1/20/2017 ExploratoryfactoranalysisWikiversity
Exploratoryfactoranalysis
Thispagesummariseskeypointsabouttheuseofexploratory
factoranalysisparticularlyforthepurposesofpsychometric Contents[hide]
[Link] 1 Assumedknowledge
involved,seeEFAtutorial. 2 Purposesoffactoranalysis
3 History
4 Assumptions
Assumedknowledge [ edit ] 5 Types(methodsofextraction)
6 Rotation
Correlation 7 Determiningthenumberoffactors
8 Criteriaforselectingitems
9 Nameanddescribethefactors
Purposesoffactoranalysis [ edit ]
10 Dataanalysisexercises
11 Pros&cons
Therearetwomainpurposesorapplicationsoffactoranalysis:
12 Glossary
[Link] 13 References
14 Seealso
[Link] 15 Externallinks
example,psychologicalquestionnairesoftenaimtomeasureseveral
psychologicalconstructs,[Link]
[Link]
severalrelateditemsisgenerallyconsideredtobemorereliableandvalidthanrelyingonresponsestoa
singleitem.
[Link]
Theoreticalquestionsabouttheunderlyingstructureofpsychologicalphenomenacanbeexploredand
[Link],isintelligencebetterunderstoodasasingle,general
factor,orasconsistingofmultiple,independentdimensions?Or,howmanypersonalityfactorsarethereand
whatarethey?
History [ edit ]
[Link],seefactor
analysishistory.
Assumptions [ edit ]
Thereareseveralrequirementsforadatasettobesuitableforfactoranalysis:
[Link]:Statisticalinferenceisimprovedifthevariablesaremultivariatenormal[1]
[Link]
bivariatescatterplots:
[Link]?
[Link]?
[Link](even(orcigarshaped)asopposedto
fanninginorout))?
[Link]
[Link],thereshouldbesomedegreeofcollinearityamongthe
[Link]
viaanyofthefollowing:
[Link] 1/6
1/20/2017 ExploratoryfactoranalysisWikiversity
[Link](correlationmatrix)arethereatleastseveralsizablecorrelationse.g.,>
.5?
[Link]>~.5.
[Link](MSAs):
KaiserMeyerOlkin(KMO)(shouldbe>~.5)and
Bartlett'stestofsphericity(shouldbesignificant)
[Link]:Thesamplesizeshouldbelargeenoughtoyieldreliableestimatesofcorrelationsamong
thevariables:
[Link],thereshouldbealargeratioofN/k(Cases/Items)e.g.,>~20:1
1.e.g.,ifthereare20itemsinthesurvey,ideallytherewouldbeatleast400cases)
[Link]>~5:1
[Link],aslowas3:1.
Formoreinformation,seetheselecturenotes.
Types(methodsofextraction) [ edit ]
Theresearcherwillneedtochoosebetweentwomaintypesofextraction:
[Link](PC):[Link]
whenthegoalisdatareduction(i.e.,toreduceasetofvariablesdowntoasmallernumberoffactors
andtocreatecompositescoresforthesefactorsforuseinsubsequentanalysis).
[Link](PAF):[Link]
preferredwhenthegoalistheoreticalexplorationsoftheunderlyingfactorstructure.
Rotation [ edit ]
Theresearcherwillneedtochoosebetweentwomaintypesoffactormatrixrotation:
[Link](VarimaxinSPSS):Factorsareindependent(i.e.,correlationsbetweenfactorsareless
than~.3)
[Link](OblimininSPSS):Factorsarerelated(i.e.,atleastsomecorrelationsbetweenfactorsare
greaterthan~.3).Theextentofcorrelationbetweenfactorscanbecontrolledusingdelta[2] .
Negativevalues"decrease"factorcorrelations(towardsfullorthogonality)
"0"isthedefault
Positivevalues(don'tgoover.8)"permit"higherfactorcorrelations.
Iftheresearcherhypothesisesuncorrelatedfactors,[Link]
hypothesisescorrelatedfactors,thenuseobliquerotation.
Inpractice,researcherswillusuallytrydifferenttypesofrotation,thendecideonthebestformofrotation
basedontherotationwhichproducesthe"cleanest"model(i.e.,withlowestcrossloadings).
Determiningthenumberoffactors [ edit ]
Thereisnodefinitive,[Link]
[Link],including:
[Link]:e.g.,Howmanyfactorswereexpected?Dotheextractedfactorsmaketheoreticalsense?
