Professional Documents
Culture Documents
American Marketing Association
American Marketing Association
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
American Marketing Association is collaborating with JSTOR to digitize, preserve and extend access to
Journal of Marketing Research.
http://www.jstor.org
Though they were introduced to marketing only re- often interfere with meaningful covariance structure
cently, structural equation models with unobservable analysis: improper solutions (i.e., solutions outside the
variables are beginning to change the conventions of admissible parameter space) and factor indeterminacy.
marketingresearch methodology (Bagozzi 1980; Fornell Wold's (1963, 1965, 1975, 1980a, b, 1982) method
1982). For social science in general, the new structural of partial least squares (PLS)1 avoids many of the re-
equations approach is closely identified with the maxi- strictive assumptions underlying maximum likelihood
mum likelihood covariance structure analysis general- (ML) techniques and ensures against improper solutions
ized by J6reskog (1970, 1973, 1979) and the associated and factor indeterminacy. As an illustration of situations
computerprogramLISREL (Joreskog and Sorbom 1978, in which LISREL is not well suited and PLS is a feasible
1981). For marketing, in particular, LISREL has been alternative, we estimate three models for a single data
used for parameterestimation in nearly every application set. The first model compares traditional LISREL ML
of structural modeling. As powerful as this method is, estimates with PLS estimates for a case in which LIS-
one may not realistically assume that all problems ame- REL produces improper solutions. Contrary to common
nable to use of structuralequation models are also suited belief, improper estimates do not necessarily stem from
to LISREL. There are other protocols of structural es- sample variance, lack of fit, or ML estimation, but may
timation which impose different assumptions about data, be due to the path-analytic fitting objective behind
theory, and the ties between unobservable variables and LISREL. We show that the removal of factor indeter-
indicators. Marketing data often do not satisfy the re- minacy via PLS provides an effective cure. In a second
quirements of multinormality and interval scaling, or model, wherein all factors are explicitly defined, LIS-
attain the sample size required by maximum likelihood REL avoids impropersolutions by giving estimates iden-
estimation. More fundamentally, two serious problems tical to those of PLS. The third model extends the sec-
ond in a direction consistent with the consumer behavior
theory it embodies. Again, the LISREL results are out-
*Claes Fornell is Associate Professor of Marketing, Graduate side the admissible parameter space.
School of Business Administration, and Fred L. Bookstein is Asso-
ciate Professor of Radiology, Medical School, University of Michi-
gan.
The authors express their gratitude to Herman Wold for his detailed
comments on a previous draft of this article. 'Previously known as NIPALS (nonlinear iterative partial least
squares) or NILES (nonlinear iterative least squares).
440
"unobserved"variance,formativeindicators(Figure1, Estimation
Mode B) would give greaterexplanatorypower. Both The PLS model is estimatedby (1) the loadings(Ay,
formative and reflective indicatorscan also be used Ax) or weights (w,q, ,t) which describehow the obser-
within a single model. For instance, if one intendsto vationsrelateto the unobservables,and(2) the structural
explain variancein the observedcriterionvariablesby relations(p, F), wherebyvaluesof unobservablesinflu-
way of the unobservables,the indicatorsof the endog- ence values of the other unobservablesin the system.
enous constructshould be reflective and those of the Insteadof optimizinga global scalarfunction,PLS es-
exogenous constructshould be formative;the result is timatesby way of a nonlinearoperatorfor which the
a mixed-modeestimation(Figure 1, Mode C). Modes vectorof all estimateditem loadings(Ay, Ax) is a fixed
A and B representtwo separateprinciples-mode A
minimizes the trace of the residual variances in the point.Sinceits introductionby Wold in 1963, the theory
of fixed point (FP) estimationhas been discussed by
"outer" (measurement)equations and mode B mini-
mizes the traceof the residualvariancesin the "inner" Lyttkens(1968, 1973) and Areskoug(1981), and in a
collectionof paperseditedby Wold (1981). Severalde-
(structural)equation,both subjectto certainsystematic velopmentsusing some form of FP can be found in the
constraints(discussedsubsequently).Mode C is a com- recentpsychometricliterature(Carroll,Pruzansky,and
promisebetweenthe two principles. Kruskal 1980; deLeeuw, Young, and Takane 1976;
Indicatormode is also shapedby an aspectof the sub-
stantivetheorybehindthe model:the way in which the Kroonenbergand deLeeuw 1980; Kruskal1980; Young
et al. 1976;PerreaultandYoung 1980;SandsandYoung
unobservableconstruct is conceptualized. Constructs 1980).
such as "personality"or "attitude"are typicallyviewed FP differs from ML models such as LISRELin its
as underlyingfactorsthat give rise to somethingthat is basicprinciplesandassumptions.In ML estimation,the
observed.Their indicatorstend to be realized, then, as
reflective.In contrast,when constructsare conceivedas probabilityof the observeddatagiven the hypothesized
model is maximized.Wold's PLS estimation,which is
explanatorycombinationsof indicators(such as "pop- a least squaresapproach,minimizesresidualvariances
ulationchange"3or "marketingmix") which are deter- underan FP constraint.ML estimatorsassume a para-
mined by a combinationof variables, their indicators metricmodel, a family of joint distributionsfor all ob-
shouldbe formative. servables;PLS operatesas a series of interdependent
Finally,the choice of indicatormode involves an em- OLS regressions,presumingno distributionalform at
piricalelement.In the formativemode, samplesize and all. FP estimation,then, bears no resemblanceto the
indicatormulticollinearity affectthe stabilityof indicator searchfor zeros of certainderivativeswhich character-
coefficients, which in this mode are based on multiple izes the estimationof ML models.
