Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

A RevisedAlcohol Expectancy Questionnaire: Factor

Structure

Confirmation

and Invariance

in a General

Population Sample
WILLIAM H. GEORGE, PH.D., MICHAEL R. FRONE, PH.D.,* M. LYNNE COOPER, PH.D.,* MARCIA RUSSELL, PH.D.,*
JEREMY B. SKINNER, PH.D.,* AND MICHAEL WINDLE, PH.D.*
Departmentof Psychology,
NI-25, Universityof Washington,Seattle,Washington98195

ABSTRACT. Objective:The alcoholexpectancyconstructhas becomeprominentin contemporary


psychosocial
approaches
to understanding alcohol use and abuse. In 1980 Brown and colleagues
developedthe90-itemAlcoholExpectancy
Questionnaire
(AEQ) to assessexpectancies.
Rohsenowmodifiedthisinstrumentto createthe 40item Alcohol EffectsQuestionnaire(AEQ-2). In the presentstudy,we
replacedthedichotomous
response
formatof theAEQ-2 with a six-point
scaleto createthe AEQ-3. The purposeof the studywasto evaluatethe
AEQ-3 for factor structureconfirmationand invarianceacrossgender
andracesubgroups.
Method.'We administered
theAEQ-3 to a largegeneral populationsample(N = 1,260).The instrumentwasself-adminis-

teredduringa structuredinterview.Results:Confirmatoryfactoranalyseswereconducted
to testa correlatedeight-factormodelconsisting
of
six positiveexpectancies
andtwo negativeexpectancies.
Fit indicesrevealedthattheeight-factormodelfit the datamoderatelywell. Furthermore,thefit of the eight-factormodelwaslargelyinvariantacrossrace
andgendersubgroups.
Nonetheless,
factorintercorrelations
andmodificationindicesrevealedinadequate
discriminantvalidity.Conclusions:
Basedonthesefindings,we cautiouslyrecommendtheAEQ-3 asa measureof alcoholexpectancies.
Specificrecommendations
andlimitations
are discussed
regardingfutureuseof the AEQ-3. (J. Stud.Alcohol56:
177-185, 1995)

drinkingcross-sectionally
(e.g., Christiansen
andGoldman,

LCOHOL
EXPECTANCIES
have
proven
tobe
useful
1983)andlongitudinally(Christiansen
et al., 1989).Among
for explaining
alcoholuseand
abuse.
A landmark
anthropological
studyby MacAndrewandEdgerton(1969) led
to the propositionthat peoplehold durable.beliefsaboutalcohol's effects--since termed alcohol expectancies.These
expectanciesare acquiredpresumablyfrom cultural messagesandpersonaldrinkingexperiences.
Laboratoryexperiments using the balanced placebo design further
substantiatedthe importanceof expectancies(Hull and
Bond, 1986; Marlatt and Rohsenow,1980). Self-reportinstrumentswere developedto assessalcoholexpectancies.
Brown andcolleagues(1980) developedthe Alcohol ExpectancyQuestionnaire(AEQ) to measurealcohol-related
outcomeexpectancies.
In theirinitial study,Brownet al. attemptedto identify the substantivecontentof expectancies
andgenerallyshowedthatindividualdifferencesexistin expectedeffectsof alcoholon moodand behavior.Sincethen
several measuresof alcohol expectancieshave surfaced
(Connorset al., 1988; Fromme et al., 1993; Georgeet al.,
1989;Leigh, 1987;Leigh andStacy,1993;Rohsenow,1983;
Southwicket al., 1981). A steadystreamof studiesrelying
on self-reportexpectancies
hasemergedrevealinga number
of fairly reliablefindings.Expectanciespredictadolescent

adults,expectancies
predictnonproblemandproblemdrinking (Brown, 1985a,b;Brownet al., 1985), alcoholabusestatus (Cooperet al., 1988) and posttreatment
relapse(Brown,
1985c).Expectancies
alsomoderatetherelationshipbetween
stressanddrinking(Cooperet al., 1992;Froneet al., 1993a).
Furthermore,in a prospectivestudyof collegestudents,
expectancies
predictedfuturedrinkingindirectlyvia intended
future drinking(Stacyet al., 1990). In short,expectancies
have becomea prominentconstructin contemporarypsychosocialapproachesto understandingalcohol use and
abuse,a constructthatis likely to haveenduringimportance
in the alcohol field (Abrams and Niaura, 1987; Goldman
et al., 1987; Wilson, 1987).

Althoughthis initial wave of work on alcoholexpectancies deservessomeapplause,it hasbeenthe targetof criticism (Collins et al., 1990;ConnorsandMaisto, 1988;Leigh,
1989a,b;Stacyet al., 1990) and debate(cf., Goldmanet al.,
1991;Leigh andStacy,1991). Amongthe notableproblems
(Leigh, 1989b),the presentstudyaddresses
two concerns:
factorstructureverificationand subgroup(race andgender)
differences.

First,we examinea factorstructurederivedfrom the original AEQ usingconfirmatoryfactoranalyses.Goldmanet al.


(1991) questionedthe superiorityof confirmatoryfactor
analysesover exploratoryfactor analysesfor expectancy
scale development.Indeed, the AEQ factor structurewas
originallyidentifiedwith exploratoryanalyses,and it since

Received:February15, 1994. Revision:July 26, 1994.


*MichaelR. Frone,MarciaRussell,JeremyB. SkinnerandMichaelWindie arewith the ResearchInstituteon Addictions,Buffalo,N.Y.M. Lynne
Cooperis with the Departmentof Psychology,
StateUniversityof New
York, Buffalo, N.Y.

177

178

JOURNAL

OF STUDIES

ON ALCOHOL

hasbeenadvancedasa conceptuallymeaningfulpartitioning
of the alcoholexpectancyconstruct.The AEQ is now well
beyondthe scaledevelopmentstage,and its putativefactor
structurequalifiesas an a priori model warrantingthe more
stringentverificationaffordedby confirmatoryanalyses.
To date,only one publishedstudyhasusedconfirmatory
factoranalysisto examinethefit of the factorstructureproposed to underlie the AEQ. Leigh (1989a) evaluatedthe
AEQ andtwo otheralcoholexpectancyinstruments:
Southwick et al.'s (1981) AlcoholEffectsScaleandLeigh's(1987)
Effects of Drinking Alcohol scale. For each instrument,
Leigh founda poorfit betweenthefactorstructureproposed
by its creatorsand the observeddata. She concludedthat,
contraryto the claimsof the respectivecreators,eachinstrument failed to reflect a multidimensional

model of alcohol

expectancies
characterized
by distinctsubthemes.
However,
this conclusionwas lessdefinitive in the caseof the AEQ becauseLeigh usedthe originaldichotomous
responseformat.
As shenoted in anotherarticle (Leigh, 1989b), the use of
variablesassessed
with a dichotomous
response
formattypically violatesassumptions
underlyingfactor analysisand
may lead to variablesloadingtogetherbecausethey sharea
commondistribution(i.e., difficultyfactors)ratherthanbecausethey assessa conceptuallyhomogeneous
construct.
Therefore,Leigh's (1989a) failure to observeadequatediscriminationamong the six subscalesof the original AEQ
may be primarilya resultof usingdichotomous
indicatorsof
eachlatentvariable.The presentstudyaddresses
thislimitation by usinga six-pointagree/disagree
responseformatto
assess
eachexpectancyitem andthereforeconstitutes
a more
definitive evaluationof the AEQ's factor structure.
Second, we examine whether the AEQ-derived factor
structureis invariant acrossgenderand race. Researchers

