factor analysis.ppt

Upload: saif-ali

Post on 20-Feb-2018

245 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/24/2019 Factor Analysis.ppt

    1/28

    2007 Prentice Hall 19-1

    Factor Analysis

    2007 Prentice Hall

    Advanced Statistical Modeling

    Adnan ButtAssistant Professor!ra "niveristy# $arac%i

  • 7/24/2019 Factor Analysis.ppt

    2/28

    2007 Prentice Hall 19-2

    Factor Analysis Factor analysisis a general na&e denoting a class of

    'rocedures 'ri&arily used for data reduction andsu&&ari(ation)

    Factor analysis is an interdependence techniquein t%at anentire set of interde'endent relations%i's is e*a&ined +it%out&a,ing t%e distinction et+een de'endent and inde'endent

    variales) Factor analysis is used in t%e follo+ing circu&stances.

    /o identify underlying di&ensions# or factors# t%at e*'laint%e correlations a&ong a set of variales)

    /o identify a ne+# s&aller# set of uncorrelated variales tore'lace t%e original set of correlated variales in suse!uent&ultivariate analysis regression or discri&inant analysis)

    /o identify a s&aller set of salient variales fro& a larger setfor use in suse!uent &ultivariate analysis)

  • 7/24/2019 Factor Analysis.ppt

    3/28

    2007 Prentice Hall 19-

    Factors UnderlyingSelected Psychographics

    and Lifestyles

    Factor 2

    Footall Baseall

    3vening at %o&e

    Factor 1

    Ho&e is est 'lace4o to a 'arty

    Plays Movies

  • 7/24/2019 Factor Analysis.ppt

    4/28

    2007 Prentice Hall 19-5

    Statistics Associated with FactorAnalysis

    Bartlett's test of sphericity)Bartlett6s test of s'%ericityis a test statistic used to e*a&ine t%e %y'ot%esis t%at t%evariales are uncorrelated in t%e 'o'ulation) n ot%er

    +ords# t%e 'o'ulation correlation &atri* is an identity

    &atri* eac% variale correlates 'erfectly +it% itself r8 1ut %as no correlation +it% t%e ot%er variales r8 0)

    Correlation matrix)A correlation &atri* is a lo+ertriangle &atri* s%o+ing t%e si&'le correlations#r#

    et+een all 'ossile 'airs of variales included in t%eanalysis) /%e diagonal ele&ents# +%ic% are all 1# are

    usually o&itted)

  • 7/24/2019 Factor Analysis.ppt

    5/28

    2007 Prentice Hall 19-

    Communality):o&&unality is t%e a&ount of variance

    a variale s%ares +it% all t%e ot%er variales eingconsidered) /%is is also t%e 'ro'ortion of variancee*'lained y t%e co&&on factors)

    igen!alue)/%e eigenvalue re'resents t%e total

    variance e*'lained y eac% factor)

    Factor loadings)Factor loadings are si&'lecorrelations et+een t%e variales and t%e factors)

    Factor loading plot)A factor loading 'lot is a 'lot of

    t%e original variales using t%e factor loadings ascoordinates)

    Factor matrix) A factor &atri* contains t%e factorloadings of all t%e variales on all t%e factors e*tracted)

    Statistics Associated with FactorAnalysis

  • 7/24/2019 Factor Analysis.ppt

    6/28

    2007 Prentice Hall 19-;

    Factor scores) Factor scores are co&'osite scoresesti&ated for eac% res'ondent on t%e derived factors)

    "aiser#$eyer#%l&in "$%( measure of samplingadequacy)/%e $aiser-Meyer-erences et+een t%e oserved

    correlations# as given in t%e in'ut correlation &atri*# and t%ere'roduced correlations# as esti&ated fro& t%e factor &atri*)

    Scree plot)A scree 'lot is a 'lot of t%e 3igenvalues againstt%e nu&er of factors in order of e*traction)

