MODULEFORQUANTI TATIVEMETHODESINECONOMI

CSWI
THAPPLI
CATI
ON.
COMPI LEDBY:
1.GETACHEW PETROS(ASSISTANCEPROFESSOR)
2.LAMBAMOAREGA( ASSISTANCEPROFESSOR)

WOLAI
TASODO,
ETHI
OPI
A

Cont ents
1.1Nat ureandI mpor tanceofMat hemat ical Economics 5
1.3.Adv antagesoft heMat hemat ical Appr oaches5
1.4.Rev iewofEconomi cModel andAnal y sis6
1.5.St ati
c(Equi l
ibrium)Anal ysisinEconomi cs 6
1.5.1.Par ti
al Mar ketEqui l
ibr i
um -ALi nearModel6
1.5.2.Par ti
al Mar ketEqui l
ibr i
um -ANonl inearModel7
1.5.3.Gener al Mar ketEqui librium 8
1.5.4.Equi l
ibr i
um i nNat ional -IncomeAnal ysis 8
CHAPTERTWO 9
REVI SIONONCALCULUSANDLI NEARALGEBRA9
2.1.Di ff
erent ial Cal culus: Fundament alTechni ques 9
2.1.1.TheConceptoft heDer ivati
ve. 9
2.1.2.TheRul esorTheor emsofDer iv ates 10
a)Theconst antr ule 10
b)TheSi mpl ePowerr ules10
c)Thecoef f i
ci entr ules 10
d) Thesum ordi fferencer ule10
e)Thepr oductr ules 10
g) Di ff
erent iationofacomposi tef unct ion 11
i
) Der i
vati
v esofLogar i
thmi candExponent ialFunct
ions 12
2.2.Integral cal cul us: Techni quesandappl i
cationsofint
egral
cal
cul
us 14
2.2.1.TheI ndef initeI ntegr al 14
2.2.2.Techni quesofI nt egr ation&t heGener alPowerRul e14
2.2.3.Def i
niteI ntegr als 16
2.2.4.Cont inuousMoneySt reams 17
2.2.4.1.Cont i
nuousMoneySt reamswi t
hi naPer i
od 17
2.2.5.Impr operI nt egr al
s 18
2.3.1.Mat ri
xOper at i
ons 20
2.3.2.Det ermi nant sandI nv erseofaMat rix 26
2.3.2.1.TheConceptofDet ermi nant s, Mi norandCof actor26
2.3.2.2.TheI nv erseofaMat rix31
2.3.3.Mat ri
xRepr esent at ionofLi nearEquat ions 32
2.3.4.Inputoutputanal ysi s( Leont i
efmodel )33
2.3.5.LinearPr ogr ammi ng35
3.2.2.LinearAppr oximat ion 39
3.2.6TheI nter medi at eVal ueTheor em 48
3.3.2TheMul tivariat eChai nRul e 53
CHAPTERFOUR 55
UNCONSTRAI NEDOPTI MI ZATI ON 55
4.1.Funct ionswi t
honev ar i
abl e 55
4.1.1.Theconceptofopt i
mum ( ext r
eme)v alue 55
4.3Impl icitf unct ionsandunconst rainedEnv elopetheory 71
CHAPTERFI VE 74
5.CONSTRAI NEDOPTI MI ZATI ON 74
CHAPTERSI X 96
COMPARATI VESTATI CANALYSI S 96
CHAPTERSEVEN108
DYNAMI COPTI MI ZATI ON. 108
7.1.1Fi rst–Or derLi nearDi fferent i
al Equat ions 109
7.2.TheDy nami cSt abili
tyofEqui librium113
ECONOMETRI CSONE121
1.1Def initionandscopeofeconomet rics 121
1.5Met hodol ogyofeconomet rics 123
Chapt erTwo125
2.1THECLASSI CALREGRESSI ONANALYSI S125
2.1.St ochast i
candNon- stochast icRel ationshi ps126
2.2.Simpl eLi nearRegr essi onmodel . 126
2.2.1Assumpt ionsoft heCl assi cal LinearSt ochasticRegressi
onModel127
2.2.2Met hodsofest i
mat i
on 128
2.2.2.1Theor di nar yleastsquar e( OLS)met hod 129
2.2.2.2Est imat ionofaf unct i
onwi thzer oi nt ercept 130
2.2.2.3.St atistical Pr oper ti
esofLeastSquar eEst imators 131
2.2.2.4.St atistical testofSi gnifi
canceoft heOLSEst i
mators(FirstOrdertests) 133
2.6Conf i
denceI ntervalsandHy pot hesi sTest i
ng134
CHAPTERTHREE141
TheMul tipleLi nearRegr essi onAnal y sis141
3.1Int roduct ion 141
3.2Assumpt ionsofMul tipleRegr essi onModel s 142
3.3.AModel wi t
hTwoExpl anat or yVar iabl es143
3.3.1 Est i
mat ionofpar amet er soft wo- expl anatoryvari
ablesmodel 143
3.3.2Coef ficientofMul t
ipl eDet er mi nat i
on( R2) 144
Chapt erFour149
Violationsofbasi cCl assi cal Assumpt i
ons 149
4.3.4.1TestBasedonAuxi l
iaryRegr essi ons:172
4.3.4.3Testofmul t
icolli
near it
yusi ngEi genv aluesandcondi ti
oni ndex:174
4.3.4.4 Testofmul ticollinear i
tyusi ngTol er anceandv ar
ianceinflat
ionfactor 174
PARTTWO 176
ECONOMETRI CSTWO176
CHAPTERONE 176
1.2.3.Regr essi ononOneQuant itativeVar iabl eandTwoQual it
ativeVariables179
1.2.4.Int er actionsAmongDummyVar iables181
1.2.5.Test ingf orSt r
uct ural Stabi li
tyofRegr essionModel s182
1.3. DummyasDependentVar iabl e 182
1.3.2.TheLogi tandPr obitModel s 184
CHAPTERTWO 189
2.6ThePhenomenonofSpur i
ousRegr essi on 197
CHAPTERTHREE199
CHAPTERFOUR 208
CHAPTERONE
INTRODUCTI ONTOMATHEMATI CALECONOMI CS
1.1Nat ur eandI mpor tanceofMat hemat ical Economi cs
Mat hemat icsev ol ved t hrough t heuseofabst ract ion and l ogi calr easoni ng,f rom count ing,
calcul ation, measur ement ,andt hesy st emat i
cst udyofposi ti
ons, shapesandmot i
onsofphy sical
object s.Mat hemat icians expl or e such concept s,ai mi ng t of or mul ate new conj ectures and
establ isht hei rtrut hbyr i
gor ousdeduct ionf r
om appr opr iatelychosenaxi omsanddef ini
ti
ons.
1.2.Mat hemat ical Ver susNon- Mat hemat ical Economi cs
Sincemat hemat icaleconomi csi smer elyanappr oacht oeconomi canal ysi
s,i tshoul dnotand
doesnotdi fferf rom t henonmat hemat i
calappr oacht oeconomi canal ysisinanyf undament al
way .Thepur poseofanyt heor et i
cal anal y sis,regardl essoft heappr oach, isal way stoder iveaset
ofconcl usi onsort heor emsf r om agi vensetofassumpt i
onsorpost ulatesv iaapr ocessof
reasoni ng.
Themaj ordi ffer encebet weent heset woappr oachesi st hatint hef ormer ,theassumpt i
onsand
concl usi onsar est atedi nmat hemat i
calsy mbol srat hert hanwor dsandi nequat ionsr athert han
sent ences.Sy mbol sar emor econv enientt ousei ndeduct iv
er easoni ng,andcer tainlyaremor e
conduci v et oconci senessandpr eci senessofst atement .Thechoi cebet weenl it
erar ylogicand
mat hemat icall ogi c, isamat terofl i
ttl
ei mpor tant,butmat hemat i
cshast headv antageoff orcing
analy stst omaket hei rassumpt i
onsexpl icitatev eryst ageofr easoni ng.
Butaneconomi stequi ppedwi tht het oolsofmat hemat icsi sli
keanor malper sonwi t hmot orboat
orshi p dependi ng upon hi s per sonali nclinati
on t o each. As a r esult,mosteconomi c
resear cher sar eext ensi v
elyusi ngt het oolsofmat hemat icst oeconomi cr easoni ng.
Thet er m mat hemat icaleconomi csi sal sodi ff
er entf rom economet ricsi nt hatt hef ormeri s
concer nedwi tht heappl i
cat i
onofmat hemat icst ot hepur el
yt heor eticalaspect sofeconomi cs
analy sis,wi thlittleornoconcer naboutst ati
st i
calpr oblemsast heer rorofmeasur ementoft he
variabl esunderst udy ,whi l
et hel aterf ocusesont hemeasur ement sandanal y sisofeconomi c
data;hencei tdeal swi tht hest udyofempi ri
calobser vationsusi ng st atisticalmet hodsof
estimat ionandhy pot hesi st est i
ng.
1.3.Adv antagesoft heMat hemat ical Appr oaches
Mat hemat ical economi csi sofpar amounti mpor t
ancef oreconomi canal yses.
Speci fical l
y, t
heappr oachhast hef oll
owi ngadv antages:
a)Thel anguageusedi smor epr eciseandconci se;
b)Ther ei sweal t
hofmat hemat i
calassumpt i
onst hatmaket hingssi mpl eandl ifelesscost ly;
hencet heanal ysisi smor er i
gor ous;
c)Byf or cing us t o st ate expl i
citl
yal lourassumpt ions as a pr erequi sit
et ot he use of
mat hemat icaltheor ems,i tkeepsusf rom t hepi tfallofanuni ntent i
onaladopt i
onofunwant ed
i
mpl i
citassumpt ions
d)I thel psust ounder standr el ati
onshi psamongt woormor eeconomi cv ariablessimpl yand
neat l
ywi th whi ch t hegeomet ricand l i
teraryappr oachesar eathi gh pr obabi l
it
yofr i
skof
commi tti
ngmi stakes.
e)ital l
owsust otr eatt hegener al n-variablecase.
1.4.Rev iewofEconomi cModel andAnal ysis
Asy ouhav ecov er edi nt hecour sei ntroduct i
ont oeconomi cs,aneconomi ct heor yisnecessar ily
gener alizat ionorabst ract i
onf rom t her ealwor l
d. Thecompl exityoft her ealeconomydoesnot
makei tpossi bleforust ounder standal lthei nterr
elat i
onshi psatonce; nor ,forthatmat ter
, areal l
thesei nter rel
ati
onshi psofequali mpor t
ancef ort heunder standi ngoft hepar ti
culareconomi c
phenomenonunderst udy . Thewi sepr ocedur eis,hence,t opi ckoutwhatappealt oourr eason
tobet hepr i
mar yf actor sandr el ati
onshi psr elevantt oourpr obl em andt of ocusourat t
ent i
onon
theseal onet hemai nai m behi ndbui l
dingeconomi cmodel si st odescr i
behow t heeconomy
wor ksandt oobt ai nsomehowv ali
dpr edictionsabouteconomi cv ariables.Di fferentmodel sar e
builtfordi ffer
entpur poses.Someoft hemodel saredesi gnedt oinv estigatet heequi li
brium v alue
oft hev ariableasanul ti
mat eendwhi lesomet oinv estigat et hemov ementofav ari
able( s)ov er
ti
me. Thus, ineconomi cs, wehav ethreet ypesofanal yses:
i
.St atic( equi l
ibri
um)anal ysi
s
i
i.Compar ativ
est at icanal ysis
i
ii.Dy nami canal ysis
1.5.St atic( Equili
br ium)Anal ysisi nEconomi cs
Inessence,equi l
ibr i
um f oraspeci f
icmodeli sasi t
uat iont hati schar act erizedbyal ackof
tendencyt ochange.I tisf orthisr easont hatt heanal ysi sofequi librium i sr eferredtoasst atic.
Thef actt hatequi li
br i
um i mpl iesno t endencyt o changemayt emptonet o concludet hat
equi l
ibrium necessar il
yconst i
tut esadesi r
abl eorideal st ateofaf fairs.
Fort hepr esentanal ysiswewi llfocusont wot ypesoft heequi librium.Onei st heequi li
brium
attainedbyamar ketundergi vendemandandsuppl ycondi tions.Theot heri st heequi l
ibrium of
national incomeundergi v encondi t
ionsofconsumpt ionandi nv estmentpat ter ns.

1.5.
1.Par tialMar ketEqui libri
um -ALi nearModel
Inast atic-equili
brium model ,t hest andar dpr oblem i
st hatoff i
ndi
ngt hesetofval
uesoft
he
endogenousv ari
abl eswhi chwi ll sat i
sfytheequi l
ibri
um conditi
onsoft hemodel.
Parti
al-
Equi l
i
brium Mar ketModel --amodel ofpr i
cedeterminati
oninani solat
edmarket
.
Threevar i
abl es
Qd=t hequant i
tydemandedoft hecommodi t
y;
Qs=t hequant itysuppl iedoft hecommodi ty;
P=t hepr iceoft hecommodi t
y .
TheEqui l
ibr i
um Condi tion: Qd=Qs
Themodel is
Qd=a-bp( a,b>0)
Qs=- c+dp( c,d>0)
TheslopeofQd=- b,thev ert
ical int ercept=a.
TheslopeofQs=d, t
hev erti
cal int ercept=- c.
Notethat, cont r
aryt ot heusual pr actice,quanti
tyr at
herthanpricehasbeenpl ot
tedver
ti
cal
l
y
i
nt heFigur e.
Onewayoff indingt heequi li
brium i sbysuccessi veeliminati
onofv ar
iablesand
equati
onst hroughsubst i
tut i
on.
From Qs=Qd, wehav e
a-bp=- c+dp
andthus
(b+d)p=a+c.
Sinceb+d>0, thent heequi librium pr icei s
P=a+c/b+d
Theequi li
br ium quant i
tycanbeobt ainedbysubst it
utingpi nt oei therQsorQd:
Sincet hedenomi nat or( b+d)i sposi tive, theposi tivit
yofQr equi r est hatt henumer ator
(ad-bc) > 0.Thus,t o be economi cal l
ymeani ngf ul,t he modelshoul d cont aint he addi t
ional
restri
ct iont hatad>bc.
1.5.2.Par tial Mar ketEqui li
br ium -ANonl inearModel
Thepar ti
al mar ketmodel canbenonl inear .Supposeamodel i
sgi v
enby
Qd=4-p2
Qs=4p-1
Aspr ev i
ousl yst ated, thissy st em oft hreeequat ionscanber educedt oasi ngl eequat i
onby
subst i
t ution.4-p2=4p–1orp2+4p-5=0whi chi saquadr at icequat i
on.
Ingener al,gi venaquadr at icequat ioni nt hef orm ax2+bx+c=0( a>0) ,i
tst wor oot scanbe
obtainedf rom t hequadr aticf or mul a:
-b+_√b2- 4ac/ 2awher et he“ +"par toft he" ±"si gny i
eldsXj andt he" -"par ty i
eldsx2.
Thus, byappl yingt hequadr at i
cf or mul ast op2+4p-5=0, wehav ePl =1andp2=- 5,butonlythe
fir
stiseconomi cal l
yadmi ssibl e,asnegat ivepr icesar er uledout .
1.5.3.Gener al Mar ketEqui l
ibr i
um
Intheabov e, wehav edi scussedmet hodsofani sol atedmar ket ,wher ei ntheQdandQs
ofacommodi tyar ef unct ionsoft hepr iceoft hatcommodi tyal one.I nt her ealwor ld,there
woul dnor mal lyexi stmanysubst i
tut esandcompl ement ar ygoods.Thus, amor er eali
sticmodel
forthedemandandsuppl yf unct ionsofacommodi t
yshoul dt akei nt oaccountt heef f
ect snot
onlyoft hepr i
ceoft hecommodi tyi tsel fbutal sooft hepr icesofot hercommodi ties.
Consequent ly, theequi l
ibr ium condi tionofann- commodi tymar ketmodel wi l
linvolvenequat ions,
onef oreachcommodi ty ,int hef or m E=Qdi-Qsi=0( i=l,
2,...,n) wher eQdi&Qsi ar ethe
demand and suppl yf unct i
onsofcommodi tyi ,solv ing nequat i
onsf orP:weobt ainthen
equili
br ium pr icesP¡ —ifasol ut i
ondoesi ndeedexi st.Andt hent heQ¡maybeder ivedf r
om the
demandorsuppl yf unct ions.
1.5.4.Equi librium i nNat ional -
IncomeAnal ysis
Theequi li
br ium anal ysiscanbeal soappl iedt oot herar easofEconomi cs.Asasi mpl e
exampl e, wemayci tet hef ami l
iarKey nesi annat ional-incomemodel ,
Y=C+I o+G0( equi libri
um condi t
ion)
C=Ca+bY( theconsumpt ionf unct ion)
wher eYandCst andf ort heendogenousv ariabl e’snat ional incomeandconsumpt i
on
expendi ture,r espect ively ,Cast andsf oraut onomousconsumpt ion,andI 0andG0r epresentthe
exogenousl ydet ermi nedi nv est mentandgov ernmentexpendi tur es.
Solvingt heset wol inearequat ions, weobt ai nt heequi librium nat ional incomeand
consumpt i
onexpendi t
ur e: Y=Ca+bY+I 0+G0, bycol l
ectingl i
ket ermswehav e;
Y=Ca+I 0+G0/ 1- b
C=Ca+b( Ca+I 0+G0/ 1-b)
CHAPTERTWO
REVI SIONONCALCULUSANDLI NEARALGEBRA
2.1.Di fferent i
al Calcul us: Fundament al Techni ques
2.1.1.TheConceptoft heDer i
v ative.
Der i
v ativei sdef i
nedasader ivedf unct ionf rom anor iginalf unct ion,sayy =f(x)thatshowst he
rateofchangei ndependentv ariabl e(y)duet ov er ysmal l changei nindependentv ari
able(x).
Iffisaf unct iondef inedbyy =f( x), thent heder ivat i
veoff( x)f oranyv alueofx, isdenot edby
dy/dx, y’ orf’ ( x)i st hei nst ant aneousr ateofchangeoff( X)whi chi sgi venby :

Asy oumi ghtr

ememberf r
om theprev
iousdi
scussi
on,t
hederivati
veisalsothesl
opeoft he
tangentlinetof(x)atapoint.Bef
orewedir
ect
lygotothedi
scussionoftherul
esofder
ivat
ives,
l
et’sseesomeexampl es.
Exampl es: Forf( x)
=X2
Soluti
on
 =
=

Notet hat ,
i
.Theder iv
at i
ve,whichi soneoft hemostf undament alconcept si ncalculus,i
sthesameast he
foll
owi ngt woconcept s.
a.Slopeofal i
net angentt oacur veatx
b.Instant aneousr ateofchangeoff (
x)atx
i
i.Thepr ocessofobt ainingf ’(x)fr
om f( x)isknownasdi f
ferenti
ationandi fthederivati
veexists,
thent hef unct i
oni sdifferentiableatapoi ntorov erani nterval
.Int hef ol
lowingsection,wewi l
l
presentt her ulesofder ivativ
es, whichwi l
llargel
ysimpl i
fyourcal culati
onoft hederi
v ati
ve.
2.1.2.TheRul esorTheor emsofDer i
vates
Inthissect i
on,wef ormal lyst atethedi f
ferentrul
esofder i
vati
vest hatar eimport
anti nsev eral
problemsofget ti
ngt heder i
v at
ive.Thesear etheconst antr ule,thesi mplepowerr ule,the
coef f
icientrule,sum/ differencer ul
e,productr ul
e,quotientrule,chainrule,andothers.
a)Theconst antrule
Ifcisconst antandi ff (
x)=c,thenf ’(
x)=0
Exampl es.Iff(x)=2, thenf‘ (x)
=0
b)TheSi mpl ePowerr ules
Ifwher e t hen=nxn- 1
c)Thecoef f
icientrul
es
Foranyconst antCandf unct i
onm( x),

Exampl
e

d) Thesum ordi
ff
erencer
ule

Exampl
e I
f ,
thent
heder
ivat
ive

e)Theproductr
ules
Alt
houghmostoft hetimeafunct
ionisgi
venasasum ordi
ff
erenceoft
erms,t
her
ear
eti
mes
whenfuncti
onsarepresent
edi
nproduct
s.