/ MARCH

1995

for severalreasons.First, the AEQ-2 is a shorterinstrument


(40 vs 90 items).This givesit greaterappealto researchers
andcliniciansinterested
in assessing
expectancies
in thecon-

text of eitherlengthysurveyinterviewsor time-consuming


laboratoryprocedures.Second,it is a more efficient instrument.In theAEQ, approximately25% of theitemsweresuperfluous;only 68 of 90 itemsloadedon the six factors.In
the AEQ-2, 29 items assessthesesix factors.These 29 items

are the top five loadingitemsfrom eachof the originalsix


factors.
2 Finally, the AEQ assesses
only positiveexpectancies, whereasthe AEQ-2 also assesses
two negativeexpectancy subthemes (9 items). This modification is
importantbecausepositiveand negativeexpectanciesare
thoughtto relatedifferentlyto drinkingbehavior(Stacyet
al., 1990). In short,the increasedefficiencyand breadthof
the AEQ-2 make it more appealingfor future applications
thanits parentinstrument;
yet its derivationpermitsa legitimate evaluationof the AEQ-based factor structure.

In sum,the purposeof thepresentstudywasto revisethe


AEQ-2 and to evaluatethis revisedinstrument(henceforth,
AEQ-3). There were two specificobjectives.First, after re-

visingtheresponse
format,we soughtto confirmtheAEQ-3
factorstructure.
Second,we examinedsubgroup
differences
to evaluate whether this factor structure is invariant across

genderandrace(whitesandblacks).To pursuetheseobjectives,we useda largerandmorebroadlyrepresentative


sample thanhasbeenusedtypicallyin studiesevaluatingalcohol
expectancy
instruments.
Theseanalyseswill providea rigorousevaluationof thepsychometric
propertiesof theAEQ-3,
which is basedon the mostwidely usedinstrument(AEQ)
for assessing
alcoholexpectancies.
It is anticipated
thatthese
analyseswill moreclearlydiscernthe meritsanddrawbacks
of thisparticularexpectancyinstrument.

need to demonstrate that a similar factor structure is observed

acrossgenderandracebeforepotentialgenderandracedifferencesin meanlevelsof expectancies


and in relationships
betweenexpectancies
andtheirputativeantecedents
andoutcomescan be meaningfullyexaminedand interpreted.In
otherwords,it is importantto showthata measureis assessingthesameconstruct
in eachgenderandracialsubgroup
before one caninterpretgenderandracedifferencesin means
andcorrelationsinvolvingthat construct.To date,however,
little empiricalattentionhasbeenpaidto thisissue.Instead,
most studiesinvestigatinggenderand/orrace differencesin
alcoholexpectancies
haveassumed
an invariantfactorstructure acrossraceand gendersubgroups
(e.g., Connorset al.,
1988; Kline, 1990). The presentstudyteststhe validity of
thisassumption.
It alsoimprovesuponearlierworkby using
a broaderpopulation-based
sample,thusensuringmorerepresentativefindingsthan from studiesrelying on student
(e.g., Connorset al., 1988) or alcoholic(e.g., Kline, 1990)
samples.
To examinethe aforementionedconcerns,we employed
the Alcohol Effects Questionnaire(AEQ-2) reported by
Rohsenow(1983). The AEQ-2 was selectedover the AEQ

Method

Sample

Respondents
in thisstudytookpartin thelongitudinalfollow-upof a randomhouseholdsurveyof 1,933adultsresiding in Erie County, New York. The stratifiedthree-stage
probabilitysamplingplan was designedto yield roughly
equal representationof blacks and nonblacksand three
educationlevels(lessthanhighschoolgraduate,highschool
graduate,somecollegeor more). The overall completion
rate for Wave 1 was 78.3%. Of the 1,933 respondents
initially interviewed,1,616 (84%) were reinterviewed3 years
later.

Daa for the presentstudywere obtainedfrom a subsetof


reinterviewedrespondents
(N = 1,260) who drankat least
onceduringthe5 yearsprecedingtheWave 2 int6rview.3Respondents
in this subgroupreporteddrinking.65 drinksper
dayon averagewithinthepastyear.A thirdof theserespondents(32.8%) reportedthat they drank lessthan onceper
monthduringthe pastyear, 26.8% drank 1 to 3 timesper

GEORGE

month,30.7% drank 1 to 4 timesper week and 9.6% drank


daily or almostdaily.
In thefinal sampleof 1,260respondents,
thereweremore
women(61.2%) thanmen, abouthalf were nonwhite(54.4%)
andhalf hadcompletedat leastsomecollege(50%). The average(+_SD) age was 42.9 +_14.9 years.We usedWave 2
databecausetheresponse
scalewaschangedto improvethe
psychometriccharacteristics
of the alcohol expectancies
questions.This changewas not institutedat Wave 1 (see
the descriptionof the expectancymeasurebelow for more

ET AL.