    Statistics Associated with FactorAnalysis

  • 7/24/2019 Factor Analysis.ppt

    7/28

    2007 Prentice Hall 19-7

    Conducting Factor Analysis)SP%*+*,

    *U$B) -. -/ -0 -1 -2 -3

    . 4566 0566 3566 1566 /566 1566

    / .566 0566 /566 1566 2566 1566

    0 3566 /566 4566 1566 .566 0566

    1 1566 2566 1566 3566 /566 2566

    2 .566 /566 /566 0566 3566 /566

    3 3566 0566 3566 1566 /566 1566

    4 2566 0566 3566 0566 1566 0566

    7 3566 1566 4566 1566 .566 1566

    8 0566 1566 /566 0566 3566 0566

    .6 /566 3566 /566 3566 4566 3566

    .. 3566 1566 4566 0566 /566 0566

    ./ /566 0566 .566 1566 2566 1566

    .0 4566 /566 3566 1566 .566 0566

    .1 1566 3566 1566 2566 0566 3566

    .2 .566 0566 /566 /566 3566 1566

    .3 3566 1566 3566 0566 0566 1566

    .4 2566 0566 3566 0566 0566 1566

    .7 4566 0566 4566 1566 .566 1566

    .8 /566 1566 0566 0566 3566 0566

    /6 0566 2566 0566 3566 1566 3566

    /. .566 0566 /566 0566 2566 0566

    // 2566 1566 2566 1566 /566 1566

    /0 /566 /566 .566 2566 1566 1566

    /1 1566 3566 1566 3566 1566 4566

    /2 3566 2566 1566 /566 .566 1566

    /3 0566 2566 1566 3566 1566 4566

    /4 1566 1566 4566 /566 /566 2566

    /7 0566 4566 /566 3566 1566 0566

    /8 1566 3566 0566 4566 /566 4566

    06 /566 0566 /566 1566 4566 /566

  • 7/24/2019 Factor Analysis.ppt

    8/28

    2007 Prentice Hall 19-?

    Conducting Factor Analysis

    :onstruction of t%e :orrelation Matri*

    Met%od of Factor Analysis

    @eter&ination of u&er of Factors

    @eter&ination of Model Fit

    Prole& for&ulation

    :alculation ofFactor Scores

    nter'retation of Factors

    otation of Factors

    Selection ofSurrogate =ariales

  • 7/24/2019 Factor Analysis.ppt

    9/28

    2007 Prentice Hall 19-9

    Conducting FactorAnalysisFormulate the Pro9lem /%e oCectives of factor analysis s%ould e identiDed)

    /%e variales to e included in t%e factor analysis

    s%ould e s'eciDed ased on 'ast researc%# t%eory#and Cudg&ent of t%e researc%er) t is i&'ortant t%att%e variales e a''ro'riately &easured on aninterval or ratio scale)

    An a''ro'riate sa&'le si(e s%ould e used) As aroug% guideline# t%ere s%ould e at least four or Dveti&es as &any oservations sa&'le si(e as t%ereare variales)

  • 7/24/2019 Factor Analysis.ppt

    10/28

    2007 Prentice Hall 19-10

    Correlation$atrix

    /ale 19)2

  • 7/24/2019 Factor Analysis.ppt

    11/28

    2007 Prentice Hall 19-11

    /%e analytical 'rocess is ased on a &atri* ofcorrelations et+een t%e variales)

    Bartlett6s test of s'%ericity can e used to test t%e null%y'ot%esis t%at t%e variales are uncorrelated in t%e

    'o'ulation. in ot%er +ords# t%e 'o'ulation correlation&atri* is an identity &atri*) f t%is %y'ot%esis cannot ereCected# t%en t%e a''ro'riateness of factor analysiss%ould e !uestioned)

    Anot%er useful statistic is t%e $aiser-Meyer-

  • 7/24/2019 Factor Analysis.ppt

    12/28

    2007 Prentice Hall 19-12

    n principal components analysis#t%e total variance int%e data is considered) /%e diagonal of t%e correlation&atri* consists of unities# and full variance is roug%t intot%e factor &atri*) Princi'al co&'onents analysis isreco&&ended +%en t%e 'ri&ary concern is to deter&ine t%e&ini&u& nu&er of factors t%at +ill account for &a*i&u&

    variance in t%e data for use in suse!uent &ultivariateanalysis) /%e factors are calledprincipal components)

    ncommon factor analysis#t%e factors are esti&ated

    ased only on t%e co&&on variance) :o&&unalities areinserted in t%e diagonal of t%e correlation &atri*) /%is&et%od is a''ro'riate +%en t%e 'ri&ary concern is toidentify t%e underlying di&ensions and t%e co&&onvariance is of interest) /%is &et%od is also ,no+n as

    principal axis factoring)