Exampl
e
1.

f
) TheQuot
ientr
ule
I
fu(x)andV(
x)aretwof
unct
ionsandi
ff(
x)i
sthequot
ient
,

ExampleIfY
Solut
ion

g) Di
ff
erent
iat
ionofacomposit
efunct
ion
I
fthef
uncti
onfisafunct
ionofgandgisaf unct
ionof(
x),
then i
s:.

Example Y
Solut
ion Let

h) Deri
vati
vesofI
mpli
citf
unct
ions
Impl
ici
tfuncti
onsarethosefuncti
onswhosedependentand independentv
ari
ablesarenot
expl
i
cit
lystat
ed.Thati
s,wedon’
tknowwhetheragi
venv ar
iabl
eisacauseoranef f
ectvar
iabl
e
wit
houtanyapr i
oriinfor
mat ionwhenweencount eri
mplicitfunct
ions.
Whent hedependentandi ndependentv ar
iabl
esarenotexpli
citl
ystated,
onemaynoteasi
l
ycome
upwith. Butusingi mpli
citdiff
erenti
ati
ononecancalculate.
Supposewehav eani mplicitf
unctiongiv
enby
Then,t
otal
lydif
ferenti
ati
ngt heequat i
onweget

Rear
rangi
ngt
ermsgi
vesus:

Example:Givenafunct
ion,fi
nd usingi
mplici
tdif
fer
ent
iat
ionrul
e.
Solut
ionFir
st,l
et’
sfi
ndt hef
ir
stpar
ti
alder
ivati
vesfr
om t
hegivenfunct
ion

Then,
usi
ngt
hei
mpl
i
citdi
ff
erent
iat
ionr
ule

i
) Der i
vati
vesofLogar
ithmicandExponent ial
Functi
ons
The conceptand r ul
es gov erni
ng logari
thmic and exponenti
alfunct
ions were descr
ibed
previ
ousl
yinthefir
stchapter.Inthi
ssection,wewillr
evi
ewpoi ntouthowtofindthederi
vat
ives
oflogar
it
hmicandexponentialfuncti
ons.

Der
ivat
ivesofLogar
it
hmi
cFunct
ions
I
fY

.
Seeexpansi
onf
rom t
hepr
evi
ousdi
scussi
on

j
.Der
ivat
ivesofexponent
ial
funct
ions
I
f caneasilybefoundbychangi
ngittot
hel
ogar
it
hmi
cfunct
ion.

I
fY=aU(
x),
Usi
ngchai
nrul
e.Asar
esul
t,wehav
e

Theexponent
ial
funct
ionexi
sauni
quef
unct
ionwi
thspeci
albehav
ior
.Thati
s.

However
,incasewehaveef
(x)
,wecanuset
hechai
nrul
etoev
aluat
eit
sder
ivat
ive.Thati
slet
f(
x)=uandwewi l
lhav
eeu.
I
fY=euandU=f (
x)

Thus,
Example:
Find i
nthef
unct
ionsgi
venbel
ow

Solution
Directlyinserti
ngi nt hefor mula,weget
2.2.Integralcalculus: Techniquesandappl icati
onsofi nt
egral
calcul
us
Upt ot hi
st ime,wehav ebeenconcer nedwi t
hf i
ndi ngthederi
v at
iveofaf unct i
on. Howev er
,
therear esomepr oblemsi neconomi csthatr equi
reust ocomeupwi t
htheor i
ginalfuncti
ongiven
i
tsr ateofchange. Thesepr oblemsar ecommoni ntheareaofsoci alwelfare,dist
ri
butionof
i
ncomeofacount ryet c. Thet echniqueoff i
ndi ngoutt heor i
ginalfunctionbasedoni ts
derivativeisknownasi ntegrationorant ider i
vati
v e.Mor especif
icall
y,youwi l
llear
nt hatthe
conceptofi ntegrationist heexactopposi teofdi f
ferent i
ati
on.
2.2.1.TheI ndefiniteIntegral

Gi
vent
hepr
imi
ti
vef
unct
ionwhi
chy
iel
dst
heder
ivat
ive,
youhav
ethef
oll
owi
ngdef
ini
ti
ons.
Youhav
ethef
oll
owi
ngnot
ati
ont
hatdenot
est
hei
ntegr
ati
onof
wit
hrespectt
o,

Thesymbol i
scalledtheintegral
sign.Thepartisknownast hei
ntegr
and( thatist hef unctiontobe
i
ntegr
ated).Theparttell
sthattheoperati
onist obecarriedoutwithrespecttot hev ari
able.
Oneimpor t
antpoi ntthaty oushouldnotei sthatwhilet hepri
miti
vef uncti
oni nv ari
ablyy i
eldsa
uni
quederivati
ve, t
hederivedfuncti
oncanhav eaninf
initenumberofpossi blepr imi t
ivefunctions
thr
oughintegrat
ion.Why ?Because, i
fi
sani ntegr
alof,thensoisplusaconst antf ortheder iv
at i
ve
ofaconstantiszer o.

2.2.
2.Techni quesofI ntegr ati
on&t heGener alPowerRul e
Inthissect ion,Iwi llbrieflydiscusst herulesofi ntegrat
ionwhichshal lref
ertoincr
easingl
yfrom
ti
met oti
mewhi lewear edeal i
ngwi t
ht het echni quesofi ntegr
ation.
a.Theconst antr ul
e: Gi v
enconst antK&C,
Exampl e:
1.
b)Const antMul ti
pleRul e:for
Exampl e:
1.∫
c)Sum orDi f
ferenceRul e: For
Exampl e:
1.
d)Simpl ePowerRul e: Forf orn - 1.
e)IntegrationofLogar i
thmi c&Exponent i
al Funct ions
Likewhatwehav edonei nr elationt otheder iv at
iveofnon- algebr
aicfuncti
ons,herewewi ll
presentt hei ntegrati
onoft hel ogar i
thmi cf unct i
onsf i
rstaccompani edbyt hei nt
egrat
ionof
exponent i
alf uncti
ons.Concer ning t he int egration ofl ogar
it
hmi cf uncti
ons,let
’s begi
n wit
h
exampl e.Thegener alpowerr uleofi ntegrat ionappl i
esonlywhenn - 1;i.
e.
= iffn≠- 1.
But,thefollowi ngfor mul aappl ieswhenn=- 1.Thati s

=l
n|u|
+C

Weuseabsol
uteval
uetoprot
ectitfrom bei
ngnegat
ive&t
hewhol
elogar
it
hm undef
ined.Thi
s
for
mulai
sadi
rectr
esul
toft
hefactthat
Thus,

Simi
lar
ly,
whenUi safuncti
onofX,
.
Example:Ev
aluat
etheindef
ini
tei
ntegr
aloft
hef
oll
owi
ng.

1)
Sol
uti
on
1)Let4x+1=U

Thus,∫
Thesecondt ypeofnon-al
gebrai
cfunct
ionwesaw wasexponent
ialf
unct
ions.I
nrel
ati
ont
o
exponent
ialf
uncti
on,wehaveseenthat
Thus,
Moreover,wehaveseent
hatwhenUi safunct
ionofX,
thati
s,
giv
e .

2.
2.3.Def
ini
teI
ntegr
als

Letf
(x)beafunct
ionand“a”and“b”bereal
numbers.Thedefi
nit
eintegr
aloff(
x)overtheint
erval
fr
om X=atoX=bdenotedby isthenetchangeofananti
-der
ivat
iveoff(
x)overthei
nterv
al.Thus,
i
fF(
x)i
sanant
ider
ivat
iveoff
(x)
,thenwehav
e

Not
ethati
tisusedt
ofi
nddef
ini
tei
ntegr
alandnotant
ider
ivat
iveofaf
unct
ion.

Now i tisti
met olookatsomeexampl essothattheideaofdef
ini
tei
ntegr
alandf
undament
al
theorem ofcalcul
uswill
bewel lpl
antedi
nyourmind.
Exampl e:Eval
uatethedefi
nit
eintegr
alsoft
hefol
lowing
Solution

2.2.4.Conti
nuousMoneySt reams
You wi l
lformallybei ntroduced wi t
ht henoti
on ofconti
nuouscompounding ofint
eresti
n
monet aryeconomics. IfPdol larsisi
nvest
edatannualr
atercompoundedconti
nuousl
y,t
henit
s
valueFatt heendofty earsisgivenbyequation
. Sol vi
ngthisequati
onf orPi ntermsofFgivesthe

Wr i
tt
eni nthisfrom,theequat iont ell
sust hepresentvaluePofFdol l
arsthatwi
llberecei
vedt
yearsfrom now, assumingt hatinterestiscompoundedcont inuousl
yattheannualr
ater
.
Exampl e
Bethelknowst hatshewi llneedt oreplacehercarin3y ear. Howmuchwoul dshehavetoputin
thebankt odayat8%i nterestcompoundedcont inual
l
yi nordert ohav
et he$12,
000sheexpects
toneed3y earsfrom now?
Soluti
on
Weappl ytheboxedf ormula,remember i
ngthattheinterestrateof8% mustbewr i
tt
eninits
decimal f
orm.
.
2.2.
4.1.ContinuousMoneySt reamswi thi
naPer i
od

Therearemanysi tuati
onsi nbusi nessandi ndustrywher ei tisusef ultothi
nkofmoneyasf lowi
ng
conti
nuouslyintoanaccount . Forexampl e,theABC Companypl anst obuyanexpensi ve
machinewhi chitest i
mateswi llincreaset hecompany ’sneti ncomeby$10, 000pery ear . But
thi
sincomewon’ tcomeatt heendofeachy ear;itwil
lcomei ndribblesthroughoutt hey ear. As
amodelf orwhatwi llhappen, i
tisconv eni
entt othi
nkoft hemachi neasi fitwillproducei ncome
conti
nuously
. Ournextexampl er ai
sesani mpor t
antquest i
onf orthecompanyt oanswer .
Supposethattheannual r
ateofi ncomeofani ncomest ream i sR( t).
Supposefurtherthatthisincomest ream wi l
llastovertheper i
od0<t<T. I finterestisatr at
er
compoundedcont i
nuously,thent hepr esentvaluePVoft hisincomest ream isgivenby

Example
Supposet heABCCompanyt akesamoreopti
misti
cview andest
imatesthatthenew machi
ne
wil
lproduceincomeatt
herateR( t
)=10,000+200t. Whatisthepresentval
ueofthi
sincome
str
eam, agai
nassumi
ngint
erestat9%compoundedconti
nuousl
yandal i
fet
imeof8year
s?
Solut
ion

Letusmodi f
yt heprobl
em intheaboveexampleinanot
herwaybyassumi ngthatthemachine
wil
llasti
ndefi
nitel
y. Moregeneral
l
y,l
etusaskforthepr
esentvaluePVofaper petual(
meaning
i
nfi
nitel
ylong)incomest r
eam whichproducesincomeatanannualr ateR(t)
,assumingthat
i
nteresti
satanannual r
atercompoundedconti
nuousl
y. Theresulti
nthiscaseis

Here,
wef aceanewki
ndofi
ntegral
,onewi t
haninf
ini
teli
mit
. Suchint
egr
alsar
ecal
l
edi
mpr
oper
i
ntegr
alsandmustbegi
venacleardefi
nit
ion. Wedot hi
sbydef
ini
ng
Thispointleadsust ot heconceptknownasi mproperintegralswhichwewi llhandl
erightnow!
2.2.
5.Impr operIntegr als
Witht he count er
-revol uti
on ofnew cl assi
calmacr oeconomi cs,there has been a shiftby
economi st
st oconsi derani nfi
nitel
ylivedagent.Thisagenti swi del
yusedi nperfectforesight
-
model, r
ati
onalexpect ationmodel ,Ri
cardianequi
v al
encetheor em,endogenouseconomi cgr owth
models, amongot her s.Hence, sav i
ngandi nvest
mentdeci sionofhouseholdsi sgovernedbyt he
uti
li
tyfunctionagent smaxi mizethroughout,t
hatis,

Where
U(c)i
st heuti
li
tyfuncti
onthesocialpl
annermaximizes,
-
Bi sthediscountrat
eagent sassi
gntofutureconsumpti
on,
Cispercapitaconsumpt i
on
Listhesizeofpopul at
ion
Andδi stheinverseoftheinter
temporal
elasti
ci
tyofsubst
it
uti
on.

Nowwehav efacedsomewhatnew, butdef

ini
teint
egral
,problem i
nt hi
sfunct
ion. Suchtypeof
i
ntegrali
sknownasi mproperi
ntegr
als. Animproperi
ntegralisanintegr
alwhichisboundedby
ei
ther, , or )whereci
saconst ant
. Thisli
mitofint
egrati
oni si
nfini
te. Tofindouttheval
ue
ofthi
sintegral
usethef
ormula:

a)For

Thi
shol
dsi
ffi
scont
inuousov
ert
hei
nter
nal
[a,
)

b)I
ffi
scont
inuousont
hei
nter
val
(- b]
,then

c)I
ffi
scont
inuousoni
nter
val
(,t
hen

=
Note: Ineachthr
eecasesofabov e,i
fli
mitexi
sts,
thent
hei
mpr
operi
ntegr
ali
ssai
dtoconv
erge;
ot
herwisetheimproperi
ntegr
aldi
verges.

Exampl
es
1.Fi
ndthev
alueof anddet
ermi
nei
tsst
atusofconv
ergence.

Sol
uti
on

Sinceli
mitdoesnotexi
st,
thei
ntegr
aldi
ver
ges.
2.Find∫

Sol
uti
on

Si
ncel
i
mitdoesexi
st,
thei
ntegr
alconv
erges.

3.Nowiti
stimeforyoutocheckwhetheryouhavegraspedtheconceptornot
.Tot
hisend,
why
don’
tyouhandl
ethi
sprobl
em?Usey ourpaperandpenrightnow!

Defi
nit
ions
A matrixi
sdef i
nedasar ect
angul
ararrayofnumber s,parameters,orv ar
iabl
es.I
tisusuall
y
encl
osedinbracket
sandsometimesinparenthesesordoublever
ticalli
nes.
Themember softhearr
ayarer
efer
redtoast heelementsofthemat ri
x.
Thenumberofr owsandthenumberofcolumnsi namat ri
xtogetherdefinethedimensi
onofthe
mat
ri
x.I
famat
ri
xhasr
owsandcol
umns,
thent
hatmat
ri
xissai
dtobeofdi
mensi
on.

Considert heabov emat r

ices.Asy oucansee,mat r
ixconsi stsoft hr
eer owsandt hreecolumns.
Therefor e,itissaidt obeofdi mensi on.Comi ngtomat ri
cesand,y oucanseet hatbot har eof
dimensi on.Thati sbot hhav easi ngl
ecol umnandt hreer ows.Youwi llhaveadi scussionsuch
typeofmat ri
cesint henextsubSect i
on.Youshoul dr emembert hatwr i
ti
ngt hedi mensionofa
mat ri
x,ther ownumberal waysprecedest hecolumnnumber .
Ingener al,if
,
thenthemat r
ixissaidtohav erowsandcol umnsandi tisofdi mension.Thesubscr iptsofunder
eachel ementoft hemat r
ixshowt hecol umnandt her owwher eeachel ementbel ongs.
2.3.
1.Mat rixOperati
ons
a)Equal ityofMat ri
ces
Beforegoi ngintomat ri
xoper ati
ons( additi
on,subt
r act i
onandmul t
ipl
icati
on),l
etusf ir
ststudy
whati smeantbyequal i
tyoftwomat ri
ces.Supposey ouhav et womat ri
ces.Thet womat ri
cesare
saidt obeequali fandonl yiftheyhav et hesamedi mensi onandi denti
calelement si nthe
correspondi nglocati
onsi nthearrays.Considerthefollowingmat r
ices.