regressions. In the reflective mode, indicatorcoeffi- The distinctionbetween optimizing and fixed-point
cients are based on simple regressionsand thus are not methodsmay be comparedwith the two main models
affectedby multicollinearity.If the indicatorsare highly for solving multipleregressions.If we state as our goal
collinearbut one nonethelessdesiresoptimizationof ex- constructionof a linear form e + byxl. x1 + byx. x2
plained structuralmodel variance, one might estimate which estimatesy with least errorvariance,we face a
modeB butuse loadings,ratherthanregressionweights, problemin directminimization.But we may insteadin-
This approachis illustratedin oursub-
for interpretation. voke the vocabularyof path analysis, referringto the
sequentanalyses. total effectbyxof variablex, on variabley and attempt-
Should the considerationsinvolving study objective, ing to partition this into a direct effect b lx2 and an in-
theoryor conceptualization,andempiricalcontingencies direct effect b, by. xl mediated by xl's effect on x2. The
be contradictory,the selectionof indicatormode may be directandindirecteffects takentogethermustconstitute
difficult.For example, one may wish to minimizeresid- the total;we musthave
ual variancein the structuralportionof the model(which
suggests use of formativeindicators),even though the (6) byxl x2
+
bx2xl byx2xl- byx
constructsare conceptualizedas giving rise to the ob-
servations(which suggests use of reflectiveindicators). andsimilarlyfor x2. The resultis a systemof two equa-
In such cases the analystmight estimatetwice, once in tionswith two unknownswhichare, of course, identical
each mode. If the results correspond,there is no prob- to the normalequationsof the usual approach.In this
lem. If they differ, a compromisemight be workedout secondversion all explicit minimizationis relegatedto
using the factor structuresof the blocks separately(as the bivariate analyses, the coefficients byXl,
byx2,bX1x2
suggestedby Bookstein 1981); otherwise,a decision as being solutions of an easier optimum problem, the
to the overridingconcernmust be made. simple regression.Multipleregressionmay be thought
of, then, as a revision by joint constraintsof simple
regressioncoefficientsindependentlyarrivedat.
A similardistinctionseparatesPLS from LISRELin
3"Populationchange" is presumedto be determinedby natality,
mortality,andmigration.This exampleis fromHauser's(1973) crit- the structuralanalysis of systems of latent variables.
icism of sociologists'overrelianceon reflectiveindication. LISRELposes and solves the global optimizationprob-
lem (maximization of likelihood) explicitly. PLS limits The fixed-point estimation addresses the problem of un-
its explicit optimization computations to the now-famil- known unobservables by substituting the proxy estimates
iar case of ordinary multiple regression. The separate in an iterative manner (as described before for a simple
simple analyses are jointly adjusted by nonlinear alge- model).
braic constraints according to the specific model speci- Because PLS estimation involves no assumptions
fication. For example, consider the specification of Fig- about the population or scale of measurement, there are
ure 1, Mode B. We estimate this model by giving no distributional requirements. Residual variances are
arbitrarystarting values to the weights mr, and xr12, nor- minimized to enhance optimal predictive power. In con-
malizing rTto unit variance and regressing it on x1 and trast, residual covariances in LISREL are minimized for
x2. Let t be the predicted value of rqin this regression. optimal parameter accuracy.
f1 is estimated by a corresponding procedure as the pre-
dicted value of t. That is, rTT1and nr, are given arbitrary ANALYSES
starting values and t is normalized to unit variance and We estimate LISREL and PLS models from a small
regressed on Yl and Y2. rqand i are replaced by i and data set, with collinear indicators, in the context of a
|, each normalized to unit variance, and the procedure study of consumer exit and voice caused by dissatisfac-
is repeated k times until k equals _k-i and (k equals tion. The theoretical models and the data set are delib-
_k-1. At this point, the estimated 'l is regressed on the
estimated |k to obtain the structuralrelationship between erately chosen to illustrate a case to which LISREL (or
other covariance structure models) is not well suited.
the two. The regression of Tk on y1 and Y2gives the es-
The analysis was suggested by Hirschman's (1970)
timated criterion weights (iTr,, 'S 2) and the regression
exit-voice theory of consumer reaction to dissatisfaction.
of k on x, and x2 gives the estimated predictor weights
This theory states that individuals react to discrepancies
(TTre, 7Ty).To estimate a model with reflective indicators between desired and actual states of affairs by the market
(e.g., Figure 1, Mode A), we follow a similar procedure mechanism of "exiting" or by expressing their dissat-
with the exception that the observed variables (i.e., x's
isfaction through "voice." The exiting consumer makes
and y's) are predictands and the unobserved (i.e., Tj and
use of the market by switching brands, terminating
t) are predictors.4
usage, or by shifting patronage-all economic actions.