179

disagreemoderately,disagreestrongly).For the presentadministration,severalAEQ-2 itemswere slightlyrewrittento


enhancereadabilityfor a primarily noncollegehousehold
sample.
The expectancyquestionnaire
was self-administered.
Its
placementin the interviewscheduleoccurredafterthe interviewer administeredquestionspertainingto drinkingmotives,usualconsumption
anddrinkingproblems.
Results

details).
Procedure

Data for this studywere collectedby 20 professionally


trainedinterviewersduringthe springand summerof 1989
aspart of a largerstudyof stress,alcoholuseandhypertension.Interviewswere conductedin respondents'homesusing a highly structuredinterview schedulethat required
approximately90 minutesto administer.Respondents
were
paid$25 for participating.
Measuringalcoholexpectancies

Alcohol expectancieswere assessedwith Rohsenow's


(1983) 40-item Alcohol Effects Questionnaire (AEQ-2),

which is a modifiedversionof the Alcohol Expectancy


Questionnaire(AEQ) (Brown et al., 1980). Rohsenow's
(1983) modificationof the AEQ includedshortening
the six
positiveexpectancysubscales
(globalpositive,5 items;social andphysicalpleasure,5 items;socialexpressiveness,
5
items;sexualenhancement,
5 items;powerand aggression,
6 items; and tensionreductionand relaxation,5 items) and
addingtwo negativeexpectancysubscales(cognitiveand
physicalimpairment,5 items;carelessunconcern,4 items).
The itemsarelistedin theAppendix;thecorresponding
subscaleassignments
are indicatedin Table 1. In addition,we
made two modificationsto Rohsenow's AEQ-2. First, we
usedthe original instructions(Brown et al., 1980) for completingtheitemsratherthanRohsenow's (1983) instructions.
These instructionsemphasizerespondingin termsof personalbeliefs aboutalcoholas opposedto Rohsenow'sinstructionsthat ask respondents
to indicatetheir agreement
with items based on the actual effects of a few alcoholic

drinks. The verbatim instructions that were used are: "I want

you to respondaccordingto your own personalthoughts,


feelings and beliefs about alcohol now. I am interestedin
whatyou think aboutalcohol,regardless
of whatotherpeople might think." Second,basedon Leigh's (1989b) argumentthatthe two-pointagree/disagree
response
formatused
by both Brown et al. (1980) and Rohsenow(1983) fails to
provideinformationconcerning
the strengthof endorsement
andposesa varietyof psychometric
problems,we expanded
theresponse
formatto a six-pointagree/disagree
scale(agree
strongly,agreemoderately,agreeslightly,disagreeslightly,

First, the factorstructureof the expectancyitemswasexaminedusingconfirmatoryfactoranalytictechniques


in the
full sample.Second,the resultingmodelwas examinedfor
invarianceacrossmales(n -- 489) andfemales(n = 771) and
acrossblacks(n = 626) andwhites(n = 575). Fifty-ninerespondents
who wereneitherblacknor white wereexcluded
from analysesinvolvingcross-racecomparisons.
Confirmationof thefactor structure

The confirmatoryfactoranalysiswasconductedwith the


EQS structuralequations
modelingprogram(Bentier,1989).
Scalingmetricsfor thelatentvariableswerefixedby setting
factorvariancesto 1.0. Followingprocedures
recommended
by a numberof authors(Cliff, 1983; J6reskogand S6rbom,
1983; Mulaik et al., 1989), threefit indiceswere usedto evaluateoverall model fit: Normed Fit Index (NFI) (Bentlerand
Bonett, 1980), ComparativeFit Index (CFI) (Bentier, 1989)
and standardizedroot-mean-square
residual(RMSR). Becauseeach fit index has differentlimitations,consistency
acrossindicesmaybe regardedasthemostreliableindicator
of goodness-of-fit(Mulaik et al., 1989). Both the NFI and
CFI rangein value from 0 to 1, with valuesof .90 or higher
indicatinga goodfit (BentlerandBonett,1980). The RMSR
is the standardized
averageabsolutedifferencebetweenthe
originalandreproduced
covariancematrices(MarshandHocevar,1985);relativelysmallvalues(e.g., .05, .06) indicate
minimaldiscrepancies
betweenoriginalandreproducedmatrices.Values reportedhere are for off-diagonalelements
only, whichprovidea moreconservativeestimatethanthose
includingdiagonalelements.
For the hypothesizedcorrelatedeight-factormodel, the
valueof the X2goodness-of-fit
testwas5,462.22 (N = 1,260,
40 df, p < .001). In addition,the valuesof the NFI, CFI and
RMSR were .860, .876 and .05, respectively.Collectively,
the goodness-of-fit
measuresindicatea marginallevel of
overallfit for thehypothesized
model.An examinationof the
univariatemodificationindices(i.e., La GrangeMultiplier
tests)revealedthatthemarginallevel of fit wasdueto many
correlatederrorsof measurement
and cross-factorloadings

that wereconstrained
to equalzero,but whichwouldhave
beenstatisticallysignificanthad theybeenfreely estimated.
The significantmodificationindicessuggestthat all of the
expectancyitems were factorially complex. We did not,

180

JOURNAL

OF STUDIES

ON ALCOHOL

/ MARCH

1995

TABLE1. Standardized
factorloadingsfor theeight-factormodels
SUBSCALES

Item no.a

Global
positive

.77

17

.69

22

.75

29

.70
.76

40
13

Socialand
physical
, pleasure

Powerand
aggression

Tension
reduction
andrelaxation

Cognitive
andphysical
impairment

Careless
unconcern

.77

15

.59

21

.81

24

.67
.78

27

Social
Sexual
expressiveness enhancement

.75

20
35

.87
.80

38

.88

39

.89

.84

12

19

.87
.84

28

.90

31

.82

.43

.69

.77

16

.81

32

.71

37

.73

.59

.68

11

25

.69
.81

34

.70

.74

10

.79

18

.78

23

.83

26

.85

14

.8O

30

.77

33

.67

36

.75

Note: N = 1,260.

aForitem content,seeAppendix.
ball factorloadingsweresignificantatp < .001 (z's = 15.6to 40.6).

however,attemptto free any of thecorrelatederrorsof measurement


or cross-factor
loadingsbecausewe didnotwantto
improvethe fit of the modelby capitalizingon chanceand
becausedoingso may have resultedin an underidentified
model(e.g., Bollen, 1989;MacCallumet al., 1992).
Table 1 presentsthe standardized
factorloadingsfor the
correlatedeight-factormodel.All itemsloadedsignificantly
on their hypothesized
factors(z valuesrangedfrom 15.6 to
40.6; p's < .001). Table 2 summarizesthe distributional
properties
andtheinternalconsistency
reliabilityfor eachof
theeightexpectancy
subscales.
An examination
of thesedata
suggests
thatfor eachsubscale
thedistribution
of valueswas
approximately
normal,therewasno restriction
in range(the
actualrangefor all subscales
variedfrom 1 to 6), andtheestimatedreliabilitywashigh.Zero-ordercorrelations
among

the latent expectancysubscales


are reportedin Table 3. As
shownin thistable,the eightsubscales
weresubstantially
intercorrelated(meanr = .78). However,themagnitudeof the
correlationsamongthe six positive(meanr = .81) andbetweenthe two negative(r = .92) expectancysubscales
was
greaterthanthemagnitudeof thecorrelations
betweenthese
two setsof scales(meanr = .61). This patternof correlations
suggeststhat individualscan better discriminatebetween
the two general categoriesof positive and negative expectancies
thanamongsubstantive
typesof positiveor negative expectancies.
It shouldbe noted,however,that the two negativesubscaleshad not been demonstratedpreviouslyto constitute
separatefactors.No factoranalyticresultswerepublishedin
theRohsenow(1983) study.Becausethe negativesubscales

GEORGE

ET AL.