Itcan beconcl udet hat becauset het wo mat riceshav et hesamedi mension and i dent
ical
element s.Though,mat ri
cesandhav et hesamedi mensi on,theyar enotequalsi ncetheel ements
i
nt heirfir
strowaredifferent.Thus, weconcl ude.
b)Addi t
ionandSubt r
act ionofMat rices
Twomat r
icescanbeaddedi fandonl yiftheyhav ethesamedi mensi on(i.
e.t
heyhav et hesame
numberofr owsandcol umns) .Ifthi sconditi
onissat i
sfi
ed,thent het womat ri
cesar esai dtobe
conformabl eforaddition.Thesum oft hemat ricesisobt ainedbyaddi ngthecor responding
element softhemat r
ices.
Similarl
y,subtr
acti
onofamat ri
xf rom anot hermat ri
xispossi bl
ei fandonl yi
fbothmat r i
cesare
ofthesamedi mension( i.e.theyhav ethenumberofr owsandcol umns) .Thediff
erencebet ween
twomat ri
cesisobtai
nedbysubt ract i
ngeachcor respondingelement .

Gener
all
y,i
f
and
Then

Exampl es
Comput eandgiven
and
Solut
ion
Sincebot hmatr
iceshav
ethesamedi
mensi
on,t
heyar
econf
ormabl
eforaddi
ti
onand
subtracti
on.Thus,
wecancomput
eand.

Hence,

Findi
f
and
Solut
ion
Sincet
hemat
ri
ceshav
eident
ical
dimensi
on,
youcancomput
e.Hence,

b)Mat
rixMulti
pli
cat
ion
i
)Mult
ipli
cat
ionbyaScalar
I
fisar
ealnumberandi samat
ri
x,t
henmul
ti
ply
ingt
hemat
ri
xbyt
henumber
isr
efer
redt
oast
he
multi
pli
cat
ionofmat
ri
xbyscal
ar.I
nthi
scase,ev
eryel
ementoft
hemat
ri
xismul
ti
pli
edbyt
he
const
ant.

Examples
I
fand, t
henf
ind.
Solut
ion

I
fand,
find.
Sol
uti
on

i
i)Mul t
ipl
i
cati
onofMat rices
Whereasascal arcanbeusedt omul ti
plyamat rixofanydi mension,themul t
ipl
icati
onof
matri
cesi sdependentont hesati
sfactionofar equi r
ement .Supposet hatyouhavetwomat r
ices,
and,andy ouwantt of
indt hepr
oduct .Then,thepr oductoft hetwomat r
icesi
sdefi
nedi fandonly
i
fthenumberofcol umnsofmat r
ixist hesameast henumberofr owsofmat ri
x.Inotherwor ds,
i
nor dertomul t
ipl
ytwomat ri
ces,thef ir
stmat r
ix( the“ l
ead”mat ri
x)musthav etheasmany
col
umnsast herowsoft hesecondmat ri
x(the“lag”mat r
ix).Matr
icesthatful
fi
ll
thi
scondi t
ionare
sai
dt obeconf or
mablef ormulti
pli
cation.
Ifmat r
ixisofdi mensionandmat rixisofdi mensi on,thenmul tipl
icati
onofmat ri
xbymat ri
xis
possibl
ebecauset henumberofcol umnsofmat ri
x( thati
s)equalst henumberofrowsmat ri
x.

Whatist hedimensi
onoftheresulti
ngmatri
x?Thedi mensi
onoftheresul
ti
ngmatr
ixi
s.This
i
mpliesthati
thasthesamenumberofr owsasthefi
rstmatri
x(t
hati
s),
andthesamenumberof
col
umnsast hesecondmatr
ix(t
hatis).
Symboli
call
y, wher et
hesubscr
iptsdenot
edimension

Howcany oumul ti
plytwomat ri
ces?Sofar ,youhaveseenthecondit
ionst hatmustbef ul
fi
ll
edf or
mat r
ixmul t
ipli
cati
on.Now,y ouwi l
ldiscusst hepr ocedur
eofmul tipli
cati
on ofmat r
ices.I n
multipl
yi
ngt womat r
ices,youaddt hepr oductsofelementsoft herowsoft he1stmat rixand
elementsoft hecolumnsoft he2ndmat ri
x.Thesum oft heproductsofel ementsofthe1str ow
ofthe1stmat r
ixandelement softhe1stcol umnofthe2ndmat ri
xy i
eldst he1strow-
1stcol umn
elementforthepr oductmat r
ix.Si
milar
ly,thesum ofproductsofelement sof1str owoft he1st
mat r
ixandelement soft he2ndcolumnoft he2ndmat r
ixfor
mst he1str ow-2ndcolumnel ement
fortheproductmat r
ixandsoon.I notherwor ds,matr
ixmulti
pli
cati
onisr owbycol umn.

Forinstance,youhavetwomat ri
ces
and,
Si
ncet hecolumnnumberofmat r
ixequalstherow numberofmat ri
x,theyareconf
ormablef
or
multipl
icati
on.Hence,
=
Thepr oductyiel
dsanot hermat r
ixwithdimension.Theelement soft heproductmatri
xare
obtainedasf ol
lows.
(thatismul t
ipl
yi
ngt hefi
rstrow elementsofmat r
ixwit
ht hef i
rstcol
umncor r
espondi
ng
el
ement sofmat r
ixandthenaddingtheproducts)
(thatismulti
plyi
ngthefi
rstrowelement sofmatri
xwiththesecondcol umncorr
espondi
ng
el
ement sofmat r
ixandthenaddingtheproducts)
(thatismulti
plyi
ngthesecondr owelementsofmat r
ixwi
t hthefir
stcolumncorr
espondi
ng
el
ement sofmat r
ixandthenaddingtheproducts)
(thatismulti
plyi
ngthesecondr owelementsofmat r
ixwi
t hthefir
stcolumncorr
espondi
ng
el
ement sofmat r
ixandthenaddingtheproducts)

Examples
Given
 and, f
ind.
Solut
ion
i
sdef i
nedsincethecondi
ti
onformul
ti
pli
cati
onissat
isf
ied.Thati
st hecol
umnnumberofthe
fi
rstmat r
ix(t
hatis)equal
stherownumberofthesecondmat r
ix(t
hatis).
Theproductmat
ri
xwil
l
haveadi mension.Ther
efor
e

Given
and,fi
nd.
Solut
ion
Themat
ri
cesar
econf
ormabl
eformul
ti
pli
cat
ionandt
her
esul
ti
ngmat
ri
xhasadi
mensi
on.
Thus,

Given
andfi
nd,.
Solut
ion
Theconf
ormabi
l
itycondi
ti
onf
ormul
ti
pli
cat
ioni
ssat
isf
ied.Hence,

Gi
ven
,f
ind.
Solut
ion
Theconfor
mabi
l
itycondi
ti
oni
ssat
isf
ied.Hence,

i
i
i)Mul
tipli
cat
ionofVectors
I
fi
sancolumnv ect
orandisarowv
ect
or,
thent
hei
rpr
oducty
iel
dsamat
ri
xofdi
mensi
on.

Example
Given
and,fi
nd.
Solut
ion

Ther efor e, t
hepr oductofvectorsandy i
eldsanothermat rixwithdimensi on.
2.3.2.Det erminant sandInverseofaMat r
ix
Inthepr ev ioussect i
on,youhav ebeeni nt
r oducedtothebasi cconcept sofmat ri
xal gebra.Int
his
section,y ouwi l
lstudytheconceptofadet erminantandi tsusei ndet ermini
ngt hei nver
seofa
mat ri
x.
2.3.2.1.TheConceptofDet erminants,MinorandCof actor
a)Det ermi nants
Thedet er mi nantofasquar emat ri
xi sauni quelydefi
nedscal ar(
number )associatedwi t
hthat
mat ri
x.Thedet erminantofamat ri
xisdenot edby .Det erminantsar edef i
nedonl yf orsquare
mat ri
ces.Thedet erminantofanmat rixisknownasdet er minantofor der .How canwef indthe
determi nant ?Herear etheanswer s.
i
)Det ermi nant sofOr derOne
Letmat rix.Thati sitmat r
ixwithonlyoneel ement.Thent hedet erminantofmat ri
xis.

i
i)Deter
minant
sofOrderTwo
Supposematri
xisasquar
ematr
ixandi
tisgi
venas
Thedet
ermi
nantoft
hismat
ri
xisdef
inedast
hedi
ff
erenceoft
wot
ermsasshownbel
ow:

Thatisi
tisobtai
nedbymul ti
ply
ingthetwoel ementsint
heprinci
paldi
agonalofmat
ri
xandthen
subtr
act
ingtheproductoft
hetwor emaini
ngel ement
s.Si
ncemat r
ixi
samatri
x,t
hedet
erminanti
s
cal
ledasecondorderdeter
minant(det
erminantofordert
wo).

Exampl
es
a)Fi
ndt
hedet
ermi
nantoft
hemat
ri
x

Sol
uti
on
Usingt
hef
ormul
agi
venabov
e,t
hedet
ermi
nantofmat
ri
xi
s

i
ii
)Determinant
sofOrderThr
ee
Adeterminantofor
derthr
eei
sassoci
atedwi
thamat
ri
x.Gi
ven

t
hedet
ermi
nanti
sdef
inedas

Here,y ouhav easum threeterms.Thev erti

calbracketinthefi
rstt
ermi st
hedet erminantofthe
mat r
ixobt ainedaft
err emovingt he1str ow andt he1stcol umn.Thesecondbr acketisthe
determi nantoft hemat ri
xobt ai
nedbyr emov i
ngt he1str ow andthe2ndcol umn.Thev ert
ical
bracketi nt hethi
rdtermi st
hedet er
minantofthemat ri
xobtai
nedbyr emovi
ngt he1str owand
the3r dcol umn.Theel ementsi neachterm arethe1st ,2ndand3rdelementsoft hef i
rstr
owof
mat r
ix.Inot herwords,eachtermi saproductofaf ir
strowelementandapar t
icul
arsecondor der
determi nant.Theref
ore,youhav e

Af
terr
ear
rangi
ng,
youget

LaplaceExpansi
on
Determinantsofamatri
xcanbeobtai
nedalternati
vel
ybyusingtheLapl
aceexpansion.Inthi
s
approach,anytwor owsofthemat r
ixarewr it
tenagai
nandf rom anexpandedmat ri
x.The
i
ll
ustrat
ionthatfol
l
owsshowshowthedeter
mi nantcanbefoundusi
ngthi
sapproach.Consi
dera
matrixgivenbel
ow.

Wr
it
ingdownt
hef
ir
stt
wocol
umnsagai
n,y
ouget

Then,youmul
ti
plytheelementsalongthethr
eelinesf al
l
ingtot
heright(
alongt
hedownwar
d
poi
nti
ngarr
ows)andgiv
ealltheseproduct
saplussign.Thus,y
ouhav
e

Si
milar
ly,
youmul
ti
plyelementsal
ongt
hethreel
inesr
isi
ngtotheri
ght(
alongt
heupwar
dpoi
nti
ng
ar
rows)andgi
veal
ltheseproduct
sanegat
ivesi
gn.Asaresul
t,y
ouget

Thesum oftheabovet
wot
ermsi
sexact
lyt
hesameasexpr
essi
onf
ort
hedet
ermi
nantofmat
ri
x
shownabove.Thati
s

Examples
Given

,
comput
ethedet
ermi
nant
.
Sol
uti
on

b)MinorandCofact
or
Letusbeginwi
ththeconceptofmi
nor
.
Supposematri
xi
sgivenas

Hence,

Del
eti
ngthe1strowandthe1stcol
umnofthedeter
minantofmatr
ixgi
vesthesubdeter
minant.
Thi
ssubdeter
minanti
scall
edt
heminorofel
ementanditi
sdenot
edby . Ther
efor
e,theminorof
el
ementi
swrit
tenas

I
ngener al,ifmat ri
xisasquarematrixofdimensi onormor e,t
hencanceli
ngt herowandt hecolumn
ofthedet erminantoft hatmatri
xgi vest hemi noroft heelementwhi chi sasubdet er
mi nant
denotedby .
Hav i
ngdi scussedwhatami noris,itisti
met omov eont otheconceptofcof actor.Thecof actor
denotedby i
sobt ai
nedby
wher eisthemi noroftheel ement obtai
nedbycancel i
ngther ow andt hecol
umnoft he
determinantofapar ti
cul
armatri
x.andr eferrespecti
velytotherowandt hecolumn.
Therefore,thissuggeststhatacofactorisami norwit
hapr escr
ibedalgebrai
csi gnattachedt oit.
Forinstance, consider
ingmatri
xgivenabov e,thecofactor is

Si
mil
arl
y,t
hecof
act
or

Notet hatthemi nori

sobtai
nedbycanceli
ngt he2ndrowandt he3rdcolumnoft hedet er
mi nant
mat r
ix.
Thedet erminantofamat ri
xcanbefoundbyusi ngtheconceptoft hecof actor.I
nt hi
scase,t he
determinantofamat r
ixi
sobtai
nedbyf i
rsttaki
ngeit
herar ow oracol umnoft hatmatrixand
multi
plyingeachel ementi
nthatroworcolumnbyi t
scofact
or.Thenaddi ngt heseproductsyields
thedet erminantoft hematri
x.Fori
nstance,youcancalculatethedet erminantofmat r
ixgiven
aboveasf ol
lows.Supposeyoutaket
hef i
rstr
ow.Thus,t
hedet erminantofmat rixi
s

Si
nce

Ther
efor
e,

Af
terr
ear
rangi
ng,
youget

Thi
sexact
lywhaty
ouhadwhenwedef
inedt
hedet
ermi
nantofamat
ri
x.

Exampl
es
Comput
ethecof
act
oroft
heel
ementatt
he2ndr
ow-
2ndcol
umnofmat
ri
x
Soluti
on
Fir
styouneedtodetermi
netheminoroft
heelementatt
he2ndrow-2ndcol
umn.Thus,t
he
minorinthi
scaseisobt
ainedbycancel
i
ngthe2ndr
ow-2ndcol
umn.Ther
efore,

Hence,
thecof
act
orwi
l
lbe

2.3.2.2.TheInver
seofaMat r
ix
Ifmat ri
xisasquar emat ri
x,thent heinver
seofthatmatri
xdenot edbyexi st
sifitsat
isf
iesthe
proper t
yt hat
.Ifthei
nv er
seexists,thenmatri
xi
ssaidtobeinvert
ible.Matr
ixisal
soaninverseof
andhencebot hmat ri
cesar ei
nversesofeachother.Not
ethatiti
sonl ysquarematri
cesthatcan
havei nverses.
How doy oufindtheinverseofamat ri
x?Theanswerfoll
ows.Supposey ouhav ethefoll
owing
mat r
ix:

Infi
ndingtheinver
seofmat r
ix,t
hefir
ststepistoobt ai
nthedeterminant
.Thati
s.Secondl
y,y
ou
needtofindthecofactorofallel
ement sofmat r
ix.Ar
rangi
ngthem asamat ri
xgivesyouwhatis
knownasacof actormat r
ixdenotedby.Thatis
wheredenotethecofactorsoftheelementsinmatri
x.
Thir
dly,youtakethet r
ansposeoft hecof act
ormat ri
x.Thi
stransposemat r
ixi
sknownast he
adj
ointofmatrixwhichisdenotedby.Hence, wehav e

Thelaststepinfindi
ngthei
nver
seofmat
ri
xi
sdi
vi
dingt
headj
ointofmat
ri
xbyt
hedet
ermi
nant
.
Ther
efore,
itfol
lowsthat

Thei nverseofamat rix,say, exitsifandonl yifthedet ermi

nantoft hatmatr
ixi
sdiff
erentfr
om
zeroi.e.,
.Otherwise, with,thei nversebecomesundef i
nedandhence, noneedtoproceedwiththe
subsequentst epsofmat ri
xi nv ersi
on.
Herei sthesummar yoft hest epsi nv ol
vedi nmat r
ixinversi
on.Gi
venmat r
ix
fi
ndt hedet erminantofandpr oceedt ot henextstepsi f
fi
ndt hecof actormat rixfor
obtai
nt headj ointofmat ri
x
multi
pl ytheadj oi
ntofmat rixby andt hisgivesy outheinverseofmatri
x
Exampl es a)Gi ven
,
findthei nverse.
Soluti
on
First,f
indthedet erminant .Thus, y
ouhav e
(
Since,thei nverseformat ri
xisdef i
nedandhence, youcanproceedt ot
henextsteps)
Then, determi ni
ngt hecof act ormat ri
x,youget

Notet
hatsi
ncethemat
rixi
sa,themi
norofeachelementi
nthi
scasei
sadet
ermi
nant
.
Tr
ansposi
ngthecof
act
ormatr
ix,wegett
headjoi
ntmatri
x

Last
ly,
div
idi
ngt
headj
ointmat
ri
xbyt
hedet
ermi
nantgi
vesust
hei
nver
semat
ri
x.Thati
s

2.3.
3.Mat rixRepresent
ati
onofLinearEquat ions
Hav i
ngstudi edthebasicsofmatri
xalgebraandmat ri
xinversion,i
tisnowt i
met oreviseandyou
comeacr osst heappli
cati
onofmat r
ixalgebr ainsol
vingsy stemsofl i
nearequations.Youare
expectedt or evi
set heInver
semet hod,t heCr amer’
sRul e,andt heGauss- Jor
danEl imi
nat
ion
techniqueasy ourreadi
ngassignmentsfrom l i
nearal
gebracour se.
Mat r
ixrepresentati
on
Supposey ouhav ethef
oll
owingsystem oflinearequat
ions.
Thesy
stem canbeexpandedi
ntomat
ri
xfor
m as:

Now,youcancal
cul
atet
heunknownv
aluesbyt
hef
oll
owi
ngequat
ionusi
nganyt
heabov
ethr
ee
method:

2.3.4.I nputoutputanal ysis( Leont i

efmodel )
Input–out putanal y sisat tempt st oest ablishequi l
ibri
um condi ti
onsunderwhi chindustri
esaan
economyhav ej ustoutputt osat i
sf yeachot herdemandsi nadditi
ont of inal(outside)demand.
Commonl yspeaki ngt hepr oduct ionofonegoodr equi r
edt heinputofmanyot hergoodsas
i
nt ermedi atei nputsi nt hepr oduct ionpr ocess.Hence,t he,totaldemandf orapr oductist he
summat i
onal lthei nt ermedi atedemandf orthepr oductpl usthef inaldemandf orthepr oduct
arisi
ngf r
om consumer s, i
nv estors, gover nment ,expor ters,etcasult
imat euser s.
Assumei n hy pothet icaleconomyt her ear en-number sofsect orst hatpr oducet heirown
homogenouspr oduct sx1x2, -
--
--,Xn.
Ther ef ore,totaldemand( x1)=demandf rom intermedi ateusersintheeconomy+ul ti
mat eusers
demandf orthei rfi
nal consumpt ion.
i
,e.x1 = x11 +x12+ ……. .+x1n+f 1
X2=x21+x22+…….+X2n+f 2
“ “ “ “
“ “ “ “
Xn= xn1xn2 +….+xnn+f nn
Wher e, xi -r eferst otal outputpr oducedbyf irmi
Xij –r eferst otal out putpar tpr oducedbysect ori andusedbysect orj or
t
hedel iveryofsect ori’soutputt osect orjasi
nput
fi–r efer sout putr equi rementofsect ori forfi
naldemand.
Let ‘
si ntroduceanewv ar i
abl eintot hemodel .
aij=, wher eai j–r eferscoef fi
cientmat ri
xel ement sor
ani nputr equi rementbysul furj from sectorj topr oduceoneuni tBxj
Ther ef ore,from t heabov erelationshi p

Whenwecondenset
heabov
eequat
ion,
weget

X=( I
-A)
-1F
Sobyusi ngtheaboveformulawecancalculat
et hetotaloutputpr
oducedbyeachsect
orst
hat
meetneedsofi nt
eri
ndustry(
int
ermedi
ate)andfinalusers.
Example
Givenatechnical
coef
fi
cientmatri
xandthevectoroffinaldemandas

A=,
F= Fi
ndt
heequi
l
ibr
ium out
putl
evel
soft
hesect
orsout
put
s
Sol
uti
on

Sot hedet ermi

nant( I–A)=0. 18
Adj( I–A)=cof actorstr
anspose( CT).
2.3.
5.Li nearProgrammi ng
Upunt ilnow wehav ebeendeal ingwi t
ht heclassicalapproachinor dertosol v
eopt i
mi zation
problems – concept si n cal
culus Here up on we i ntr
oduce non – cl assi
calappr oach of
opti
mi zati
oncalledmat hemat i
calpr ogr
ammi ng.Linearpr ogrammingdeal swi t
hmat hemat ical
progr ammi ngproblemsi nwhichbot hobjecti
vefunctionandconstrai
ntsar ealll
i
near.
Theobj ecti
veofsuchpr ogrammi ngi st
odet erminet heoptimalall
ocationofscar cer esour ces
amongcompet i
ngact i
v i
tiesintheeconomy .Itdeal scommonl ywi thopt i
mizationofl inear
funct i
onsubjecttomor et hanonel i
nearinequali
tyconstrai
nts.
Inanyl i
nearpr
ogr ammingpr obl
em, therearethreeessent i
ali
ngredi
ents.
Theobj ect
ivef
unct i
on
Asetoft echni cal const rai
nts,and
Asetofnon- negat ivityr estri
ctions.
I
nsol v ingLi nearpr ogr ammi ngpr oblems, wehav ethr eecommonl yusedmet hods
Graphicappr oach
Simplexmet hod
Dualityappr oach
I
tisknownt hatGr aphi cappr oachi sappl i
cabl eonlyf orthecaseoft wounknowns,butsi mpl ex
met hodi sanal gebr aicmet hodoff i
ndi ngt heextremepoi ntsofnon–gr aphableconstrai
nt s
whichar esupposedt or esultinanopt imal v alueoftheobj ecti
vef unction.
Dualityappr oachi sf requentlyusedmet hodt osolv
emi nimizati
onpr oblems.Ev erymaximizati
on
/minimi zat i
onpr obl em hasacor respondi ngmi ni
mi zati
on/maxi mizati
onpr oblem.Theor i
ginal
problem i scal l
edt hepr imal andt hecor respondi ngpr oblem i
scal l
edt hedual .
Everydual problem canbet r
ansf or medi ntoi t
spri
mal counterpartvicev ersa.
Thus, forgi venl inearpr ogrammi ngpr obl em wecanuseei theroft heappr oachestocalcul
atethe
optimal values.
DearDi stancel ear ner s!Inor dert or ef r
eshy ourmemor ysolvet hefollowingl i
nearpr
ogrammi ng
problem.
Exampl e:-
GivenaPr imal pr obl em
Mi nC=20Z1+30Z2+16Z3
Subj ectto: 2.5Z1+3Z2+Z3>=3
Z1+3Z2+2Z3>=4
Z1, Z2, Z3>=0
Formt hedual count erpar toft hepr obl em
Soluti
on
Thedual probl em i s=MaxC*=3x1+4x2
Subj ectt o 2.5X1+X2<=20
3X1+3X2<=30
X1+X2<=16
X1, X2>=0.
CHAPTERTHREE
DERI VATI VEI NUSE.

3.1.Hi gherOr derDer i

v ati
ves:
3.2.1.Concav i
tyandConv exityofaFunct ion
Sincet heder i
vativeofaf unct ionisal soaf unct ion,wecanwr it
eandi tsderi
vati
ve,.Wecal lthe
fi
rstder ivativef unct i
onoft hef unct i
onandt heder i
vati
veofthefir
stder i
vati
ve,
ort hesecondder ivativef unct i
on.Si ncethesecondder i
vat
ivei
sal soaf uncti
on,wecanal sof i
nd
i
tsder i
vat iv
e.
Wecal lt het hirdder ivativef unctionandwr iteitas or
toindi catet hatt hi
sf unct i
oni sfoundbyt hreesuccessi veoperationofdi f
ferenti
ation,starti
ng
witht hef unction.Ofcour se,t hispr ocessmaycont inueindef
ini
telyandsoweuset hegener al
notat i
on t oi ndicatet he der ivat
iveoft hef unction.
Exampl e: Findthef irstf ourder ivati
v esoft hefunct i
on.
i
ft hef irstt wo der ivativesofaf uncti
on exi st,wesayt hef unct i
on istwicedi ffer
ent i
able.
Concent ratingatf i
rstandsecondder ivat
iveleadsust otheconceptofconcav it
yandconv exity.
Consi dert hef unct i
onf orthedomai n.On, thisfunct i
onisupwardsloppi ngandwecanseet hatits
sl
opei ncr easesasxi ncr eases.Thi smeanst hatt hefir
stderi
vati
vef uncti
onisincreasinginx, and
sot heder i
vativeoft his,thef unction’ssecondder i
vati
ve,mustbeposi ti
vevaluedforal l
.Upon
fi
ndi ngt heder i
v ati
v es,weget and.Thust hesecondder i
vati
veisi ndeedpositi
vef oranyv al
ue
ofx.
Figur e3.4: Funct ionfx=x^ 2( x≥0)andi tsDer i
vati
v es
Now considert hegraphoft hesamef uncti
ondef i
nedon.Onthi
ssetofvaluest
hef
unct
ioni
s
negati
vel
ysl oped.
on.Thegr eaterthevalueof,thel
esssteepisthecurvei
nabsol
utev
alues.
Figur
e3.5:Funct i
onfx=x^2forx<0anditsderi
vati
ves

Definingt hisf unctionont hedomai nR,weseet hatt hesecondder i

vati
veisposit
ivethroughout
.
Afunct i
onwi tht hisshape, asdet er
mi nedbyt hesecondder ivati
vebeingpositi
ve,i
sconv ex.
Definiti
on:At wi cedi f
fer
entiablefunctionisconv exi fatallpointsoni t
sdomai n.
Accordingtothe
abov edef init
ion, alinearfunctionisconv ex.Toexcl udel i
nearf uncti
ons,wecomet otheconcept
str
ictconv exity.Thisisdonebyr eplacingtheweeki nequalit
y wi t
hstri
ctinequal
it
y(>).
Definiti
on: At wicedi ffer
entiabl
ef unctionisstri
ctl
yconv exif exceptpossi bl
yatasinglepoint
.
Noticet hatt hef unctionhast hesecondder i
vativewhi chisposi tiv
ef orallxexceptwher ethe
second der ivativebecomeszer o.Thi sfunction is,howev er,stri
ctl
yconv ex,and hencet he
qualif
icationi nt heabov edef i
niti
onr equir
ementexceptpossi blyatonepoi nt.
Wemayhav est ri
ctlyconvexf uncti
onswi thmonot onicincreasing,monot oni
cdecreasingornon-
monot onicfunct i
ons.

Fi
gur
e3.
6:I
ll
ust
rat
ionsofMonot
oni
candNon-
monot
oni
cFunct
ions

Def i
niti
on:
At wicedi
fferentiabl
ef unct
ioni sconcaveifonallpoi
ntsofisdomain.
Asf orthecaseofconv exf unctions,alinearfuncti
onsatisfi
esthedefi
niti
onofconcavi
tyforall
).
Toexcl udeit
,weshouldr edef i
neitasst ri
ctconcavit
y.
Definiti
on:Atwicediff
erentiablefuncti
oni sstrict
lyconcaveifonallpoint
sofitsdomainexcept
possi bl
yatasinglepoi
nt.
Alt
er nati
vel
y,si
ncemul ti
plyi
ngaf uncti
onby( -
1)reversestheinequal
i
ty,wecouldsaythatf(x
)is
(st
rictly
)concaveif–f(
x)i s(stri
ctl
y)conv ex.
Af unct i
onwhosesecondder ivat
iveissomet i
mesposi ti
veandsomet i
mesnegativeisneit
her
convexnorconcaveev er
ywhere.Howev
er,wecansomet
imesf
indi
nter
val
sov
erwhi
cht
he
funct
ionisconv
exorconcave.
Example: on
Solut
ion
and

Thus,sincefor and f or,thef unctionisconv exon[0,

3]andconcav e[3,+].
Activi
ty:Show t hatthepr oduct ionf unctionf x=-
2/3x^3+10x^ 2+5xhasbot haconcav eanda
convexsect i
on.
Note:Pr oductionf uncti
onsar econcav emeansaddi ti
onall aborunitsy i
eldlesserandl esser
productandhencear ether eflecti
onofdi minishi
ngmar ginalproductlow.I fproducti
onfunctions
areconv ex,i
tbecomesdi f
fi
cul ttoexplainiti neconomicbackgr ound.
Exampl e:Asingl einput,x,i susedt opr oduceout puty.show t hati
ft hepr oduct
ionfunct
ioni s
thenthecostf unct i
on,c(y)isconv exwhilet heproducti
onf unctionisconcav e.
Soluti
on
Productionfunct ion
Fir
stder i
vati
ve
Secondder iv
at i
v e
Thus,ifxisgreat erthanzer o,sot hepr oduct i
onfuncti
onisconcav e.
Taket heshort-runcostf unction
intermsoff ixedcost andv ari
ablecostxneededt opr oduceout puty .Butsince andwecan
writ
et hecostf unctionintermsofy .

for
And, sot hecostfuncti
onisconvex.
Exampl e Given , det
erminet
heslopeofmarginal
costatx=7.
Solution:
c’
(x)=3x2- 18x,measurestheslopeoftot
alcostcur
v eandt ofi
ndt heslopeofmarginalcost
curvet aket hederi
vati
veofMC,whichi
sthesecondor
derder i
vat
iveoftot
alcostf
unct
ion.
C”(x)=6X- 18
C”(7)=6[ 7]-18=24.

3.2.2.LinearApproximat
ion
Toav oi
dwor kingwithacompli
cat
edfunct
ionwesomet i
metryt
ofindasimpl
erfunct
ionthatcan
appr oximatetheorigi
nal
one.
Sincel inearfuncti
onsareeasytowor kwith,wefir
sttryt
of i
ndal i
nearapproxi
mationtoa
funct i
on.
Supposet hefuncti
on isdif
fer
ent
iabl
eat,whatist
heequati
onofthet
angenttothegraphofthe
funct i
onatthepoint.

Fi
gur
e3.
7:I
ll
ust
rat
ingGr
aphf
orLi
nearAppr
oxi
mat
ion

Giv
en and
wher
ebissl
ope
Thi
sistherel
ati
onforthet
angentstr
aightli
ne.
Theequati
onofastrai
ghtl
inepassi
ngt hr
oughandhavi
ngaslope
i
s.
Ther
efore,
theequat
ionofatangentt
ot hegraphof t
othepointi
s

I
fweappr oximatethegr
aphofbyit
stangentl
i
neat,
ther
esul
ti
ngappr
oxi
mat
ioni
scal
l
edLi
near
Approximati
on.
Examples
Fi
ndt heli
nearappr
oximat
ionofaf
uncti
on about
Soluti
on.

I
f

Butt
heact
ual
val
ueoft
heor
igi
nal
funct
ionat i
s.

Findt hel inearappr oxi mationofaf unction

at
Answer :
Findt hel inearappr oxi mationv aluefor
Itcanbewr itt
enasaf uncti
onof as;
at
Tol i
near l
yappr oxi mat eitweuset hef ormulaf x=fa+f^'
a(x-a)
fx=x^ 50⇒f 1=1^ 50
f^'
x=50x^ 49⇒f ^'
1=501^ 49=50
fx=1+50x- 1=50x- 49
Evaluat ingt hisr esul taroundx=1suchasx=0. 001, wegetal i
nearesti
mationof
fx=50x- 49
f1.001=501. 001- 49=1. 05.
But ,
t heact ual valueort hecor r
ectv al
ueoff x=x^ 50atx=1i s1.0512.
TheDi fferent i
al Funct ion
GivenY=f x,thenweknowt hatdY/ dx=f^'
x. I tf oll
owst hatdY=f ^
'xdx-
thi
sisthediffer
ent
ialoft
he
funct i
onf (x).Itmeasur esthechangei nt hev al ueofYr esult
ingfrom asmallchangeintheval
ue
ofx.
Let’scompar edYandΔY.
dYanddxar epr opor t
ionaltoeachot hersincedY=f ^'
xdx.
Letxchangesf rom x→x+dx, t
henΔY=f x+dx-fx.
Givent hel i
nearAppr oximat ionformul afx≅fa+f ^ '
ax-axcl osetoa,
Subst i
tutingabyxandxbyx+dx, wehav e
fx+dx=f x+f^'x(x+dx- x)
fx+dx- fx ΔY=f ^ '
xdx dY.
.
Ther efore, inlinearappr oximation,weappr oximat e(esti
mat e)ΔYbydx.

Whenxchangesbydx,t
heact
ualchangei
nYi
sgi
venbyΔY(
mov
emental
ongt
hecur
ve)
.But
whenweconsi deratangentorli
nearapproxi
mationori
fweappr oxi
matethefunct
ionbyits
tangentatpointP,thechangeinthevalueofYbecomedY( movementalongthetangentl
ine
from PtoR.
dY≅ Y
ThelinesegmentRQmeasur esthedist
ancethatshowst
hedi
ff
erencebetweenΔYanddY( er
ror
term).Theer
rortermtendstobezeroasdY→0.

