Key terms

Cbapter : Introduction: Metbods and model building

A|pha ()
See type l ettot
8eta ()
See 1ype ll ettot
8|var|ate part|a| corre|at|on
Slmple (Lwovarlable) correlaLlon beLween Lwo seLs of reslduals (unexplalned varlances) LhaL remaln
afLer Lhe assoclaLlon of oLher lndependenL varlables ls removed
8ootstrapp|ng
An approach Lo valldaLlng a mulLlvarlaLe model by drawlng a large number of subsamples and
esLlmaLlng models for each subsamples LsLlmaLes from all Lhe subsamples are Lhen comblned
provldlng noL only Lhe besL" esLlmaLed coefflclenLs (eg means of each esLlmaLed coefflclenL across
all Lhe subsample models) buL Lhelr expecLed varlablllLy and Lhus Lhelr llkellhood of dlfferlng from
zero LhaL ls are Lhe esLlmaLed coefflclenLs sLaLlsLlcally dlfferenL from zero or noL? 1hls approach
does noL rely on sLaLlsLlcal assumpLlons abouL Lhe populaLlon Lo assess sLaLlsLlcal slgnlflcance buL
lnsLead makes lLs assessmenL based solely on Lhe sample daLa
Compos|te measure
See sommoteJ scoles
ependent techn|que
ClasslflcaLlon of sLaLlsLlcal Lechnlques dlsLlngulshed by havlng a varlable or seL of varlables ldenLlfled
as Lhe JepeoJeot votloble(s) and Lhe remalnlng varlables as loJepeoJeot 1he ob[ecLlve ls predlcLlon
of Lhe dependenL varlable(s) by Lhe lndependenL varlable(s) An example ls regresslon analysls
ependent var|ab|e
resumed effecL of or response Lo a change ln Lhe loJepeoJeot votloble(s)
ummy var|ab|e
Noomettlcolly measured varlable Lransformed lnLo a mettlc varlable by asslgnlng a 1 or 0 Lo a sub[ecL
dependlng on wheLher lL possesses a parLlcular characLerlsLlc


ect s|ze
LsLlmaLe of Lhe degree Lo whlch Lhe phenomenon belng sLudled (eg correlaLlon or dlfference ln
means) exlsLs ln Lhe populaLlon
Independent var|ab|e
resumed cause of any change ln Lhe JepeoJeot votloble
Ind|cator
Slngle varlable used ln con[uncLlon wlLh one or more varlables Lo form a composlte meosote
Interdependence techn|que
ClasslflcaLlon of sLaLlsLlcal Lechnlques ln whlch Lhe varlables are noL dlvlded lnLo JepeoJeot and
loJepeoJeot seLs raLher all varlables are analyzed as a slngle seL (eg facLor analysls)
,easurement error
lnaccuracles of measurlng Lhe Lrue" varlable values due Lo Lhe falllblllLy of Lhe measuremenL
lnsLrumenL (le lnapproprlaLe response scales) daLa enLry errors or respondenL errors
,etr|c data
Also called quanLlLaLlve daLa lnLerval daLa or raLlo daLa Lhese measuremenLs ldenLlfy or descrlbe
sub[ecLs (or ob[ecLs) noL only on Lhe possesslon of an aLLrlbuLe buL also by Lhe amounL or degree Lo
whlch Lhe sub[ecL may be characLerlzed by Lhe aLLrlbuLe lor example a person's age and welghL are
meLrlc daLa
,u|t|co|||near|ty
LxLenL Lo whlch a varlable can be explalned by Lhe oLher varlables ln Lhe analysls As mulLlcolllnearlLy
lncreases lL compllcaLes Lhe lnLerpreLaLlon of Lhe votlote because lL ls more dlfflculL Lo ascerLaln Lhe
effecL of any slngle varlable owlng Lo Lhelr lnLerrelaLlonshlps
,u|t|var|ate ana|ys|s
Analysls of mulLlple varlables ln a slngle relaLlonshlp or seL of relaLlonshlps
,u|t|var|ate measurement
use of Lwo or more varlables as loJlcotots of a slngle composlte meosote lor example a personallLy
LesL may provlde Lhe answers Lo a serles of lndlvldual quesLlons (lndlcaLors) whlch are Lhen
comblned Lo form a slngle score (sommoteJ scole) represenLlng Lhe personallLy LralL
-onmetr|c data
Also called quallLaLlve daLa Lhese are aLLrlbuLes characLerlsLlcs or caLegorlcal properLles LhaL
ldenLlfy or descrlbe a sub[ecL or ob[ecL 1hey dlffer from mettlc Joto by lndlcaLlng Lhe presence of an
aLLrlbuLe buL noL Lhe amounL Lxamples are occupaLlon (physlclan aLLorney professor) or buyer
sLaLus (buyer nonbuyer) Also called nomlnal daLa or ordlnal daLa

9ower
robablllLy of correcLly re[ecLlng Lhe null hypoLhesls when lL ls false LhaL ls correcLly flndlng a
hypoLheslzed relaLlonshlp when lL exlsLs ueLermlned as a funcLlon of (1) Lhe sLaLlsLlcal slgnlflcance
level seL by Lhe researcher for a 1ype l ettot () (2) Lhe sample slze used ln Lhe analysls and (3) Lhe
effect slze belng examlned
9ract|ca| s|gn||cance
Means of assesslng mulLlvarlaLe analysls resulLs based on Lhelr subsLanLlve flndlngs raLher Lhan Lhelr
sLaLlsLlcal slgnlflcance Whereas sLaLlsLlcal slgnlflcance deLermlnes wheLher Lhe resulL ls aLLrlbuLable
Lo chance pracLlcal slgnlflcance assesses wheLher Lhe resulL ls useful (le subsLanLlal enough Lo
warranL acLlon) ln achlevlng Lhe research ob[ecLlves
ke||ab|||ty
LxLenL Lo whlch a varlable or seL of varlables ls conslsLenL ln whaL lL ls lnLended Lo measure lf
mulLlple measuremenLs are Laken Lhe rellable measures wlll all be conslsLenL ln Lhelr values lL
dlffers from vollJlty ln LhaL lL relaLes noL Lo whaL should be measured buL lnsLead Lo how lL ls
measured
pec||cat|on error
CmlLLlng a key varlable from Lhe analysls Lhus affecLlng Lhe esLlmaLed effecLs of lncluded varlables
ummated sca|es
MeLhod of comblnlng several varlables LhaL measure Lhe same concepL lnLo a slngle varlable ln an
aLLempL Lo lncrease Lhe telloblllty of Lhe measuremenL Lhrough moltlvotlote meosotemeot ln mosL
lnsLances Lhe separaLe varlables are summed and Lhen Lhelr LoLal or average score ls used ln Lhe
analysls
@reatment
lndependenL varlable Lhe researcher manlpulaLes Lo see Lhe effecL (lf any) on Lhe dependenL
varlable(s) such as ln an experlmenL (eg LesLlng Lhe appeal of color versus blackandwhlLe
adverLlsemenLs)
@ype I error
robablllLy of lncorrecLly re[ecLlng Lhe null hypoLhesls ln mosL cases lL means saylng a dlfference or
correlaLlon exlsLs when lL acLually does noL Also Lermed olpbo () 1yplcal levels are 3 or 1 percenL
Lermed Lhe 003 or 001 level respecLlvely
@ype II error
robablllLy of lncorrecLly falllng Lo re[ecL Lhe null hypoLhesls ln slmple Lerms Lhe chance of noL
flndlng a correlaLlon or mean dlfference when lL does exlsL Also Lermed beto () lL ls lnversely
relaLed Lo Lhe 1ype l ettot 1he value of 1 mlnus Lhe 1ype ll error (1) ls deflned as powet


Dn|var|ate ana|ys|s o var|ance (A-IA)
SLaLlsLlcal Lechnlque used Lo deLermlne on Lhe basls of one dependenL measure wheLher samples
are from populaLlons wlLh equal means
Ia||d|ty
LxLenL Lo whlch a measure or seL of measures correcLly represenLs Lhe concepL of sLudy Lhe degree
Lo whlch lL ls free from any sysLemaLlc or nonrandom error valldlLy ls concerned wlLh how well Lhe
concepL ls deflned by Lhe measure(s) whereas telloblllty relaLes Lo Lhe conslsLency of Lhe measure(s)
Iar|ate
Llnear comblnaLlon of varlables formed ln Lhe mulLlvarlaLe Lechnlque by derlvlng emplrlcal welghLs
applled Lo a seL of varlables speclfled by Lhe researcher



Cbapter : Cleaning and transforming data

A||ava||ab|e approach
lmpototloo meLhod for mlsslng daLa LhaL compuLes values based on allavallable valld observaLlons
also known as Lhe palrwlse approach
8oxp|ot
MeLhod of represenLlng Lhe dlsLrlbuLlon of a varlable A box represenLs Lhe ma[or porLlon of Lhe
dlsLrlbuLlon and Lhe exLenslons called whlskers reach Lo Lhe exLreme polnLs of Lhe dlsLrlbuLlon
1hls meLhod ls useful ln maklng comparlsons of one or more varlables across groups
Censored data
CbservaLlons LhaL are lncompleLe ln a sysLemaLlc and known way Cne example occurs ln Lhe sLudy
of causes of deaLh ln a sample ln whlch some lndlvlduals are sLlll llvlng Censored daLa are an
example of lqootoble mlssloq Joto
Compar|son group
See tefeteoce coteqoty
Comp|ete case approach
Approach for handllng mlssloq Joto LhaL compuLes values based on daLa from compleLe cases LhaL
ls cases wlLh no mlsslng daLa Also known as Lhe llsLwlse approach
ata transormat|ons
A varlable may have an undeslrable characLerlsLlc such as nonnormallLy LhaL deLracLs from lLs use ln
a mulLlvarlaLe Lechnlque A LransformaLlon such as Laklng Lhe logarlLhm or square rooL of Lhe
varlable creaLes a Lransformed varlable LhaL ls more sulLed Lo porLraylng Lhe relaLlonshlp
1ransformaLlons may be applled Lo elLher Lhe dependenL or lndependenL varlable or boLh 1he need
and speclflc Lype of LransformaLlon may be based on LheoreLlcal reasons (eg Lransformlng a known
nonllnear relaLlonshlp) or emplrlcal reasons (eg problems ldenLlfled Lhrough graphlcal or sLaLlsLlcal
means)
ummy var|ab|e
Speclal meLrlc varlable used Lo represenL a slngle caLegory of a nonmeLrlc varlable 1o accounL for l
levels of a nonmeLrlc varlable l 1 dummy varlables are needed lor example gender ls measured
as male or female and could be represenLed by Lwo dummy varlables (x1 and x2) When Lhe
respondenL ls male x11 and x2 0 Llkewlse when Lhe respondenL ls female x10 and x21
Powever when x11 we know LhaL x2 musL equal 0 1hus we need only one varlable elLher x1 or
x2 Lo represenL Lhe varlable gender lf a nonmeLrlc varlable has Lhree levels only Lwo dummy
varlables are needed We always have one dummy varlable less Lhan Lhe number of levels for Lhe
nonmeLrlc varlable 1he omlLLed caLegory ls Lermed Lhe tefeteoce coteqoty

ects cod|ng
MeLhod for speclfylng Lhe tefeteoce coteqoty for a seL of Jommy votlobles where Lhe reference
caLegory recelves a value of mlnus one (1) across Lhe seL of dummy varlables WlLh Lhls Lype of
codlng Lhe dummy varlable coefflclenLs represenL group devlaLlons from Lhe mean of all groups
whlch ls ln conLrasL Lo loJlcotot coJloq
eteroscedast|c|ty
See bomosceJotlclty
|stogram
Craphlcal dlsplay of Lhe dlsLrlbuLlon of a slngle varlable 8y formlng frequency counLs ln caLegorles
Lhe shape of Lhe varlable's dlsLrlbuLlon can be shown used Lo make a vlsual comparlson Lo Lhe
ootmol Jlsttlbotloo
omoscedast|c|ty
When Lhe varlance of Lhe error Lerms (e) appears consLanL over a range of predlcLor varlables Lhe
daLa are sald Lo be homoscedasLlc 1he assumpLlon of equal varlance of Lhe populaLlon error
(where ls esLlmaLed from e) ls crlLlcal Lo Lhe proper appllcaLlon of many mulLlvarlaLe Lechnlques
When Lhe error Lerms have lncreaslng or modulaLlng varlance Lhe daLa are sald Lo be
betetosceJostlc Analysls of teslJools besL lllusLraLes Lhls polnL
Ignorab|e m|ss|ng data
Mlssloq Joto ptocess LhaL ls expllclLly ldenLlflable and/or ls under Lhe conLrol of Lhe researcher
lgnorable mlsslng daLa do noL requlre a remedy because Lhe mlsslng daLa are expllclLly handled ln Lhe
Lechnlque used
Imputat|on
rocess of esLlmaLlng Lhe mlssloq Joto of an observaLlon based on valld values of Lhe oLher varlables
1he ob[ecLlve ls Lo employ known relaLlonshlps LhaL can be ldenLlfled ln Lhe valld values of Lhe
sample Lo asslsL ln represenLlng or even esLlmaLlng Lhe replacemenLs for mlsslng values
Ind|cator cod|ng
MeLhod for speclfylng Lhe tefeteoce coteqoty for a seL of Jommy votlobles where Lhe reference
caLegory recelves a value of zero across Lhe seL of dummy varlables 1he dummy varlable coefflclenLs
represenL Lhe caLegory dlfferences from Lhe reference caLegory Also see effects coJloq
urtos|s
Measure of Lhe peakedness or flaLness of a dlsLrlbuLlon when compared wlLh a ootmol Jlsttlbotloo A
poslLlve value lndlcaLes a relaLlvely peaked dlsLrlbuLlon and a negaLlve value lndlcaLes a relaLlvely
flaL dlsLrlbuLlon


|near|ty
used Lo express Lhe concepL LhaL Lhe model possesses Lhe properLles of addlLlvlLy and homogenelLy
ln a slmple sense llnear models predlcL values LhaL fall ln a sLralghL llne by havlng a consLanL unlL
change (slope) of Lhe dependenL varlable for a consLanL unlL change of Lhe lndependenL varlable ln
Lhe populaLlon model b0+b1\l+e Lhe effecL of a change of 1 ln \1 ls Lo add b1 (a consLanL) unlLs Lo