TABLE2. Distributionalpropertiesand reliabilitiesfor the expectancy

Tables4 and5. A X2differencetestcomparingthe X2values


for the two between-group
modelsis providedon line 5 of
Tables4 and5. In bothtables,the X2 differencetestwas statisticallysignificant,suggesting
thattheassumption
of completefactorloadinginvarianceacrossmen and womenand
acrosswhitesandblackswasnot supported
in a strictlysta-

subscales

Subscale
a

Mean (_+SD) Skewhess Kurtosis

Globalpositive
Socialandphysical
pleasure
Socialexpressiveness

1.94 _+1.00

1.19

3.23 _+1.25
2.62 _+1.37

-.28
.38

Sexual enhancement

2.47 _+ 1.35

Powerandaggression

Alpha

1.28

.85

-.73
-.92

.85
.92

.46

-.96

.93

2.32 _+1.10

.68

-.17

.83

2.69 _+ 1.23

.26

-.70

.83

tistical sense.

Examination

Cognitiveand physical
impairment

3.26 +_1.49

Careless unconcern

2.65 _+ 1.36

-.08

- 1.06

.89

-.95

.83

.29

Note: All statistics are based on manifest variables.

aPotentialrangeof valuesfor all subscales


is 1.00 to 6.00.

werenot established
as differentin exploratoryfactoranalyses,theclaim thatthesetwo subscales
are separatefrom each
othermay havebeenoverstated.
This may partiallyaccount
for theirhighintercorrelation.
Invarianceacrossgenderand race

To determinethe extent to which the correlatedeightfactormodelwasinvariantacrossgenderandracegroups,a


seriesof within- and between-group
modelswas specified
andindependently
testedamongmenandwomenandamong
whites and blacks. Goodness-of-fit

of the modification

indices from the con-

strainedbetween-group
analysisbasedon genderrevealed
thatthemagnitudeof onlyfourfactorloadingsoutof 40 differedsignificantly
(p's = .04 to .01) acrossmenandwomen.
The specificexpectancyitems were: Item 24 (social and
physicalpleasurescale), Item 5 (power and aggression
scale),Item 10 (cognitiveand physicalimpairmentscale)
and Item 34 (tensionreductionand relaxationscale). Althoughthe genderdifferenceswere statisticallysignificant,
thewithin-groupanalyses
revealedthateachof thesefourexpectancyitemsloadedsignificantlyon their respectivefactors in both gendergroups(z valuesrangefrom 14.66 to
17.70 and from 20.57 to 27.70 for men and women,respectively;p' s < .001). Moreover,theabsolutemagnitudeof the
genderdifferencefor thesefour unstandardizedfactor
loadingswas small(averageabsolutedifference= .21, range=

Tension reduction and

relaxation

181

16 to .30).
Examination

of the modification

indices from the con-

strainedbetween-group
analysisbasedon racerevealedthat
the magnitudeof only six out of 40 factorloadingsdiffered
significantly(p's = .048 to .002) acrosswhitesandblacks.
The specificexpectancy
itemswere:Item 29 (globalpositive
scale),Item 35 (socialexpressiveness
scale),Item 28 (sexual
enhancement
scale),Item 5 (power and aggressionscale),
Item 11 (tensionreductionand relaxationscale)and Item 30
(carelessunconcernscale). Although the race differences
were statisticallysignificant,the within-groupanalysesrevealedthateachof thesesix expectancy
itemsloadedsignificantly on their respectivefactorsin both racial groups(z
valuesrangefrom 17.56to 28.81 for whitesandfrom 17.84
to 27.34for blacks;p' s < .001).Moreover,theabsolute
magnitudeof theracedifferencefor thesesixunstandardizedfactor loadingswas small(averageabsolutedifference= . 10,
range= .02 to .22).
In light of substantial
similaritiesacrossbothgenderand
race groupsand the generallyinconsequential
effects of

statistics for the within-

groupmodelsareprovidedon lines 1 and2 of Tables4 and


5 for genderandracecomparisons,
respectively.
Analysisof
the fit indicessuggests
that the overallfit of the correlated
eight-factormodelwasvirtuallythe samefor bothmen and
women and for whites and blacks. Values for NFI and CFI

rangedfrom .814 to .872 for all within-groupmodels,andthe


RMSRs rangedfrom .05 to .06.
In additionto the within-groupmodels,two simultaneous,
between-group
modelswere specified.First, a model was
testedin whichthe correlatedeight-factormodelwasspecifiedacrossgenderandracegroups,butthemagnitude
of factor loadingswas allowed to vary. The overall X2 for this
modelis providedon line 3 of Tables4 and5. Second,a factor-loadingequivalentmodel was then testedin which the
factor loadings were constrainedto equivalenceacross
groups.The overallX2for thismodelis providedon line 4 of

TABLE3. Zero-ordercorrelationsamongtheexpectancysubscales
Subscale

1. Globalpositive
2. Socialand physicalpleasure
3. Socialexpressiveness

.78
.88

.79

4. Sexual enhancement

.80

.74

.81

5. Powerandaggression

.86

.67

.87

.74

6. Tension reduction and relaxation

.92

.83

.87

.78

.86

7. Cognitiveand physicalimpairment

.42

.48

.54

.48

.63

.60

8. Careless unconcern

.65

.58

.73

.62

.88

.76

Note: Correlations are for the latent factors.