Rul esofDi ffer ent i

als
Allt her ul esofdi fferentiationcanbeappl i
edi nt hecaseofdi fferenti
als.
Giv ent wodi ffer ent i
alfunct i
ons( fandg)t hef oll
owi ngrul
eshol dtrue:
df∓g=df ∓dg
daf +bg=adf +bdg
dfg=f dg+gdf
df/g=gdf -fdg/ g^ 2f org≠0.
Allot herr ulesofdi ff
er ent i
ati
ons,di scussedi ny ourCalculusf orEconomi stscourseyouhav e
takenear lier,ar eal lappl icablei nt hi
scont extt oo.
Exampl e:
Giv enY=( 2x-5) ^2, then
dY=f ^'xdx
f^'
x=22x- 52=42x- 5
dY=42x -5dx.
Pol ynomi al Appr oximat ion
Appr oxi mat ionbyal inearf unctionmaybei nsuf f
ici
entl
yaccur ate,soi ti
snaturaltotr
yquadrati
c
appr oxi mat ionorappr oximat ionbypol ynomi al ofhigherorder .
Quadr aticAppr oximat ion
Wehav eagener alquadr aticformul a
Px=A+Bx- a+C( x-a)^ 2 (forxcl oset oa)
Int heabov eequat i
onswehav et hreeunknowncoef fici
ent sA,B andC.Wei mposethree
condi tionsont hepol ynomi alt
odet erminet hev alueofthet hreeunknownv ari
ables.
Assumeatx=a,
fa=Pa
f^'
a=P^ 'a
f^'
'a=P^ '
'(a)
Giv enPx=A+Bx- a+Cx- a^2
P^ 'x=B+2C( x- a)
P^ ''
x=2C
Whenx=a,
fa=A
f^'
a=B
f^'
'a=2C⇒C=1/ 2f ^''
(a)
Hence, thequadr aticappr oximat ionoff (x)isgi venby
fx≅Px=A+Bx- a+Cx- a^2
=fa+f^'ax-a+1/2f^
''
(a)(x-a)^
2
Thi si sj ustsi mi lart ot hel inearappr oximat i
onexceptoneaddi ti
onal t
erm
1/2f ^ '
'(a)( x-a)^ 2
Exampl es
Findt hequadr aticappr oxi mationt ofx=3&xabout x=1.
Sol ution
fx=x^ 1/ 3
f1=1
f^'
1=1/ 3x^ -2/ 3=1/ 3
f^'
'1=- 2/ 9x^ -5/ 3
=- 2/9
Then, thequadr at i
cappr oxi mat i
onwi llbe
Px=1+1/ 3x- 1-1/ 9( x-1)^2
I
fx=1.03,then3&1. 03isquadrat i
call
yappr oxi
matedtobeequal to
3&1.03=1+1/31.03-1-
1/9(1.03-1)^2=1.0099whichisnear ertotheexactval
uethantheprev
ious
l
inearapproximation.
HigherOrderAppr oxi
mat i
on
Functi
onsinhi gherorderderiv
at i
vescanbebet t
erapproximatednearapoi
ntbyusi
ngpoly
nomi al
ofhigherdegree.
Supposewewantt oapproximat eaf uncti
onfxaroundx=aoft heform
Px=A_0+A_ 1x-
a+A_ 2(
x-a)^2+A_ 3(
x-a)^3+…+A_n(x-
a)^n n+1Coeffi
cients

I
thasn+1coef ficient s,sowei mposen+1condi ti
onsonthepolynomialt
odeterminet
hev al
ueof
n+1unknowncoef fi
ci ent s.
fa=Pa
f^'
a=P^ 'a
f^'
'a=P^ '
'
a
f^na=P^ n( a)
Thei mpl icat ioni sf (
x)andi tsn- deri
vat i
vesareequalt
oP(x)anditsn-
deri
vat
ives.
Px=A_ 0+A_ 1x- a+A_ 2( x-a)^2+A_ 3(
x-a)^3+…+A_n(x-
a)^
n
P^ '
x=A_ 1+2A_ 2x- a+3A_ 3(x-a)^2+…+nA_ n(x-
a)^
n-1
P^ '
'x=2A_ 2+6A_ 3( x-a) +…+n( n-1)A_n(x-a)^n-
2
P^ '
''
x=6A_ 3+…+nn- 1(n- 2)A_ n(x-a)^n-
3.
P^ (n)x=nn- 1n- 2…2( 1) A_ n
Whenx=a,
Pa=A_ 0=0! A_ 0
P^ '
a=A_ 1=1! A_ 1
P^ '
'a=2A_ 2=2! A_ 2
P^ '
''
a=6A_ 3=3! A_ 3
P^ na=nn- 1n- 2…. 21A_ n=n! A_ n.
Ther ef ore,
Pa=0! A_ 0=f a⇒A_ 0=f (a)/0!
P^ '
a=1! A_ 1=f ^'
( a)⇒A_ 1=f '(
a) /
1!
P^ '
'a=2! A_ 2=f ''
a⇒A_ 2=f '
'(a)/2!
P^ '
''
a=3! A_ 3=f '
''
a⇒A_ 3=f '
''
(a)/3!
Ty peequat ionher e.
.
P^ na=n! A_ n=f ^(n)a⇒A_ n=f ^(
n) (
a)/n!
Subst i
tut i
ngt hev aluesont hecoef fi
cientswegetthefoll
owingapproxi
mati
ont ofxatx=abya
n^thdegr eepol y nomi al.

f
x≅P(
x)=f
af^
'a(
x-a)
/1!
+f^
''
a(x-
a)^
2/2!
+f^
''
'
a(x-
a)^
3/3!
+…+f
^(n)
a(x-
a)^
n/n!

Exampl e:Findthethir
ddegr eepol ynomialapproximati
onof
fx=&1+xaboutx=0.
Soluti
on:
f0=1
f^'
0=1/2
f^'
'
0=-1/4
f^'
'
'0=3/8
Px=faf^'
ax-a/1!+f
^'
'ax-a^2/2!
+f ^'
^'
'ax-a^
3/3!
=1+1/2x+-1/ 4/
2x^2+3/ 8/6x^3
Therefore, Px=1+1/ 2x-1/8x^ 2+1/16x^3
EstimationofFunct i
ons( Mcl auri
nandTay l
orseries.
)
Thisisexpansi onofaf unctionf (x)
byMcl auri
nandTay lorser
ies.

Mclaur
inSeri
es
Thi
si sexpansi
onoft hefunct
ionf
xar
oundx=0. Let
’sconsi
deranexpansi
onofn^
thdegr
ee
pol
ynomialbyMclaur
inser
ies.
Px=a_ 0+a_ 1x+a_ 2x^ 2+a_ 3x^ 3+…+a_ nx^ n
Successi vedi ffer ent iat i
onofaf unct iony i
elds:
P^ '(x)=a_ 1+2a_ 2x+3a_ 3x^ 2+…+na_ nx^ n-1
P^ ''
(x)=2a_ 2+6a_ 3x+…+n( n-1) a_nx^ n- 2
P^ ''
'(x)=6a_ 3+…+nn- 1( n-2) a_ nx^ n-3
P^ (n) (x)=nn- 1n-2n- 3…32( 1) a_ n
Ev al uat i
ngt heMacl aur i
nexpansi onar oundx=0y ields
P0=a_ 0=0! a_ 0⇒a_ 0=P( 0) /0!
P^ '0=a_ 1=1! a_ 1⇒a_ 1=P' (0) /1!
P^ ''
0=a_ 2=2! a_2⇒a_ 2=P' '
(0) /2!
P^ ''
'0=a_ 3=3! a_ 3⇒a_ 3=P' ''
( 0)/3!
P^ n0=n! a_ n⇒a_ n=P^ (n) (
0) / n!
Ther ef ore, theabov epol y nomi al f
unct ioncanber ewr i
tt
enas
Px=P0/ 0!+P^ '
0/1! x+P^ '
'0/2! x^ 2+P^ ''
'0/3! x^3+…+P^ (n)0/n!x^
n
Thi spowerser i
esr epr esent at i
oni scal ledMcl aurinSer i
es.
Not e: Expansi onusi ngMcl aur i
nser iesy i
eldsexact lythesamefunct
ionastheor
igi
nal
one.
Exampl e1: Findt heMcl aur i
nser i
esf ort hef ol
lowingf uncti
on
fx=8+36x+12x^ 2+x^ 3
Sol ut i
on
Giv ent heabov ef unct ion, ev aluatingi tar oundx=0basedonMcl auri
nser
ies,
weexpandunt i
l thet hirdder ivativ
e, becauset hehi ghestpoweri
s3.
f0=8
f^'0=36+240+30^ 2=36
f^''
0=24+60=24
f^''
'0=6
Int heMacl aur inser ies
fx=f 0/ 0!+f ^'0/1!x+f ^'
'0/ 2!x^ 2+f ^'
'
'0/3! x^3
=8/ 0!+36/ 1!x+24/ 2!x^ 2+6/ 3!x^3
fx=8+36x+12x^ 2+x^ 3.
Tay lor’sSer ies
Thi si sexpansi onofaf unct i
onar oundx=x_ 0.Supposenow wewantt oexpandannthdegr
ee
pol y nomi al atsomear bitrar ypoi ntx=x_ 0.
fx=a_ 0+a_ 1x+a_ 2x^ 2+a_ 3x^ 3+…+a_ nx^ n
Rewr i
tingt hef unct ioni nt er msoft hev ariable(x-x_0),weget
fx=a_ 0+a_ 1( x-x_ 0)+a_ 2x- x_ 0^ 2+a_ 3x- x_0^ 3+…+a_ n(x-x_0)^
n
Let ’sf indt hesuccessi veder ivativesasf oll
ows
f'
x=a_ 1+2a_ 2(x- x_ 0)+3a_ 3x- x_ 0^2+…+na_ n(x-
x_0)^n-1
f'
'x=2a_ 2+6a_ 3( x-x_ 0) +…+n( n-1)a_ n( x-x_0) ^
n-2
f'
''
x=6a_ 3+…+nn- 1(n- 2) a_ n( x-x_0) ^
n- 3
.
f^nx=nn- 1n- 2…. (3) 21a_ n

Wet henf indtheder ivativesatx=x_ 0asf ol

lows:
fx_0=a_ 0=0! a_
0⇒a_ 0=f x_0/ 0!
f'
x_0=a_ 1=1!a_1⇒a_ 1=f 'x_0/ 1!
f'
'x_0=2a_ 2=2!a_ 2⇒a_ 2=f '
'x_ 0/2!
f'
''
x_0=6a_ 3=3!a_3⇒a_ 3=f''
'x_ 0/3!
.
f^nx=nn- 1n-2….321a_ n=n! a_ n⇒a_ n=f^nx_0/
n!
Let’ssubst i
tut
et hev aluesoft hecoef fi
ci
entsont heorigi
nalequati
on:
fx=f(x_0) /
0!+f^'x_0/
1!x-x_ 0+f ^''
(x_
0) /
2!(x-
x_0)^2+f^
''
'
(x_0)/
3!(
x-x_0)^
3+…+f ^(n)
(x_
0)/
4!(
x-x_
0)^
n
Thispowerser iesrepresent at i
oni sknownasTay l
or’sseri
esrepresentati
on.
Exampl e:Findt heTay lor’sser i
esexpansi onforfx=e^xaroundthepoi ntx_0=0.
Solut i
on
Given f x=e^x ev aluat i
ngt hefunct i
onat x_ 0=0,weget
 f0=e^ 0=1
Theder i
v ativeofe^ x=e^ xandnev erchangesf orsuccessi veder ivati
vessot hat
f^'
0=e^ 0=1, f ^
''
0=1andcont i
nues.
TheTay lor ’
sser iesexpansi onwi l
lbe
fx=f(x_ 0)/ 0!+f^'
x_ 0/ 1!x-x_ 0+f ^''
(x_0)/2!(x-x_0)^2+…+f ^
(n)(x_0)/4!(x-x_0)
^n
=f0/ 0!+f^'0/ 1!x+f^''
0/ 2!x^ 2+…+f ^n0/4!x^n
=1+x+x^ 2/ 2+x^ 3/ 6+…+x^ n/ n!
Expansi onofAr bitrar yFunct ion( Tay l
or’sTheor em)
Accor dingt oTay lor ’
st heor em,gi venanar bi
traryfunction∅( x),whi chiscontinuouswi t
haf ini
te
powern, itsexpansi onar oundt hepointx_ 0y i
elds.
∅x=∅( x_ 0)/0!+∅^ 'x_ 0/1!x- x_0+∅^ '
'(x_0)/2!(x-
x_ 0)
^2+…+∅^ (n)(x_0) /
4!(
x-x_0)^
n+R_ n
∅x=P_ n+R_ nwher eP_ ni spol ynomi alappr oximat i
onandR_ nisr emainder.
Ifthear bi tr
aryf unct ion∅xi sann^ thdegr eepol ynomialandt heexpansi onisintoanot hern^th
degr eepol ynomi ali st obeexpandedi nt oal esserdegr eet hel attercanonlybeconsi deredasan
appr oxi mat ionandt her emai nderi sboundt oappear .
Activ i
ty: Expandt hef unct ionf x=2+3x+4x^ 2+3x^ 2byusi ngTay l
or^ '
sseries
toat hirddegr eepol ynomi al ar oundt hepoi ntx=1.
3.2.6TheI ntermedi at eVal ueTheor em
Letf (x)beaf unct i
ont hati scont i
nuousf orall xintheinterval [
a,b]andassumet hatf(
a)≠f(b)
.Asx
variesbet weenaandb, sof (
x)t akesonev eryvaluebetweenf (a)andf (
b).

Thet heorem i mpliest hatthegr aphoft heconti

nuousf uncti
onf (x)musti nter
sectt
hel i
ne
y=m=f(c)atleastatonepoi nt(c,m).
Animpor tantresul
toft het heorem isl
etf(
x)beaf unct
ioncontinuousin(a,b)andassumethatf(
a)
andf(b)hav ediff
erentsi gns,t
hent her
eisatleastonec∈a,bsucht hatfc=0.
Newton’sMet hod
Theintermediatev aluetheorem showst hatagivenequati
onf(x)hasasol uti
oninagiveni
nterv
al;
howev er
,itdoesn’tpr ov
ideaddi ti
onali
nformati
onaboutthelocationofthesoluti
on.
Byconst ruct ingat angentat( x_ 0,fx_ 0),weobt ainx_ 1.Repeatt hepr ocedur ebyconst ructing
anot hert angentl i
neat( x_1, fx_ 1)andt henequat i
onx_ 1der i
vedbyt heNewt on’sMet hod.
Whati st heequat ionoft het angentat( x_0,f x_ 0)?
y=f x_ 0; y_ 0=f x_ 0
y-y_0=m( x- x_ 0)
Atx_ 0, fx_ 0, wehav e
y-fx_ 0=mx- x_ 0 butm=f ^'x_ 0sot hatwecanr ewr it
ei tas
y-fx_ 0=f ^ '
x_ 0x- x_ 0
Andatx=x_ 1, thet angentl inei nt ersectst hexaxi sandhencet hev alueofyassumeszer o,sot hat
i
tcanbewr ittenas
0-f x_0=f ^'x_ 0x_ 1-x_ 0
-fx_ 0=x_ 1f ^'x_ 0- x_0f 'x_ 0
x_1f ^'x_ 0=-f x_ 0+x_ 0f 'x_ 0
x_1=x_ 0- fx_ 0/ f'x_ 0
Ifwecomput ef orx_ 2, x_ 3andcont i
nue i nt hesameway , weget
x_2=x_ 1- f(x_ 1) /f^ '
(x_ 1)
x_3=x_ 2- f(x_ 2) /f^ '
(x_ 2)
Ingener al thepoi ntsgener at edbyNewt on’smet hodar egi venby
x_n+1=x _ n-f( x_n) /f
^ '
( x_ n)
Exampl e: Fi ndanappr oximat ev alueoft hef unct ionx^ 15=2usi ngNewt on’smet hodt wice.
Sol ution
Firstofal l,forani nt er medi at esol utionweneedt oequat ef x=0sot hatourr eformul atedequat ion
i
swr i
tt enas
fx=x^ 15- 2=0
Besi des,wehav et ot rymanual lytof indt henear esti nt egerthatmakesourf unct i
ont obe
appr oxi mat elyequal tozer oandt hatint egeri s1sot hatourx_ 0=1.
Findi ngandev al uat i
ngt hef unct ionandi t
sf irstder ivativeatt heiniti
al v
aleofx( x_0=1) ,
weget :
f1=x^ 15- 2=1^ 15- 2=- 1
f^'
x=15x^ 14⟹f ^ '
1=15( 1) ^14=15
Hence, wef indf orx_ 1usi ngt heNewt on’ smet hodf ormul aonce:
x_1=x_ 0- fx_ 0/ f'x_ 0=1- -1/15=16/ 15
Findi ngandev al uat i
ngt hef unct ionandi t
sf irstder ivativeat( x_1=16/ 15),weget :
f16/ 15=x^ 15- 2=16/ 15^ 15- 2=0. 633
f^'
x=15x^ 14⟹f ^ '
16/ 15=15( 16/ 15) ^14=37. 025
Oncewef i
ndx_ 1i tiseasi ert of indx_ 2usi ngNewt on’smet hodf ormul atwiceasf oll
ows:
x_2=x_ 1- fx_ 1/ f'x_ 1=16/ 15- 0.633/ 37.025=1. 05
Ther ef or e,t heappr oxi mat ev alueofxi ntheequat i
onx^ 15=2usi ngNewt on’smet hodt wicei s
foundt obeequal to1. 05t ot hehundr edt hdi gi t.
3.3.Mul t ivar iateCal cul us
Cal cul usf orMul ti
v ariat eFunct ions
Though f unct i
ons ofsi nge v ar iable hav el arge si mpl icity,theyar e notcommonl yused i n
economi cs. Theyar enotcommonsi ncet hef unct ionst hatt her ealeconomi cwor ldconsi sted
ofar emul tiv ariat e. Forexampl e,ourr ati
onalconsumerwi l
lnotconsumeonecommodi tyonl y;
rather ,s/ heconsumesdi ffer entcombi nationst omaxi mi zeher /hisut i
lit
y. Ther epresentat i
ve
product ionf unct i
onofaf irmi snotonl yaf unct i
onofl aborbuti ti salsoaf unct i
onofcapi tal
.
Theov er al li ncomeofnat ionsi snotaf unct ionofphy si calcapi talstocksonl y. I tisalsoa
funct ionofhumancapi t
al,f act orpr oduct i
vity ,inst i
tut i
ons,andgov ernmentpol i
cies. Hence,we
hav et oext endt heconcept sofuni variat ecal cul ust omul ti
v ari
ateones. Fort het i
mebei ng, let’
s
begi nwi thaf unct i
onoft woi ndependentv ar iabl essot hatwecansi mplyextendi tto
n-v ariabl escase.
Let ’
st aket heCobbDougl aspr oduct i
onf unct ionofahy pothet i
calfirm.
Wher e:
Y-i
st hel ev el ofout put
A-isthest ateoft echnol ogy,whi chref l
ectspr oduct i
vi
tyoff act ors
L&Kr epresentt heuni tsofl aborandcapi tal
areel ast i
ci tyofywi t
hr espectt ol aborandcapi talrespect iv
el y
Exampl e
Supposef oraf ir
m , then
a.Howmanyuni tsout putwi llbepr oducedi fthefirm uses81l aborand16capi tal
b.Showt hatt heproduct i
onf unctionisconst antreturnst oscal e
Solution
a.Y=f( L, K) =60( 81) 3/4(16)1/ 4=3240
b.Tocheckf orreturnst oscal e,addt heexponent sofl aborandcapi t
al.I
fthesum i sgreaterthan
onei ti si ncr easingr eturnst oscal e,l esst hanonei ti sdecr easingr eturnst oscaleandi fitis
equalt oonei ti
sconst antr eturnst oscal e.Since3/ 4+1/ 4=1,i tisconst antreturnstoscale.For
thepr oof, youcanseet hef ollowingcl ari
ficati
on.
Proof :Foragener alpr oduct i
onf unct i
on,whenwechangeal lthei nputsbyagi venproporti
onif
outputchangesbyequalpr opor t
ion,thent heproduct ionf unct i
oni ll
ustr
atesconst antretur
nst o
scale. I fi t changes by gr eater/smal ler propor ti
on,t he pr oduction funct i
on il
lustr
ates
i
ncr easing/ decr easi ngret ur
nst oscal er espect i
vel
y .
ForYand, LetYo=
Assumel aborandcapi t
alhav echangedbyλ,t henew l aborandcapi talwillbeλL,andλK
respect i
vel y.
Sot henewpr oduct ionf unctionwi ll
be,

=AλαLαλβKβ
=A(λα+β)
=(λα+β)
A
=(λα+β)
Y0
I
fα+β>1, i
tshowsincr
easingret
urnst
oscal
e,i
fα+β<1,
itshowsdecr
easi
ngr
etur
nst
oscal
e,and
i
fα+β=1, i
tisconst
antr
eturnstoscal
e.