    Conducting Factor Analysis+etermine the $ethod of Factor

    Analysis

  • 7/24/2019 Factor Analysis.ppt

    13/28

    2007 Prentice Hall 19-1

    )esults of PrincipalComponents Analysis

    /ale 19)

  • 7/24/2019 Factor Analysis.ppt

    14/28

    2007 Prentice Hall 19-15

    )esults of PrincipalComponents Analysis

    /ale 19)#cont)

  • 7/24/2019 Factor Analysis.ppt

    15/28

    2007 Prentice Hall 19-1

    )esults of PrincipalComponents Analysis

    /ale 19)#cont)

  • 7/24/2019 Factor Analysis.ppt

    16/28

    2007 Prentice Hall 19-1;

    /%e lo+er-left triangle contains t%e re'roducedcorrelation &atri* t%e diagonal# t%e co&&unalitiest%e u''er-rig%t triangle# t%e residuals et+een t%e

    oserved correlations and t%e re'roducedcorrelations)

    )esults of PrincipalComponents Analysis

    /ale 19)#cont)

  • 7/24/2019 Factor Analysis.ppt

    17/28

    2007 Prentice Hall 19-17

    A Priori +etermination5 So&eti&es# ecause of 'rior,no+ledge# t%e researc%er ,no+s %o+ &any factors toe*'ect and t%us can s'ecify t%e nu&er of factors to ee*tracted efore%and)

    +etermination Based on igen!alues5 n t%isa''roac%# only factors +it% 3igenvalues greater t%an 1)0are retained) An 3igenvalue re'resents t%e a&ount ofvariance associated +it% t%e factor) Hence# only factors

    +it% a variance greater t%an 1)0 are included) Factors +it%variance less t%an 1)0 are no etter t%an a single variale#since# due to standardi(ation# eac% variale %as a varianceof 1)0) f t%e nu&er of variales is less t%an 20# t%isa''roac% +ill result in a conservative nu&er of factors)

    Conducting Factor Analysis+etermine the *um9er of

    Factors

  • 7/24/2019 Factor Analysis.ppt

    18/28

    2007 Prentice Hall 19-1?

    +etermination Based on Scree Plot5 A scree 'lot isa 'lot of t%e 3igenvalues against t%e nu&er of factorsin order of e*traction) 3*'eri&ental evidence indicatest%at t%e 'oint at +%ic% t%e scree egins denotes t%etrue nu&er of factors) 4enerally# t%e nu&er offactors deter&ined y a scree 'lot +ill e one or a fe+&ore t%an t%at deter&ined y t%e 3igenvalue criterion)

    +etermination Based on Percentage of -ariance5n t%is a''roac% t%e nu&er of factors e*tracted isdeter&ined so t%at t%e cu&ulative 'ercentage ofvariance e*tracted y t%e factors reac%es a satisfactorylevel) t is reco&&ended t%at t%e factors e*tracteds%ould account for at least ;0 of t%e variance)

    Conducting Factor Analysis+etermine the *um9er of

    Factors

  • 7/24/2019 Factor Analysis.ppt

    19/28

    2007 Prentice Hall 19-19

    Scree Plot

    0)

    2 5

    ;

    :o&'onent u&er

    0)0

    2)0

    )0

    3igenva

    lue

    1)0

    1)

    2)

    1

    Fig) 19)

  • 7/24/2019 Factor Analysis.ppt

    20/28

    2007 Prentice Hall 19-20

    +etermination Based on Split#;alf )elia9ility5/%e sa&'le is s'lit in %alf and factor analysis is

    'erfor&ed on eac% %alf)

  • 7/24/2019 Factor Analysis.ppt

    21/28

    2007 Prentice Hall 19-21

    Alt%oug% t%e initial or unrotated factor &atri* indicatest%e relations%i' et+een t%e factors and individualvariales# it seldo& results in factors t%at can einter'reted# ecause t%e factors are correlated +it%