Assumptions In contrast, voice is a political action-a verbal protest
From equation 1 it follows that E (Irt') = E (') directed at the seller, and, if remedy is not obtained,
=
E(;) = 0, where g = q - E(rl) is a vector of residuals. possibly via third parties. Hirschman posits that when
From equations 2 and 3 we have E(E) = E(8) = E(ye') the exit option is blocked or when cross-elasticities are
= E(xS') = 0. We standardize such that E () = E(g) low, voice will increase. By this reasoning, exit should
= 0 and Var (-i) = Var dominate in highly competitive markets, whereas the
(tj) = Var (xk) = Var (Yr) = more a market resembles a monopoly, the more voice
1, all i, j, k, r, and E(x) = E(y) = 0. The standard- would be expected. Because seller concentration is a
ization of the observed variables is not an essential as-
measure of monopoly power, we hypothesize that con-
sumption. It is made here because we are working with centration is related negatively to exit and positively to
variables of different scales. For predictive purposes,
location parameters can be estimated and the preceding voice.
standardizationdropped. Data
The residual covariance structure is not restricted in
Both census data and nationwide sample data were
PLS. We define E(ee') = O,, E(88') = 0a, and E(t')
= I. In contrast to covariance structure models (such used. Data on consumer exit and voice were obtained
from Best and Andreasen (1977). From telephone inter-
as LISREL) in which the objective is to minimize the
views in 24 cities during Februaryand March 1975, they
residual covariance matrix (by reproducing the observed
generated a total of 2419 responses (a response rate of
covariances), PLS aims to minimize the trace (sum of
the diagonal elements) of It and, with reflective speci- 80.3%). Data on concentration were gathered from the
1972 Census of Manufacturers' four-digit concentration
fication, also 0, and 0,. Because the off-diagonal ele- ratios. These ratios were adjusted with respect to breadth
ments are not among the unknowns of the model and
of product category and geographic scope.5
because the unobservables are explicitly estimated, there
In total, seven variables were used: four measures of
are no identification problems for recursive PLS models.
concentration, two measures of voice (aided and unaided
recall), and one measure of exit. The voice and exit
measures were expressed as the proportion of respond-
4Theinterestedreaderis referredto Fornelland Bookstein(1982)
for a completedescriptionof the algorithmsof the models in the ar-
ticle. BecausePLSestimationis performedby iterationsof regression,
any PLS modelcan be estimatedeasily with suchcomputerpackages 'The adjustmentsappliedto the concentrationratiosare described
as MIDAS,SAS, andTROLL.An excellentspecializedPLS program in detail by Fornell and Robinson (1982). Results from PLS and
availablefromLohmoller(1981) was used to estimatethe models in LISRELmodels on unadjustedconcentrationratios are reportedby
this article. Fornelland Bookstein(1982).
Table 2
MODEL1: PLSAND LISREL
ESTIMATES
WITHREFLECTIVE
INDICATORS
t2 = X2
x1
(10) 'l1 = [71 'Y2 Y3 Y4]
t3 = X3
x2
(14) t = [lrr1 %62 'rr3 lTT4]
= X3
t4
_,x X4
Yi
Y1 k1I
Ay,l El
E X4
Y3 E3
Figure 3
Xy3
MODEL2: LISREL
VERSION
with O6symmetric. The LISREL model is illustrated in
Figure 3 and its estimates are reported in Table 3. For
purposes of interpretationwe make reference to the load-
ings for the x-block, computed as EI
Ax= RXr, E2
Table 3
MODEL2: PLS AND LISRELESTIMATES
.88 .95
08=
.84 .93 .96
The results, accordingto a two-constructmode B esti- those from a two-constructPLS model with formative
mation(see Figure4), also appearin Table 3. Clearly, indicators.The formativespecificationwith ; = 0-that
the PLS and LISREL solutions have identical x-weights unobservablesbe exact linearcombinationsof their in-
(y's in LISREL, 'r's in PLS). Further, the loadings for dicators-is not as restrictiveas it may appear.When
the y-side (hy's in LISREL) are equal to the loadings of the errorsof they-variablesarecorrelated,the errorterm
the PLS results up to a factor of 1/'YPLS-7 5 of the unobservable is distributed instead throughout
Thus, if LISREL is specified in a MIMIC version with the elements of 0, (see Hauser and Goldberger 1971).
0e symmetric, it will always produce results identical to The PLS estimate of the structural parameter YPLS in
Model 2 is larger than either of the two estimates Y1,y2
in Model 1. This is because the objective of a formative-
mode model is to minimize the trace of the residual
7The y-loadings in PLS are
matrix V (the variance-covariance matrix of [), so that
the measurement portion of Model 2 absorbs the largest
Ay = [Xyl y2Xy3] possible part of the total residual, subject to the con-
YPLS
straintof being a fixed point. The formative formulation,
where \,ps is the estimated structural parameter from PLS and Kyl
. . . y3are the estimated loadings from LISREL. Numerically, we
have Figure 4
MODEL2: PLSVERSION
- = [.75 .14 - .31]
Ay = .62 [.47 .08 - .19]
Table4
MODEL3: PLS AND LISRELESTIMATES
.97 .99
08= 08 not estimated by LISREL
.92 .97 .98
analogue to the traditional F-test in regression (see For- (1975) F-test. Both these tests suggest a significant re-
nell and Larcker 1981a, p. 48). Average variance ex- lationship.