3.3.1.Par tialDi fferentiation

Thef ormatY=f (x)r epresent saf unct ionofonev ariabl
e.Gi v
enZ=f(
x,y)
,whichisaf uncti
onoft wo
var i
ables.
Firstorderpar tial derivativesar egi v enas:
f_x=∂Z/ ∂x=f _
x( x, y)
f_y=∂Z/ ∂y =f_y(x, y)
Secondor derpar tialder i
v ati
ves
Wehav etwot y pesofsecondor derpar ti
alderiv
at iv
es
DirectSecondOr derPar t
ial Derivatives
f_xx=∂f _x/∂x=f _ xx(x,y)
f_yy =∂f_y/∂y =f_y y(x,y)
Mi xed( Cross)SecondOr derPar tial Derivat
ives
f_xy =∂f_x/∂y =f_xy (x,y)
f_yx=∂f _y/∂x=f _y x(x,y)
Young’ sTheor em
Themi xed( cross)par ti
alsf oragi v enf uncti
onwi l
lalwaysbeequal i
fbothcross-part
ial
sexistand
arecont i
nuous.
Giv enaf unct i
onf (x_1,x_2),theny oung’st heorem asser tsthatf
_12=f_21oralt
ernativ
elycanbe
wr i
ttenasf _x_ 1x_ 2=f_ x_2x_ 1.
Exampl e:Gi venaf unct i
onU=f x_1,x_ 2=x_ 1l
nx_2, showt hatYoung’
stheorem i
st r
ue.
Sol uti
on
Thef ir
stt askwi llbet of indt hef irstandsecondder ivati
vesofthefuncti
onsonebyoneand
secondl yt ocheckt heequal it
yoft hesecondmi xedpar ti
als.
U_ 1=lnx_2 U_ 2=x_1/x_ 2
U_ 11=0 U_22=-x_1/?x_ 2^2
U_ 12=-1/x_2 U_21=1/x_ 2
So,asYoung’ st heorem pr edictsthatU_ 12=U_ 21andourr esul
talsoconfi
rmedt hatisthe
case. Acti
vity1:Gi v
ent hepr oduct i
onf unct i
oni sgi venby Y=fL,
K=L^αK^β.Pr
ovethatYoung’s
theor em i
st rue.
Act i
vity:Fort hef uncti
onY=fx_ 1,
x_2=x_ 1e^x_1+?x_1^ 2,showthatY_12=Y_21.
3.3.2TheMul tivariat
eChainRul e
Manyeconomi cmodel si nvolvecomposi t
ef unctions.Diff
erenti
ati
onsofcomposi
t efuncti
ons
requiretheappl icati
onofchai nr ul
e.
Forf uncti
onsofonev ar
iableY=f xandx=f tthendy /dtisgivenby
dy/dt=dy/ dx.dx/dt

Mul tiv ar i
ateFunct i
ons( asi mpl ecase)
SupposeZ=f x,y ,x=f tandy =f t,t
hent hemul ti
v ariatechainr ulei sgiv enas
dZ/ dt =dz/ dx. dx/ dt+dz/ dy .dy /dt
Exampl e:SupposeY=l n ( x_ 1+x_ 2)wher ex_ 1=2tandx_ 2=t
^2,finddy /dt .
Sol ution
dy /dt=dy /dx_ 1. dx_ 1/ dt +dy /dx_ 2. dx_2/ dt
=1/ x _1+x_22+1/ x_1+x_ 2(2t)
=1/ x_ 1+x_ 22+2t
Towr itet her esul tof dy /dti nt ermsofonl yt ,subst i
tutetforx’ s:
=1/ 2t +t ^2(2+2t )
Anot heral ter nat iv emet hodt of indt hi sresul tisbydi rectsubstitution,thati s:
Y=l n ( 2t +t ^2)
Then, wecandi r
ect l
yf i
nddy /dtusi ngdi rectdi fferent iati
onasf ol l
ows:
dy /dt=1/ 2t +t ^2( 2+2t )
Act ivity: SupposeZ=f x,y=x^ 2e^ yandx=l nt,y=t ^
2.Fi nd dy /
dtusi ngt hechai nrul
e.
HomogeneousFunct ionandEul er ’
sTheor em
Def inition:Af unct i
onY=f (x_ 1,
x_ 2,…, x_n)i ssai dt obehomogeneousofdegr eeri fandonlyi f
ftx_1, tx_ 2,…,t x_ n=t ^rfx_ 1,x_ 2,…,x_ n=t^r y.
Mul tipl yingal lthei ndependentv ariabl ebyaf act ortwi l
lmul t
iplyt hev al ueofthef unct
ionbythe
fact ort ^ r.
Exampl e1: Gi venf x,y,z=x/ y+2z/ 3x, checkwhet hert hef uncti
onishomogeneousornot .
Sol ution
First, l
et ’smul tipl yt hef unct i
onbyt :
ftx,
t y,tz=t x/t y+2t z/3t x
=t/tx/ y +2z/ 3x=t ^0f x,y ,
z
Ther ef or e, thef unct ioni shomogeneousofdegr eezer o.
Exampl e2: Consi dert heCobb- Dougl aspr oduct ionf unct i
onY=f L,K=L^ αK^ 1-
α.
ftL,tK=t L^ αt K^ 1- α
=t^αL^ αt ^1-αK^ 1-α
=t^α+1- αL^ αK^ 1- α
=t^1L^ αK^ 1-α
=t^1f (L, K)
Ther ef or e,t hisf unct i
oni shomogeneousofdegr ee1.Suchf unct ionsar ealsocal l
edli
near
homogeneousf unct ions.
Act ivity: Checkt hehomogenei tyoff x,y=3x^ 2y -
y^ 3.
I
neconomi cs,t hedegr eeofhomogenei t
yofaf unct i
onhasonei mpor tanti
mplicati
on.Thatis,
givenf tx_1,t x_ 2,…, t
x_ n=t^rfx_ 1,x_ 2,…,x _n
I
fr =1⇒ Const antr et urnst oscal e
I
fr >1⇒ I ncr easi ngr et urnst oscal e
Ifr <1 ⇒ Decr easi ngr et urnst oscal e
Euler ’sTheor em
Euler ’st heor em st at est hati fahomogeneousf unct i
onhasadegr eer,t hen
∂f/∂x_ 1. x_1+∂f /∂x_ 2.x_2+…+∂f /∂x_ nx_n=rf(
x_1,x_2,…,
x_ n)
Wher e, rmeasur est hedegr eeofhomogenei ty.
Exampl e:
Supposef x,y=x^ 4+x^ 2y ^
2, t
hencheckt heEuler’stheor em.
Solution:
Accor dingt oEul er’st heorem f _ x.x+f _
y.yisequaltosomev alue(r)ti
mest heorigi
nalfuncti
onand
thequest ioni swhati sthisv alue.
∂f/∂x=f _x=4x^ 3+2xy ^2
∂f/∂y =f_ y=2x^ 2y
Then, f_x.x+f _y .
y=4x^ 3+2xy ^2x+2x^ 2yy
=4x^ 3+2x^ 2y ^2+2x^ 2y ^
2
=4x^ 3+4x^ 2y ^2
=4x^ 3+x^ 2y ^2
=4(fx,y )
Hence, sincet hef unct i
oni shomogeneousofdegr eeone, wecal li
tli
nearhomogeneousf uncti
on
CHAPTERFOUR
UNCONSTRAI NEDOPTI MI ZATI ON
4.1.Funct i
onswi thonev ariable
Gi venanyf unct i
onYasaf unctionofx=f (
x) ,t
hent heder i
vati
vev er
sionof
thef unct ion( Fir
stOr derCondi ti
onorFOC), asdescr i
bedear l
i
er,i
sstatedin
thef or m dy /dx, y’orf’ (
x)i stheinstant
aneousr ateofchangeoff( X)which
isgi venby :
4.1.1.Theconceptofopt i
mum ( extreme)v al
ue
Anext r
emev alueofaf unct i
oni st hemaxi maormi nimav alueoft hatfuncti
on. Wecanhav e
twot y pesofext remev alues.
a.Rel at iveext remev alues( Local ext r
eme)
b.Absol uteext remev alues(Globalext r
eme)
Relat iveExt remeVal ue
Relativ eext remev alue: i sanext r
emev al
uewhencompar edwi ththev al
uesneart oit
.I tcan
ber elat ivemaxi ma( wheni ti shi ghert hananyot herpoi ntneart oit
)orr el
ati
vemi ni
ma( whenitis
l
owert hananyv alueneart oi t)
Forexampl e, consi dert hef unct ion .
It
sgr aphi s:

Ingener al,ifaf unct i

onfhasr elat i
vemaxi maat i fther eexi st
sani ntervalaroundx0onwhi ch
foral lxint hei nter val
Ifaf unct i
onfhasr elat
ivemi nimaat i ft hereexistani ntervalaroundonwhi ch f orallxinthe
i
nter val
Iti
seasyt of i
ndoutr elativ
eext remev aluesusi ngtheder i
vative. Prev iouslywehav eseent hata
funct i
onchangesf rom increasingt odecr easingandv i
cev ersa,wheni t
sf i
rstder i
vativeiszeroor
undef ined. Si nceaf unctionchangef rom decr easingt oani ncr easi
ngatar el
ativemi nimum, we
have i sundef inedatr elativ
emi nimaormaxi ma.
Hence,wecanf indr elat
ivemaxi maandmi nimaf oraf unct i
onbyf indingt hev aluesofxf or
which i sundef i
ned. Oncewegett hecr i
ti
calv al
ue, wecanuset hebehav i
oroff‘ (
x)nearthe
extremepoi nttoknowwhet heritisar elat i
vemi ni
maorr elati
vemaxi mapoi nt.
Tof indoutr elativ emi nimaormaxi ma,
1st Fi nd
2nd Sol vef ort hev al
ueofxwhi chwi llmake
3rdUset heabov ev aluet ogett heext remev alueoft hef unct i
on( i
.e,insertthev aluey ougotin
the2ndst epinf (x)togett heext remev al ue).
4thEv aluat et hef i
rstder i
vati
v ef unctionatsomev aluesofxt othet eftandt her i
ghtofeach
cri
tical point
a.Iftot hel eftand t ot her i
ghtoft hecr iticalvalueoff( x),thecriti
cal pointisrelati
vemaxi mum
b.Ift
ot hel
eftandt
otheri
ghtoft
hecrit
icalv
alueoff(x)
,thecr
it
ical
poi
nti
srel
ati
vemi
nima.
Example:
Fi
ndt herel
ati
veext
remaasofthefol
l
owi ngfunct
ions

Sol
uti
on

=x-

=-
7

Thus,t
hecr
it
icalv
aluesar
e(. Tof
indoutwhi
choner
elat
ivemi
nimaormaxi
ma,
uset
henumber
l
ine.

Thus, i
srel
ati
vemaxi
ma.

Thus,
(isr
elat
ivemi
nima.

b.

Cri
ticalvaluesar e( 1,
4)and( -
1,8).
+ 0 - - 0 +
-
1
1
Relati
vemaxi maat (-
1,8) Rel
ati
vemini
ma
at(1,4)
Notet hatinsteadoft akingt henumberl i
net ocheckf ormaximaormi ni
ma,wecanuset he2nd
deri
vat i
ve. I fi st hecr i
ti
cal val
uewhi chmakes, t
henthepoint(,
f(
))isrel
ati
vemaximaif and
rel
ativemi nimai f .Butwhen wecan’ tuset he2ndder i
vati
vetestandwehavet ousethe
behav i
oroft hef i
rstder i
vati
v etocheckf orr el
ati
veoptima.
Exampl e:Uset he2ndder ivati
v etestofthepr eviousfunct
iont oknowwhethert
heyarerel
ati
ve
maxi maorr el
ativ
emi nima.
a.
b.
c. .
a. cri
ticalvaluesar e(-1,11/3)and( 3,
-7)
isr el
ativ
emaxi ma
i
srelativ
emi nima.
b.

cr
it
ical
val
uesx=1 (
1,3)

Uset
henumberl
i
nei
nthi
scase.

C. ,
Cri
ti
cal
val
ues(
1,4) and(
-1,
8)
 Uset
hef
ir
stder
ivat
ivet
esti
nthi
scase.

Absol uteExt r
ema
Absol uteext remev al ue: i sanext remev aluet hatt hefunct i
onat t
ainst hroughouti t
sdomai n.
Thepossi blelargestv aluet hatthefunct ionat tainsi ni tsdomai niscal ledabsol ut emaxi maoft he
funct ion. Si mi l
arly,thesmal l
estpossi bleval uet hatt hefunct i
onat t
ainst hr oughouti t
sdomai nis
knownast heabsol utemi nimaoft hef unction. Not ethataf unct i
onmay( ormaynot )have
absol ut eext r
emev aluel i
ket heremayormaynotexi str el
ativeext remev alue.
Forexampl econsi deraf unct i
ony =2x+1. Aswei ncr easethev alueofx, thev al ueofyi ncr
eases
cont inuousl y. Agai n,when we decr ease t he v alue ofx,t he v alue ofyal so decreases
cont inuousl y. Howev er,ifthedomai ni sr est r
ictedt oacl osedi nterval[ a,b]whi chmeansand
thef unct i
oni scont inuousev erywher ewi thint hei nterval
,t hef unctionwi llhav eanabsol ute
mi nimaandabsol utemaxi mawi thinthei nterv al
. Theymaybel ocat edei ther :
a.Att heendpoi nts; i
.e. ,aandbor
b.Att hei nteri
orpoi nts
Tof i
ndabsol uteext remawi thi
nani nter val
a.Fi ndt hecr it
icalv aluesandcr iti
calpoi ntsandcheckwhet hertheyar ewi thint hei nterval
b.Cal culatef(a)andf (b)
c.Thel argestv alueandsmal l
estv alueobt ainedf rom t heabov est epsr epresentt heabsolute
maxi maandt heabsol utemi ni
maoft hef unct ionov ert heinterval respectivel y.
Exampl e
Findt heabsol utemaxi maandmi nimaof ov ert hei nt er
val
Sol ution
f’
(x) =0

Thereisabsol
utemaximaatx=0andabsolut
emi ni
maatx=3wi
thi
n[0,
5].
2. Findtheabsol
uteext
remaof Ov ertheint
erval
[
-4,2]
Sol
ution

Absol utemaxi maatx=2( 2,24)andabsol utemi nimaatx=- 4(-

4, -
12)
4.2.Opt imi zati
onofaFunct i
onwi t
hMor eThanOneVar i
able
Dearst udent !Youhav est udiedopt i
mizationofaf uncti
oni nsection4. 1inwhi chcaset he
discussi on was r est r
icted toaf unct i
on with a si ngle explanat
oryv ariable.But ,economi c
phenomenai nr ealit
yinv olvethest udyandanal ysisofl argenumberofv ar
iablessi mul t
aneousl y
.
Ther efore, i
nt hissect i
onweext endt hediscussi onofopt i
mizati
ontoaf unct i
onwi thmor ethana
singleexpl anat oryv ari
abl e.
4.2.1.Fi r
standSecondOr derCondi tionsforaFunct ionwi t
hTwoExpl anator yVariables
Thi ssect ioni sasi mpl eextensi ont oy ourst udyonopt i
mizationofaf unct i
onwi t
hasi ngle
variable.Youwi llbeginy ourstudywi ththef i
rstordernecessar ycondit
ionf oropt i
mi zati
onofa
funct i
onoft woexpl anat oryv ari
abl esandt hendi scusst hesecondor dersuf fi
cientcondi ti
on.
Bef oregoi ngi ntot hedi scussion, youshoul dnot ethatamaxi mum andami nimum i nt hissecti
on
refertoar elativemaxi mum andar elat i
vemi ni
mum.
FirstOr derCondi ti
onf oraFunct ionwi thTwoExpl anat or
yVar i
ables
Whatdoy out hinkaboutt hefirstandsecondor dercondi ti
onsforaf unct
ionwi tht woexpl anatory
vari
abl es?Youcanuset hespacebel owtowr i
tey ouranswer .
Consi dert hef unct ion.Int hiscase,t hef i
rstor dernecessar yconditionf oranext remum ( eit
hera
maxi mum orami nimum)r equirest hef irstpar tialder ivati
vesoft hef unct i
onequalt oz ero
simul taneousl y.
Thisi nv olvesf indingt hev aluesof andsot hatt hef irstpartialderivati
vesar ezer oatsamet i
me.
Thati sor .
Cany oumakeaconcl usionbasedupont hi
sf i
rstor dercondi t
ionf oranext remum?Or ,isi t
possi blef ory out oconcl udewhet herapar ticularpoi ntisamaxi mum orami nimum ont hebasi s
ofthef irstor dercondi ti
on?Theansweri snobecauseoft hepr esenceofpoi ntswhi chsat isf
yt he
fi
rstor dercondi t
ionbutnei theramaxi mum norami nimum.Onegoodexampl ei sthecaseofan
i
nf l
ect i
onpoi nt. Ther efor e,simi lart owhaty ouhav estudiedi nt hecaseofasi nglev ari
able
optimi zat i
on,t hef i
rstor dercondi tioni sonl yanecessar ybutnotasuf fi
cientcondi ti
on.Asa
result,y ouneedt omakeuseofaddi ti
onalcr i
teria.Wi ththiscomest hesecondor dert est.
SecondOr derCondi t
ion.
Thesecondor dercondi tionf orthef unct i
oninv olvesr estr
ictionsont hesi gnsoft hesecondpar tial
derivativ es:,and.Thesecondor dercondi tionwhi chi sal soknownast hesuf fi
cientcondition
foll
owsbel ow.
Letbet hecr i
ticalv al
uesoft hef unct i
on( i.
e.,wher e)andt hesecondpar tialder ivat
ivesar e
evaluat edatt hesepoi nts.
i
sar elat iv
emaxi mum i fandonl yi fand.
i
sar elat iv
emi ni mum i fandonl yifand.
i
sasaddl epoi nti f,andandhav ediff erentsigns.
i
sani nf lectionpoi ntif,andandhav et hesamesi gns.
couldbear elativemaxi mum,ar elat iv
emi nimum,orasaddl epointif.Thati st hesecondor der
testf ailst ot ellthebehav i
oroft hef uncti
on. Thet abl ebelowsummar izesbot ht hefir
standt he
secondor dercondi ti
onsf orr el
ativeext r
emum oft hef unction.
Table4. 1Condi ti
onsf orRel at i
veExt remum
Condi tion
Maxi mum
Minimum