,|ss|ng at random (,Ak)
ClasslflcaLlon of mlssloq Joto appllcable when mlsslng values of depend on \ buL noL on buL noL
on When mlsslng daLa are MA8 observed daLa for are a Lruly random sample for Lhe \ values ln
Lhe sample buL noL a random sample of all values due Lo mlsslng values of \
,|ss|ng comp|ete|y at random (,CAk)
ClasslflcaLlon of mlssloq Joto appllcable when mlsslng values of are noL depend on \ When mlsslng
daLa are MCA8 observed values of are a Lruly random sample of all values wlLh no underlylng
process LhaL lends blas Lo Lhe observed daLa
,|ss|ng data
lnformaLlon noL avallable for a sub[ecL (or case) abouL whom oLher lnformaLlon ls avallable Mlsslng
daLa ofLen occur when a respondenL falls Lo answer one or more quesLlons ln a survey
,|ss|ng data process
Any sysLemaLlc evenL exLernal Lo Lhe respondenL (such as daLa enLry errors or daLa collecLlon
problems) or any acLlon on Lhe parL of Lhe respondenL (such as refusal Lo answer a quesLlon) LhaL
leads Lo mlssloq Joto
,u|t|var|ate graph|ca| d|sp|ay
MeLhod of presenLlng a mulLlvarlaLe proflle of an observaLlon on Lhree or more varlables 1he
meLhods lnclude approaches such as glyphs maLhemaLlcal LransformaLlons and even lconlc
represenLaLlons (eg faces)
-orma| d|str|but|on
urely LheoreLlcal conLlnuous probablllLy dlsLrlbuLlon ln whlch Lhe horlzonLal axls represenLs all
posslble values of a varlable and Lhe verLlcal axls represenLs Lhe probablllLy of Lhose values occurrlng
1he scores on Lhe varlable are clusLered around Lhe mean ln a symmeLrlcal unlmodal paLLern known
as Lhe bellshaped or normal curve
-orma| probab|||ty p|ot
Craphlcal comparlson of Lhe form of Lhe dlsLrlbuLlon Lo Lhe ootmol Jlsttlbotloo ln Lhe normal
probablllLy ploL Lhe normal dlsLrlbuLlon ls represenLed by a sLralghL llne angled aL 43 degrees 1he
acLual dlsLrlbuLlon ls ploLLed agalnsL Lhls llne so LhaL any dlfferences are shown as devlaLlons from
Lhe sLralghL llne maklng ldenLlflcaLlon of dlfferences qulLe apparenL and lnLerpreLable

-orma||ty
uegree Lo whlch Lhe dlsLrlbuLlon of Lhe sample daLa corresponds Lo a ootmol Jlsttlbotloo
u|t|er
An observaLlon LhaL subsLanLlally dlfferenL from Lhe oLher observaLlons (le lL has an exLreme value)
on one or more characLerlsLlcs (varlables) AL lssue ls lLs represenLaLlveness of Lhe populaLlon
keerence category
1he caLegory of a nonmeLrlc varlable LhaL ls omlLLed when creaLlng Jommy votlobles and acLs as a
reference polnL ln lnLerpreLlng Lhe dummy varlables ln loJlcotot coJloqLhe reference caLegory has
values of zero (0) for all dummy varlables WlLh effects coJloq Lhe reference caLegory has values of
mlnus one (1) for all dummy varlables
kes|dua|
orLlon of a dependenL varlable noL explalned by a mulLlvarlaLe Lechnlque AssoclaLed wlLh
dependence meLhods LhaL aLLempL Lo predlcL Lhe dependenL varlable Lhe resldual represenLs Lhe
unexplalned porLlon of Lhe dependenL varlable 8eslduals can be used ln dlagnosLlc procedures Lo
ldenLlfy problems ln Lhe esLlmaLlon Lechnlque or Lo ldenLlfy unspeclfled relaLlonshlps
kobustness
1he ablllLy of a sLaLlsLlcal Lechnlque Lo perform reasonably well even when Lhe underlylng sLaLlsLlcal
assumpLlons have been vlolaLed ln some manner
catterp|ot
8epresenLaLlon of Lhe relaLlonshlp beLween Lwo meLrlc varlables porLraylng Lhe [olnL values of each
observaLlon ln a Lwodlmenslonal graph
ewness
Measure of Lhe symmeLry of a dlsLrlbuLlon ln mosL lnsLances Lhe comparlson ls made Lo a ootmol
Jlsttlbotloo A poslLlvely skewed dlsLrlbuLlon has relaLlvely few large values and Lalls off Lo Lhe rlghL
and a negaLlvely skewed dlsLrlbuLlon has relaLlvely few small values and Lalls off Lo Lhe lefL Skewness
values falllng ouLslde Lhe range of 1 Lo +1 lndlcaLe a subsLanLlally skewed dlsLrlbuLlon
Iar|ate
Llnear comblnaLlon of varlables formed ln Lhe mulLlvarlaLe Lechnlque by derlvlng emplrlcal welghLs
applled Lo a seL of varlables speclfled by Lhe researcher

Cbapter : Factor analysis

Ant||mage corre|at|on matr|x
MaLrlx of Lhe parLlal correlaLlons among varlables afLer facLor analysls represenLlng Lhe degree Lo
whlch Lhe facLors explaln each oLher ln Lhe resulLs 1he dlagonal conLalns Lhe meosotes of somplloq
oJepoocy for each varlable and Lhe offdlagonal values are parLlal correlaLlons among varlables
8art|ett test o spher|c|ty
SLaLlsLlcal LesL for Lhe overall slgnlflcance of all correlaLlons wlLhln a cottelotloo mottlx
C|uster ana|ys|s
MulLlvarlaLe Lechnlque wlLh Lhe ob[ecLlve of grouplng respondenLs or cases wlLh slmllar proflles on a
deflned seL of characLerlsLlcs Slmllar Lo O foctot ooolysls
Common actor ana|ys|s
lacLor model ln whlch Lhe facLors are based on a reduced correlaLlon maLrlx 1haL ls commooolltles
are lnserLed ln Lhe dlagonal of Lhe cottelotloo mottlx and Lhe exLracLed facLors are based only on Lhe
commoo votlooce wlLh Lhe speclflc and ettot votlooce excluded
Common var|ance
varlance shared wlLh oLher varlables ln Lhe facLor analysls
Communa||ty
1oLal amounL of varlance an orlglnal varlable shares wlLh all oLher varlables lncluded ln Lhe analysls
Component ana|ys|s
lacLor model ln whlch Lhe facLors are based on Lhe LoLal varlance WlLh componenL analysls unlLles
(1s) are used ln Lhe dlagonal of Lhe cottelotloo mottlx Lhls procedure compuLaLlonally lmplles LhaL all
Lhe varlance ls common or shared
Compos|t measure
See sommoteJ scoles
Conceptua| de|n|t|on
SpeclflcaLlon of Lhe LheoreLlcal basls for a concepL LhaL ls represenLed by a facLor
Content va||d|ty
AssessmenL of Lhe degree of correspondence beLween Lhe lLems selecLed Lo consLlLuLe a sommoteJ
scole and lLs cooceptool Jefloltloo


Corre|at|on matr|x
1able showlng Lhe lnLercorrelaLlons among all varlables
Cronbach's a|pha
Measure of telloblllty LhaL ranges from 0 Lo 1 wlLh values of 060 Lo 070 deemed Lhe lower llmlL of
accepLablllLy
Cross|oad|ng
A varlable has Lwo more foctot looJloqs exceedlng Lhe Lhreshold value deemed necessary for
lncluslon ln Lhe facLor lnLerpreLaLlon process
ummy var|ab|e
8lnary meLrlc varlable used Lo represenL a slngle caLegory of a nonmeLrlc varlable
|genva|ue
Column sum of squared loadlngs for a facLor also referred Lo as Lhe loteot toot lL represenLs Lhe
amounL of varlance accounLed for by a facLor
;DI,Ak
Cne of Lhe ottboqoool foctot tototloo meLhods LhaL ls a compromlse" beLween Lhe vA8lMAx and
CuA81lMAx approaches buL ls noL wldely used
rror var|ance
varlance of a varlable due Lo errors ln daLa collecLlon or measuremenL
Iace va||d|ty
See cooteot vollJlty
Iactor
Llnear comblnaLlon (varlaLe) of Lhe orlglnal varlables lacLors also represenL Lhe underlylng
dlmenslons (consLrucLs) LhaL summarlze or accounL for Lhe orlglnal seL of observed varlables
Iactor |ndeterm|nacy
CharacLerlsLlc of commoo foctot ooolysls such LhaL several dlfferenL foctot scotes can be calculaLed
for a respondenL each flLLlng Lhe esLlmaLed facLor model lL means Lhe facLor scores are noL unlque
for each lndlvldual
Iactor |oad|ngs
CorrelaLlon beLween Lhe orlglnal varlables and Lhe facLors and Lhe key Lo undersLandlng Lhe naLure
of a parLlcular facLor Squared facLor loadlngs lndlcaLe whaL percenLage of Lhe varlance ln an orlglnal
varlable ls explalned by a facLor

Iactor matr|x
1able dlsplaylng Lhe foctot looJloqs of all varlables on each facLor
Iactor pattern matr|x
Cne of Lhe Lwo facLor maLrlces found ln an obllpoe tototloo LhaL ls mosL comparable Lo Lhe facLor
maLrlx ln an ottboqoool tototloo
Iactor rotat|on
rocess of manlpulaLlon or ad[usLlng Lhe facLor axes Lo achleve a slmpler and pragmaLlcally more
meanlngful facLor soluLlon
Iactor score
ComposlLe measure creaLed for each observaLlon on each facLor exLracLed ln Lhe facLor analysls 1he
facLor welghLs are use ln con[uncLlon wlLh Lhe orlglnal varlable values Lo calculaLe each observaLlon's
score 1he facLor score Lhen can be used Lo represenL Lhe facLor(s) ln subsequenL analyses lacLor
scores are sLandardlzed Lo have a mean of 0 and a sLandard devlaLlon of 1
Iactor structure matr|x
A foctot mottlx found ln an obllpoe tototloo LhaL represenLs Lhe slmple correlaLlons beLween
varlables and facLors MosL researchers prefer Lo use Lhe foctot potteto mottlx when lnLerpreLlng an
obllque soluLlon
Ind|cator
Slngle varlable used ln con[uncLlon wlLh one or more oLher varlables Lo form a composlte meosote
atent root
See elqeovoloe
,easure o samp||ng adequacy (,A)
Measure calculaLed boLh for Lhe enLlre correlaLlon maLrlx and each lndlvldual varlable evaluaLlong
Lhe approprlaLeness of applylng facLor analysls values above 030 for elLher Lhe enLlre maLrlx or an
lndlvldual varlable lndlcaLe approprlaLeness
,easurement error
lnaccuracles ln measurlng Lhe Lrue" varlable due Lo Lhe falllblllLy of Lhe measuremenL lnsLrumenL
(le lnapproprlaLe response scales) daLa enLry errors or respondenL errors
,u|t|co|||near|ty
LxLenL Lo whlch a varlable can be explalned by Lhe oLher varlables ln Lhe analysls


b||que actor rotat|on
loctot tototloo ls compuLed so LhaL Lhe exLracLed facLors are correlaLed 8aLher Lhan arblLrarlly
consLralnlng Lhe facLor roLaLlon Lo an ottboqoool soluLlon Lhe obllque roLaLlon ldenLlfles Lhe exLenL
Lo whlch each of Lhe facLors ls correlaLed
rthogona|
MaLhemaLlcal lndependence (no correlaLlon) of facLor axes Lo each oLher (le aL rlghL angles or 90
degrees)
rthogona| actor rotat|on
loctot tototloo ln whlch Lhe facLors are exLracLed so LhaL Lhelr axes are malnLalned aL 90 degrees
Lach facLor ls lndependenL of or ottboqoool Lo all oLher facLors 1he correlaLlon beLween Lhe facLors
ls deLermlned Lo be 0
; actor ana|ys|s
lorms groups or respondenLs or cases based on Lhelr slmllarlLy on a seL of characLerlsLlcs (also see
Lhe dlscusslon of clostet ooolysls ln ChapLer 9)
;DAk@I,Ak
A Lype of ottboqoool foctot tototloo meLhod focuslng on slmpllfylng Lhe rows of a facLor maLrlx
Cenerally consldered less effecLlve Lhan Lhe vA8lMAx roLaLlon
k actor ana|ys|s
Analyzes relaLlonshlps among varlables Lo ldenLlfy groups of varlables formlng laLenL dlmenslons
(facLors)
ke||ab|||ty
LxLenL Lo whlch a varlable or seL of varlables ls conslsLenL ln whaL lL ls lnLended Lo measure lf
mulLlple measuremenLs are Laken rellable measures wlll all be conslsLenL ln Lhelr values lL dlffers
from vollJlty ln LhaL lL does noL relaLe Lo whaL should be measured buL lnsLead Lo how lL ls
measured
keverse scor|ng
rocess of reverlng Lhe scores of a varlable whlle reLalnlng Lhe dlsLrlbuLlonal characLerlsLlcs Lo
change Lhe relaLlonshlps (correlaLlons) beLween Lwo varlables used ln sommoteJ scole consLrucLlon
Lo avold a cancellng ouL beLween varlables wlLh poslLlve and negaLlve foctot looJloqs on Lhe same
facLor
pec||c var|ance
varlance of each varlable unlque Lo LhaL varlable and noL explalned or assoclaLed wlLh oLher varlables
ln Lhe facLor analysls

ummated sca|es
MeLhod of comblnlng several varlables LhaL measure Lhe same concepL lnLo a slngle varlable ln an
aLLempL Lo lncrease Lhe rellablllLy of Lhe measuremenL ln mosL lnsLances Lhe separaLe varlables are
summed and Lhen Lhelr LoLal or average score ls used ln Lhe analysls
urrogate var|ab|e
SelecLlon of a slngle varlable wlLh Lhe hlghesL foctot looJloq Lo represenL a facLor ln Lhe daLa
reducLlon lnsLead of uslng a sommoteJ scole or foctot scote
@race
8epresenLs Lhe LoLal amounL of varlance on whlch Lhe facLor soluLlon ls based 1he Lrace ls equal Lo
Lhe number of varlables based on Lhe assumpLlon LhaL Lhe varlance ln each varlable ls equal Lo 1
Dn|que var|ance
See speclflc votlooce
Ia||d|ty
LxLenL Lo whlch a measure or seL of measures correcLly represenLs Lhe concepL of sLudy Lhe degree
Lo whlch lL ls free from any sysLemaLlc or nonrandom error valldlLy ls concerned wlLh how well Lhe
concepL ls deflned by Lhe measure(s) whereas telloblllty relaLes Lo Lhe conslsLence of Lhe
measure(s)
Iar|ate
Llnear comblnaLlon of varlables formed by derlvlng emplrlcal welghLs applled Lo a seL of varlables
speclfled by Lhe researcher
IAkI,Ak
1he mosL popular ottboqoool foctot tototloo meLhods focuslng on slmpllfylng Lhe columns ln a foctot
mottlx Cenerally consldered superlor Lo oLher orLhogonal facLor roLaLlon meLhods ln achlevlng a
slmpllfled facLor sLrucLure