[Link]:
[Link] 2/6
1/20/2017 ExploratoryfactoranalysisWikiversity
[Link]'scriterion:Howmanyfactorshaveeigenvaluesover1?Note,however,thatthiscutoff
isarbitrary,soisonlyageneralguideandotherconsiderationsarealsoimportant.
[Link]:[Link]'elbow'minus1(i.e.,wherethereisanotabledrop)
therestis'scree'.Extractthenumberoffactorsthatmakeupthe'cliff'(i.e.,whichexplainmost
ofthevariance).
[Link]:Ideally,trytoexplainapproximately50to75%ofthevarianceusing
theleastnumberoffactors
[Link]:Areallfactorsinterpretable?(especiallythelastone?)Inotherwords,canyou
reasonablynameanddescribeeachsetofitemsasbeingindicativeofanunderlyingfactor?
[Link]:Tryseveraldifferentmodelswithdifferentnumbersoffactorsbeforedecidingona
[Link],examine,say,
2,3,4,5,6and7factormodelsbeforedeciding.
[Link]'tbelong:Havingdecidedonthenumberoffactors,itemswhichdon'tseemto
belongshouldberemovedbecausethiscanpotentiallychangeandclarifythestructure/numberof
[Link]'tseemto
belong,[Link]
maybethatadifferentnumberoffactors(probablyoneortwofewer)[Link]
moreinformation,seecriteriaforselectingitems.
[Link]:Themoreitemsperfactor,thegreaterthereliabilityofthefactor,butthe
[Link],afactorcould,intheory,beindicatedbyaslittle
asasingleitem.
[Link]?Iftheyaretoohigh(e.g.,over
~.7),thensomeofthefactorsmaybetoosimilar(andthereforeredundant).Considermergingthetwo
relatedfactors(i.e.,runanEFAwithonelessfactor).
[Link],isthefactorstructureconsistentfor
malesandfemales?(e.g.,inSPSSthiscanbedoneviaDataSplitfileCompareGroupsorOrganise
OutputbyGroupsSelectacategoricalvariabletosplittheanalysesby(e.g.,Gender)Paste/Runor
OKThenreruntheEFAsyntax)
Criteriaforselectingitems [ edit ]
Ingeneral,aimforasimplefactorstructure(unlessthereisaparticularreasonwhyacomplexstructurewould
bepreferable).Inasimplefactorstructureeachitemhasarelativelystrongloadingononefactor(target
loadinge.g.,>|.5|)andrelativelysmallloadingsonotherfactors(crossloadingse.g.,<|.3|).
[Link]
[Link]
itemswhichcontributetoameaningfulmeasureofanunderlyingfactorandtoremoveitemsthatweaken
measurementoftheunderlyingfactor(s).Inmakingthesedecisions,consider:
[Link],above.5
foreachitem.
[Link](target)factorloadingindicateshowstronglyeachitemloadsoneachfactorshouldgenerally
beabove|.5|foreachitempreferablyabove|.6|.
[Link](nontarget)[Link]
beagapofatleast~.[Link]
loadingsabove.3areworrisome.
[Link] 3/6
1/20/2017 ExploratoryfactoranalysisWikiversity
[Link]
extenttowhicheachitemappearstomakeameaningfulanduseful(nonredundant)contributiontothe
underlyingtargetfactor(i.e.,assessitsfacevalidity)
[Link]'s
alphaandcheckthe"Alphaifitemremoved"optiontodeterminewhetherremovalofanyadditional
itemswouldimprovereliability)
[Link]:HowdoIeliminateitems?(lecturenotes)
Nameanddescribethefactors [ edit ]
Oncethenumberoffactorshasbeendecidedandanyitemswhichdon'tbelonghavebeenremoved,then
[Link]
[Link]
dotheyrepresent?
[Link],emphasisethetoploadingitemsinnamingthefactor
[Link]
[Link]
Dataanalysisexercises [ edit ]
Dataanalysistutorial
Pros&cons [ edit ]
Advantages(Wikipedia)
Disadvantages(Wikipedia)
Glossary [ edit ]
Basicterms(fornovices/beginners)
[Link]:[Link]
areusedasameasureofsamplingadequacy(MSA).Note:Becarefulnottoconfusethiswiththeanti
imagecovariancematrix.
[Link]'stestofsphericity:Statisticaltestfortheoverallsignificanceofallcorrelationswithina
[Link](MSA).
[Link]:Varianceinavariablethatissharedwithothervariables.