Fi
rstor
dernecessar
ycondi
ti
on

Secondor
dersuf
fi
cientcondi
ti
on

and

and

Gener all
y,i
nf i
ndingther el
ativemaxi mum/ mi
nimum ofaf unctionwi thtwoexpl anatoryvari
abl
es,
youneedt of i
ndf i
rstthecrit
icalv aluesofthatfuncti
onusingt hef i
rstor dernecessarycondit
ion
andt hencar ryont hesecondor dert estt
odetermineifthecr it
icalvaluesar eeithermaximum,
minimum, saddleorinflect
ionpoi nts.
Exampl es
a)Findt hecrit
icalvaluesofthef unct i
onanddeterminewhethert heyar emaxi mum ormi nimum.
Soluti
on
Usingt hef i
rstordercondi ti
on,y oucanf i
ndthecr i
ti
calvalues.Thus,y ouneedt of i
ndt hefir
st
parti
al deri
vati
ves,and.
Then,fi
ndthevaluesofand(t
hecrit
icaloropt
imal
val
ues)whi
chsat
isf
ythecondit
ion.Hence,
f
orand,i
.e.
,thecr
it
icalpointi
nthiscaseis
Now,yougot othesecondor dert
est.Here,y
ouneedt
of i
ndt
hesecondorderpar
tial
deri
vati
ves.

Ev
aluat
ingt
hesecondpar
ti
alder
ivat
ivesatt
hecr
it
ical
val
ue,
youhav
e

Therefore,and,and.Hence,y
ouconcl
udethatt
hecr
it
ical
valueisar
elat
ivemini
mum.
b)Gi v
ent hefuncti
on,findthest
ati
onar
yv al
uesandex aminewhethertheyar
emaxi
mum or
minimum.
Solut
ion
Fi
rstor dercondit
ion
Thefirstparti
alderi
vat
ivesare

Thef
ir
stor
dercondi
ti
onr
equi
res.Thus,
youhav
e

Solv
ingtheaboveequati
onssimult
aneousl
y,y
ouget
andandhence,
thecr
it
ical
val
uei
s.
Secondordercondi
ti
on
Thesecondparti
alder
ivat
ivesar
e

Since and, and( becauseand) ,thecr iti

calvalueisar elati
vemi nimum.
4.2.2.Det eri
ment al TestforSi gnDef i
niteness:
SecondOr derCondi t
ionDet erminantTest
Af orm ofapol y
nomi alexpr essioni nwhi cheachcomponentt er
m hasauni quedegree
i
.et hesum ofcomponent si neacht ermi suniform
e.g.W =f (x,y ,z)=ax–by+z-i sl i
nearf unction
w=f (x,y,z)=ax2+bxy+cz2–quadr atiquedi f
ferential
d2z=f xxdx2+2f xydxdy +fyydy2-i stotaldef f
erential
Letg=d2z, U=dx, V=dy ,a=f xx, b=f yy,h=f xy=fyx
g=au2+2huv+bv 2
Differentcases
If g>o –I fposi tivedef i
nite, Suf f
icientcondi ti
onformi ni
mum
If g>o -Posi ti
vesemi definite, Necessar ycondi ti
onf ormi nimum
If g<o -I fNegat ivedef i
nite,Suf fi
cientcondi ti
onf ormaxi mum
If g<o -I fnegat i
v esemi def init
e, necessar ycondi ti
onf ormaxi mum
Ifgchangessi ngswhent hev ari
ablesassumedi ff
erentv al
ues, gissai
dt obeindef
inite
We can check whet hert he f unction ( g)is def i
nite ori ndefini
te usi
ng deter
mi nantt
est–
deter minantt estf orsigndef i
niteness.
g=au2+2hu+bv 2
=au2+huv +bv 2
=u( au+hv )+v (hu+bv )
Not ef ortwomat ricest obemul ti
plied,conf ormabi l
it
ycondi t
ionmusthol dtr
ueThus,

Therefore,
theabovedet
erminantiscal
ledHessi
anDeter
minant
.
Thef i
rstpr i
nci
palminorofHessian( H1)isthemat r
ixwhichcont
ainst
hef i
rstonepri
nci
pal
diagonalelement;t
hesecondpri
ncipalminorofHessi
an(H2)isamatri
xwhichcont
ainst
hefi
rst
twopr i
ncipaldi
agonalel
ements,
etc.
 Notes
thesecondor derdifferent
iald2z>0iff/H1/>o,/H2/>o, itisposit
ivedef
ini
teand showst he
mi nimum poi nt
d2z<o, if
f/ H1/<o, but/ H2/>o, -negativ
edefini
teanditr
efersmaximum point
d2zi sindef i
nit
eiff/ H2/<o
Exampl e
Opt imizet hef ol
l
owi ngf unctionbyusingtheHessi andet
erminantandstatesigndef
ini
teness.
Findt hemaxi mum ormaxi mum value.
Z=f (x,y
)=6x2–9x- 3xy-
7y+5y 2
Sol uti
on
F.O. C: Zx=12x- 9-3y=0
Zy=- 3x-7+10y=0
BYsol v
ingsi multaneously,thecri
ti
calval
ueswi l
lbe:
x=1, y=1
S.O. C: Zxx=12
Zxy=Zy x=-3
Zyy=10

Sinceboth/H1/ and/
H2/ aregreaterthanZer o/thefuncti
onisatminimum point
.
Thesigndefini
tenessisPositi
ve,because/ H1/>0,/H2/>0
Themi ni
mum v alue=6(1)2-9(1)- 3(1x1)–7( 1)+5( 1)
2=- 8
Thecaseoft hreevar
iablesQuadr ati
cf or
m
Thisapproachcanal sobeappl iedt omor et hanthreechoicevar
iabl
es.GivenZ=f
(X1,X2,X3)
,
thentheform ofHessianDetermi nantwillbe:.

-
isposi
ti
vedefi
nit
eiff/H1/>0,/H2/>0,/
H3/>0
Showsminimum point
-
isnegat
ivedef
ini
teif
f/H1/ <0,/H2/>0,/H3/<0
Showsmaximum poi nt
Fort
heCaseofn-v ari
ables,Z=f (
x1,x2,
x3...
Xn)

-i
sposi t
ivedef i
ni t
ei ff
, /
H1/>0,/H2/>0, /
H3/ >0…./Hn/ =/H/
Thef unct i
oni sati tsmi nimum point
-i
snegat i
vedef init
ei ff,/H1/<0,/H2/>0,/H3/<0…. /Hn/=/ H/
Wher ethef unct ionisatmaxi mum
Exampl e
Afirm pr oducest hr eet ypesofproduct sQ1, Q2, andQ3.Andapr ofi
tfunctionisgiv
enby :
Л=F( Q1, Q2,Q3)=180Q1+200Q2+150Q3- 3Q1Q2- 2Q2Q3-2Q1Q3-4Q12- 5Q22-4Q32.
Trytoopt imizet hef unct i
onandf indthecriticalval
ueswhi chmaket hefunctionatmaximum or
minimum poi ntandal sothemaxi mum ormi nimum v alue.
Tellt
y peofdef i
niteness.
Soluti
on
1st–f indFOC
Л1=180- 3Q2- 2Q3- 8Q1=0 88Q1–3Q2- 2Q3=- 180
Л2=200- 3Q1- 2Q3- 10Q2=0 3Q1–10Q2- 2Q3=-200
Л3=150- 2Q2- 2Q1- 8Q3=0 2Q1–2Q2- 8Q3=- 150
2nd_For m acœf ficient mat ri
x,A
A=
Herei fisdi ff
icultt ousesi mul
taneousmet hodt osolv et
hecrit
icalvalues.Thus,inst
eadweuse
Cramer ’
sr ul
e.
3rd_f indt hedet ermi nantofcoef f
icientmatrix, A.
Ther
efor
etheopt
imalval
uest
hatopti
mizet
heprof
itf
uncti
onar
e
Q1=14.58,Q2=13,27andQ3=11.79.
4th–S,O,C–tocheckthepr
ofi
tisr
eall
ymaxi
mizedornot

5t
h-For
m Hessi
anmat
ri
x

Since/ H1/<0, /H2/ >0and/ H3/<0, wecanconcl udet hatprof itisatmaxi mum whenev aluatedat
optimal values&Q1=14. 58Q2=13. 27andQ3=11. 79andi mpl i
esnegat iv
edef i
niteness.
Themaxi mum pr ofitisobt ainedbysubst it
utingt heopt i
malv aluesi ntheorigi
nalpr of
itf uncti
on
(π)
π=180( 14.58) +200( 13.27)+150( 11.79) -
3(14.58x13. 27)-2(13.27x11. 79) 2(14.58x11. 79)-
4(14.58) 2- 5(13. 27) 2-4(11.79) 2
=2624. 4+2654+1768. 5-580. 43-312.91-343.80-850.31- 880.46- 556.02
=7046. 90-3523. 93
=3522. 97.
4.3Impl i
citfunct ionsandunconst r
ainedEnv elopet heor y
Implicitfunct ions
Consi dert wo choi cev ariablesopt imizati
onmodel :Y =f (x1,x2)– expl i
citl
yshowst hatthe
optimi zationofYdependsonl yonx1andx2.Howev er ,i
nreal ityopt i
mizati
onofYi saf f
ect edby
somei mpl icitpar amet ersandi tispossi bl
et oopt imizeYbyassumi ngthatthesepar amet er
sare
heldconst ant .
Optimi zationpr oblemsi neconomi csusual lyinv olvef unctionst hatdependonanumberof
paramet ersl ikepr ices,taxr ates,et c.Although t hesepar amet ersar ehel d constantdur i
ng
optimi zationt heydov ar
yaccor dingt otheeconomi csi tuati
ons.
Let’
sconsi dertheef fectofchangei nonepar amet eront heopt imal soluti
onofagi v enequat ionY
=f (
x1x2, α)wher ex1x2ar echoi cev ari
abl
esandαi st hepar amet er.

TheFOC( f1=0, f
2=0)y i
eldscr iticalv aluesf orx1andx2whi chr esulti
nt hest ati
onaryv alueofy .
Butt hi si spossi bleifandonl yifxi sf ixed.
Thus, thecr it
ical valuesoft hechoi cev ariablesar efunct i
onsofαar ex1*=f (α)and
x2*=f( α) .i.
e.i fwechanget hev alueof αwewi l
lfi
nddi f
fer
entopt i
mal v aluesforx1andx2
.Sot heobj ectiv efunct ioni sal soaf unct ionofpar amet er.
Thus, л( x)=f (x1*(α), x2* (
α) ,
α)bet hei ndi r
ectobj ecti
v efuncti
on.Thi sf unct i
onshowst heef f
ect
ofαont heobj ect i
vef unct ionY.
Fori nst ance, i
fαchangest oαo, wemayhav edi f
ferentoptimal valuesx1=x10and
x2=x20i fxchanges.Andwemaxi mi zef (di
rectobj ectivefuncti
on)usi ngx10, x20andα0.
Simi larly, i
f αchangesf rom α0t o α1, wewi llhaveanot hernewequi li
brium( Opti
mal)v aluesfor
x1andx2,x1 =x11andx2=x21.Theл( α)isaf unct i
onwhi chwi l
lbet angenttof (x1,x2,α)
whenx1andx2ar eusedopt imal lyforagi venα.Foragi venαiffi snotopt i
mi zed,t
het angency
willnotoccur .Thus, wehav easi ngl epoi ntwher efi sequalt oлand; hence, theinequalityf(x)<=
л( α). Thi si mpl iesл( α)i sanenv elopeoft hef uncti
onf (x)atdiff
er entvaluesofα.
Env elopTheor em
Thist heor em isusedt odet ermi net hechangei ntheopt imum v alueofanopt i
mum f unct i
ondue
tochangei nthev al
ueofapar amet er.
Let beanobj ecti
v ef unct i
onwi thapar amet erd.TheFocf ormaxi ma/mi nimaar e
Then, thest ationar yv aluesar e: andt heopt imum v alueoft heobj ecti
vefunct i
on:
Taki ngt otal der i
vativeofZ*wi thr especta, weget
,but atopt imum
Ther ef ore:-Ther ateofchangeoft heopt i
mum v alueofZ*asdchangesi sequalt ot hepar t
ial
derivat iveoft heobj ectiv efunct ionwi thr espectd.
Theabov er esul tisknownast heenv elopt heor em.
Exampl e1:Det erminet heef fectofani ncreasei nt hev al
ueofont heopt imum v alueoft he
funct ion
Foc: -
 yi
smi
nimum when

Or
Y* i ncreaseby unit
sas ai ncr
eases.
Exampl e2:Af irm pr oduci
ngacer t
ainproductQwant
stomaxi
mizei
tspr
ofi
t.Supposeat
ax,t
bir
rperuni tisimposedont hefuncti
onofα.
A.opt imizet heprofitfuncti
on
B.Findt heef fectoft hechangeintonл
Solution
A.Weknowt hatл=TR–TC
Л=TR–TC, TR=R( Q)
,C=C( Q)+tQ
=R( Q) -C( Q)–t Q=O
FOC
Л’(Q)=R’ (
Q) -
C(Q)-t=0
=MR–MC–t=0
t=MR–MC
SOC
л"(Q)=R" (Q)–C’ ’
(Q)<0
=R’ ’
(Q)<C’ ’
(Q)
=slopeofMR<SopeofMC
B.л* =R( Q*)–C( Q*)–t Q*

CHAPTERFI VE
5.CONSTRAI NEDOPTI MIZATI ON
5.1Const rainedopt imizationwi thequal ityconst raint:
Funct i
onsofTwoVar iablesandEqual it
yConst raints:
Techni quesofConst rainedOpt i
mization
TheTechni quesofconst r
ainedopt imizat ionpr esent edint hi schapt erar ebui l
tont hemet hodfor
i
dent i
fy i
ngt hest ati
onar ypoi nti nunconst rainedopt i
mi zation.
Substitut i
onMet hod
Onemet hodi nconst rainedopt i
mizationi st osubst i
tuteequal ityconst raint sint oanobj ect
ive
functi
on.Thi si st oconv ertaconst r
ainedopt i
mi zationi nt ounconst rainedopt imizationpr obl
em
byinter nali
zi ngt heconst r
aint/const raint sdirectlyint otheobj ectivef unction.
Theconst rainti si nternali
zedwhenweexpr essi tasaf unct i
onofoneoft hear gument soft he
objectivef unct ionandt hensubst i
tuti
ngt hatar gumentbyusi ngt heconst raint.Wecant hensolve
theint ernal i
zedobj ect i
vef unct ionbyusi ngt heunconst rainedopt i
mi zationt echni
que.Sucha
techniqueofsol vingconst rainedopt imi zat i
onpr oblemsi scal l
edt hesubst ituti
onmet hod.
Inordert odev elopapr ocedur eforthedet erminat ionofconst rainedext remaofaf unct i
on,we
considert hemaxi mi zationoft heut i
l
ityf unct i
on,,ofaconsumersubj ectt ot hebudgetconst rai
nt,
.
Givent hebudgetconst rai
nt ,wenot et hatnowxandycannott akev aluesi ndependent lyofeach
other.Forexampl e,if,
thent hev al
ueofymustbe12et c.
Onewayoft ackl ingt hi
spr obl em ist oel iminat eoneoft hei ndependentv ariables,xory ,f
rom the
util
i
tyf unct ion,byt heuseofgi venconst r
aint.Then,Ucanbemaxi mi zedormi ni
mi zedi nthe
usual unconst r
ai nedmanner .
Toel i
mi natey( say )from t heut i
li
tyfunct i
on,wef irstsol vet heconst raintf ory .onadj ustingthe
termst heconst raintcanbewr itt
enas subst i
tutingt hisv alueofyi nt heut i
li
tyfunct i
on,we
hav
e
Thus, (
formaxi
ma)

Substi
tuti
ngintheconstr
aint,thev alueof.Thus i smaxi mum whenconsumerpur chases10
unit
sofxand8uni t
sofy ,underthecondi t
iont hathi
sincomei s$400andpr iceofx=$20and
pri
ceofy=$25.Si nce thesecondor derconditi
oni salsosat i
sfi
ed.Thistechniqueof
constr
ainedopti
mizati
onisdubbedassubst i
tuti
onmet hod.Below anattemptwil
lbemadet o
i
ntroduceyouwiththevari
oust echniquesofconst rai
nedopt i
mizat
ion,i
ncludi
ngthesubst i
tut
ion
method.
Example1
Letusconsideraconsumerwi t
hasi mpl eut
ili
tyfuncti
on

Subj
ectt
obudgetconst
rai
nt

Theconst
rai
ntcanber
ewr
it
tenas

Subst
it
uti
ng(
1)i
ntoanobj
ect
ivef
unct
ion

……………. .
Int
ernali
zeobj
ect
ivef
unct
ion
Optimizetheinter
nali
zedobjecti
vefuncti
on.
Soluti
on
Fi
rstOr derCondit
ions[FOC]
:

Subst
it
uti
ng i
ntheconst
rai
ntf
unct
ionwewi
l
lhav
e,

SecondOr
derCondi
ti
on:

Thus:- ismaximum whenaconsumerpur chases8unit

sofx1and14uni tsofx2undert
he
condi
ti
onthathi
sincomeis$60andpri
ceofx1andx2ar e$4and$2,
respect
ivel
y.
Example2
Maximizet
hefuncti
on, subj
ectt
othel
inearconstrai
ntt
hat

Solut
ion:-
Theconst r
aintcanber ewrit
tenas
Int
ernali
zi
ngt heabov efuncti
on,
….Int
ernal
i
zedobj
ect
ivefunct
ion
Solvef orthe internal
ized obj
ect
ivef
uncti
on.Pr oceed t
hrough t
he f
ir
stand second or
der
condit
ions.
Fir
stordercondition:
-
Solvi
ngforx1y ields

Subst
it
uti
ng i
ntot
heconst
rai
ntf
unct
ion:

Secondor
dercondi
ti
on:

Ev
aluat
ingat wewi
l
lhav
ethef
oll
owi
ng.