tracted (AVE) is the mean-squared loading for each of Redundancy for voice, which has two indicators, can
the three blocks of indicators. As shown in Table 4, the also be assessed by Miller's F-test or by a jackknifing
indicators of concentration have acceptable critical ra- procedure using a Stone-Geisser test for predictive rel-
tios, whereas the critical ratios for the voice indicators evance (Geisser 1974; Stone 1974). Like the jackknifed
are low. This finding may seem surprising in view of standarderrors, the Stone-Geisser estimate is nonpara-
the high loadings for voice and the low loadings for con- metric. An adaptation of Ball's Q2 is the test criterion,
centration, but the reason is that the jackknifed standard Q2 having the form (see Lohmoller 1981)
errors for the elements in Tt, (see equation 4) are high
in relation to their jackknifed estimates and vice versa (20) Q2 = 1 - EIO
for ,T.8
Redundancy, which is the product of the squared Table5
structuralcorrelation and AVE, measures the power of MODEL3: ADDITIONALSUMMARYSTATISTICSFOR PLS
the exogenous construct for predicting the y-variables.
As exit has only one indicator, and therefore no mea- Average variance accounted for (AVA) .17"b
surement residual, its AVE is one. Thus, redundancy for Average variance extracted (AVE)
exit is equal to its squared correlation with concentra- Pv (concentration) .03c
Qc (voice) .84c
tion. The significance can be assessed by examining the
Pvc (exit) 1.00
jackknifed critical ratio of y1 (see Table 4) or by Miller's Redundancy
R2 (exit/concentration) .2ad
R2 (voice/concentration) .04e
8All critical ratios for indicator-construct relationships in the PLS "Significant at .05.
program used here were based on the jackknifing of weights (i.e., w, bSignificant at .05 (Miller's F-test).
and Ie). Alternatively, one could use the loadings (i.e., Ay and AJ) 'For corresponding critical ratios, see Table 4; see also footnote 8.
as the basis for jackknifing. Because weights typically have larger dFor corresponding critical ratio (i.e. y1), see Table 4.
standard errors, they often provide more conservative estimates. ePredictive relevance, Q2 = -.61.
where O is the sum of squares of the y-variables for each served and/or abstract level (depending on indicator
endogenous construct with more than one indicator (i.e., mode). Other major differences between the models in-
in our example, Y2 and y3) and E is the error sum of clude assumptions about factor structure, mechanisms
squares. Following Stone (1974), we estimate Q2 by of statistical inference, matters of identification, and
omitting one case from the analysis and predicting this interpretation of measurement error, as well as fre-
case from the remaining cases. As in jackknifing, the quency of convergence. Unlike ML techniques, PLS
estimation is repeated over all cases to yield an overall makes minimal demands about measurement scales,
index of Q2 as a measure of redundancy without loss of sample size, and the distribution of residuals. Small
degrees of freedom. If Q2 = 1, the observed endogenous sample sizes-sometimes fewer than the number of vari-
variables can be perfectly reconstructed by the model. ables (Wold 1980c)-can be sufficient for descriptive
The model is said to have predictive relevance if Q2 > PLS analyses. Moreover, in contrast to ML, PLS esti-
0. As shown in Table 5, the redundancy for voice is low, mation does not involve a statistical model and thus
Q2 is not positive, and the model is thus not relevant for avoids the need for assumptions about scales of mea-
predicting the observed voice indicators (y2,Y3) surement. Nominal, ordinal, and interval-scaled vari-
In sum, Model 3 indicates a negative relationship be- ables are permissible in PLS in the same ways as in or-
tween market concentration and consumer exit (as a re- dinary regression. l
sponse to dissatisfaction) and a positive relationship be- A primarydifference between PLS and LISREL is the
tween market concentration and consumer voice (as a structureof unobservables. LISREL specifies the resid-
response to dissatisfaction) as was predicted by theory. ual structures, whereas PLS specifies the estimates of
However, as is shown by the residual covariance matrix the unobservables explicitly. This difference bears im-
08, only a very small fraction of the variance of the four portantimplications which have long been debated in the
concentrationratios (x-variables) is included in the struc- psychometric literature(see the review by Steiger 1979).