Cbapter : Simple and multiple regression

Ad[usted coe|c|ent o determ|nat|on (ad[usted )
Modlfled measure of Lhe coefflcleot of Jetetmlootloo LhaL Lakes lnLo accounL Lhe number of
lndependenL varlables lncluded ln Lhe regresslon equaLlon and Lhe sample slze AlLhough Lhe
addlLlon of lndependenL varlables wlll cause Lhe coefflclenL of deLermlnaLlon Lo rlse Lhe ad[usLed
coefflclenL of deLermlnaLlon may fall lf Lhe added lndependenL varlables have llLLle explanaLory
power lf Lhe Jeqtees of fteeJom become Loo small 1hls sLaLlsLlc ls qulLe useful for comparlson
beLween equaLlons wlLh dlfferenL numbers of lndependenL varlables dlfferlng sample slzes or boLh
A||poss|b|e subsets regress|on
MeLhod of selecLlng Lhe varlables for lncluslon ln Lhe regresslon model LhaL conslders all posslble
comblnaLlon of Lhe lndependenL varlables lor example lf Lhe researcher speclfles four poLenLlal
lndependenL varlables Lhls Lechnlque would esLlmaLe all probable regresslon models wlLh one Lwo
Lhree and four varlables 1he Lechnlque would Lhen ldenLlfy Lhe model(s) wlLh Lhe besL predlcLlve
accuracy
8acward e||m|nat|on
MeLhod of selecLlng varlables for lncluslon ln Lhe regresslon model LhaL sLarLs by lncludlng all
lndependenL varlables ln Lhe model and Lhen ellmlnaLlng Lhose varlables noL maklng a slgnlflcanL
conLrlbuLlon Lo predlcLlon
8eta coe|c|ent
SLandardlzed regresslon coefflclenL (see stooJotJlzotloo) LhaL allows for a dlrecL comparlson
beLween coefflclenLs as Lo Lhelr relaLlve explanaLory power of Lhe dependenL varlable Whereas
teqtessloo coefflcleots are expressed ln Lerms of Lhe unlLs of Lhe assoclaLed varlable Lhereby maklng
comparlsons lnapproprlaLe beLa coefflclenLs use sLandardlzed daLa and can be dlrecLly compared
Coe|c|ent o determ|nat|on ()
Measure of Lhe proporLlon of Lhe varlance of Lhe dependenL varlable abouL lLs mean LhaL ls explalned
by Lhe lndependenL or predlcLor varlables 1he coefflclenL van vary beLween 0 and 1 lf Lhe
regresslon model ls properly applled and esLlmaLed Lhe researcher can assume LhaL Lhe hlgher Lhe
value or k Lhe greaLer Lhe explanaLory power of Lhe regresslon equaLlon and Lherefore Lhe beLLer
Lhe predlcLlon of Lhe dependenL varlable


Co|||near|ty
Lxpresslon of Lhe relaLlonshlp beLween Lwo (colllnearlLy) or more (mulLlcolllnearlLy) lndependenL
varlables 1wo lndependenL varlables are sald Lo exhlblL compleLe colllnearlLy lf Lhelr correlaLlon
coefflclenL ls 1 and compleLe lack of colllnearlLy ls Lhelr correlaLlon coefflclenL ls 0 Moltlcollloeotlty
occurs when any slngle lndependenL varlable ls hlghly correlaLed wlLh a seL of oLher lndependenL
varlables An exLreme case of colllnearlLy/mulLlcolllnearlLy ls sloqolotlty ln whlch any dependenL
varlable ls perfecLly predlcLed (le correlaLlon of 10) by anoLher lndependenL varlable (or more Lhan
one)
Corre|at|on coe|c|ent ()
CoefflclenL LhaL lndlcaLes Lhe sLrengLh of Lhe assoclaLlon beLween any Lwo meLrlc varlables 1he slgn
(+ or ) lndlcaLes Lhe dlrecLlon of Lhe relaLlonshlp 1he value can range from +1 Lo 1 wlLh +1
lndlcaLlng a perfecL poslLlve relaLlonshlp 0 lndlcaLlng no relaLlonshlp and 1 lndlcaLlng a perfecL
negaLlve or reverser relaLlonshlp (as one varlable grows larger Lhe oLher varlable grows smaller)
Cr|ter|on var|ab|e ()
See JepeoJeot votloble
egrees o reedom ()
value calculaLed from Lhe LoLal number of observaLlons mlnus Lhe number of esLlmaLed potometets
1hese parameLer esLlmaLes are resLrlcLlons on Lhe daLa because once made Lhey deflne Lhe
populaLlon from whlch Lhe daLa are assumed Lo have been drawn lor example ln esLlmaLlng a
regresslon model wlLh a slngle lndependenL varlable we esLlmaLe Lwo parameLers Lhe lotetcept (b0)
and a teqtessloo coefflcleot for Lhe lndependenL varlable (b1) ln esLlmaLlng Lhe random error
deflned as Lhe sum of Lhe pteJlctloo ettots (acLual mlnus predlcLed dependenL values) for all cases
we would flnd (o 2) degrees of freedom uegrees of freedom provlde a measure of how resLrlcLed
Lhe daLa are Lo reach a cerLaln level of predlcLlon lf Lhe number of degrees of freedom ls small Lhe
resulLlng predlcLlon may be less generallzable because all buL a few observaLlons were lncorporaLed
ln Lhe predlcLlon Conversely a large degreesoffreedom value lndlcaLes Lhe predlcLlon ls falrly
robusL wlLh regard Lo belng represenLaLlve of Lhe overall sample of respondenLs
ependent var|ab|e ()
varlable belng predlcLed or explalned by Lhe seL of lndependenL varlables


ummy var|ab|e
lndependenL varlable used Lo accounL for Lhe effecL LhaL dlfferenL levels of a nonmeLrlc varlable have
ln predlcLlng Lhe dependenL varlable 1o accounL for l levels of a nonmeLrlc lndependenL varlable l
1 dummy varlables are needed lor example gender ls measured as male or female and could be
represenLed by Lwo dummy varlables x1 and x2 When Lhe respondenL ls male x11 and x2 0
Llkewlse when Lhe respondenL ls female x10 and x21 Powever when x11 we know LhaL x2
musL equal 0 1hus we need only one varlable elLher x1 or x2 Lo represenL Lhe varlable gender We
need noL lnclude boLh varlables because one ls perfecLly predlcLed by Lhe oLher (a sloqolotlty) and
Lhe teqtessloo coefflcleots cannoL be esLlmaLed lf a varlable has Lhree levels only Lwo dummy
varlables are needed 1hus Lhe number of dummy varlable ls one less Lhan Lhe number of levels for
Lhe nonmeLrlc varlable 1he Lwo mosL common meLhods of deLermlnlng Lhe values of Lhe dummy
values are loJlcotot coJloq and effects coJloq
ects cod|ng
MeLhod for speclfylng Lhe tefeteoce coteqoty for a seL of Jommy votlobles ln whlch Lhe reference
caLegory recelves a value of mlnus one (1) across Lhe seL of dummy varlables ln our example of
dummy varlable codlng for gender we coded Lhe dummy varlable as elLher 1 or 0 8uL wlLh effecLs
codlng Lhe value ls 1 lnsLead of 0 WlLh Lhls Lype of codlng Lhe coefflclenLs for Lhe dummy varlable
become group devlaLlons on Lhe dependenL varlable from Lhe mean of Lhe dependenL varlable
across all groups LffecLs codlng conLrasLs wlLh loJlcotot coJloq ln whlch Lhe reference caLegory ls
glven Lhe value of zero across all dummy varlables and Lhe coefflclenLs represenL group devlaLlons on
Lhe dependenL varlable from Lhe reference group
Iorward add|t|on
MeLhod of selecLlng varlables for lncluslon ln Lhe regresslon model by sLarLlng wlLh no varlables ln
Lhe model and Lhen addlng one varlable aL a Llme based on lLs conLrlbuLlon Lo predlcLlon
eteroscedast|c|ty
See bomosceJotlclty
omoscedast|c|ty
uescrlpLlon of daLa for whlch Lhe varlance of Lhe error Lerms (e) appears consLanL over Lhe range of
values of an lndependenL varlable 1he assumpLlon of equal varlance of Lhe populaLlon error
(where ls esLlmaLed from Lhe sample value e) ls crlLlcal Lo Lhe proper appllcaLlon of llnear
regresslon When Lhe error Lerms have lncreaslng or modulaLlng varlance Lhe daLa are sald Lo be
betetosceJostlc 1he dlscusslon of teslJools ln Lhls chapLer furLher lllusLraLes Lhls polnL
Independent var|ab|e
varlable(s) selecLed as predlcLors and poLenLlal explanaLory varlables of Lhe dependenL varlable


Ind|cator cod|ng
MeLhod for speclfylng Lhe tefeteoce coteqoty for a seL of Jommy votlobles where Lhe reference
caLegory recelves a value of 0 across Lhe seL of dummy varlables 1he teqtessloo coefflcleots
represenL Lhe group dlfferences ln Lhe dependenL varlable from Lhe reference caLegory lndlcaLor
codlng dlffers from effects coJloq ln whlch Lhe reference caLegory ls glven Lhe value of 1 across all
dummy varlables and Lhe regresslon coefflclenLs represenL group devlaLlons on Lhe dependenL
varlable from Lhe overall mean of Lhe dependenL varlable
In|uent|a| observat|on
An observaLlon LhaL has a dlsproporLlonaLe lnfluence on one or more aspecLs of Lhe regresslon
esLlmaLes 1hls lnfluence may be based on exLreme values of Lhe lndependenL or dependenL
varlables or boLh lnfluenLlal observaLlons can elLher be good" by relnforclng Lhe paLLern of Lhe
remalnlng daLa or bad" when a slngle or small seL of cases unduly affecLs Lhe regresslon esLlmaLes
lL ls noL necessary for Lhe observaLlon Lo be an ootllet alLhough many Llmes ouLllers can be classlfled
as lnfluenLlal observaLlons as well
Intercept ()
value on Lhe axls (dependenL varlable axls) where Lhe llne deflned by Lhe regresslon equaLlon
b0+b1\1 crosses Lhe axls lL ls descrlbed by Lhe consLanL Lerm b0 ln Lhe regresslon equaLlon ln
addlLlon Lo lLs role ln predlcLlon Lhe lnLercepL may have a managerlal lnLerpreLaLlon lf Lhe compleLe
absence of Lhe lndependenL varlable has meanlng Lhen Lhe lnLercepL represenLs LhaL amounL lor
example when esLlmaLlng sales from pasL adverLlslng expendlLures Lhe lnLercepL represenLs Lhe
level of sales expecLed lf adverLlslng ls ellmlnaLed 8uL ln many lnsLances Lhe consLanL has only
predlcLlve value because ln no slLuaLlon are all lndependenL varlables absenL An example ls
predlcLlng producL preference based on consumer aLLlLudes All lndlvlduals have some level of
aLLlLude so Lhe lnLercepL has no managerlal use buL lL sLlll alds ln predlcLlon
east squares
LsLlmaLlon procedure used ln slmple and mulLlple regresslon whereby Lhe teqtessloo coefflcleots are
esLlmaLed so as Lo mlnlmlze Lhe LoLal sum of Lhe squared teslJools
everage po|nts
1ype of lofloeotlol obsetvotloo deflned by one aspecL of lnfluence Lermed levetoqe 1hese
observaLlons are subsLanLlally dlfferenL on one or more lndependenL varlables so LhaL Lhey affecL
Lhe esLlmaLlon of one or more teqtessloo coefflcleots
|near|ty
1erm used Lo express Lhe concepL LhaL Lhe model possesses Lhe properLles of addlLlvlLy and
homogenelLy ln a slmple sense llnear models predlcL values LhaL fall ln a sLralghL llne by havlng a
consLanL unlL change (slope) of Lhe dependenL varlable for a consLanL unlL change of Lhe
loJepeoJeot votloble ln Lhe populaLlon model b0+b1\1+ Lhe effecL of changlng \1 by a value of
10 ls Lo add b1 (a consLanL) unlLs of

,easurement error
uegree Lo whlch Lhe daLa values do noL Lruly measure Lhe characLerlsLlc belng represenLed by Lhe
varlable lor example when asklng abouL LoLal famlly lncome many sources of measuremenL error
(eg relucLance Lo answer full amounL error ln esLlmaLlng LoLal lncome) make Lhe daLa values
lmpreclse
,oderator eect
LffecL ln whlch a Lhlrd lndependenL varlable (Lhe moderaLor varlable) causes Lhe relaLlonshlp
beLween a dependenL/lndependenL varlable palr Lo change dependlng on Lhe value of Lhe
moderaLor varlable lL ls also known as an lnLeracLlve effecL and ls slmllar Lo Lhe lnLeracLlve effecL
seen ln analysls of varlance meLhods
,u|t|co|||near|ty
See collloeotlty
,u|t|p|e regress|on
8egresslon model wlLh Lwo or more lndependenL varlables
-orma| probab|||ty p|ot
Craphlcal comparlson of Lhe shape of Lhe sample dlsLrlbuLlon Lo Lhe normal dlsLrlbuLlon ln Lhe
graph Lhe normal dlsLrlbuLlon ls represenLed by a sLralghL llne angled aL 43 degrees 1he acLual
dlsLrlbuLlon ls ploLLed agalnsL Lhls llne so any dlfferences are shown as devlaLlons from Lhe sLralghL
llne maklng ldenLlflcaLlon of dlfferences qulLe slmple
-u|| p|ot
loL of teslJools versus Lhe predlcLed values LhaL exhlblLs a random paLLern A null ploL ls lndlcaLlve of
no ldenLlflable vlolaLlons of Lhe assumpLlons underlylng regresslon analysls
ut||er
ln sLrlcL Lerms an observaLlon LhaL has a subsLanLlal dlfference beLween Lhe acLual value for Lhe
dependenL varlable and Lhe predlcLed value Cases LhaL are subsLanLlally dlfferenL wlLh regard Lo
elLher Lhe dependenL or lndependenL varlables are ofLen Lermed ootllets as well ln all lnsLances Lhe
ob[ecLlve ls Lo ldenLlfy observaLlons LhaL are lnapproprlaLe represenLaLlons of Lhe populaLlon from
whlch Lhe sample ls drawn so LhaL Lhey may be dlscounLed or even ellmlnaLed from Lhe analysls as
unrepresenLaLlve
9arameter
CuallLy (measure) characLerlsLlc of Lhe populaLlon lor example and r are Lhe symbols used for
Lhe populaLlon parameLers mean () and varlance (r) Lhey are Lyplcally esLlmaLed from sample daLa
ln whlch Lhe arlLhmeLlc average of Lhe sample ls used as a measure of Lhe populaLlon average and
Lhe varlance of Lhe sample ls used Lo esLlmaLe Lhe varlance of Lhe populaLlon