.92

182

JOURNAL

OF STUDIES

ON ALCOHOL

/ MARCH

1995

TABLE4. Goodness-of-fit
information
for within-andbetween-gender
analyses:
Correlated
eight-factor
modelof alcoholexpectancies
Model
1. Men (n = 489, within-group)
2. Women(n = 771, within-group)
3. Simultaneous
between-group

X2

df

NFI

CFI

RMSR

2,686.92
3,837.86

712
712

<.001
<.001

.814
.848

.856
.872

.05
.05

6,524.78

1/424

.001

.836

.866

6,591.55

1/464

.001

.834

.866

.05 (M)
.05 (F)
.06 (M)
.05 (F)

model: unconstrained

4. Simultaneous
between-group
model: constrained

5. X2difference(constrainedminus
unconstrained)

66.77

40

.005

Note:NFI = NormedFit Index,CFI -- ComparativeFit Index,RMSR = Root-Mean-Squared


Residual,M = maleandF = female.RMSRs
for thebetween-group
modelsarereportedseparately
for menandwomen.

weightingon the relationshipsof compositevariablesto

Althoughtheexpectancy
subthemes
werestatistically
discernible,the level of overallmodelfit andthe high subscale
intercorrelations
suggestthat the degreeof practicaldistinctivenessamongthe scalesmay be quitemodest.Despiteimprovedmodelfit overthatreportedby Leigh (1989a),thefit
of themodelwasstill deficient.Specifically,theNFI andCFI
fell slightlybelow the .90 cutoffcustomarilyrecognizedas
indicativeof a goodfit. An inspectionof modificationindicessuggested
thatthisshortfallwaslargelydueto all items
being factorially complex. Furthermore,intercorrelations
amongthe eight subscales
werehigh (rangingfrom r = .42
to .92, meanr = .78). Thishighdegreeof overlapamongthe
factorsis alsoevidentfrom a casualinspectionof item content.Severalitemshavelittle facevaliditywith theirrespective subscale
themesand/orseemmoreintuitivelyindicative
of anothersubscale.
For example,thebeliefthatalcohol"can
actasan anesthetic... it canstoppain"doesnot intuitively

other variables(Cohen, 1990; Wainer, 1978), we conclude

thatdifferencesof thismagnitudeareunlikelyto be of practical or substantive


significanceandthatdifferentialweighting acrossgroupsis not warranted.
Discussion

We revised and evaluated Rohsenow's (1983) 40-item


AEQ-2 instrumentwhich was derived from Brown et al.'s
(1980) originalAEQ. The revisedeight subscales
exhibited
gooddistributionalpropertiesandhighreliabilities.The con-

firmatoryfactoranalyseswerethemoststringenttestto date
of the contentionthat the AEQ/AEQ-2 reliably measuresa
multidimensional
alcoholexpectancyconstruct.Our evaluationyieldedonlymoderatesupportfor theAEQ factorstructure and demonstrated
that this factor structurewas largely
invariantacrossraceandgendersubgroups.
Confirmatoryanalysesprovidedqualified supportfor a
correlatedeight-factormodel consistingof Brown et al.'s
proposedstructureof six positive expectancyfactors and
Rohsenow'sadditionof two negativeexpectancyfactors.
The fit of this eight-factormodel was an improvementover
the fit reportedby Leigh (1989a) in her testof the original
six-factormodel.This improvementseemsmost likely attributableto the changefrom the originaldichotomousformatto a six-pointitem response
format.4 The presentresults
generallysupportedthe claim that the alcoholexpectancy
constructas measuredby the 40-item AEQ-2 can be parti-

reflect tension reduction. The belief that alcohol makes one

"morecoordinated"scoreson the globalpositivesubscale,


but intuitivelyseemsrelatedto the cognitiveand physical
impairmentsubscale.
Thus,thefactorialcomplexityproblem
emphasizedby Leigh (1989a,b) concerningthe AEQ also
plaguesthe AEQ-3.
This relativelack of distinctiveness
amongthe subscales
may be considered
from conceptualand practicalperspectives. Conceptually,Leigh (1989b) and Goldman et al.

(1991) agreethatthisproblemreflectsthe inherentoverlap


of expectancythemes."For example,it makessensethat a
belief in alcohol as a tension reducer and as a sexual lubri-

tioned into subthemes.

cant mustnecessarilybe relatedto eachother:How couldit

TABLE5. Goodness-of-fit
information
for within-andbetween-race
analyses:
Correlated
eight-factor
modelof alcoholexpectancies
Model
1. White (n = 575, within-group)
2. Blacks(n = 626, within-group)
3. Simultaneous
between-group
model:

X2

df

NFI

CFI

2,975.46

712

< .001

.837

.871

.05

3,256.00

712

<.001

.833

.864

.05

6,231.46

1/424

<.001

.835

.867

6,289.92

1/464

<.001

.834

.867

.05 (W)
.05 (B)
.06 (W)
.06 (B)

unconstrained

4. Simultaneous
between-group
model:
constrained

5. X2 difference(constrained
minus

RMSR

58.46

40

.030

unconstrained)

Note: NFI = NormedFit Index, CFI = ComparativeFit Index, RMSR = Root-Mean-Squared


Residual,B = black,andW = white. RMSRs for
thebetween-group
modelsarereportedseparately
for whitesandblacks.

GEORGE

ET AL.

183

be otherwise?"
(Leigh 1989b,p. 369). To borrowGoldman
et al.'sheuristicexample,theitem"wings"wouldloadonthe
"airplane"and"bird"subscales
of the "flying objects"construct.Accordingto this logic, an expectancyitem may legitimatelyload on multiple subscalesof the expectancy
construct.Thus, somedegreeof both factorialcomplexity
amongitemsandhigh subscale
intercorrelation
arein keeping with a plausibleconceptualaccountof the expectancy
construct.
Nevertheless,
anoptimallyconstructed
instrument
should accommodateany conceptualoverlap inherent
amongexpectancy
subthemes
andstillyieldstatistically
distinctsubscales.
In otherwords,althoughit is appropriate
for

tern of intercorrelations
that suggestthatthereis moredistinctiveness
betweenthe positiveandnegativesubscale
sets
than within either set.s Moreover, after reviewingseveral
theoriesthat arguethat positiveand negativeexpectancies
areconceptually
distinct,Stacyet al. (1990) providedempirical supportfor the importanceof maintainingthisdistinctionin expectancy
modelsof drinking.
Third, despitethe notedlimitationson linking multiple
AEQ-3 subscales
to broaddrinking-related
measures,there
is a continuingneedfor morework establishing
the linkage
betweenspecificexpectancy
subthemes
andnarrowlymeasuredalcohol-relatedoutcomes.Theoretically,an AEQ-3

subscalesto correlate, individual items should be written to

subscale should relate well to a theme-consistent indicator of

avoidfactorialcomplexity.
Froma practicalperspective,
thisproblemwith lackof distinctiveness
amongthe subscales
severelyhampersresearch
effortsaimedat establishing
the discriminantvalidity of the
subscales.
In particular,multivariateanalyseslinking expectanciesto alcohol-relatedoutcomeswill almost invariably suffer multicollinearitydifficultiesthat preventany
definitivestatements
aboutone expectancysubscalebeing
morepredictivethan anotheracrossstudies.Leigh (Leigh,
1989b;LeighandStacy,1991)hasalsoarguedthatprevious
evidenceshowingsuperiorityof onesubscale
overothersin
predictinggeneralalcohol-related
outcomes(suchasanyaggregatedindicatorof alcoholconsumption)
is morelikely the