tural portion of the model. Further, the relationship be- The main defense of the factor model is that it allows
tween concentration and the observed voice variables is for imperfect measurement by assigning surplus variance
too weak to be useful in a predictive sense. The model to the unobservables. However, such measurement error
thus provides limited empirical support for the concen- implies certain disturbing consequences. An infinite
tration/exit-voice relationships. It is not surprising that number of unobservables may bear the same pattern of
the measurement residuals are very high for the concen- correlations with observed variables and yet be only
tration construct because it is well known that its indi- weakly or even negatively correlated with each other
cators-concentration ratios-are only crude proxy (Mulaik and McDonald 1978). For exploratory analyses,
measures of monopoly power. That is not to say that such indeterminacy can be very problematic. In confir-
concentrationratios do not have some theoretical or em- matory structural equations, indeterminacy has been
pirical support.9If they are crude proxies, at least some thought to be less of a problem by reason of the pre-
systematic variance must be due to monopoly power. If sumed existence of "prior knowledge" which rules out
we assume that the exit-voice theory is correct, it is this conflicting explanations. Because the chi square statistic
variance that we retain for the structural portion of the of fit in LISREL is identical for all possible unobserv-
model and it is very small. ables satisfying the same structure of loadings, a priori
The finding that the model cannot be used to predict knowledge is necessary. However, indeterminacy can
observed voice (as opposed to observed exit) is not sur- create difficulties for confirmatory studies. In some
prising. A multitude of factors contribute to consumer cases several hypothesized models have accounted for
complaining behavior (voice).10 The potential for exit, the same data equally well (see Mulaik 1976). Thus,
in contrast, is reduced when the opportunities for brand confirmatory studies are not necessarily free from the
and product switching are restricted. As a result, there problem of having several interpretations. In this article
is less exit when concentration is high. we suggest an equally serious problem: only indeter-
DISCUSSION
Under the classical assumptions of independence and "This is not to imply that all LISREL variables require metric data
normally distributed residuals, ML and OLS estimates or multinormal distributions. Dichotomous variables can be handled
in regression analysis are identical. In structuralequation as fixed x-variables or by use of several-group analysis. If we assume
an underlying bivariate normal distribution, ordinal variables can be
modeling this is not the case. Except as applied to cer- handled by estimating the polychoric correlation instead of the product
tain MIMIC models, PLS and LISREL have different moment correlation. However, because these correlation coefficients
objectives and present systematically different results. are estimated separately, there is no guarantee that the resulting matrix
LISREL attempts to account for observed covariances, of correlations is positive definite (which means that ML estimation
whereas PLS aims to account for variances at the ob- cannot be used). Further, these correlations cannot be used to cor-
rectly estimate standard errors or the chi square goodness-of-fit mea-
sure (Joreskog and Sorbom 1981). When deviations from multinor-
mality cannot be ignored, LISREL V provides an unweighted least
9For a discussion, see Scherer (1980). squares fitting function but not the associated standard errors or the
'lSee, for example, Bearden, Teel, and Crockett (1980). ML goodness-of-fit interpretation of the chi square.
minate factors can have improper loadings. Because im- draws upon the classical true-score theory in psycho-
proper loadings lead to negative variances, such results metrics, but ignores the problem of separating measure-
do not have several interpretations, but ratherhave none ment error from specific variance. PLS, however, ig-
at all. nores classical true-score theory, but separates
For the discipline of marketing, which is often con- measurement residuals (specific and systematic) from
cerned with prediction and control, other drawbacks the structuralequation variance. As correlations between
flow from indeterminacy. As the factor scores are in- indicators of the same construct decrease, the estimated
determinate, specific case predictions are not possible association between unobservables increases in LISREL.
without prior estimation of the scores themselves. Sim- In contrast, the PLS estimate does not go "beyond the
ilarly, testing for outliers and modeling of factor scores data." If the theoretical model is correct and the indi-
cannot be done. PLS avoids factor indeterminacy by ex- cators are valid measurements of the constructs (despite
plicitly defining the unobservables. Factor scores for low correlations) the LISREL estimate would be correct
prediction or further modeling are then readily available. whereas the PLS estimate would be biased downward.
PLS estimators lack the parameter precision of ML If one had reason to doubt the accuracy of the theoretical
estimation in achieving optimal prediction. Given mul- model and/or the validity of the indicators, the LISREL
tivariate normality, LISREL estimates are efficient in estimate would be exaggerated and more credence could
large samples, and support analytic estimates of asymp- be given to the PLS estimate.
totic standard errors. In exchanging much greater a
SUMMARY
priori assertion for statistical inference, LISREL is a
model for theory testing and a general method for cov- For the marketing analyst, the choice between LIS-
ariance structure analysis in which a theoretical model REL and PLS is neither arbitrary nor straightforward.
is specified in terms of covariances and tested against Both apply to the same class of models-structural equa-
empirical data. PLS is not a model in the same sense; tions with unobservable variables and measurement er-
instead of factor analysis or covariance structure analy- ror-but they have different structures and objectives.
sis, it belongs to the same class of models as canonical
correlation, principal components, and regression anal- -LISREL attemptsto accountfor observedcovariances,
whereasPLSaimsat explainingvariances(of variables
ysis. It too can be tested against empirical data but via observedand/or unobserved).
different procedures. The assessment of fit in PLS fol- -LISREL offers statisticalprecisionin the context of
lows a logic similar to regression analysis and depends
stringentassumptions;PLS tradesparameterefficiency
on the estimation mode. For a model with reflective in- for predictionaccuracy,simplicity,andfewer assump-
dicators, Stone-Geisser's Q2 of redundancy would be tions.
appropriate. For a model with formative indicators, -Both models treatmeasurementresiduals,but in dif-
where the objective is to minimize the residuals of the ferentways. PLS separatesout "irrelevant"variance
structuralequation, Miller's F-test or the nonparametric fromthe structuralportionof the model;LISRELcom-
jackknifed standard errors can be used to assess good- bines specific varianceand measurementerrorinto a
ness of fit. The corresponding test in LISREL is the like- single estimateand adjustsfor attenuation.