9art corre|at|on
value LhaL measures Lhe sLrengLh of Lhe relaLlonshlp beLween a dependenL and a slngle lndependenL
varlable when Lhe predlcLlve effecLs of Lhe oLher lndependenL varlables ln Lhe regresslon model are
removed 1he ob[ecLlve ls Lo porLray Lhe unlque predlcLlve effecL due Lo cottelotloo coefflcleot whlch
ls concerned wlLh lncremenLal predlcLlve effecL
9art|a| corre|at|on coe|c|ent
value LhaL measures Lhe sLrengLh of Lhe relaLlonshlp beLween Lhe crlLerlon or dependenL and a slngle
lndependenL varlable when Lhe effecLs of Lhe oLher lndependenL varlables ln Lhe model are held
consLanL lor example t \2 \1 measures Lhe varlaLlon ln assoclaLed wlLh \2 when Lhe effecL of
\1 on boLh \2 and ls held consLanL 1hls value ls used ln sequenLlal varlable selecLlon meLhods of
regresslon model esLlmaLlon (eg stepwlse fotwotJ oJJltloo or bockwotJ ellmlootloo) Lo ldenLlfy
Lhe lndependenL varlable wlLh Lhe greaLesL lncremenLal predlcLlve power beyond Lhe lndependenL
varlables already ln Lhe model
9art|a| (or ) va|ues
1he parLlal lLesL ls slmply a sLaLlsLlcal LesL for Lhe addlLlonal conLrlbuLlon Lo predlcLlon accuracy of a
varlable above LhaL of Lhe varlables already ln Lhe equaLlon When a varlable (\o) ls added Lo a
regresslon equaLlon afLer oLher varlables are already ln Lhe equaLlon lLs conLrlbuLlon may be small
even Lhough lL has a hlgh correlaLlon wlLh Lhe dependenL varlable 1he reason ls LhaL \o ls hlghly
correlaLed wlLh Lhe varlables already ln Lhe equaLlon 1he parLlal l value ls calculaLed for all varlables
by slmply preLendlng LhaL each ln Lurn ls Lhe lasL Lo enLer Lhe equaLlon lL glves Lhe addlLlonal
conLrlbuLlon of each varlable above all oLhers ln Lhe equaLlon A low or lnslgnlflcanL parLlal l value for
a varlable noL ln Lhe equaLlon lndlcaLes lLs low or lnslgnlflcanL conLrlbuLlon Lo Lhe model as already
speclfled A t value may be calculaLed lnsLead of l values ln all lnsLances wlLh Lhe t value belng
approxlmaLely Lhe square rooL of Lhe l value
9art|a| regress|on p|ot
Craphlcal represenLaLlon of Lhe relaLlonshlp beLween Lhe dependenL varlable and a slngle
lndependenL varlable 1he scaLLerploL of polnLs deplcLs Lhe parLlal correlaLlon beLween Lhe Lwo
varlables wlLh Lhe effecLs of oLher lndependenL varlables held consLanL (see pottlol cottelotloo
coefflcleot) 1hls porLrayal ls parLlcularly helpful ln assesslng Lhe form of Lhe relaLlonshlp (llnear
versus nonllnear) and Lhe ldenLlflcaLlon of lofloeotlol obsetvotloos
9o|ynom|a|
1ransformaLlon of an lndependenL varlable Lo represenL a curvlllnear relaLlonshlp wlLh Lhe
dependenL varlable 8y lncludlng a squared Lerm (x) a slngle lnflecLlon polnL ls esLlmaLed A cublc
Lerm esLlmaLes a second lnflecLlon polnL AddlLlonal Lerms of a hlgher powet can also be esLlmaLed
9ower
robablllLy LhaL a slgnlflcanL relaLlonshlp wlll be found lf lL acLually exlsLs ComplemenLs wlLh Lhe
more wldely used slqolflcooce level olpbo ()


9red|ct|on error
ulfference beLween Lhe acLual and predlcLed values of Lhe dependenL varlable for each observaLlon
ln Lhe sample (see teslJool)
9red|ctor var|ab|e (O)
See loJepeoJeot votloble
9k stat|st|c
valldaLlon measure obLalned by ellmlnaLlng each observaLlon one aL a Llme and predlcLlng Lhls
dependenL value wlLh Lhe regresslon model esLlmaLed from Lhe remalnlng observaLlons
keerence category
1he omlLLed level of a nonmeLrlc varlable when a Jommy votloble ls formed from Lhe nonmeLrlc
varlable
kegress|on coe|c|ent ()
-umerlcal value of Lhe parameLer esLlmaLe dlrecLly assoclaLed wlLh an lndependenL varlable for
example ln Lhe model b0+b1\1 Lhe value b1 ls Lhe regresslon coefflclenL for Lhe varlable \1 1he
regresslon coefflclenL represenLs Lhe amounL of change ln Lhe dependenL varlable for a oneunlL
change ln Lhe lndependenL varlable ln Lhe mulLlple predlcLor mode (eg b0+b1\1+b2\2) Lhe
regresslon coefflclenLs are parLlal coefflclenLs because each Lakes lnLo accounL noL only Lhe
relaLlonshlps beLween and \1 and beLween and \2 buL also beLween \1 and \2 1he coefflclenL ls
noL llmlLed ln range because lL ls based on boLh Lhe degree of assoclaLlon and Lhe scale unlLs of Lhe
lndependenL varlable lor lnsLance Lwo varlables wlLh Lhe same assoclaLlon Lo ? would have
dlfferenL coefflclenLs lf one lndependenL varlable was measured on a 7polnL scale and anoLher was
based on a 100polnL scale
kegress|on var|ate
Llnear comblnaLlon of welghLed lndependenL varlables used collecLlvely Lo predlcL Lhe dependenL
varlable
kes|dua| ( or )
Lrror ln predlcLlng our sample daLa Seldom wlll our predlcLlons be perfecL We assume LhaL random
error wlll occur buL we assume LhaL Lhls error ls an esLlmaLe of Lhe random error ln Lhe populaLlon
() noL [usL Lhe error ln predlcLlon for our sample (e) We assume LhaL Lhe error ln Lhe populaLlon we
are esLlmaLlng ls dlsLrlbuLed wlLh a mean of 0 and a consLanL (bomosceJostlc) varlance
amp||ng error
1he expecLed varlaLlon ln any esLlmaLed parameLer (lotetcept or teqtessloo coefflcleot) LhaL ls due Lo
Lhe use of a sample raLher Lhan Lhe populaLlon Sampllng error ls reduced as Lhe sample slze ls
lncreased and ls used Lo sLaLlsLlcally LesL wheLher Lhe esLlmaLed parameLer dlffers from zero


|gn||cance |eve| (a|pha)
Commonly referred Lo as Lhe level of sLaLlsLlcal slgnlflcance Lhe slgnlflcance level represenLs Lhe
probablllLy Lhe researcher ls wllllng Lo accepL LhaL Lhe esLlmaLed coefflclenL ls classlfled as dlfferenL
from zero when lL acLually ls noL 1hls ls also known as 1ype l error 1he mosL wldely used level of
slgnlflcance ls 003 alLhough researchers use levels ranglng from 001 (more demandlng) Lo 010 (less
conservaLlve and easler Lo flnd slgnlflcance
|mp|e regress|on
8egresslon model wlLh a slngle lndependenL varlable also known as blvarlaLe regresslon
|ngu|ar|ty
1he exLreme case of collloeotlty or moltlcollloeotlty ln whlch an lndependenL varlable ls perfecLly
predlcLed (a correlaLlon of +/ 10) by one or more lndependenL varlables 8egresslon models cannoL
be esLlmaLed when a slngularlLy exlsLs 1he researcher musL omlL one or more of Lhe lndependenL
varlables lnvolved Lo remove Lhe slngularlLy
pec||cat|on error
Lrror ln predlcLlng Lhe dependenL varlable caused by excludlng one or more relevanL lndependenL
varlables 1hls omlsslon can blas Lhe esLlmaLed coefflclenLs of Lhe lncluded varlables as well as
decrease Lhe overall predlcLlve power of Lhe regresslon model
tandard error
LxpecLed dlsLrlbuLlon of an esLlmaLed regresslon coefflclenL 1he sLandard error ls slmllar Lo Lhe
sLandard devlaLlon of any seL of daLa values buL lnsLead denoLes Lhe expecLed range of Lhe
coefflclenL across mulLlple samples of Lhe daLa lL ls useful ln sLaLlsLlcal LesLs of slgnlflcance LhaL LesL
Lo see wheLher Lhe coefflclenL ls slgnlflcanLly dlfferenL from zero (le wheLher Lhe expecLed range of
Lhe coefflclenL conLalns Lhe value of zero aL a glven level of confldence) 1he t value of a teqtessloo
coefflcleot ls Lhe coefflclenL dlvlded by lLs sLandard error
tandard error o the est|mate (e)
Measure of Lhe varlaLlon ln Lhe predlcLed values LhaL can be used Lo develop confldence lnLervals
around any predlcLed value lL ls slmllar Lo Lhe sLandard devlaLlon of a varlable around lLs mean buL
lnsLead ls Lhe expecLed dlsLrlbuLlon of predlcLed values LhaL would occur lf mulLlple samples of Lhe
daLa were Laken
tandard|zat|on
rocess whereby Lhe orlglnal varlable ls Lransformed lnLo a new varlable wlLh a mean of 0 and a
sLandard devlaLlon of 1 1he Lyplcal procedure ls Lo flrsL subLracL Lhe varlable mean from each
observaLlon's value and Lhen dlvlde by Lhe sLandard devlaLlon When all Lhe varlables ln a teqtessloo
votlote are sLandardlzed Lhe b0 Lerm (Lhe lotetcept) assumes a value of 0 and Lhe teqtessloo
coefflcleots are known as beto coefflcleots whlch enable Lhe researcher Lo compare dlrecLly Lhe
relaLlve effecL of each lndependenL varlable on Lhe dependenL varlable


tat|st|ca| re|at|onsh|p
8elaLlonshlp based on Lhe correlaLlon of one or more lndependenL varlables wlLh Lhe dependenL
varlable Measures of assoclaLlon Lyplcally correlaLlons represenL Lhe degree of relaLlonshlp because
Lhere ls more Lhan one value of Lhe dependenL varlable for each value of Lhe lndependenL varlable
tepw|se est|mat|on
MeLhod of selecLlng varlables for lncluslon ln Lhe regresslon model LhaL sLarLs by selecLlng Lhe besL
predlcLor of Lhe dependenL varlable AddlLlonal lndependenL varlables are selecLed ln Lerms of Lhe
lncremenLal explanaLory power Lhey can add Lo Lhe regresslon model lndependenL varlables are
added as long as Lhelr parLlal cottelotloo coefflcleots are sLaLlsLlcally slgnlflcanL lndependenL
varlables may also be dropped lf Lhelr predlcLlve power drops Lo a nonslgnlflcanL level when anoLher
lndependenL varlable ls added Lo Lhe model
tudent|zed res|dua|
1he mosL commonly used form of sLandardlzed teslJool lL dlffers from oLher meLhods ln how lL
calculaLes Lhe sLandard devlaLlon used ln stooJotJlzotloo 1o mlnlmlze Lhe effecL of any observaLlon
on Lhe sLandardlzaLlon process Lhe sLandard devlaLlon of Lhe resldual for observaLlon l ls compuLed
from regresslon esLlmaLes omlLLlng Lhe lLh observaLlon ln Lhe calculaLlon of Lhe regresslon esLlmaLes
um o squared errors (e)
Sum of Lhe squared predlcLlon errors (teslJools) across all observaLlons lL ls used Lo denoLe Lhe
varlance ln Lhe dependenL varlable noL yeL accounLed for by Lhe regresslon model lf no lndependenL
varlables are used for predlcLlon lL becomes Lhe squared errors uslng Lhe mean as Lhe predlcLed
value and Lhus equal Lhe totol som of spootes
um o squares regress|on (r)
Sum of Lhe squared dlfferences beLween Lhe mean and predlcLed values of Lhe dependenL varlable
for all observaLlon lL represenLs Lhe amounL of lmprovemenL ln explanaLlon of Lhe dependenL
varlable aLLrlbuLable Lo Lhe lndependenL varlable(s)
uppress|on eect
1he lnsLance ln whlch Lhe expecLed relaLlonshlps beLween lndependenL and dependenL varlables are
hldden or suppressed when vlewed ln a blvarlaLe relaLlonshlp When addlLlonal lndependenL
varlables are enLered Lhe moltlcollloeotlty removes unwanLed" shared varlance and reveals Lhe
Lrue" relaLlonshlp
@o|erance
Commonly used measure of collloeotlty and moltlcollloeotlty 1he Lolerance of varlable l (1CLl) ls
1 k*l where k*l ls Lhe coefflclenL of deLermlnaLlon for Lhe predlcLlon of varlable l by Lhe oLher
lndependenL varlables ln Lhe teqtessloo votlote As Lhe Lolerance value grows smaller Lhe varlable ls
more hlghly predlcLed by Lhe oLher lndependenL varlables (colllnearlLy)


@ota| sum o squares (r)
1oLal amounL of varlaLlon LhaL exlsLs Lo be explalned by Lhe lndependenL varlables 1hls basellne
value ls calculaLed by summlng Lhe squared dlfferences beLween Lhe mean and acLual values for Lhe
dependenL varlable across all observaLlons
@ransormat|on
A varlable may have an undeslrable characLerlsLlc such as nonnormallLy LhaL deLracLs from Lhe
ablllLy of Lhe cottelotloo coefflcleot Lo represenL Lhe relaLlonshlp beLween lL and anoLher varlable A
LransformaLlon such as Laklng Lhe logarlLhm or square rooL of Lhe varlable creaLes a new varlable
and ellmlnaLes Lhe undeslrable characLerlsLlc allowlng for a beLLer measure of Lhe relaLlonshlp
1ransformaLlons may be applled Lo elLher Lhe dependenL or lndependenL varlables or boLh 1he
need and speclflc Lype of LransformaLlon may be based on LheoreLlcal reasons (such as Lransformlng
a known llnear relaLlonshlp) or emplrlcal reasons (ldenLlfled Lhrough graphlcal or sLaLlsLlcal means)
Iar|ance |n|at|on actor (III)
lndlcaLor of Lhe effecL LhaL Lhe oLher lndependenL varlables have on Lhe sLandard error of a
teqtessloo coefflcleot 1he varlance lnflaLlon facLor ls dlrecLly relaLed Lo Lhe toletooce value
(vlll 1/1CLl) Large vll values also lndlcaLe a hlgh degree of collloeotlty or moltlcollloeotlty among
Lhe lndependenL varlables