    &any variales) /%erefore# t%roug% rotation t%e factor&atri* is transfor&ed into a si&'ler one t%at is easier tointer'ret)

    n rotating t%e factors# +e +ould li,e eac% factor to %avenon(ero# or signiDcant# loadings or coeGcients for only

    so&e of t%e variales) i,e+ise# +e +ould li,e eac%variale to %ave non(ero or signiDcant loadings +it%only a fe+ factors# if 'ossile +it% only one)

    /%e rotation is called orthogonal rotationif t%e a*esare &aintained at rig%t angles)

    Conducting FactorAnalysis

    )otate Factors

  • 7/24/2019 Factor Analysis.ppt

    22/28

    2007 Prentice Hall 19-22

    /%e &ost co&&only used &et%od for rotation is t%e!arimax procedure) /%is is an ort%ogonal &et%odof rotation t%at &ini&i(es t%e nu&er of variales

    +it% %ig% loadings on a factor# t%erey en%ancing t%einter'retaility of t%e factors)

  • 7/24/2019 Factor Analysis.ppt

    23/28

    2007 Prentice Hall 19-2

    Factor $atrix Before and After)otation

    Factors

    a

    Hig% oadingsBefore otation

    Fig) 19)5

    Hig% oadingsAfter otation

    Factors

    =ariales

    1

    2

    5

    ;

    1

    I

    II

    I

    I

    2

    I

    I

    I

    I

    1

    I

    I

    I

    2

    I

    I

    I

    =ariales

    1

    2

    5

    ;

  • 7/24/2019 Factor Analysis.ppt

    24/28

    2007 Prentice Hall 19-25

    A factor can t%en e inter'reted in ter&s oft%e variales t%at load %ig% on it)

    Anot%er useful aid in inter'retation is to 'lott%e variales# using t%e factor loadings as

    coordinates) =ariales at t%e end of an a*isare t%ose t%at %ave %ig% loadings on onlyt%at factor# and %ence descrie t%e factor)

    Conducting FactorAnalysis=nterpret Factors

  • 7/24/2019 Factor Analysis.ppt

    25/28

    2007 Prentice Hall 19-2

    Factor LoadingPlot

    Fig) 19)

    .56

    652

    656

    #652

    #.56

    Component.

    Component Plotin)otated Space

    .56 652 656 #652 #.56

    -.

    -0

    -3

    -/

    -2

    -1

    Component-aria9le . /

    -. 6583/ #/533#

    6/

    -/ #254/#6/ 65717

    -0 65801 #65.13

    -1 #8570#6/ 65721

    -2 #65800 #7516#6/

    -3 75004#6/ 65772

    )otated Component $atrix

    Component /

  • 7/24/2019 Factor Analysis.ppt

    26/28

    2007 Prentice Hall 19-2;

    /%e factor scoresfor t%e it% factor &aye esti&ated

    as follo+s.

    Fi= Wi1X1+ Wi2X2+ Wi3X3+ . . . + WikXk

    Conducting FactorAnalysisCalculate Factor Scores

  • 7/24/2019 Factor Analysis.ppt

    27/28

    2007 Prentice Hall 19-27

    By e*a&ining t%e factor &atri*# one could selectfor eac% factor t%e variale +it% t%e %ig%est

    loading on t%at factor) /%at variale could t%en eused as a surrogate variale for t%e associatedfactor)

    Ho+ever# t%e c%oice is not as easy if t+o or &ore

    variales %ave si&ilarly %ig% loadings) n suc% acase# t%e c%oice et+een t%ese variales s%oulde ased on t%eoretical and &easure&entconsiderations)

    Conducting Factor

    AnalysisSelect Surrogate-aria9les

  • 7/24/2019 Factor Analysis.ppt

    28/28

    2007 Prentice Hall 19-2?

    /%e correlations et+een t%e variales can e

    deduced or re'roduced fro& t%e esti&atedcorrelations et+een t%e variales and t%efactors)

    /%e di>erences et+een t%e oserved correlationsas given in t%e in'ut correlation &atri* and t%ere'roduced correlations as esti&ated fro& t%efactor &atri* can e e*a&ined to deter&ine&odel Dt) /%ese di>erences are called residuals)

    Conducting FactorAnalysis+etermine the $odel Fit