lihood ratio chi square which evaluates goodness of fit -LISREL requiresrelativelylarge samples'3for accu-
rate estimationand relativelyfew variablesand con-
with respect to the covariance matrix. In contrast to the structsfor convergence;PLS is applicableto small
PLS measures of fit, a good covariance fit does not sup-
samplesin estimationas well as testing(via jackknif-
port any conclusions beyond those related specifically ing andthe Stone-Geissertest) andappearsto converge
to accounting for observed covariance. In fact, it is easy quicklyeven for largemodelswith manyvariablesand
to show that the strength of observed variable relation- constructs(Lohmoller1982).
ships is negatively associated with goodness of fit (For-
nell and Larcker 1981a, b). Therefore, other statistics Only the general research setting can determine the
(e.g., critical ratios) must be used for statistical evalu- appropriatemodeling approach. It is within this context
ation of construct relationships and loadings in LISREL. that the LISREL benefits of statistical parameter effi-
The factor model underlying LISREL allows more ciency and "disattenuated" relationships should be
errors in measurement than the components model in- weighed against the problems of indeterminacy and the
voked by PLS. The measurement residuals in LISREL
group specific variance with measurement error; no dis- x-variables.This second-orderconstructwas estimatedin termsof its
tinction is made between the two. The measurement re-
siduals in PLS are specific if the off-diagonal terms in relationshipto a second-orderconstructfor voice (exit has only one
indicator).The resultingrelationshipwas, not surprisingly,weak.
0 and 08 are statistically indistinguishable from zero; '3Notethateven thoughthe data used in this articleare based on
otherwise they are systematic which means the construct a very largenumberof observations,the unit of analysismakesthe
in question is not unidimensional.12 Thus, LISREL numberof cases small. ThoughLISRELwouldprobablynot be used
underthesecircumstances,the questionof samplesize is not germane
to our analysis.It is the patternof the covariancematrixthatcauses
'2See, for example, 08 of Model 3 (Table 4). The large residual theimpropersolutions.Unlessthispatternwouldchangesubstantially
covariancesindicatethata secondconstructcouldbe formedfromthe in a largersample,the impropersolutionswould remain.
necessity of substantial theoretical knowledge. Our anal- mensional Analysis of Many-Way Arrays with Linear Con-
yses suggest that when relevant correlation submatrices straints on Parameters," Psychometrika, 45, 3-24.
(block by block) are not of the appropriatereduced rank, deLeeuw, J., F. W. Young, and Y. Takane (1976), "Additive
factor indeterminacy is more serious than generally ac- Structurein Qualitative Data: An Alternating Least Squares
Method with Optimal Scaling Features," Psychometrika,
knowledged. The frequent occurrences of improper and 41, 471-503.
uninterpretablesolutions advise against the use of LIS- Driel, Otto van (1978), "On Various Causes of Improper So-
REL unless (1) its objectives are consistent with the ob- lutions in Maximum Likelihood Factor Analysis," Psy-
jectives of the study and (2) its assumptions are verifi- chometrika, 43, 225-43.
ably true; if they are not, PLS is a feasible alternative. Forell, Claes, ed. (1982), A Second Generation of Multivar-
iate Analysis, Vols. I and II. New York: Praeger, in press.
REFERENCES and Fred L. Bookstein (1982), "A Comparative Anal-
Agren, A. (1972), Extensions of the Fixed-Point Method, pub- ysis of Two StructuralEquation Models: LISREL and PLS
lished doctoraldissertation,Departmentof Statistics,Uni- Applied to Market Data," in A Second Generation of Mul-
versityof Uppsala,Sweden, tivariate Analysis: Methods, Vol. I, C. Forell, ed. New
Apel, H. (1977), Simulation sozio-okonomischer Zusamman- York: Praeger, in press.
hange-Kritik and Modification von Systems Analysis, doc- and David F. Larcker (1981a), "Evaluating Structural
toraldissertation,J. W. von GoetheUniversity,Frankfurt- Equation Models with Unobservable Variables and Mea-
am-Main,Germany. surement Error," Journal of Marketing Research, 18 (Feb-
Areskoug,B. (1981), "Some AsymptoticPropertiesof PLS ruary), 39-50.
Estimatorsand a SimulationStudy for ComparisonsBe- and (1981b), "Structural Equation Models
tween LISREL and PLS," in Systems under Indirect Ob- with Unobservable Variables and Measurement Error: Al-
servation: Causality, Structure, Prediction, K. G Joreskog gebra and Statistics," Journal of Marketing Research, 18
and H. Wold, eds. Amsterdam:NorthHolland. (August), 382-8.
, H. Wold, andE. Lyttkens(1975), "Six Modelswith and William T. Robinson (1982), "Industrial Orga-
Two Blocks of Observablesas Indicatorsfor One or Two nization and Consumer Satisfaction/Dissatisfaction," work-
LatentVariables,"ResearchReportNo. 6, Departmentof ing paper, Graduate School of Business Administration,
Statistics,Universityof Goteborg,Sweden. University of Michigan.
Bagozzi, Richard P. (1980), Causal Models in Marketing. Geisser, S. (1974), "A Predictive Approach to the Random
New York:JohnWiley & Sons, Inc. Effect Model," Biometrika, 61, 101-7.
and Claes Fornell (1982), "TheoreticalConcepts, Green, Bert F. (1976), "On the Factor Score Controversy,"
Measurements, and Meaning," in A Second Generation of Psychometrika, 41, 263-6.