ity andsamplevariability,thanrealdiscriminative
power.
Together,thepresentconfirmatoryanalysesandtheabove
considerations
suggestsometentativeconclusions
and recommendations
regardingfuturework with theAEQ-3. First,
further improvementsin the existinginstrumentcould be
soughtthroughskillful rewritingof factoriallycomplexor
ambiguousitems.Suchwork couldconceivablyfurtherimprovethefit indicesof thecorrelated
eight-factor
AEQ-based
modelandenhancethediscriminant
validityof thesubscales.
Thisdirectionmustbe weighedagainstthealternativedirection of creatingnew instruments
thataredevelopedfrom the
outsetwith an eyetowardsurmounting
theproblemsexperiencedwith the AEQ (e.g., Frommeet al., 1993;Leigh and
Stacy, 1993). A potentialadvantageto refining the AEQ-3
furtheris retentionof theoriginalsubstantive
subdomains
in
orderto maintainsomecontinuitywith previouswork.
Second,in studieslinking the expectancyconstructwith
broadlyaggregated
drinkingmeasures,
investigators
would
be well advisednotto treattheAEQ-3 subscales
asseparate
variablesin thesameanalyticmodelor evenin separatemod-

drinkingor postdrinkingoutcome.For instance,the sexual


enhancement
subscale
shouldrelatestronglyto drinkingin
sexualsituations
andto postdrinking
sexualresponding.
This
basicvalidationof theconstructs
presumably
assessed
by the
separateAEQ subscales
hasnot beenestablished.
However,
effortsto establishthislevel of construct
validitywith the
AEQ-3 undoubtedly
will be hamperedby the low discriminantvalidityamongthesesubscales.
In general,apartfrom the propertiesof the AEQ-3, careful discriminantvalidity work of this sortwith alcoholexpectancyinstrumentshas been scarce.Such work makesa
strongercasefor the meritsof any instrument
purportingto
measure
subdomains
of alcoholexpectancies.
Usinga newly
developedmeasureof sex-relatedalcoholexpectancies
(Dermen andCooper,in press-a),DermenandCooper(in pressb) foundthatsex-related
expectancies
werebetterpredictors
of drinkingin sexualsituations
thanweregeneralalcoholexpectancies.
Furthermore,
theyfoundthereverseto be truefor
drinkingin socialsituations.
Demonstration
thatthe expectancy
modelwaslargelyinvariantfor raceandgenderis a new finding.Within-group
confirmatoryanalysessupportedthe fit of the correlated
eight-factor
modelacross
raceandgendergroups.Betweengroupanalyses
partiallysupported
themorestringentcriteria
of factorloadingequivalence.Theseanalysessupportthe
conclusion
thattheAEQ-3 assesses
the alcoholexpectancy
constructin the sameway acrossgenderandrace.However,
thisfindingimpliesthatlack of discriminant
validityis also
a problemin eachrace and gendersubgroup.(Unreported
supplemental
analysesshowedthisto be true.)Nonetheless,
thesefindingslegitimizesubstantive
interpretation
of mean
differences
in expectancies
betweenmenandwomenandbetweenblacksandwhites,andof differences
in themagnitude
of correlations
involvingalcoholexpectancies
acrossgender

els. Instead,it would be bettereither to (1) combinethe sub-

and race.

result of chancevariations in the data, due to multicollinear-

scale scoresinto a single multivariateanalytic unit by


evaluatingan overall summarystatistic(an examplewould
be to includeall subscalescoreswithin a singlehierarchical
regression
stepandto evaluatetheAR2butnotindividualbetas),or (2) computea positiveexpectancyscorebasedon the
six positivesubscales
and/ora negativescorebasedon the
twonegativesubscales.
The latterpointis supported
by a pat-

Summaryand conclusions

The findingsleadusto cautiouslyrecommend


theAEQ-3
(revisedAEQ-2) as a measureof alcoholexpectancies.
The
strengths
of AEQ-2 havebeenretained:It is briefandeasyto
administer
andscore;it providesa multifactorialassessment

184

JOURNAL

OF STUDIES

ON ALCOHOL

acrossa rangeof expectancysubthemes;


it subsumes
the
originalsix positiveAEQ factors,thusretainingsomecomparabilitywith the numerousAEQ studies;and it includes
negativeaswell aspositivefactors.Additionaladvantages
of
theAEQ-3 arethatthedistributional
properties
andreliabilities of the subscales
are strong;the correlatedeight-factor
structurehasbeenconfirmed(thoughthe fit is still lessthan
adequate);
the factorstructure
hasbeenshownto be largely
invariantacrossraceandgender,andthe AEQ-3 is generalJzableacrossa broaderpopulationthanpreviouslydemonstrated.Concernsabout the discriminantvalidity of the
subscales
remainunresolvedand requirefurtherresearch
(e.g., Froneet al., 1993b).

Appendix
1.
2.
3.
4.

Drinkingmakesme feel warmandflushed.


Alcohol lowersmuscletensionin my body.
A few drinksmakeme feel lessshy.
Alcoholhelpsme to fall asleepmoreeasily.

5. I feelpowerfulwhenI drink,asif I canreallymakeotherpeople do asI want.


6. I'm moreclumsyaftera few drinks.
7. I am more romantic when I drink.

8. Drinkingmakesthe futureseembrighterto me.


9. If I havehada coupleof drinks,it is easierfor meto tell someone off.

10. I can't act asquicklywhenI've beendrinking.


11. Alcoholcanactasan anesthetic
for me,thatis, it canstoppain.
12. I often feel sexier after I've had a few drinks.

13. Drinkingmakesme feel good.


14. Alcohol makesme carelessaboutmy actions.

15. Somealcoholhasa pleasant,cleansing,


tinglytasteto me.
16. Drinkingmakesme moreaggressive.
17. Alcohol seemslike magicto me.
18. Alcohol makes it hard for me to concentrate.
19. I'm a better lover after a few drinks.

20. WhenI'm drinking,it iseasierto openupandexpressmy feelings.

21. Drinkingaddsa certainwarmthandfriendliness


to socialoccasions for me.

22. If I'm feelingtied downor frustrated,a few drinksmake me


feel better.

23. I can't think asquicklyafterI drink.

24. Havinga few drinksis a niceway for meto celebrate


special
occasions.

25. Alcohol makesme worry less.


26. Drinkingmakesme lessefficient.

in the mood for sex.

I feel morephysicallycoordinated
afterI drink.
I'm morelikely to sayembarrassing
thingsafterdrinking.
I enjoyhavingsexmoreif I've hadsomealcohol.
I'm morelikely to getintoan argumentif I've hadsomealcohol.

