficients may take on the wrong sign—that
scale scores are in the path analysis models.) Moreover, the per- is, the coefficients take on negative values when theory or
centage of the variance explained in the dependent variable, as common sense suggests a positive relationship exists
measured by adjusted r', can increase to levels that lack credibility. between the independent and dependent variables,

PROBLEMS WITH MEASURING CONSUMER • The standard errors of the coefficients are likely to be large
ATTITUDES AND PERCEPTIONS and, consequently, a coefficient can be insignificant (i,e,,
Why this occurs relates to a fundamental problem that often not significantly different from zero), even though the true
happens when measuring consumer attitudes and perceptions in value of beta is not zero.
survey research. By their very nature, measures of these attitudes
and perceptions are correlated Consider the following: Further evidence of the effects of multicollinearity can be
shown in the inflated standard errors of the models' coefficients
• To some extent a halo effect operates when consumers are in the structural equation model vs, the path analysis mode! by
evaluating products through rating scales on a quantita- using a bootstrapping technique.'
tive survey instrument. For example, if a respondent is rat- According to lames Arbuckle and Werner Wothke, in their
ing two products and one product is preferred over the book Afflfl5 4.0 User's Guide, bootstrapping "is a versatile method
other, the more preferred product will generally receive for evaluating the empirical sampling distribution of parameter
more favorable ratings than the less preferred product, estimates. In particular, bootstrapping can be used to obtain
regardless of the attribute, empirical standard errorestimates ofthe model parameters,"
Bootstrapping is a computationally intensive simulation
• This halo effect can be exacerbated when respondents do approach. It treats the original sample as a stand-in for the data
not carefully consider how a product performs on a specific population and draws many bootstrap samples from it by sam-
attribute and, perhaps because of interview length, become pling with replacement. Each bootstrap sample is used to obtain
lazy in assigning numerical ratings to product attributes. a new set of parameter estimates. The distribution of each
parameter estimate across bootstrap samples is used to obtain
empirical estimates of its estimation bias, standard error, and
EXHIBIT 3 Total Effect of Exogenous Variables on confidence intervals, i
Loyalty: Consumer Electronic Product Exhibit 4 shows the relatively larger standard errors in the
structural equation model compared to the path model. The stan-
dard errors for the structural equation models range from 1,13 to
Analysis " 1,71 times as large as for the path analysis model,' What does this
Variable Coefficient Rank Coefficient Rank mean to the research practitioner? Consider the following:
Durabiiily ,65 1 ,67 1
• Path analysis |or for that matter, simple or multiple regres-
Ease of Use
1 .07
2
i ° ,17
4
2-3
sion equations] may disguise muiticollinearity s effects,
Sound
Price
1
F
.14
.09 3-4 ,18 2-3
(We, of course, do not include situations where regression is
used in controlled experiments where the independent vari-
Adjusted .71 ables are orthogonal to each other,) This is because regres-

18 hall 2000
EXHIBIT 4 Bootstrapped Estimates; Path Analysis vs CONCLUSIONS
Researchers use regression-based path analysis and SEM to
Structural Equation Models .
understand how consumer attitudes and perceptions affect
behavior Graphically, path and structural equation models look
Bootstrapped
similar Moreover both models produce comparable, although
Estifnates of
not identical results. If the models are not too complex, it's quite
Coefficients'
Hkely that path analysis and structural equation models will
Standard Errors
identify the same "most important" and "least important" vari-
(2) ables, although the rank order of variable importance may not be
(1) Structural completely the same.
Model variable variables Path equation Path analysis, however assumes that survey measures are
analysis models
made without random measurement error As a result, this
Consumef method can partially disguise multicollinearity's effects.
Electronics Loyalty Durability .035 049 1.40 Although SEM explicitly takes random measurement error into
Ease of Use .036 ,056 1,56 account, the results can produce less stability in the model's
Sound .038 .065 1.71 estimated coefficients due to higher coefficient standard errors
.031 ,035 1.13 caused in part by multicollinearity. Consequently, the research
.039 ,045 1.15 practitioner may be well advised to consider alternative
research designs that control for multicollinearity as well as
other factors that can affect the reliability and validity of
sion analysis assumes the independent variables are mea- research findings. •
sured without random error, which in nearly all instances in
applied survey research is an incorrect assumption. Author's Note: The author thanks Dr. Werner Wothke,
principal of Chicago-based SmallWaters Corp. and creator of
• SEM takes into account random measurement error, and Amos, for helpful comments on early drafts of this article.
as a result, more reliably exposes multicollinearity s effects.
This as well as other factors (refer to endnote#I) unfortu-
nately can cause an increase in most, if not all, of the stan- Terry Grapentlne is the principal of Ankeny, Iowa-based
dard error estimates of the model's coefficients. Increased Grapentine Company Inc. He is also a member of Marketing
standard errors mean less stable coefficients. Researci] magazine's editorial review board.