Cbapter : Con|oint analysis

Adapt|ve con[o|nt method
MeLhodology for conducLlng a con[olnL analysls LhaL relles on respondenLs provldlng addlLlonal
lnformaLlon noL ln Lhe acLual coojolot tosk (eg lmporLance of aLLrlbuLes) 1hls lnformaLlon ls Lhen
used Lo adapL and slmpllfy Lhe coojolot tosk
Add|t|ve mode|
1echnlque for slmpllfylng con[olnL analysls by comblnlng Lhe selfexpllcaLed model and LradlLlonal
con[olnL analysls
8a|anced des|gn
roflle Jeslqo ln whlch each level wlLhln a foctot appears an equal number of Llmes across Lhe
proflles of Lhe coojolot tosk
8r|dg|ng des|gn
roflle Jeslqo for a large number of foctots (aLLrlbuLes) ln whlch Lhe aLLrlbuLes are broken lnLo a
number of smaller groups Lach aLLrlbuLe group has some aLLrlbuLes conLalned ln oLher groups
enabllng Lhe resulLs from each group Lo be comblned or brldged
Cho|ce s|mu|ator
rocedure LhaL enables Lhe researcher Lo assess many whaLlf" scenarlos Cnce Lhe con[olnL pott
wottbs have been esLlmaLed for each respondenL Lhe cholce slmulaLor analyzes a seL of ptoflles and
predlcLs boLh lndlvldual and aggregaLe cholces for each proflle ln Lhe seL MulLlple seLs of proflles can
be analyzed Lo represenL any scenarlo (eg preferences for hypoLheLlcal producL or servlce
conflguraLlons or Lhe compeLlLlve lnLeracLlons among proflles assumed Lo consLlLuLe a markeL)
Compos|t|on ru|e
8ule used Lo represenL how respondenLs comblne aLLrlbuLes Lo produce a [udgmenL of relaLlve value
or otlllty for a producL or servlce lor lllusLraLlon leL us suppose a person ls asked Lo evaluaLe four
ob[ecLs 1he person ls assumed Lo evaluaLe Lhe aLLrlbuLes of Lhe four ob[ecLs and Lo creaLe some
overall relaLlve value for each 1he rule may be as slmple as creaLlng a menLal welghL for each
percelved aLLrlbuLe and addlng Lhe welghLs for an overall score (oJJltlve moJel) or lL may be a more
complex procedure lnvolvlng lotetoctloo effects
Compos|t|ona| mode|
Class of mulLlvarlaLe models LhaL esLlmaLes Lhe dependence relaLlonshlp based on respondenL
observaLlons regardlng boLh Lhe dependenL and Lhe lndependenL varlables Such models calculaLe or
compose" Lhe dependenL varlable from Lhe respondenLsupplled values for all of Lhe lndependenL
varlables rlnclpal among such meLhods are regresslon analysls and dlscrlmlnanL analysls 1hese
models are ln dlrecL conLrasL Lo Jecomposltloool moJels

Con[o|nt tas
1he procedure for gaLherlng [udgmenLs on each proflle ln Lhe con[olnL Jeslqo uslng one of Lhe Lhree
Lypes of presenLaLlon meLhod (le follptoflle poltwlse compotlsoo or ttoJeoff)
Con[o|nt var|ate
ComblnaLlon of lndependenL varlables (known as foctots) speclfled by Lhe researcher LhaL consLlLuLe
Lhe LoLal worLh or otlllty of Lhe proflle
ecompos|t|ona| mode|
Class of mulLlvarlaLe models LhaL decompose Lhe lndlvldual's responses Lo esLlmaLe Lhe dependence
relaLlonshlp 1hls class of models presenLs Lhe respondenL wlLh a predeflned seL of ob[ecLs (eg
hypoLheLlcal or acLual producL or servlce) and Lhen asks for an overall evaluaLlon or preference of Lhe
ob[ecL Cnce glven Lhe evaluaLlon/preference ls decomposed by relaLlng Lhe known aLLrlbuLes of Lhe
ob[ecL (whlch become Lhe lndependenL varlables) Lo Lhe evaluaLlon (dependenL varlable) rlnclpal
among such models ls con[olnL analysls and some forms of mulLldlmenslonal scallng (see ChapLer 10)
es|gn
Speclflc seL of con[olnL ptoflles creaLed Lo exhlblL Lhe sLaLlsLlcal properLles of ottboqooollty and
bolooce
es|gn e|c|ency
uegree Lo whlch a Jeslqo maLches an ottboqoool deslgn 1hls measure ls prlmarlly used Lo evaluaLe
and compare oeotly ottboqoool deslgns ueslgn efflclency values range from 0 Lo 100 whlch denoLes
an optlmol Jeslqo
nv|ronmenta| corre|at|on
See lotetotttlbote cottelotloo
Iactor
lndependenL varlable Lhe researcher manlpulaLes LhaL represenLs a speclflc aLLrlbuLe ln con[olnL
analysls Lhe facLors are nonmeLrlc lacLors musL be represenLed by Lwo or more values (known as
levels) whlch are also speclfled by Lhe researcher
Iactor|a| des|gn
MeLhod of deslgnlng ptoflles by generaLlng all posslble comblnaLlons of levels lor example a Lhree
facLor con[olnL analysls wlLh Lhree levels per facLor (3x3x3) would resulL ln 27 comblnaLlons LhaL
would acL as proflles ln Lhe coojolot tosk


Iract|ona| actor|a| des|gn
MeLhod of deslgnlng proflles (le an alLernaLlve Lo foctotlol Jeslqo) LhaL uses only a subseL of Lhe
posslble proflles needed Lo esLlmaLe Lhe resulLs based on Lhe assumed composlLlon rule lLs prlmary
ob[ecLlve ls Lo reduce Lhe number of evaluaLlons collecLed whlle sLlll malnLalnlng ottboqooollty
among Lhe levels and subsequenL pottwottb esLlmaLes lL achleves Lhls ob[ecLlve by deslgnlng
proflles LhaL can esLlmaLe only a subseL of Lhe LoLal posslble effecLs 1he slmplesL deslgn ls an
oJJltlve moJel ln whlch only molo effects are esLlmaLed lf selecLed lotetoctloo tetms are lncluded
Lhen addlLlonal proflles are creaLed 1he deslgn can be creaLed elLher by referrlng Lo publlshed
sources or by uslng compuLer programs LhaL accompany more con[olnL analysls packages
Iu||pro||e method
MeLhod of gaLherlng respondenL evaluaLlons by presenLlng ptoflles LhaL are descrlbed ln Lerms of all
foctots lor example leL us assume LhaL a candy was descrlbed by Lhree facLors wlLh Lwo levels each
prlce (13 cenLs or 23 cenLs) flavor (clLrus or buLLerscoLch) and color (whlLe or red) A full proflle
would be deflned by one level of each facLor Cne such proflle would be a red buLLerscoLch candy
cosLlng 13 cenLs
o|dout pro||es
See vollJotloo ptoflles
Interact|on eects
LffecLs of a comblnaLlon of relaLed feaLures (lndependenL varlables) also known as lotetoctloo tetms
ln assesslng value a person may asslgn a unlque value Lo speclflc comblnaLlons of feaLures LhaL runs
counLer Lo Lhe addlLlve composltloo tole lor example leL us assume a person ls evaluaLlng
mouLhwash producLs descrlbed by Lhe Lwo facLors (aLLrlbuLes) of color and brand LeL us furLher
assume LhaL Lhls person has an average preference for Lhe aLLrlbuLes red and brand x when
consldered separaLely 1hus when Lhls speclflc comblnaLlon of levels (red and brand x) ls evaluaLed
wlLh Lhe addlLlve composlLlon rule Lhe red brand x producL would have an expecLed overall
preference raLlng somewhere ln Lhe mlddle of all posslble proflles even above oLher comblnaLlons of
aLLrlbuLes (color and brand) LhaL had hlgher evaluaLlons of Lhe lndlvldual feaLures Lhen an
lnLeracLlon ls found Lo exlsL 1hls unlque evaluaLlon of a comblnaLlon LhaL ls greaLer (or could be less)
Lhan expecLed based on Lhe separaLe [udgmenLs lndlcaLes a Lwoway lnLeracLlon Plgherorder
(Lhreeway ore more) lnLeracLlons can occur among more comblnaLlons of levels


Interattr|bute corre|at|on
Also known as eovltoomeotol cottelotloo lL ls Lhe correlaLlon among aLLrlbuLes LhaL makes
comblnaLlons of aLLrlbuLes unbellevable or redundanL A negaLlve correlaLlon deplcLs Lhe slLuaLlon ln
whlch Lwo aLLrlbuLes are naLurally assumed Lo operaLe ln dlfferenL dlrecLlons such as horsepower
and gas mlleage As one lncreases Lhe oLher ls naLurally assumed Lo decrease 1hus because of Lhls
correlaLlon all comblnaLlons of Lhese Lwo aLLrlbuLes (eg hlgh gas mlleage and hlgh horsepower) are
noL bellevable 1he same effecLs can be seen for poslLlve correlaLlons where perhaps prlce and
quallLy are assumed Lo be poslLlvely correlaLed lL may noL be bellevable Lo flnd a hlghprlce low
quallLy producL ln such a slLuaLlon 1he presence of sLrong lnLeraLLrlbuLe correlaLlons requlres LhaL
Lhe researcher closely examlne Lhe proflles presenLed Lo respondenLs and avold unbellevable
comblnaLlons LhaL are noL useful ln esLlmaLlng Lhe pottwottbs
eve|
Speclflc nonmeLrlc value descrlblng a foctot Lach facLor musL be represenLed by Lwo or more levels
buL Lhe number of levels Lyplcally never exceeds four or flve lf Lhe facLor ls orlglnally meLrlc lL musL
be reduced Lo a small number of nonmeLrlc levels lor example Lhe many posslble values of slze and
prlce may be represenLed by a small number of levels slze (10 12 or 16 ounces) or prlce (t119
t139 or t199) lf Lhe facLor ls nonmeLrlc Lhe orlglnal values can be used ln Lhese examples color
(red or blue) brand (x ? or Z) or fabrlc sofLener addlLlve (presenL or absenL)
,a|n eects
ulrecL effecL of each foctot (lndependenL varlable) on Lhe dependenL varlable May be
complemenLed by lotetoctloo effectsln speclflc slLuaLlons
,onoton|c re|at|onsh|p
1he assumpLlon by Lhe researcher LhaL a preference order among levels should apply Lo Lhe pott
wottb esLlmaLes Lxamples may lnclude ob[ecLlve facLors (closer dlsLance preferred over farLher
dlsLance Lraveled) or more sub[ecLlve facLors (more quallLy preferred over lower quallLy) 1he
lmpllcaLlon ls LhaL Lhe esLlmaLed parLworLhs should have some orderlng ln Lhe values and vlolaLlons
(known as tevetsols) should be addressed
-ear|y orthogona|
CharacLerlsLlc of a proflles deslgn LhaL ls noL ottboqoool buL Lhe devlaLlons from orLhogonallLy are
sllghL and carefully conLrolled ln Lhe generaLlon of Lhe proflles 1hls Lype of deslgn can be compared
wlLh oLher proflles deslgns wlLh measures of Jeslqo efflcleocy
pt|ma| des|gn
roflles deslgn LhaL ls ottboqoool and bolooceJ


rthogona||ty
MaLhemaLlcal consLralnL requlrlng LhaL Lhe pottwottb esLlmaLes be lndependenL of each oLher ln a
con[olnL analysls ottboqooollty refers Lo Lhe ablllLy Lo measure Lhe effecL of changlng each aLLrlbuLe
level and Lo separaLe lL from Lhe effecLs of changlng oLher aLLrlbuLe levels and from experlmenLal
error
9a|rw|se compar|son method
MeLhod of presenLlng a palr of ptoflles Lo a respondenL for evaluaLlon wlLh Lhe respondenL selecLlng
one proflle as preferred
9artworth
LsLlmaLe from con[olnL analysls of Lhe overall preference or otlllty assoclaLed wlLh each level or each
foctot used Lo deflne Lhe producL or servlce
9reerence structure
8epresenLaLlon of boLh Lhe relaLlve lmporLance or worLh of each foctot and Lhe lmpacL of lndlvldual
levels ln affecLlng otlllty
9ro||e
8y Laklng one level from each foctot Lhe researcher creaLes a speclflc ob[ecL" (also known as a
tteotmeot) LhaL can be evaluaLed by respondenLs lor example lf a sofL drlnk was belng deflned by
Lhree facLors each wlLh Lwo levels (dleL versus regular cola versus noncola and caffelnefree or
noL) Lhen a proflle would be a caffelnefree dleL cola a regular caffelnefree cola or a dleL caffelne
free noncola 1here can be as many proflles as Lhere are unlque comblnaLlons of levels Cne meLhod
of deflnlng proflles ls Lhe foctotlol Jeslqo whlch creaLes separaLe proflles for each comblnaLlon of all
levels lor example Lhree facLors wlLh Lwo levels each would creaLe elghL (2x2x2) proflles Powever
ln many con[olnL analyses Lhe LoLal number of comblnaLlons ls Loo large for a respondenL Lo
evaluaLe Lhem all ln Lhese lnsLances some subseLs of proflles are creaLed accordlng Lo a sysLemaLlc
plan mosL ofLen a ftoctloool foctotlol Jeslqo
9roh|b|ted pa|r
A speclflc comblnaLlon of levels from Lwo foctots LhaL ls prohlblLed from occurrlng ln Lhe creaLlon of
proflles 1he mosL common cause ls lotetotttlbote cottelotloo among Lhe facLors
kespondent heterogene|ty
1he varlaLlon ln pottwottbs across unlque lndlvlduals found ln dlsaggregaLe models When aggregaLe
models are esLlmaLed modlflcaLlons ln Lhe esLlmaLlon process can approxlmaLe Lhls expecLed
varlaLlon ln pottwottbs


keversa|
A vlolaLlon of a moootoolc telotloosblp where Lhe esLlmaLed pottwottb for a level ls greaLer/lower
Lhan lL should be ln relaLlon Lo anoLher level lor example ln dlsLance Lraveled Lo a sLore closer
sLores would always be expecLed Lo have more uLlllLy Lhan Lhose farLher away A reversal would be
when a farLher dlsLance has a larger parLworLh Lhan a closer dlsLance
t|mu|us
See ptoflle
@radeo method
MeLhod of presenLlng proflles Lo respondenLs ln whlch foctots (aLLrlbuLes) are deplcLed Lwo aL a Llme
and respondenLs rank all comblnaLlons of Lhe levels ln Lerms of preference
@rad|t|ona| con[o|nt ana|ys|s
MeLhodology LhaL employs Lhe classlc prlnclples of con[olnL analysls ln Lhe coojolot tosk uslng an
oJJltlve moJel of consumer preference and poltwlse compotlsoo or follptoflle metboJs of
presenLaLlon
Dt|||ty
An lndlvldual's sub[ecLlve preference [udgmenL represenLlng Lhe hollsLlc value or worLh of a speclflc
ob[ecL ln con[olnL analysls uLlllLy ls assumed Lo be formed by Lhe comblnaLlon of pottwottb
esLlmaLes for any speclflc seL of levels wlLh Lhe use of an oJJltlve moJel perhaps ln con[uncLlon wlLh
lotetoctloo effects
Ia||dat|on pro||es
SeL of ptoflles LhaL are noL used ln Lhe esLlmaLlon of pottwottbs LsLlmaLed pottwottbs are Lhen
used Lo predlcL preference for Lhe valldaLlon proflles Lo assess valldlLy and rellablllLy of Lhe orlglnal
esLlmaLes Slmllar ln concepL Lo Lhe valldaLlon sample of respondenLs ln dlscrlmlnanL analysls