Multivariate Analysis, Vol. II: Measurement and Evalua- Hauser, Robert M. (1973), "Disaggregating a Social-Psycho-
tion, C. Forell, ed. New York:Praeger,in press. logical Model of Educational Attainment," in Structural
, and David F. Larcker(1981), "Canonical Equation Models in the Social Sciences, A. S. Goldberger
CorrelationAnalysis as a Special Case of a StructuralRe- and O. D. Duncan, eds. New York: Seminar Press, 255-84.
lations Model," Multivariate Behavioral Research, 16 (Oc- and Arthur S. Goldberger (1971), "The Treatment of
tober), 437-54. Unobservable Variables in Path Analysis," in Sociological
Bearden,W. O., J. E. Teel, andM. Crockett(1980), "A Path Methodology, H. L. Costner, ed. San Francisco: Jossey-
Model of ConsumerComplaintBehavior," in 1980 Edu- Bass, 81-117.
cators' Proceedings. Chicago:AmericanMarketingAsso- Hirschman, Albert O. (1970), Exit, Voice, and Loyal-
ciation, 101-4. ty-Responses to Decline in Firms, Organizations, and
Benter, P. M. (1976), "Multistructure
StatisticalModel Ap- States. Cambridge: Harvard University Press.
plied to Factor Analysis," Multivariate Behavioral Re- Hui, B. S. (1978), "The Partial Least Squares Approach to
search, 11, 3-25. Path Models of Indirectly Observed Variables with Multiple
Bergstrom,R. (1972), "An Investigationof the Reduced Indicators," unpublished doctoral thesis, University of
Fixed-PointMethod," seminarpaper, Departmentof Sta- Pennsylvania.
tistics, Universityof Uppsala,Sweden. Jagpal, HarsharanjeetS. (1981), "Measuring Joint Advertis-
Best, Arthurand Alan R. Andreasen(1977), "ConsumerRe- ing Effects in Multiproduct Firms," Journal of Advertising
sponseto UnsatisfactoryPurchases:A Surveyof Perceiving Research, 21 (1), 65-9.
Defects, Voicing ComplaintsandObtainingRedress,"Law Joreskog, Karl G. (1970), "A General Method for Analysis
& Society, 11, 701-42. of Covariance Structures," Biometrika, 57, 293-51.
Bodin, L. (1974), Recursive Fixed-Point Estimation: Theory (1973), A General Method for Estimating a Linear
and Application,publisheddoctoral dissertation,Depart- StructuralEquation System," in Structural Equation Models
mentof Statistics,Universityof Uppsala,Sweden. in the Social Sciences, A. S. Goldberger and 0. D. Duncan,
Bookstein,Fred L. (1980), "Data Analysis by PartialLeast eds. New York: Seminar Press, 85-112.
Squares," in Evaluation of Econometric Models, J. Kmenta (1979), "Structural Equation Models in the Social
andJ. B. Ramsey,eds. New York:AcademicPress,75-90. Sciences: Specification, Estimation and Testing," in Ad-
(1981), "The GeometricMeaningof Soft Modeling vances in Factor Analysis and Structural Equation Models,
with Some Generalizations," in Systems under Indirect Ob- Karl G. Joreskog and Dag Sorbom, eds. Cambridge, MA:
servation: Causality, Structure, Prediction, K. G. Joreskog ABT Books, 105-27.
and H. Wold, eds. Amsterdam:NorthHolland,2:55-74. (1981), "Basic Issues in the Application of LISREL,"
Carroll,J. Douglas, SandraPruzansky,and Joseph Kruskal Data, 1 (1), 1-6.
(1980), "CANDELINC:A GeneralApproachto Multidi- and Arthur S. Goldberger (1975), "Estimation of a
Model with Multiple Indicators and Multiple Causes of a Least Squares-Part II," Scandinavian Journal of Educa-
Single Latent Variable," Journal of the American Statistical tional Research, 24, 1-24.
Association, 70, 631-9. Perreault, William D. and Forrest W. Young (1980), "Alter-
and Dag Sorbom (1978), LISRELIV: Analysis of Lin- nating Least Squares Optimal Scaling: Analysis of Non-
ear Structural Relationships by the Method of Maximum metric Data in Marketing Research," Journal of Marketing
Likelihood. Chicago: National Educational Resources. Research, 17 (February), 1-13.
and (1981), LISREL V: Analysis of Linear Pijper, de W. M. and W. E. Saris (1981), "The Effect of
Structural Relationships by Maximum Likelihood and Least IdentificationRestriction on the Test Statistic in Latent Vari-
Squares Methods. Chicago: National Education Resources. able Models," in Systems under Indirect Observation:
and Herman Wold, eds. (1981), Systems under Indi- Causality, Structure, Prediction, K. G. Joreskog and H.
rect Observation: Causality, Structure, Prediction. Amster- Wold, eds. Amsterdam: North Holland.
dam: North Holland. Sands, R. and F. W. Young (1980), "Component Models for
Kowalski, B. R., R. W. Gergerlach, and H. Wold (1981), Three-Way Data; An Alternating Least Squares Algorithm
"Chemical Systems under Indirect Observation," in Sys- with Optimal Scaling Features," Psychometrika, 45, 39-87.
tems under Indirect Observation: Causality, Structure, Pre- Scherer, F. M. (1980), Industrial Market Structure and Eco-
diction, K. G. Joreskog and H. Wold, eds. Amsterdam: nomic Performance. Chicago: Rand-McNally.