33. Alcohol makesme lessworriedaboutdoingthingswell.


34. Alcohol helpsme sleepbetter.

35. Drinkinggivesmemoreconfidence
in myself.
36. Alcohol makesme more irresponsible.
37. After a few drinksit is easierfor me to pick a fight.
38. A few drinksmakeit easierfor me to talk to people.

1995

39. If I havea coupleof drinks,it is easierto expressmy feelings.


40. Alcohol makesme moreinteresting.
Notes

1. Both Brown et al. (1980) andRohsenow(1983) usedtheacronymAEQ


to refer to their respectiveinstruments.We will use AEQ to refer to
Brown et al.'s Alcohol ExpectancyQuestionnaire,
AEQ-2 to refer to
Rohsenow'sAlcohol Effects Questionnaireand AEQ-3 to refer to the
currentlyrevisedAEQ-2.
2. One item loadedon two separatefactors.The item "If I'm feeling restrictedin any way, a few drinks make me feel better"(Item 22 in
AEQ-3) loadedon boththe Global Positiveand Powerand Aggression
subscales.
For thepresentconfirmatoryfactoranalyses,
we allowedthis
item to load only on the Global Positivesubscale.
3. BetweenWave 1 and Wave 2, 318 subjectswere lost to attrition.These
attritionsubjectswere notdifferentfrom reinterviewedsubjectswith respectto genderdistribution,
racedistribution,
Wave I averagedrinksper
day in the pastyear andWave I expectancyscalescores.However,on
average,the attritionsubjectswere older (F = 31.51, 1/1,575 df, p <
.001) and lesseducated(F -- 13.75, 1/1,574 df, p < .001) thanreinterviewedsubjects.Althoughstatisticallysignificant,the effectsizeswere
smallfor botheducation(12.98 vs 12.42years;12= .009) andage(45.80
vs40.16 years;'13
2 = .020).
4. There are two otherdifferencesbetweenthe presentstudyand Leigh's
studythatmay haveaffectedthedifferencesin fit. The AEQ-3 hadeight
factors rather than six and there were fewer indicators for each latent vari-

able in the AEQ-3 than in the AEQ. To examinethe influenceof the number of factors, the overall fit indices were computedfor a model

containingonlythe sixpositiveexpectancy
factors.The fit of thismodel
(NFI = .869, CFI = .881, RMSR -- .04) wasvirtuallyidenticalto thefit
of theeight-factormodel(NFI = .860,CFI = .876,RMSR = .05). Thus,
numberof factorsdoesnot explainthis study's improvementin fit relativeto Leigh's study.Nonetheless,
thedifferencein thenumberof items
is still a plausiblealternativeexplanation.
5. To evaluatethis impressionof therebeingsuperordinate
positiveand
negativefactors,we attemptedto conducta higherorderfactoranalysis.
However,the modelwasunderidentifiedbecausethe second-order
negativeexpectancy
factorhadonlytwo first-orderindicators.
At leastthree
indicators
areneededto identifya factor(first-or second-order).
As an
indicationof underidentification,we obtaineda negativeerrorvariance
(Heywoodcase)for oneof the first-orderindicatorsof the second-order
negativeexpectancyfactor.In addition,the standarderrorfor the negative errorvariancerevealedthatit wassignificantlydifferentfrom zero,
therebyrulingout thepossibilitythatit wasdueto samplingfluctuation.
In sum,a higherorderfactoranalysiswouldhaveinformedour subjective impression
of positiveandnegativesuperordinate
factors;but it is
impossible
to estimatesucha modelwithouta thirdnegativeexpectancy
scale.

27. Drinkingis pleasurable


because
it's enjoyable
for metojoin in
with peoplewhoareenjoyingthemselves.
28. After a few drinks,I ammoresexuallyresponsive,
thatis, more
29.
30.
31.
32.

/ MARCH

References
ABRAMS,D.B. ANDN1AURA,R.S. Sociallearningtheory.In: BLANE,H.T.
ANDLEONARD,
K.E. (Eds.)Psychological
Theoriesof DrinkingandAlcoholism,New York: Guilford Press,1987, pp. 131-178.
BENTLER,P.M. EQS StructuralEquationsProgramManual, Los Angeles:
BMDP Statistical Software, 1989.

BENTLER,
P.M. ANDBONETT,D.G. Significancetestsandgoodness
of fit
in the analysisof covariancestructures.
Psychol.Bull. 88: 588-606,
1980.

BOLLEN,K.A. StructuralEquationswith LatentVariables,New York: John


Wiley & Sous,Inc., 1989.
BROWN,S.A. Contextof drinkingandreinforcement
from alcohol:Alcoholicpatterns.Addict.Behav.10:191-195, 1985a.

GEORGE

BROWN,S.A. Expectancies
versusbackground
in the predictionof college
drinkingpatterns.
J. Cons.Clin. Psychol.53: 123-130,1985b.
BROWN,S.A. Reinforcement
expectancies
and alcoholismtreatmentoutcomeaftera one-yearfollow-up.J. Stud.Alcohol,46: 304-308, 1985c.
BROWN,S.A., GOLDMAN,
M.S., INN, A. ANDANDERSON,
L.R. Expectancies of reinforcement from alcohol: Their domain and relation to drink-

ing patterns.
J. Cons.Clin. Psychol.48: 419-426, 1980.
BROWN,S.A., GOLDMAN,M.S. AND CHRISTIANSEN,
B.A. Do alcohol ex-

pectancies
mediatedrinkingpatternsof adults.J. Cons.Clin. Psychol.
53: 512-519, 1985.

CHRISTIANSEN,
B.A. AND GOLDMAN,M.S. Alcohol-relatedexpectancies
versusdemographic/background
variablesin the predictionof adolescentdrinking.J. Cons.Clin. Psychol.51: 249-257, 1983.
CHRISTIANSEN,B.A., SMITH, G.T., ROEHLING,P.M. AND GOLDMAN,M.S.

Usingalcoholexpectancies
to predictadolescent
drinkingbehaviorafter oneyear.J. Cons.Clin. Psychol.57: 93-99, 1989.
CLIFF,N. Somecautionsconcerningthe applicationof causalmodeling
methods.MultiMar. Behav. Res. 18:115-126, 1983.

COHEN,J. ThingsI learned(so far). Amer. Psychologist


45: 1304-1312,
1990.