Cbapter : Multiple discriminant analysis and logistic regression

Categor|ca| var|ab|e
See ooomettlc votloble
Centro|d
Mean value for Lhe Jlsctlmlooot 2 scotes of all ob[ecLs wlLhln a parLlcular caLegory or group lor
example a Lwogroup dlscrlmlnanL analysls has Lwo cenLrolds one for Lhe ob[ecLs ln each of Lhe Lwo
groups
|scr|m|nant unct|on
A varlaLe of Lhe lndependenL varlables selecLed for Lhelr dlscrlmlnaLory power used ln Lhe predlcLlon
of group membershlp 1he predlcLed value of Lhe dlscrlmlnanL funcLlon ls Lhe Jlsctlmlooot 2 scote
whlch ls calculaLed for each ob[ecL (person flrm or producL) ln Lhe analysls
|scr|m|nant |oad|ngs
MeasuremenL of Lhe slmple llnear correlaLlon beLween each lndependenL varlable across Lhe groups
of Lhe dependenL varlable lndependenL varlables wlLh large dlscrlmlnaLory power usually have large
welghLs and Lhose wlLh llLLle dlscrlmlnaLory power usually have small welghLs Powever
mulLlcolllnearlLy among Lhe lndependenL varlables wlll cause excepLlons Lo Lhls rule Also called Lhe
Jlsctlmlooot coefflcleot
xponent|ated |og|st|c coe|c|ent
AnLllog of Lhe loqlstlc coefflcleot whlch ls used for lnLerpreLaLlon purposes ln loglsLlc regresslon 1he
exponenLlaLed coefflclenL mlnus 10 equals Lhe percenLage change ln Lhe oJJs lor example an
exponenLlaLed coefflclenL of 020 represenLs a negaLlve 80 percenL change ln Lhe odds (02010
080) for each unlL change ln Lhe lndependenL varlable (Lhe same as lf Lhe odds were mulLlplled by
020) 1hus a value of 10 equaLes Lo no change ln Lhe odds and values above 10 represenL lncreases
ln Lhe predlcLed odds
|e||hood va|ue
Measure used ln loqlstlc teqtessloo Lo represenL Lhe lack of predlcLlve flL Lven Lhough Lhls meLhod
does noL use Lhe leasL squares procedure ln model esLlmaLlon as ls done ln mulLlple regresslon Lhe
llkellhood value ls slmllar Lo Lhe sum of squared error ln regresslon analysls
og|st|c coe|c|ent
CoefflclenL ln Lhe loqlstlc teqtessloo model LhaL acLs as Lhe welghLlng facLor for Lhe lndependenL
varlables ln relaLlon Lo Lhelr dlscrlmlnaLory power Slmllar Lo a regresslon welghL or Jlsctlmlooot
coefflcleot


og|st|c curve
An Sshaped curve formed by Lhe loqlt ttoosfotmotloo LhaL represenLs Lhe probablllLy of an evenL
1he Sshaped form ls nonllnear because Lhe probablllLy of an evenL musL approach 0 and 1 buL never
fall ouLslde Lhese llmlLs 1hus alLhough Lhe mldrange lnvolves a llnear componenL Lhe probablllLles
as Lhey approach Lhe lower and upper bounds of probablllLy (0 and 1) musL flaLLen ouL and become
asympLoLlc Lo Lhese bounds
og|st|c regress|on
Speclal form of regresslon ln whlch Lhe dependenL varlable ls a nonmeLrlc dlchoLomous (blnary)
varlable AlLhough some dlfferences exlsL Lhe general manner of lnLerpreLaLlon ls qulLe slmllar Lo
llnear regresslon
og|t ana|ys|s
See loqlstlc teqtessloo
og|t transormat|on
1ransformaLlon of Lhe values of Lhe dlscreLe blnary dependenL varlable of loqlstlc teqtessloo lnLo an
Sshaped curve (loqlstlc cotve) represenLlng Lhe probablllLy of an evenL 1hls probablllLy ls Lhen used
Lo form Lhe oJJs totlo whlch acLs as Lhe dependenL varlable ln loglsLlc regresslon
,etr|c var|ab|e
varlable wlLh a consLanL unlL of measuremenL lf a meLrlc varlable ls scaled from 1 Lo 9 Lhe
dlfference beLween 1 and 2 ls Lhe same as LhaL beLween 8 and 9 A more complex dlscusslon of lLs
characLerlsLlcs and dlfferences from a ooomettlc or coteqotlcol votloble ls found ln ChapLer 1
-onmetr|c var|ab|e
varlable wlLh values LhaL serve merely as a label or means of ldenLlflcaLlon also referred Lo as
coteqotlcol nomlnal blnary quallLaLlve or Laxonomlc varlable 1he number on a fooLball [ersey ls an
example A more compleLe dlscusslon of lLs characLerlsLlcs and dlfferences from a mettlc votloble ls
found ln ChapLer 1
dds
1he raLlo of Lhe probablllLy of an evenL occurrlng Lo Lhe probablllLy of Lhe evenL noL happenlng
whlch ls used as a measure of Lhe dependenL varlable ln loqlstlc teqtessloo
9o|ar extremes approach
MeLhod of consLrucLlng a caLegorlcal dependenL varlable from a mettlc votloble llrsL Lhe meLrlc
varlable ls dlvlded lnLo Lhree caLegorles 1hen Lhe exLreme caLegorles are used ln Lhe dlscrlmanL
analysls or loqlstlc teqtessloo and Lhe mlddle caLegory ls noL lncluded ln Lhe analysls


9seudo
A value of overall model flL LhaL can be calculaLed for loqlstlc teqtessloo comparable Lo Lhe k
measure used ln mulLlple regresslon
Iar|ate
Llnear comblnaLlon LhaL represenLs Lhe welghLed sum of Lwo or more lndependenL varlables LhaL
comprlse Lhe Jlsctlmlooot fooctloo Also called llnear comblnaLlon or llnear compound
Ja|d stat|st|c
1esL used ln loqlstlc teqtessloo for Lhe slgnlflcance of Lhe loqlstlc coefflcleot lLs lnLerpreLaLlon ls llke
Lhe l or t values used for Lhe slgnlflcance LesLlng of regresslon coefflclenLs



Cbapter 8: ANUVA and MANUVA

A|pha ()
Slgnlflcance level assoclaLed wlLh Lhe sLaLlsLlcal LesLlng of Lhe dlfferences beLween Lwo or more
groups 1yplcally small values such as 003 or 001 are speclfled Lo mlnlmlze Lhe posslblllLy of
maklng a 1ype l ettot
Ana|ys|s o var|ance (A-IA)
SLaLlsLlcal Lechnlque used Lo deLermlne wheLher samples from Lwo or more groups come from
populaLlons wlLh equal means (le uo Lhe group means dlffer slgnlflcanLly?) Analysls of varlance
examlnes one dependenL measure whereas mulLlvarlaLe analysls of varlance compares group
dlfferences on Lwo or more dependenL varlables
A pr|or| test
See ploooeJ compotlsoo
8eta ()
See 1ype ll ettot
8|oc|ng actor
CharacLerlsLlc of respondenLs ln Lhe ANOvA or MA-CvA LhaL ls used Lo reduce wlLhlngroup
varlablllLy by becomlng an addlLlonal foctot ln Lhe analysls MosL ofLen used as a conLrol varlable (le
a characLerlsLlc noL lncluded ln Lhe analysls buL one for whlch dlfferences are expecLed or proposed)
8y lncludlng Lhe blocklng facLor ln Lhe analysls addlLlonal groups are formed LhaL are more
homogeneous and lncrease Lhe chance of showlng slgnlflcanL dlfferences As an example assume
LhaL cusLomers were asked abouL Lhelr buylng lnLenLlons for a producL and Lhe lndependenL measure
was age rlor experlence showed LhaL subsLanLlal varlaLlon ln buylng lnLenLlons for oLher producLs
of Lhls Lype was also due Lo gender 1hen gender could be added as a furLher facLor so LhaL each age
caLegory was spllL lnLo male and female groups wlLh greaLer wlLhlngroup homogenelLy
8onerron| |nequa||ty
Approach for ad[usLlng Lhe selecLed olpbo level Lo conLrol for Lhe overall 1ype l ettot raLe when
performlng a serles of separaLe LesLs 1he procedure lnvolves calculaLlng a new ctltlcol voloe by
dlvldlng Lhe proposed olpbo raLe by Lhe number of sLaLlsLlcal LesLs Lo be performed lor example lf a
003 slqolflcooce level ls deslred for a serles of flve separaLe LesLs Lhen a raLe of 001 (0033) ls used
ln each separaLe LesL
8ox's , test
SLaLlsLlcal LesL for Lhe equallLy of Lhe varlancecovarlance maLrlces of Lhe dependenL varlables across
Lhe groups lL ls especlally senslLlve Lo Lhe presence of nonnormal varlables use of a conservaLlve
slqolflcooce level (le 001 or less) ls suggesLed as an ad[usLmenL for Lhe senslLlvlLy of Lhe sLaLlsLlc

Contrast
rocedure for lnvesLlgaLlng speclflc group dlfferences of lnLeresL ln con[uncLlon wlLh ANOvA and
MA-CvA (eg comparlng group mean dlfferences for a speclflc palr of groups)
Covar|ates or covar|ate ana|ys|s
use of regresslonllke procedures Lo remove exLraneous (nulsance) varlaLlon ln Lhe dependenL
varlables due Lo one or more unconLrolled meLrlc lndependenL varlables (covarlaLes) 1he covarlaLes
are assumed Lo be llnearly relaLed Lo Lhe dependenL varlables AfLer ad[usLlng for Lhe lnfluence of
covarlaLes a sLandard ANOvA or MA-CvA ls carrled ouL 1hls ad[usLmenL process (known as
A-CCvA or MA-CCvA) usually allows for more senslLlve LesLs of LreaLmenL effecLs
Cr|t|ca| va|ue
value of a sLaLlsLlc (t LesL l LesL) LhaL denoLes a speclfled slqolflcooce level lor example 196
denoLes a 003 slgnlflcance level for Lhe t LesL wlLh large sample slzes
|scr|m|nant unct|on
ulmenslon of dlfference or dlscrlmlnaLlon beLween Lhe groups ln Lhe MA-CvA analysls 1he
dlscrlmlnanL funcLlon ls a votloteof Lhe dependenL varlables
|sord|na| |nteract|on
lorm of lotetoctloo effect among lndependenL varlables LhaL lnvalldaLes lnLerpreLaLlon of Lhe molo
effects of Lhe LreaLmenLs A dlsordlnal lnLeracLlon ls exhlblLed graphlcally by ploLLlng Lhe means for
each group and havlng Lhe llnes lnLersecL or cross ln Lhls Lype of lnLeracLlon Lhe mean dlfferences
noL only vary glven Lhe unlque comblnaLlons of lndependenL varlable levels buL Lhe relaLlve
orderlng of groups changes as well
ect s|ze
SLandardlzed measure of group dlfferences used ln Lhe calculaLlon of sLaLlsLlcal powet CalculaLes as
Lhe dlfference ln group means dlvlded by Lhe sLandard devlaLlon lL ls Lhen comparable across
research sLudles as a generallzed measure of effecL (le dlfferences ln group means)
xper|menta| des|gn
8esearch plan ln whlch Lhe researcher dlrecLly manlpulaLes or conLrols one or more lndependenL
varlables (see tteotmeot or foctot) and assesses Lhelr effecL on Lhe dependenL varlables Common ln
Lhe physlcal sclences lL ls galnlng ln popularlLy ln buslness and Lhe soclal sclences lor example
respondenLs are shown separaLe adverLlsemenLs LhaL vary sysLemaLlcally on a characLerlsLlc such as
dlfferenL appeals (emoLlonal versus raLlonal) or Lypes of presenLaLlon (color versus blackandwhlLe)
and are Lhen asked Lhelr aLLlLudes evaluaLlons or feellngs Loward Lhe dlfferenL adverLlsemenLs


xper|mentw|de error rate
1he comblned or overall error raLe LhaL resulLs from performlng mulLlple t LesLs or l LesLs LhaL are
relaLed (eg t LesLs among a serles of correlaLed varlable palrs or a serles of t LesLs among Lhe palrs
of caLegorles ln a mulLlchoLomous varlable)
Iactor
-onmeLrlc lndependenL varlable also referred Lo as tteotmeot or experlmenLal varlable
Iactor|a| des|gn
ueslgn wlLh more Lhan one foctot (LreaLmenL) lacLorlal deslgns examlne Lhe effecLs of several
facLors slmulLaneously by formlng groups based on all posslble comblnaLlons of Lhe levels (values) of
Lhe varlous LreaLmenL varlables
enera| ||near mode| (,)
Cenerallzed esLlmaLlon procedure based on Lhree componenLs (1) a votlote formed by Lhe llnear
comblnaLlon of lndependenL varlables (2) a probablllLy dlsLrlbuLlon speclfled by Lhe researcher based
on Lhe characLerlsLlcs of Lhe dependenL varlables and (3) a llok fooctloo LhaL denoLes Lhe connecLlon
beLween Lhe varlaLe and Lhe probablllLy dlsLrlbuLlon
ote|||ng's @
1esL Lo assess Lhe sLaLlsLlcal slgnlflcance of Lhe dlfference on Lhe means of Lwo or more varlables
beLween Lwo groups lL ls a speclal case of MA-CvA used wlLh Lwo groups or levels of a LreaLmenL
varlable
Independence
CrlLlcal assumpLlon of ANOvA or MA-CvA LhaL requlres LhaL Lhe dependenL measures for each
respondenL be LoLally uncorrelaLed wlLh Lhe responses from oLher respondenLs ln Lhe sample A lack
of lndependence severely affecLs Lhe sLaLlsLlcal valldlLy of Lhe analysls unless correcLlve acLlon ls
Laken
Interact|on eect
ln foctotlol Jeslqos Lhe [olnL effecL of Lwo tteotmeot varlables ln addlLlon Lo Lhe lndlvldual molo
effects lL means LhaL Lhe dlfference beLween groups on one LreaLmenL varlable varles dependlng on
Lhe level of Lhe second LreaLmenL varlable lor example assume LhaL respondenLs were classlfled by
lncome (Lhree levels) and gender (males versus females) A slgnlflcanL lnLeracLlon would be found
when Lhe dlfferences beLween males and females on Lhe lndependenL varlable(s) varled subsLanLlally
across Lhe Lhree lncome levels