What does this mean for the marketing research analyst? My ENDNOTES
answer is that research practitioners should more often consider 'other factors that can cause intlatlon in these standard error estimates are (11
using alternative research methods that control for multi- more parameters estimated with the same sample siie, resulting in lower effi-
collinearity as opposed to always collecting and analyzing cross- ciency of each estimate and (21 non-linear item to factor relations, resulting in
sectional data and then blindly using path analysis or SEM to larger bootstrapped standard errors
analyze it, which is how most survey research is conducted today. •' These as well other findings are not unique to the data set on which this article is
Examples of alternative methods are conjoint analysis and based Earlier drafts of this article demonstrated similar findings in two additional
research designs incorporating control and test groups where studies from two different product categories. In some cases, the standard errors
stimuli are manipulated in ways to eliminate the effect multi- for the structural equation models were over 3 times as large as for the corre-
colllnearity or other factors may have on research findings. sponding variables in the path models.
Admittedly, experimental designs are not always feasible
in a practical setting. If one, therefore, is limited to using cross- ADDITIONAL READING
sectional data and path analysis or SEM, one needs to insure the Agarwal, Sanjeev and R, Kenneth Teas 11997). "Quality Signals
model is grounded in a solid theoretical framework. (See my arti- and Perceptions of Quality, Sacrifice. Value, and Willingness-
cle entitled "Practical Theory," from the Summer 1998 issue of to-Buy; An Examination of Cross-National Applicability." iowa
Marketing Research: /K Magazine of Mamgemenl and fiipplications, for fur- State University Working Paper#37-I6.
ther explanation.)
Arbuckle, lames L, and Werner Wothke (1999), Amos 4,0 User's
Additionally, if you know a priori that two highly correlated
Guide. Chicago, SmallWaters Corporation,
variables are critically important in affecting brand loyalty, you
might consider an altogether different research design to assess Baron, Ruben M. and David A. Kenny (1986). "The Moderator-
which is more 'important" to customers. For example, you could Mediator Variable Distinction in Social Psychological
use a series of direct questions to determine where a company Research. Conceptual, Strategic, and Statistical
should seek to improve performance to maximize customer Considerations," \ournal of Marketing Research. 19 (May), 229-39.
value. (Although in practice, this too can be a tricky task and
Churchill, Gilbert A,, Ir (19991, Marketing Researcfi. Melhdoiogical
raises the topic of what one means by the term importance, which
Foundations, 7th Edition, New York, Dryden Press,
is d topic best addressed in a future article,)

m:irivi'lim;rt'warch 19
Grapentine, Terry {lW^l, "Dimensions of an Attribute," Marketing , Rick H., editor (1995), Structural Equation Modeling: Concepts.
Research: A Magazine of Management and Applications, 7 (Summer), Issues, and Applications. Thousand Oaks, California. Sage
19-27. Publications.
1 i998), "Practical Theory," Marketing Research: A Magazineof Schumacker. Randall E. and Richard G. Lomax (1996), A
Management and AppUcatiom. 6 (Summer), 5-12. Guide to Structural Equation Modeling. Mahwah, New lersey,
Lawrence Erlbaum Associates.

P A T H A N A L Y S T S A N D L I S R EL
An interview with Iowa State University's
distinguished protessor of business Dr. R. Kenneth Teas.
What are some of the things a marketing research practitioner needs to of exploratory factor analysis. From a psychometric
keep in mind when using structure equation modeling? theory standpoint, using this approach is backward
Teas: One important issue involves the degree to which an a priori model has thinking. The ideal way to proceed is to have well-
been well specified. Maximum likelihood estimation methods, such as those defined, theoretically meaningful concepts in the
used in LISREL and Amos software, that include confirmatory factor analysis first place and then develop items for those concepts. Confirmatory factor
and structural equation modeling, are confirmatory in nature. They require a analysis, then, can be used in which those a priori defined concepts are spec-
well thought-out a priori specified model and well-defined theoretical con- ified in a model and then hypotheses can be tested to assess the degree to
cepts. When the researcher doesn't have a good theoretical model and theo- which the measurement items are valid indicators of the concepts. This
retically meaningful concepts, he or she may be tempted to use structural reverses the process of measurement. We now start with having a good
equation estimation methods to explore the data for a model. This can be understanding of what needs to be measured and then proceed to measure it
expected to result in finding models with, perhaps, good fits, but are non- rather than obtaining data we hope is measuring something and then trying
sense in terms of logical internally consistent explanatory models. to discover what we are measuring with exploratory analysis.

What do you mean hy having well-detined concepts? As this article discusses, path analysis and structural equation modeling
give similar results. What then is the advantage of structural equation
Teas: We have many vague concepts in applied marketing research, and even
modeling?
in academic research that are poorly defined. Now these are the concepts that
are identified inside those ovals in a structural equation mode! (refer to Teas: One advantage is that using maximum likelihood estimation allows us
Exhibit 2). When using LISREL, the researcher specifies multiple indicators of to examine complex models such as non-recursive models. An example of a
those concepts. If the researcher doesn't know the meaning of the concept non-recursive model would be a situation where there is reciprocal causation.
specified, it's impossible to develop valid indicators of the concept. To appro- In other words, the causal arrow goes In both directions between concepts.
priately use structural equation modeling, the researcher must have ade- For example, I have seen a model specified in an applied research project
quately defined concepts. Based upon those definitions, the researcher devel- where two concepts were specified that shared a reciprocal causal relation-
ops multiple indicators that adequately reflect the specified concepts. ship. One construct was called "Technologically Advanced" and another con-
Although LISREL is designed to estimate models under the assumption of struct was called "Product Quality." The arrow for this model indicated recip-
measurement error, the measurement error we are talking about is random rocal causation.
measurement error—not systematic error resulting from a lack of measure-
ment validity. Technologically
advanced
So what are the positive benefits ot programs like LISREL or Amos, given
all the caveats you have mentioned?
These kinds of models cannot be estimated using ordinary least-squares.
Teas: I think one of the most beneficial aspects of programs like Amos—for
regression, or standard path analysis. However, when we have these kinds of
the practitioner—is confirmatory factor analysis. In marketing, researchers
complex models, we often can specify models that are Identified and, there-
need to develop valid measures of the concepts specified in marketing mod-
fore, can be estimated via LISREL- or Amos-type approaches.
els. One way to do this is to create summated scales based upon the results

20 Tall 2000
View publication stats