North Holland. Stapelton, D. C. (1978), "Analyzing Political Participation
Kroonenberg, Pieter M. and Jan deLeeuw (1980), "Principal Data with a MIMIC Model," in Sociological Methodology.
Component Analysis of Three-Mode Data by Means of Al- San Francisco: Jossey-Bass, 52-74.
ternating Least Squares Algorithms," Psychometrika, 45, Steiger, James H. (1979), "Factor Indeterminacy in the
69-97. 1930's and the 1970's-Some Interesting Parallels," Psy-
Kruskal, Joseph (1980), "An Elegant New/Old Approach to chometrika, 44, 157-67.
Estimating Path Models (Structural Equation Models) with Stone, M. (1974), "Cross-Validatory Choice and Assessment
Unobserved Variables," technical report. Murray Hill, NJ: of Statistical Predictions," Journal of the Royal Statistical
Bell Laboratories. Society, B36, 111-33.
Lohmoller, Jan-Bernd (1981), LVPLS 1.6: Latent Variables Wold, Herman (1963), "Toward a Verdict on Macroeconomic
Path Analysis with Partial Least-Squares Estimation, Uni- Simultaneous Equations," in Semaine d'e'tude sur le role
versity of the Federal Armed Forces, Munich, Federal Re- de l'analyse econometrique dans la formulation des plans
public of Germany. de developpement, P. Salviucci, ed., Scripta Varia 28 (Pon-
(1982) "An Overview of Latent Variables Path Anal- tifical Academy of Science, Vatican City). Cited in Wold,
ysis," paper presented at the Annual Meeting of the Amer- H., ed. (1981), The Fixed Point Approach in Interdependent
ican Educational Research Association, New York, March, Systems. Amsterdam: North Holland.
19-23. (1965), "A Fixed-Point Theorem with Econometric
Lyttkens, E. (1968), "On the Fixed-Point Property of Wold's Background, I-II," Arkiv for Matematik, 6, 209-40.
Iterative Estimation Method for Principal Components," in (1974), "Causal Flows with Latent Variables: Partings
Multivariate Analysis, P. R. Krishnaiah, ed. New York: of the Ways in the Light of NIPALS Modeling," European
Academic Press, 335-50. Economic Review, 5, 67-86.
(1973),"The Fixed-Point Method for Estimating In- (1975), "Path Models with Latent Variables: The
terdependent Systems with the Underlying Model Specifi- NIPALS Approach," in Quantitative Sociology: Interna-
cation," Journal of the Royal Statistical Society, A136, tional Perspectives on Mathematical and Statistical Model
353-94. Building, H. M. Blalock et al., eds. New York: Academic
Meissner, W. and M. Uhle-Fassing (1981), "PLS-Modeling Press, 307-57.
and Estimation of Politimetric Models," in Systems under (1980a), "Model Construction and Evaluation When
Indirect Observation: Causality, Structure, Prediction, K. Theoretical Knowledge is Scarce-Theory and Application
G. Joreskog and H. Wold, eds. Amsterdam: North Holland. of Partial Least Squares," in Evaluation of Econometric
Miller, John K. (1975), "The Sampling Distribution and a Models, J. Kmenta and J. G. Ramsey, eds. New York:
Test for the Significance of the Bimultivariate Redundancy Academic Press, 47-74.
Statistic: A Monte Carlo Study," Multivariate Behavioral (1980b), "Soft Modelling: Intermediate Between Tra-
Research (April), 233-44. ditional Model Building and Data Analysis," Mathematical
Mulaik, Stanley A. (1976), "Comments on the Measurement Statistics, 6, 333-46.
of Factorial Indeterminacy," Psychometrika, 41, 249-62. (1980c), "Factors Influencing the Outcome of Eco-
and Roderick P. McDonald (1978), "The Effect of nomic Sanctions: An Application of Soft Modeling," paper
Additional Variables on Factor Indeterminacy in Models presented at the Fourth World Congress of the Econometric
with a Single Common Factor," Psychometrika, 43, 177-92. Society, Aix-en-Provence, France.
Noonan, Richard (1980), "School Environments and School , ed. (1981), The Fixed Point Approach to Interdepen-
Outcomes: An Empirical Comparative Study Using IEA dent Systems. Amsterdam: North Holland.
Data," Working Paper Series No. 26, Institute of Interna- (1982), "Systems under Indirect Observation Using
tional Education, University of Stockholm, Sweden. PLS," in A Second Generation of Multivariate Analysis:
and Herman Wold (1977), "NIPALS Path Modelling Methods, Vol. I, C. Fornell, ed. New York: Praeger, in
with Latent Variables: Analyzing School Survey Data Using press.
Nonlinear Iterative Partial Least Squares," Scandinavian Young, F. W., J. deLeeuw, and Y. Takane (1976), "Regres-
Journal of Educational Research, 21, 33-61. sion with QualitativeVariables:An AlternatingLeast Squares
and (1980), "PLS Path Modelling with Latent Method with Optimal Scaling Features," Psychometrika,
Variables: Analyzing School Survey Data Using Partial 41, 505-29.