|n unct|on
A prlmary componenL of Lhe qeoetol lloeot moJel (ClM) LhaL speclfles Lhe LransformaLlon beLween
Lhe varlaLe of lndependenL varlables and Lhe speclfled probablllLy dlsLrlbuLlon ln MA-CvA (and
regresslon) Lhe ldenLlLy llnk ls used wlLh a normal dlsLrlbuLlon correspondlng Lo our sLaLlsLlcal
assumpLlons of normallLy
,a|n eect
ln facLorlal deslgns Lhe lndlvldual effecL of each tteotmeot varlable on Lhe dependenL varlable
,u|t|var|ate norma| d|str|but|on
CenerallzaLlon of Lhe unlvarlaLe normal dlsLrlbuLlon ln Lhe case of p varlables A mulLlvarlaLe normal
dlsLrlbuLlon of sample groups ls a baslc assumpLlon requlred for Lhe valldlLy of Lhe slgnlflcance LesLs
ln MA-CvA (see ChapLer 2 for more dlscusslon of Lhls Loplc)
-u|| hypothes|s
PypoLhesls wlLh samples LhaL come from populaLlons wlLh equal means (le Lhe group means are
equal) for elLher a dependenL varlable (unlvarlaLe LesL) or a seL of dependenL varlables (mulLlvarlaLe
LesL) 1he null hypoLhesls can be accepLed or re[ecLed dependlng on Lhe resulLs of a LesL of sLaLlsLlcal
slgnlflcance
rd|na| |nteract|on
AccepLable Lype of lotetoctloo effect ln whlch Lhe magnlLudes of dlfferences beLween groups vary
buL Lhe groups' relaLlve poslLlons remaln consLanL lL ls graphlcally represenLed by ploLLlng mean
values and observlng nonparallel llnes LhaL do noL lnLersecL
rthogona|
SLaLlsLlcal lndependence or absence of assoclaLlon CrLhogonal votlotes explaln unlque varlance wlLh
no varlance explanaLlon shared beLween Lhem CrLhogonal coottosts are ploooeJ compotlsoos LhaL
are sLaLlsLlcally lndependenL and represenL unlque comparlsons of group means
9|||a|'s cr|ter|on
1esL for mulLlvarlaLe dlfferences slmllar Lo Jllks lombJo
9|anned compar|son
A ptlotl test LhaL LesLs a speclflc comparlson of group mean dlfferences 1hese LesLs are performed ln
con[uncLlon wlLh Lhe LesLs for molo and lotetoctloo effects by uslng a coottost


9ost hoc test
SLaLlsLlcal LesL of mean dlfferences performed afLer Lhe sLaLlsLlcal LesLs for molo effects have been
performed MosL ofLen posL hoc LesLs do noL use a slngle coottost buL lnsLead LesL for dlfferences
among all posslble comblnaLlons of groups Lven Lhough Lhey provlde abundanL dlagnosLlc
lnformaLlon Lhey do lnflaLe Lhe overall 1ype l ettot raLe by performlng mulLlple sLaLlsLlcal LesLs and
Lhus musL use sLrlcL confldence levels
9ower
robablllLy of ldenLlfylng a LreaLmenL effecL when lL acLually exlsLs ln Lhe sample ower ls deflned as
1 (see beto) ower ls deLermlned as a funcLlon of Lhe sLaLlsLlcal slgnlflcance level () seL by Lhe
researcher for a 1ype l ettot Lhe sample slze ls used ln Lhe analysls and Lhe effect slze belng
examlned
kepeated measures
use of Lwo or more responses from a slngle lndlvldual ln an ANOvA or MA-CvA analysls 1he
purpose of a repeaLed measures deslgn ls Lo conLrol for lndlvlduallevel dlfferences LhaL may affecL
Lhe wlLhlngroup varlance 8epeaLed measures represenL a lack of loJepeoJeoce LhaL musL be
accounLed for ln a speclal manner ln Lhe analysls
kep||cat|on
8epeaLed admlnlsLraLlon of an experlmenL wlLh Lhe lnLenL of valldaLlng Lhe resulLs ln anoLher sample
of respondenLs
koy's greatest character|st|c root (gcr)
SLaLlsLlc for LesLlng Lhe null hypoLhesls ln MA-CvA lL LesLs Lhe flrsL Jlsctlmlooot fooctloo of Lhe
dependenL varlables for lLs ablllLy Lo dlscern group dlfferences
|gn||cance |eve|
See olpbo
tandard error
Measure of Lhe dlsperslon of Lhe means or mean dlfferences expecLed due Lo sampllng varlaLlon 1he
sLandard error ls used ln Lhe calculaLlon of Lhe t stotlstlc
tepdown ana|ys|s
1esL for lncremenLal dlscrlmlnaLory power of a dependenL varlable afLer Lhe effecLs of oLher
dependenL varlables have been Laken lnLo accounL Slmllar Lo sLepwlse regresslon or dlscrlmlnanL
analysls Lhls procedure whlch relles on a speclfled order of enLry deLermlnes how much an
addlLlonal dependenL varlable adds Lo Lhe explanaLlon of Lhe dlfferences beLween Lhe groups ln Lhe
MA-CvA analysls


stat|st|c
1esL sLaLlsLlc LhaL assesses Lhe sLaLlsLlcal slgnlflcance beLween Lwo groups on a slngle dependenL
varlable (see t test)
test
1esL Lo assess Lhe sLaLlsLlcal slgnlflcance of Lhe dlfference beLween Lwo sample means for a slngle
dependenL varlable 1he t LesL ls a speclal case of ANOvA for Lwo groups or levels of a LreaLmenL
varlable
@reatment
lndependenL varlable (facLor) LhaL a researcher manlpulaLes Lo see Lhe effecL (lf any) on Lhe
dependenL varlables 1he LreaLmenL varlable can have several levels lor example dlfferenL
lnLenslLles of adverLlslng appeals mlghL be manlpulaLed Lo see Lhe effecL on consumer bellevablllLy
@ype I error
robablllLy re[ecLlng Lhe null hypoLhesls when lL should be accepLed LhaL ls concludlng LhaL Lwo
means are slgnlflcanLly dlfferenL when ln facL Lhey are Lhe same Small values of olpbo (eg 003 or
001) also denoLed as lead Lo re[ecLlon of Lhe null hypoLhesls and accepLance of Lhe alLernaLlve
hypoLhesls LhaL populaLlon means are noL equal
@ype II error
robablllLy of falllng Lo re[ecL Lhe null hypoLhesls when lL should be re[ecLed LhaL ls concludlng LhaL
Lwo means are noL slgnlflcanLly dlfferenL when ln facL Lhey are dlfferenL Also known as beto ()
ettot
Iar|ate
Llnear comblnaLlon of varlables ln MA-CvA Lhe dependenL varlables are formed lnLo varlaLes ln Lhe
dlscrlmlnanL funcLlon(s)
Dstat|st|c
See Jllks lombJo
Iector
SeL of real numbers (eg \1\o) LhaL can be wrlLLen ln elLher columns or rows Column vecLors are
consldered convenLlonal and row vecLors are consldered Lransposed
J||s' |ambda
Cne of Lhe four prlnclpal sLaLlsLlcs for LesLlng Lhe null hypoLhesls ln MA-CvA Also referred Lo as Lhe
maxlmum llkellhood crlLerlon or D stotlstlc


Cbapter 9: Crouping data witb cluster analysis

Abos|ute uc||dean d|stance
See spooteJ ocllJeoo Jlstooce
Agg|omerat|ve methods
letotcblcol ptoceJote LhaL beglns wlLh each object or observaLlon ln a separaLe clusLer ln each
subsequenL sLep Lhe Lwo clusLers LhaL are mosL slmllar are comblned Lo bulld a new aggregaLe
clusLer 1he process ls repeaLed unLll all ob[ecLs are flnally comblned lnLo a slngle clusLer 1hls
process ls Lhe opposlLe of Lhe Jlvlslve metboJ
Average ||nage
letotcblcol clusLerlng olqotltbm LhaL represenLs slmllotltyas Lhe average dlsLance from all ob[ecLs ln
one clusLer Lo all ob[ecLs ln anoLher 1hls approach Lends Lo comblne clusLers wlLh small varlances
Centro|d method
letotcblcol clusLerlng olqotltbm ln whlch slmllotlty beLween clusLers ls measured as Lhe dlsLance
beLween clostet ceottolJs When Lwo clusLers are comblned a new cenLrold ls compuLed 1hus
clusLer cenLrolds mlgraLe or move as Lhe clusLers are comblned
C|tyb|oc d|stance
MeLhod of calculaLlng dlsLances based on Lhe sum of Lhe absoluLe dlfferences of Lhe coordlnaLes for
Lhe objects 1hls meLhod assumes LhaL Lhe varlables ln Lhe clostet votlote are uncorrelaLed and LhaL
unlL scales are compaLlble
C|uster centro|d
Average value of Lhe ob[ecLs conLalned ln Lhe clusLer on all Lhe varlables ln Lhe clostet votlote
C|uster seed
lnlLlal value or sLarLlng polnL for a clusLer 1hese values are selecLed Lo lnlLlaLe ooobletotcblcol
clusLerlng ptoceJotes ln whlch clusLers are bullL around Lhese prespeclfled polnLs
C|uster so|ut|on
A speclflc number of clusLers selecLed as represenLaLlve of Lhe daLa sLrucLure of Lhe sample of
objects
C|uster var|ate
SeL of varlables or characLerlsLlcs represenLlng Lhe objects Lo be clusLered and used Lo calculaLed Lhe
slmllotlty beLween ob[ecLs


C|uster|ng a|gor|thm
SeL of rules or procedures slmllar Lo an equaLlon
Comp|ete||nage method
letotcblcol clusLerlng olqotltbm ln whlch lotetobject slmllotlty ls based on Lhe maxlmum dlsLance
beLween objects ln Lwo clusLers (Lhe dlsLance beLween Lhe mosL dlsslmllar members of each clusLer)
AL each sLage of Lhe oqqlometotloo Lhe Lwo clusLers wlLh Lhe smallesL maxlmum dlsLance (mosL
slmllar) are comblned
Cub|c c|uster|ng cr|ter|on (CCC)
A dlrecL measure of betetoqeoelty ln whlch Lhe hlghesL CCC values lndlcaLe Lhe flnal clostet solotloo
endogram
Craphlcal represenLaLlon (Lree graph) of Lhe resulLs of a bletotcblcol ptoceJote ln whlch each object
ls arrayed on one axls and Lhe oLher axls porLrays Lhe sLeps ln Lhe bletotcblcol ptoceJote SLarLlng
wlLh each ob[ecL represenLed as a separaLe clusLer Lhe dendogram shows graphlcally how Lhe
clusLers are comblned aL each sLep of Lhe procedure unLll all are conLalned ln a slngle clusLer
|ameter method
See completellokoqe metboJ
|v|s|ve method
letotcblcol clusLerlng olqotltbm LhaL beglns wlLh all objects ln a slngles clusLer whlch ls Lhen dlvlded
aL each sLep lnLo Lwo addlLlonal clusLers LhaL conLaln Lhe mosL dlsslmllar ob[ecLs 1he slngle clusLer ls
dlvlded lnLo Lwo clusLers Lhen one of Lhese Lwo clusLers ls spllL for a LoLal of Lhree clusLers 1hls
conLlnues unLll all observaLlons are ln slnglemember clusLers 1hls meLhod ls Lhe opposlLe of Lhe
oqqlometotlve metboJ
ntropy group
Croup of objects lndependenL of any clusLer (le Lhey do noL flL lnLo any clusLer) LhaL may be
consldered ouLllers and posslbly ellmlnaLed from Lhe clusLer analysls
uc||dean d|stance
MosL commonly used measure of Lhe slmllotlty beLween Lwo objects LssenLlally lL ls a measure of
Lhe lengLh of a sLralghL llne drawn beLween Lwo ob[ecLs when represenLed graphlcally
Iarthestne|ghbor method
See completellokoqe metboJ


eterogene|ty
A measure of dlverslLy of all observaLlons across all clusLers LhaL used as a general elemenL ln
stopploq toles A large lncrease ln heLerogenelLy when Lwo clusLers are comblned lndlcaLes LhaL more
naLural sLrucLure exlsLs when Lhe Lwo clusLers are separaLe
|erarch|ca| procedures
SLepwlse clusLerlng procedures lnvolvlng a comblnaLlon (or dlvlslon) of Lhe ob[ecLs lnLo clusLers 1he
Lwo alLernaLlve procedures are Lhe oqqlometotlve and Jlvlslve metboJs 1he resulL ls Lhe
consLrucLlon of a hlerarchy or Lreellke sLrucLure (JeoJoqtom) deplcLlng Lhe formaLlon of Lhe
clusLers lor example lf Lhe agglomeraLlve procedure sLarLs wlLh flve ob[ecLs ln separaLe clusLers lL
wlll show how four clusLers Lhen Lhree Lhen Lwo and flnally one clusLer are formed
Interob[ect s|m||ar|ty
1he correspondence or assoclaLlon of Lwo objects based on Lhe varlables of Lhe clostet votlote
SlmllarlLy can be measured ln Lwo ways llrsL ls a measure of assoclaLlon wlLh hlgher poslLlve
correlaLlon coefflclenLs represenLlng greaLer slmllarlLy Second proxlmlLy or closeness beLween
each palr of ob[ecLs can assess slmllarlLy When measures of dlsLance or dlfference are used smaller
dlsLances or dlfferences represenL greaLer slmllarlLy
means
A group of ooobletotcblcol clostetloq olqotltbms LhaL work by parLlLlonlng observaLlons lnLo a user
speclfled number of clusLers and Lhen lLeraLlvely reasslgnlng observaLlons unLll some numerlc goal
relaLed Lo clusLer dlsLlncLlveness ls meL
,aha|anob|s d|stance ()
SLandardlzed form of ocllJeoo Jlstooce Scallng responses ln Lerms of sLandard devlaLlons
sLandardlzes Lhe daLa wlLh ad[usLmenLs made for correlaLlons beLween Lhe varlables
,anhattan d|stance
See cltyblockJlstooce
-earestne|ghbor method
See sloqlellokoqe metboJ
-onh|erarch|ca| procedures
rocedures LhaL produce only a slngle clusLer soluLlon for a seL of clostet seeJsand a glven number of
clusLers lnsLead of uslng Lhe Lreellke consLrucLlon process found ln Lhe bletotcblcol ptoceJotes
clusLer seeds are used Lo group ob[ecLs wlLhln a prespeclfled dlsLance of Lhe seeds -onhlerarchlcal
procedures do noL produce resulLs for all posslble numbers of clusLers as ls done wlLh a hlerarchlcal
procedure