[Link]:Theproportionofavariable'[Link]
communalityestimatesarethesumofsquaredloadingsforavariableinanorthogonalfactormatrix.
[Link]:Avariablewhichhasnotableloadings(e.g.,>.4)ontwoormorefactors.
[Link]:ThePearsonorproductmomentcorrelationcoefficient.
[Link]:Avariablewhichrepresentscombinedresponsestomultipleothervariables.A
[Link]
foreachcaseforeachfactor.
[Link]:Atableshowingthelinearcorrelationsbetweenallpairsofvariables.
[Link]:Reducingthenumberofvariables(e.g.,byusingfactoranalysistodeterminea
smallernumberoffactorstorepresentalargersetoffactors).
[Link] 4/6
1/20/2017 ExploratoryfactoranalysisWikiversity
[Link]:[Link]
whichisaccountedforbyaspecificfactor.
[Link]:Afactoranalysistechniqueusedtoexploretheunderlyingstructureofa
collectionofobservedvariables.
[Link]:[Link]
(constructs)thatsummariseoraccountfortheoriginalsetofobservedvariables.
[Link]:Astatisticaltechniqueusedtoestimatefactorsand/orreducethedimensionalityofa
largenumberofvariablestoafewernumberoffactors.
[Link]:Correlationbetweenavariableandafactor,andthekeytounderstandingthenature
[Link]
variableisexplainedbyafactor.
[Link]:[Link]
presentedascolumnsandthevariablesarepresentedasrows.
[Link]:Aprocessofadjustingthefactoraxestoachieveasimplerandpragmaticallymore
meaningfulfactorsolutionthegoalisausuallyasimplefactorstructure.
[Link]:Compositescorecreatedforeachobservation(case)foreachfactorwhichusesfactor
weightsinconjunctionwiththeoriginalvariablevaluestocalculateeachobservation'[Link]
scoresarestandardisedtoaccordingtoazscore.
[Link](MSA):Measureswhichindicatetheappropriatenessofapplying
factoranalysis.
[Link]:[Link]
arbitrarilyconstrainingthefactorrotationtoanorthogonal(90degreeangle),theobliquesolutionallows
[Link],thisiscalledObliminrotation.
[Link]:[Link]
factorisindependentof,ororthogonalto,[Link],thisiscalledVarimaxrotation.
[Link]:Whentwoormoretheoriesexplainthedataequallywell,selectthesimplest
theorye.g.,ifa2factoranda3factormodelexplainaboutthesameamountofvariance,interpretthe
2factormodel.
[Link](PAF):Amethodoffactoranalysisinwhichthefactorsarebasedona
[Link],communalitiesareinsertedin
thediagonalofthecorrelationmatrix,andtheextractedfactorsarebasedonlyonthecommon
variance,withuniquevarianceexcluded.
[Link](PCorPCA):Thefactorsarebasedonthetotalvarianceofall
items.[3] [4]
[Link]:[Link]
[Link]
(ordrops)fromclifftoscree.
[Link]:Apatternoffactorloadingresultssuchthateachvariableloadshighlyontooneand
onlyonefactor.
[Link]:Theproportionofavariable'svariancethatisnotsharedwithafactorstructure.
Uniquevarianceiscomposedofspecificanderrorvariance.
Other
[Link]:Afactoronwhichtwoormorevariablesload.
[Link]:Astatisticaltechniquewhichusesthecorrelationsbetweenobserved
variablestoestimatecommonfactorsandthestructuralrelationshipslinkingfactorstoobserved
variables.
[Link] 5/6
1/20/2017 ExploratoryfactoranalysisWikiversity
[Link]:[Link]
independentofcommonvariance,andacomponentoftheuniquevarianceofavariable.
[Link]:[Link]:
antiimageofavariable.
[Link]:Ifitisimpossibletoestimatepopulationfactorstructuresexactlybecauseaninfinite
numberoffactorstructurescanproducethesamecorrelationmatrix,thentherearemoreunknowns
thanequationsinthecommonfactormodel,andwesaythatthefactorstructureisindeterminate.
[Link]:Atheoreticalunderlyingfactorhypothesisedtoinfluenceanumberofobserved
[Link].
[Link]:(1)Varianceofeachvariableuniquetothatvariableandnotexplainedor
associatedwithothervariablesinthefactoranalysis.[5] (2)Thecomponentofuniquevariancewhich
isreliablebutnotexplainedbycommonfactors.[6]
[Link] 6/6