185

(Eds.)Psychological
Theoriesof DrinkingandAlcoholism,New York:
GuilfordPress,1987,pp. 181-226.
GOLDMAN,M.S., BROWN,S.A, CHRISTIANSEN,
B.A. AND SMITH, G.T. Al-

coholismandmemory:Broadening
thescopes
of alcohol-expectancy
research.Psychol.Bull. 110: 137-146,1991.
HULL,J.G. ANDBOND,C.F., JR. Socialandbehavioralconsequences
of alcoholconsumption
andexpectancy:
A recta-analysis.
Psychol.Bull. 99:
347-360, 1986.

JORESKOG,
K.G. ANDSORBOM,
D. LISREL 6: Analysisof LinearStructural
Relationships
by the Method of Maximum Likelihood,Mooresville,
Ind.: Scientific Software, Inc., 1983.

KLINE, R.B. The relation of alcoholexpectanciesto drinking patterns


amongalcoholics:
Generalization
acrossgenderandrace.J. Stud.Alcohol 51: 175-182, 1990.
LEIGH, B.C. Beliefs about the effects of alcohol on self and others. J. Stud.
Alcohol 48: 467-475.

1987.

LEIGH,B.C. Confirmatoryfactoranalysisof alcoholexpectancyscales.J.


Stud. Alcohol 50: 268-277, 1989a.
LEIGH, B.C. In search of the seven dwarves: Issues in measurement and

meaningin alcoholexpectancyresearch.Psychol.Bull. 105: 361-373,

COLLINS, R.L., LAPP, W.M., EMMONS,K.M. AND ISAAC,L.M. Endorse-

mentandstrength
of alcoholexpectancies.
J. Stud.Alcohol51: 336-342,
1990.

CONNORS,G.J. AND MAISTO, S.A. The alcohol expectancyconstruct:


Overviewandclinicalapplications.
Cog.Ther.Res.12: 487-504, 1988.
CONNORS,G.J., MAISTO, S.A. AND WATSON,D.W. Racial factors influenc-

ing collegestudents'ratingsof alcohol'susefulness.


Drug AlcoholDepend.21: 247-252, 1988.
COOPER,
M.L., RUSSELL,
M. ANDGEORGE,
W.H. Coping,expectancies,
and alcoholabuse:A testof sociallearningformulations.
J. Abhorre.
Psychol.97: 218-230, 1988.
COOPER,M.L., RUSSELL,M., SKINNER,J.B., FRONE,M.R. ANDMUDAR, P.

Stressandalcoholuse:Moderatingeffectsof gender,coping,andalcohol expectancies.


J. Abhorre.Psychol.101:139-152, 1992.
DERMEN,
K.H. ANDCOOPER,
M.L. Sex-related
alcoholexpectancies
among
adolescents:
Scaledevelopment.
Psychol.Addict.Behav.,in press-a.
DERMEN,
K.H. ANDCOOPER,
M.L. Sex-related
alcoholexpectancies
among
adolescents:
Prediction
of drinkingin socialandsexualsituations.
Psychol.Addict. Behav.,in press-b.
FROMME,
K., STROOT,
E.A. ANDKAPLAN,D. Comprehensive
effectsof alcohol:Development
andpsychometric
assessment
of a newexpectancy
questionnaire.
Psychol.Assess.5: 19-26, 1993.
FRONE,M.R., RUSSELL,M. AND COOPER,M.L. Relationshipof workfamilyconflict,gender,andalcoholexpectancies
to alcoholuse/abuse.
J. Organ.Behav.14: 545-558, 1993a.
FRONE, M.R., RUSSELL,M. AND COOPER,M.L.

ET AL.

Individual differences in

differentiationamong alcohol expectancydomains.Addictions88:


1119-1126, 1993b.

GEORGE,
W.H., DERMEN,
K.H. ANDNOCHAJSKI,
T.H. Expectancy
set,selfreportedexpectancies
and predispositional
traits:Predictinginterestin
violence and erotica. J. Stud. Alcohol 50: 541-551, 1989.

GOLDMAN,
M.S., BROWN,S.A. ANDCHRISTIANSEN,
B.A. Expectancy
theory: Thinkingaboutdrinking.In: BLANE,H.T. ANDLEONARD,
K.E.

1989b.

LEIGH,B.C. ANDSTACY,A.W. On the scopeof alcoholexpectancyresearch:Remainingissuesof measurement


andmeaning.Psychol.Bull.
110: 147-154, 1991.

LEIGH,B.C. ANDSTACY,A.W. Alcoholoutcomeexpectancies:


Scaleconstructionandpredictiveutility in higherconfirmatorymodels.Psychol.
Assess.5: 216-229, 1993.

MACANDREW,
C. ANDEDGERTON,
R.B. DrunkenComportTent:A Social
Explanation,Chicago:AldinePublishingCo., Inc., 1969.
MACCALLUM, R.C., ROZNOWSKI,M. AND NECOWITZ, L.B. Model modifi-

cationsin covariance
structure
analysis:The problemof capitalization
onchance.Psychol.Bull. 111:490-504, 1992.
MARLATT,G.A. ANDROHSENOW,
D.J. Cognitiveprocesses
in alcoholuse:
Expectancy
andthebalancedplacebodesign.In: MELLO,N.K. (Ed.) Advancesin Substance
Abuse:Behavioraland BiologicalResearch,Vol.
1, Greenwich,Conn.:JAI Press,Inc., 1980,pp. 159-199.
MARSH,H.W. ANDHOCEVAR,
D. Applicationof confirmatory
factoranalysisto thestudyof self-concept:
First-andhigherorderfactormodelsand
theirinvariance
acrossgroups.Psychol.Bull. 97: 562-582,1985.
MULAIK, S.A., JAMES,L.R., VAN ALSTINE,J., BENNETT,N., LIND, S. AND
STILWELL,C.D. Evaluationof goodness-of-fit
indicesfor structural
equationmodels.Psychol.Bull. 105: 430-445, 1989.
ROHSENOW,
D.}. Drinkinghabitsandexpectancies
aboutalcohol'seffects
for selfversusothers.J. Cons.Clin. Psychol.51: 752-756, 1983.
SOUTHWICK,L., STEELE,C., MARLATT, A. AND LINDELL, M. Alcohol-

relatedexpectancies:
Definedby phaseof intoxication
anddrinkingexperience.J. Cons.Clin. Psychol.49: 713-721, 1981.

STACY,
A.W., WIDAMAN,
K.F. ANDMARLATT,
G.A. Expectancy
modelsof
alcoholuse.J. Pets.SocialPsychol.58: 918-928, 1990.
WAINER,H. On thesensitivity
of regression
andregressors.
Psychol.Bull.
85: 267-273, 1978.

WILSON,G.T. Cognitivestudiesin alcoholism.


J. Cons.Clin. Psychol.55:
325-331, 1987.

You might also like