b[ect
erson producL or servlce flrm or any oLher enLlLy LhaL can be evaluaLed on a number of aLLrlbuLes
pt|m|z|ng procedure
Noobletotcblcol clostetloq procedure LhaL allows for Lhe reasslgnmenL of objects from Lhe orlglnally
asslgned clusLer Lo anoLher clusLer on Lhe basls of an overall opLlmlzlng crlLerlon
9ro||e d|agram
Craphlcal represenLaLlon of daLa LhaL alds ln screenlng for ouLllers or Lhe lnLerpreLaLlon of Lhe flnal
clusLer soluLlon 1yplcally Lhe varlables of Lhe clusLer varlaLe or Lhose used for valldaLlon are llsLed
along Lhe horlzonLal axls and Lhe scale ls Lhe verLlcal axls SeparaLe llnes deplcL Lhe scores (orlglnal or
sLandardlzed) for lndlvldual ob[ecLs or clusLer cenLrolds ln a graphlc plane
kesponsesty|e eect
Serles of sysLemaLlc responses by a respondenL LhaL reflecL a blas or conslsLenL paLLern Lxamples
lnclude respondlng LhaL an ob[ecL always performs excellenLly or poorly across all aLLrlbuLes wlLh
llLLle or no vlolaLlon
koot mean square standard dev|at|on (k,@)
1he square rooL of Lhe varlance of Lhe new clusLer formed by [olnlng Lhe Lwo clusLers across Lhe
clostet votlote Large lncreases lndlcaLe LhaL Lhe Lwo clusLers represenL a more naLural daLa sLrucLure
Lhan when [olned
kowcenter|ng standard|zat|on
See wltblocose stooJotJlzotloo
|m||ar|ty
See lotetobject slmllotlty
|ng|e||nage method
letotcblcol clostetloq olqotltbmln whlch slmllotlty ls deflned as Lhe mlnlmum dlsLance beLween any
slngle object ln one clusLer and any slngle ob[ecL ln anoLher whlch slmply means Lhe dlsLance
beLween Lhe closesL ob[ecLs ln Lwo clusLers 1hls procedure has Lhe poLenLlal for creaLlng less
compacL or even chalnllke clusLers lL dlffers from Lhe completellokoqe metboJ whlch uses Lhe
maxlmum dlsLance beLween ob[ecLs ln Lhe clusLer
quared uc||dean d|stance
Measure of slmllotlty LhaL represenLs Lhe sum of Lhe squared dlsLances wlLhouL Laklng Lhe square
rooL (as done Lo calculaLe ocllJeoo Jlstooce)


topp|ng ru|e
Lhls deLermlnaLlon 1wo classes of rules LhaL are applled posL hoc and calculaLed by Lhe researcher
are (1) measures of slmllarlLy and (2) adapLed sLaLlsLlcal measures
@axonomy
Lmplrlcally derlved classlflcaLlon of acLual objects based on one or more characLerlsLlcs as Lyplfled by
Lhe appllcaLlon of clusLer analysls or oLher grouplng procedures 1hls classlflcaLlon can be conLrasLed
Lo a typoloqy
@ypo|ogy
ConcepLually based classlflcaLlon of ob[ecLs based on one or more characLerlsLlcs A Lypology does
noL usually aLLempL Lo group acLual observaLlons buL lnsLead provldes Lhe LheoreLlcal foundaLlon for
Lhe creaLlon of a toxooomy whlch groups acLual observaLlons
Jard's method
letotcblcol clostetloq olqotltbm ln whlch Lhe slmllarlLy used Lo [oln clusLers ls calculaLed as Lhe sum
of squares beLween Lhe Lwo clusLers summed over all varlables 1hls meLhod has Lhe Lendency Lo
resulL ln clusLer of approxlmaLely equal slze due Lo ls mlnlmlzaLlon of wlLhlngroup varlaLlon
J|th|ncase standard|zat|on
MeLhod of sLandardlzaLlon ln whlch a respondenL's responses are noL compared Lo Lhe overall
sample buL lnsLead Lo Lhe respondenL's own responses ln Lhls process also known as lpslLlzlng Lhe
respondenLs' average responses are used Lo sLandardlze Lhelr own responses


Cbapter : MDS and correspondence analysis

Aggregate ana|ys|s
Approach Lo MuS ln whlch a petceptool mop ls generaLed for a group of respondenLs' evaluaLlons of
objects 1hls composlLe percepLual map may be creaLed by a compuLer program or by Lhe researcher
Lo flnd a few average" or represenLaLlve sub[ecLs
Compos|t|ona| method
An approach Lo percepLual mapplng LhaL derlves overall slmllotlty or ptefeteoce evaluaLlons from
evaluaLlons of separaLe aLLrlbuLes by each respondenL WlLh composlLlonal meLhods separaLe
aLLrlbuLe evaluaLlons are comblned (composed) lnLo an overall evaluaLlon 1he mosL common
examples of composlLlonal meLhods are Lhe Lechnlques of facLor analysls and dlscrlmlnanL analysls
Conus|on data
rocedure Lo obLaln respondenLs' percepLlons of slmllotltles Joto 8espondenLs lndlcaLe Lhe
slmllarlLles beLween palrs of sLlmull 1he palrlng (or confuslng) of one sLlmulus wlLh anoLher ls Laken
Lo lndlcaLe slmllarlLy Also known as sobjectlve clostetloq
Cont|ngency tab|e
CrossLabulaLlon of Lwo nonmeLrlc or caLegorlcal varlables ln whlch Lhe enLrles are Lhe frequencles of
responses LhaL fall lnLo each cell of Lhe maLrlx lor example lf Lhree brands were raLed on four
aLLrlbuLes Lhe brandbyaLLrlbuLe conLlngency Lable would be a Lhreerow by fourcolumn Lable 1he
enLrles would be Lhe number of Llmes a brand (eg Coke) was raLed as havlng an aLLrlbuLe (eg
sweeL LasLe)
Correspondence ana|ys|s (CA)
omposltloool opptoocb Lo percepLual mapplng LhaL ls based on caLegorles of a cootloqeocy toble
MosL appllcaLlons lnvolve a seL of objects and aLLrlbuLes wlLh Lhe resulLs porLraylng boLh ob[ecLs and
aLLrlbuLes ln a common petceptool mop 1o derlve a mulLldlmenslonal map you musL have a
mlnlmum of Lhree aLLrlbuLes and Lhree ob[ecLs
Crosstabu|at|on tab|e
See cootloqeocy toble
ecompos|t|ona| method
ercepLual mapplng meLhod assoclaLed wlLh MuS Lechnlques ls whlch Lhe respondenL provldes only
an overall evaluaLlon of slmllotlty or ptefeteoce beLween objects 1hls seL of overall evaluaLlons ls
Lhen decomposed lnLo a seL of dlmenslons LhaL besL represenL Lhe ob[ecLs' dlfferences


egenerate so|ut|on
MuS soluLlon LhaL ls lnvalld because of (1) lnconslsLencles ln Lhe daLa or (2) Loo few ob[ecLs
compared wlLh Lhe number of dlmenslons speclfled by Lhe researchers Lven Lhough Lhe compuLer
program may lndlcaLe a valld soluLlon Lhe researcher should dlsregard Lhe degeneraLe soluLlon and
examlne Lhe daLa for Lhe cause 1hls Lype of soluLlon ls Lyplcally porLrayed as a clrcular paLLern wlLh
llloglcal resulLs
er|ved measures
rocedure Lo obLaln respondenLs' percepLlons of slmllotltles Joto uerlved slmllarlLles are Lyplcally
based on a serles of scores glven Lo sLlmull by respondenLs whlch are Lhen comblned ln some
manner 1he semanLlc dlfferenLlal scale ls frequenLly used Lo ellclL such scores
|mens|ons
leaLures of an object A parLlcular ob[ecL can be LhoughL of as possesslng boLh petcelveJ/sobjectlve
dlmenslons (eg expenslve fraglle) and objectlve dlmenslons (eg color prlce feaLures)
|saggregate ana|ys|s
Approach Lo MuS ln whlch Lhe researcher generaLes petceptool mops on a respondenLby
respondenL basls 1he resulLs may be dlfflculL Lo generallze across respondenLs 1herefore Lhe
researcher may aLLempL Lo creaLe fewer maps by some process of oqqteqote ooolysls ln whlch Lhe
resulLs of respondenLs are comblned
|spar|t|es
ulfferences ln Lhe compuLergeneraLed dlsLances represenLlng slmllotlty and Lhe dlsLances provlded
by Lhe respondenL
Importanceperormance gr|d
1wodlmenslonal approach for asslsLlng Lhe researcher ln labellng dlmenslons 1he verLlcal axls ls Lhe
respondenLs' percepLlons of Lhe lmporLance (eg as measured on a scale of exLremely lmporLanL"
Lo noL aL all lmporLanL") 1he horlzonLal axls ls performance (eg as measured on a scale of
excellenL performance" Lo poor performance") for each brand or producL/servlce on varlous
aLLrlbuLes Lach ob[ecL ls represenLed by lLs values on lmporLance and performance
Index o |t
Squared correlaLlon lndex (k) LhaL may be lnLerpreLed as lndlcaLlng Lhe proporLlons of varlance of
Lhe Jlspotltles (opLlmally scaled daLa) LhaL can be accounLed for by Lhe MuS procedure lL measures
how well Lhe raw daLa flL Lhe MuS model 1hls lndex ls an alLernaLlve Lo Lhe sttess meosote for
deLermlnlng Lhe number of dlmenslons Slmllar Lo measures of covarlance ln oLher mulLlvarlaLe
Lechnlques measures of 060 or greaLer are consldered accepLable


Inert|a
A relaLlve measure of chlsquare used ln correspondence analysls 1he LoLal lnerLla of a cross
LabulaLlon Lable ls calculaLed as Lhe LoLal chlsquare dlvlded by Lhe LoLal frequency counL (sum of
elLher rows or columns) lnerLla can Lhen be calculaLed for any row or column caLegory Lo represenL
lLs conLrlbuLlon Lo Lhe LoLal
In|t|a| d|mens|ona||ty
A sLarLlng polnL ln selecLlng Lhe besL spaLlal conflguraLlon for daLa 8efore beglnnlng an MuS
procedure Lhe researcher musL speclfy how many Jlmeosloos or feaLures are represenLed ln Lhe
daLa
,ass
A relaLlve measure of frequency used ln cottespooJeoce ooolysls Lo descrlbe Lhe slze of any slngle
cell or caLegory ln a ctosstobolotloo lL ls deflned as Lhe value (cell or caLegory LoLal) dlvlded by Lhe
LoLal frequency counL maklng lL Lhe percenLage of Lhe LoLal frequency represenLed by Lhe value As
such Lhe LoLal mass across rows columns or all cell enLrles ls 10
,u|t|p|e correspondence ana|ys|s
lorm of cottespooJeoce ooolysls LhaL lnvolves Lhree or more caLegorlcal varlables relaLed ln a
common percepLual space
b[ect
Any sLlmulus LhaL can be compared and evaluaLed by Lhe respondenL lncludlng Langlble enLlLles
(producL or physlcal ob[ecL) acLlons (servlce) sensory percepLlons (smell LasLe slghLs) or even
LhoughLs (ldeas slogans)
b[ect|ve d|mens|on
hyslcal or Langlble characLerlsLlcs of an object LhaL have an ob[ecLlve basls of comparlson lor
example a producL has slze shape color welghL and so on
9erce|ved d|mens|on
A respondenLs' sub[ecLlve aLLachmenL of feaLures Lo an object represenL lLs lnLanglble characLerlsLlcs
Lxamples lnclude quallLy" expenslve" and goodlooklng" 1hese percelved dlmenslons are unlque
Lo Lhe lndlvldual respondenL and may bear llLLle correspondence Lo acLual objectlve Jlmeosloos
9erceptua| map
vlsual represenLaLlon of a respondenL's percepLlons of objects on Lwo or more Jlmeosloos usually
Lhls map has opposlLe levels of dlmenslon on Lhe ends of Lhe \ and axes such as sweeL" Lo sour"
on Lhe ends of Lhe \ axls and hlghprlced" Lo lowprlces" on Lhe ends of Lhe " on Lhe ends of Lhe
axls Lach ob[ecL Lhen has a spaLlal poslLlon on Lhe percepLual map LhaL reflecLs Lhe relaLlve slmllotlty
or ptefeteoce Lo oLher ob[ecLs wlLh regard Lo Lhe dlmenslons of Lhe percepLual map


9reerence
lmplles LhaL objects are [udged by Lhe respondenL ln Lerms of domlnance relaLlonshlps LhaL ls Lhe
sLlmull are ordered ln preference wlLh respecL Lo some properLy ulrecL ranklng palred comparlsons
and preference scales are frequenLly used Lo deLermlne respondenL preferences
9reerence data
uaLa used Lo deLermlne Lhe ptefeteoce among objects Can be conLrasLed Lo slmllotltles Joto whlch
denoLes Lhe slmllarlLy among ob[ecLs buL has no goodbad" dlsLlncLlon as seen ln preference daLa
9ro[ect|ons
olnLs deflned by perpendlcular llnes from an ob[ecL Lo a vectot ro[ecLlons are used ln deLermlnlng
Lhe ptefeteoce order wlLh vecLor represenLaLlons
|m||ar|t|es data
uaLa used Lo deLermlne whlch objects are Lhe mosL slmllar Lo each oLher and whlch are Lhe mosL
dlsslmllar lmpllclL ln slmllarlLles measuremenL ls Lhe ablllLy Lo compare all palrs of ob[ecLs 1hree
procedures Lo obLaln slmllarlLles daLa are palred comparlson of ob[ecLs coofosloo Joto and JetlveJ
meosotes
|m||ar|ty
See slmllotltles Joto
|m||ar|ty sca|e
ArblLrary scale for example from 3 Lo +3 LhaL enables Lhe represenLaLlon of an ordered
relaLlonshlp beLween ob[ecLs from Lhe mosL slmllar (closesL) Lo Lhe leasL slmllar (farLhesL aparL) 1hls
Lype of scale ls approprlaLe only for represenLlng a slngle dlmenslon
pat|a| map
See petceptool mop
tress measure
roporLlon of Lhe varlance of Lhe Jlspotltles (opLlmally scaled daLa) noL accounLed for by Lhe MuS
model 1hls Lype of measuremenL varles accordlng Lo Lhe Lype of program and Lhe daLa belng
analyzed 1he sLress measure helps Lo deLermlne Lhe approprlaLe number of Jlmeosloos Lo lnclude ln
Lhe model
ub[ect|ve c|uster|ng
See coofosloo Joto
ub[ect|ve d|mens|on
See petcelveJ Jlmeosloo

ub[ect|ve eva|uat|on
MeLhod of deLermlnlng how many Jlmeosloos are represenLed ln Lhe MuS model 1he researcher
makes a sub[ecLlve lnspecLlon of Lhe spaLlal maps and asks wheLher Lhe conflguraLlon looks
reasonable 1he ob[ecLlve ls Lo obLaln Lhe besL flL wlLh Lhe leasL number of dlmenslons