Thus,themaximum v aluefort
hefuncti
onsubj
ectt
otheconstr
aint
b).TheLagrangeMultipli
erMethod
Whent heconstr
aintisacompl i
catedfunct
ionorwhent her
ear esev
eralconst
rai
ntsunder
consideration,t hesubst ituti
onmet hodcoul dbecomev erycumber some.Thi shasl edt ot he
developmentofanot hersi mpl ermet hodoff i
ndingextremaofaf unct i
on,thelagrangeanmet hod.
Thismet hodi nvolvesf ormi ngal agrangeanf uncti
ont hatincludest heobj ect
ivef uncti
on,t he
constraintfunct i
onandt hev ari
abl e,
,calledtheLagr angemul t
ipli
er.Theessenceoft hi
smet hod
i
st oconv ertaconst rainedex tremum pr oblem intoaf orm sucht hatt hefir
stordercondi t
ionsof
theunconst rainedopt imi zat i
onpr oblem canst i
llbeapplied.
Wemaynot eher ethatt henecessar yconditi
on,obtainedabov eundert hesubst i
tutionmet hod,
canal sobeobt ainedf rom anauxi li
aryfunction,tobetermedasLagr angeFunct i
on.Thi sfuncti
on
i
sf ormed by t he used ofobj ectivef uncti
on and t he const raints.The Lagr ange f uncti
on
correspondi ngt ot heobj ect ivefunct ionandt heconst r
aint , canbewr it
tenas
i
sanundet er mi nedmul t i
pli
erknownas
LagrangeMul ti
plier.
Wenot et hat ,since,t hev alueofLatapoi ntissameast hev alueofatt hatpoi nt.Thus,t he
exteremaof andVoccur satt hesamepoi nt.Thenecessar ycondi tions[fi
rstordercondi t
ions]
forextermaofLar e:

…………………………………. .(1)
……………………………………( 2)
………………………………………( 3)
Thesi mult
aneoussol utionsof( 1),(
2)and( 3)gi
vesthestat
ionary(orcrit
icalorequi
li
bri
um)poi nt.
Not ethatthest ationarypoi ntobtainedabov ewil
lsati
sfyequation( 3)
,i.e.t
heconst r
aint.Thus,
theunconst r
ainedext r
emaofLi sequiv
al enttotheext
remaof subj ecttotheconstrai
nt.
Onel i
mi nating f rom (1)and( 2),weget
……………………………………………….( 4)
Thisissamecondi ti
onasobt ainedearlier.Equati
on(3)and( 4)canbesi mult
aneouslysolvedto
getthecoor di
nat esoft hest ati
onarypoint .
c).TotalDif
ferentialAppr oach
Whi l
ewedi scussunconst rainedopt i
mi zationofZ=f( x,y),theFocmaybest at
edint ermsof
totaldi
ffer
ential
.

Thi
sst
atementr
emai
nsv
ali
daf
teraconst
rai
nt.

Subst
it
uti
ng i
ntot
heobj
ect
ivef
unct
ion

Or i
sthe
f
ir
stor
dercondi
ti
on

Not
e:Thi
stot
aldiff
erent
ialappr
oachy
iel
dst
hesamer
esul
t(f
ir
stor
dercondi
ti
on)asoft
he
l
agr
angi
anmethod.

Whatisthedi
fference:-
The tot
aldif
ferentialappr
oach y
iel
ds onl
ythe v
alue of
,the Lagrange method al
so gi
ves
prov
idesameasur eofthesensi
ti
vit
yofZt ot
heshiftoft
heconstraint
,theLagrangemethod
otheranadv ant
ageofcer t
ainbuilt–incompar at
ivest
aticinformati
onofthesolut
ion.
I
nterpret
ationoftheLagr angeMul t
ipl
ier
I
ti sameasur eoft heef fectofachangei nt heconstrai
ntv iathepar
ameter“C”ont heoptimal
valuetheobject
ivefunction.
Wecandemonst r
ateitbyusi ngthechai nrul
e.
Considertheoptimalvaluesoft hear gument soffunct
ionf( x,y)are
The“ C”i
nt hebracketrefl
ectsthef actthattheoptimalvaluesarefunct
ionsoftheconstr
aint
.

Taki
ngdi
ff
erent
ial
ofeachsi
dewi
thr
espectt
o “
C”.

Al
so,
tot
all
ydi
ff
erent
iat
etheobj
ect
ivef
unct
ion

Equ(
1)

TheFocr
equi
ret
hat

Subst
it
uti
ngt
heFoci
ntoequat
ion(
1)

But
,thet
ermi
nthebr
acketi
sequal
to1.Thus.

i
.etheLagr
angemulti
pli
errepr
esentsthechangei
ntheopt
imum v
alueoft
heobj
ect
ive
f
unct
ionwit
hasmall
changeintheconst
rai
nt.

Generali
zat
ion:
-
LangrangeFunctionwit
hMul t
ipl
eConstr
aints
Wecansol vet hesolut
iont oaconstrai
nedoptimizat
ionproblem wi thmorethanoneconst
rai
nt
byusingthelagrangi
anf uncti
onwithaLagrangemul t
ipl
ier
scorr espondi
ngtoeachconst
rai
nt.
Letan“n”vari
ablefuncti
onbesubj ect
,si
mul t
aneously
, t
othetwoconst r
aint
s
Thenadopt i
ng and ast het wounconst r
ainedmul t
ipl
i
er s,wemayconductt heLagrange
funct
ionas:-

Foc:
-

Theextensi
onofthi
smodeltothecaseofmorethantwoar
gument
sint
heobj
ect
ivef
unct
ioni
s
str
aightf
orward.I
nthi
sthel
agrangeanf
unct
iontakest
hefor
m

TheLagrangi
anfunct
ionforaconstrainedopt
imizat
ionpr
obl
em wi
thanobj
ect
ivef
unct
ionof“
n”
var
iabl
esubjectt
om equali
tyconst
raintsiswr
it
tenas:-

Wher e representthei
thconstrai
nt.
5.1.2.Suffi
cientCondi
tionsforopt
imizati
on
Substi
t uti
onmet hod
Exampl e:-Consumpt ionProblem
Supposey ouhav e$6. 00t
ospendonal unchofsoupandsal
ad.Ar
est
aur
antatwhi
chy
oudi
ne
sell
sbothsoupandsaladbyweight
.Anounceofsoapcost
s$0.25andanounceofsal
adcosts
$0. 50.How manyounceofeachwi l
lyoupur
chasetomaximi
zeyoursat
isf
act
ioni
ftheuti
li
ty
functi
onis,

Sol
uti
on:-
Theconstr
aintf
unct
ioni
sgi
venby
,
…………………… budget const
rai
nt.It
showst hedi
ff
erentcombi
nat
ionofsal
adandsoup(bywei
ght
)thatcost
syouj
ust$6.I
norderto
solvetheopti
mizati
onprobl
em thr
oughthesubst
it
uti
onmethod,l
etusrewr
it
etheconstrai
nt
functi
on as

I
nter
nal
i
zingt
hisi
ntot
heobj
ect
ivef
unct
iony
iel
ds,

Sol
vi
ngt
hei
nter
nal
i
zedobj
ect
ivef
unct
ion,
wehav
e

Subst
it
uti
ngandsol
vi
ngt
heconst
rai
ntf
unct
ionf
orx2,
wehav
e

SecondOr
derCondi
ti
on:

Subst
it
uti
ng f orx2inthesecondorderderi
vati
vewewi l
lgetthefol
l
owi ng.
Ther
efore,t
heopti
malamountofsal ad(x1)andsoup( x2)tobepur
chasedi
s8and
8ounces,andthemaxi mum sat
isf
acti
onfrom consumingtheseis

Gr
aphi
cal
l
y:-

Thereare3i ndi ff
erencecur vesconsistentwitht heutil
i
tyfunct i
on:
A:Showsal lcombi nationofx1andx 2thatpr ovidealevel ofutil
i
tyequal o1.04.
Allcombinat ionsar eat tainabl
ebutsub–opt imalatwhi cht heconsumerwi llnotmaximizehis
uti
li
ty.
C: Showsal lcombi nati
onofx1andx2t hatprov i
deal evelofut i
l
ityequalto2.29.
Giventhebudgetconst raint,al
l combinati
onsar enotaf fordable.
B: Al l
combi nati
onsexcept“ T”isnotaffordableundert hebudgetconst r
aint
.
Point“T”isnotonl yat tainablebutalsoanopt imal combination.
Proof
Weknowt hat,theconsumerwi llmaximum hi sut i
li
tyatapoi ntwher ethebudgetl i
neistangent
tothehighestpossi bleindi f
fer
encecur ve,wher e
Then

Orattheopti
malpoint
,anydecr
easeinutil
i
tyduet oasmallr
educt
ioni
ntheconsumptionofone
goodwouldbejustmatchedbythei
ncreaseinutil
i
tyduetoanincr
easei
nconstr
aintoftheot
her
good.
I
fsaladconst
rai
ntincr
easebydx2,t
hent heuti
l
itydecr
easeby

Whenev
aluat
edatopti
malpointwhereMRS[themar
ginalr
ateofsubst
it
uti
on]i
s-2,
ther
educt
ion
i
nsal
adall
owstheconstr
aintofsoupt
oincr
easeby because

And,
uti
l
ityr
iseby

Fr
om t
hef
ir
stor
dercondi
ti
ons,
aft
ersomer
ear
rangementwecanhav
e,

Wemaynot eher ethatwhent heconst rai

ntisacompl icatedf unction,orwhenther eareseveral
equalit
yconstraintsto beconsi dered,thesubst it
ution met hod mi ghtbedi f
ficultorsimpl y
i
mpossi bl
etocar ryoutinpr
actice.
Inthatcase,youneedt ouseot hert echni
ques.Thishasl edt ot hedev el
opmentofal ter
nati
ve
met hodsoff
indingextremaofaf uncti
onunderconst raints.
Onesuchmet hodi sthemet hodoft heLagrangeMul ti
plierwhi chwi llbediscussedi nthe
fol
lowingsubsect i
on.Fornow,st udyt hefol
lowingexampl est hatdemonst r
atet heuseoft he
substit
uti
ontechniquetosolveconst r
ainedoptimizati
on.

b)Second–OrderTotalDif
fer
ent
ial
Takingdi
ff
erent
ialof,
wehave

Fur
ther
,

Here,theconst
rai
ntg(x,y) =c
………….( 1)
I
nt hi
scase,Bothdxanddyarenol ongerarbi
tr
ary
.
Wemayt akedxasanarbi
trar
ychangebutdymustberegar
dedasdependentondx(equat
ion(
1)
).
Oncedxi sspeci
fi
ed,dywil
ldependongxandgybutthel
att
erder
ivat
ivesi
ntur
ndependonxand
y.hence,dywil
lal
sodependonx&y .
Tof i
ndanappropri
atenewexpressionford2z:-

Becausedyi
s v
ari
abl
eal
so.

But :
Then:-d2Z=fxx dx2 + 2 f xy dx dy + f y
y dy 2 + fy d(dy
)
=fxx dx2 + 2 f xy dx dy + f yy dy2 + f y d2 y …………( 2)
Here:d2y dy2,t
husthepr
esenceofd2yi
ntheaboveequat
iondi
squal
if
iesd2zasaquadr
ati
c
for
m.
Howev
er,
d2zcanbet
ransf
ormedi
ntoaquadr
ati
cfor
m byv
irt
ueoft
heconst
rai
ntg(
x,y
)=c.
d2g=gxx dx2 + 2 gxy dx dy + gy y dy2 + gy dy 2 = 0

…………………(3)
Subst
it
uti
ngequat
ion(
2)i
ntoequat
ion(
3):
-

But
,
And

……………………(
4)

(Herewehav et
aken asanar bit
rar
yincr
ementandyasaf
unct
ionofxbecauseoft
he
constr
aint.
…………………………(5)

Wenot et
hat isdiff
erentf
rom . canbeeli
minat
edfr
om t
heabov
eequat
ionwi
tht
hehel
pof
secondor
derdi
ff
erenti
aloft
heconst
rai
ntasshownbel
ow.

Fol
l
owi
ngt
hepr
ocedur
eusedt
oobt
ainequat
ion(
5),
wecanwr
it
e

Substit
uti
ngthi
sv al
uei
n(5),
wehav
e
…...
.(
6)
From equati
on(2)wehave,

C)TheSecond–Or derCondi ti
onf orLagr angi anf unct i
on
Thedi scussi onoft heLagr angemul tipliermet hodhasf ocusedonnecessar ycondi t
ionsf or
i
dent ifyinganext remepoi nt.But,thenecessar ycondi ti
onsdonotdi stingui shbet weenmaxi ma
andmi ni ma.
Inthi ssect ion,wepr esentsuf fi
ci entcondi ti
onsf ordet erminingwhet heranopt imum setof
val
uesr epresentamaxi mum /mi nimum oft heobj ectivefunctionint hecont extoft heLagr ange
mul ti
pliermet hod.
Thesuf f ici
entcondi t
ionsf orcasingt hest ati
onar ypoi ntofamul t
ivariatef unctioni nthef ree–
extremum casei sbasedonwhet hert heHessi anmat ri
xoft hef unct ioni sposi ti
v edefiniteor
negat ivedef ini t
ewhenev al
uatedatt hest ationar ypoi nt.
Similarly ,thesuf f
icientcondi t
ionf oraconst rainedopt imizati
onpr obl em dependuponwhet her
thel agrangi anf unct i
oni sposi ti
vedef i
niteornegat i
vedef i
nit
ewhenev aluat edatt hestationary
point.
But,wemustbecar efulnott oappl ythesocdev elopedf ortheunconst rainedpr oblem.Aswe
shall be, thenewcondi ti
onsmustbest atedi nt ermsoft hesecondor dert otal dif
fer
ent i
al,d2z.
Foraconst rainedext remum ofZ=f( x,y ),subj ectt og( x,y)=c,t hesecondor dernecessar y
andsuf ficientcondi ti
onsst i
l
l r
evolvear oundt hesi ngofd2z, evaluatedatt hest ati
onar ypoint.
But, i
nt hepr esentcont ext,wear econcer nedwi t
ht hesi gndef init
eness/semidef i
nitenessofd2
z,notf oral lpossi blev aluesofdxanddybutonl yf orthosedxanddyv alues(notbot hzer o)
sati
sf yingt hel i
nearconst rai
nt:-

Thesecond–ordernecessar
ycondit
ionsare:-
Formaximum ofz:d2z isnegat
ivesemi def
ini
te

Formini
mum ofz:-d2z isposit
ivesemi
def
ini
te
Thesecondor
dersuff
ici
entcondi
ti
on:-
Formaximum:-d2znegati
vedefi
nit
esto dg =0
 Formini
mum: -d2zposi t
ivedef
ini
te
5.1.
3Theconcept sofBorderHessi
anandJacobi
anDet
erminant
s
TheBorderHessian(H)
Todeterminet
hecondi t
ionsunderwhich ,
weconsi
derthequadrat
icf
orm:
subj
ectto

Let a =

Tosi
nce t
hesi
gnof i
sdet
ermi
nedbyt
hesi
ngof whi
chcanbewr
it
teni
ndet
ermi
nant:
-

N.B .
Thus:

Sincethedeter
mi nantcanbeobt
ainedbyborder
ingt
hedet
ermi
nantoft
hesocs,f
or(t
heHessi
an
determinant
sunconst rai
nedext
rema,byar ow andacol
umn,iti
scall
edBorderHessi
anand
denotedby .
GiventheLagrangefuncti
on wecanwr it
e

Thus,

Notethat
:Lagrangiancanbesy mboli
zedbyL,VorZ
Wer ecal
lthatif,thestat
ionarypointwil
lbeamaxima.But,i
f,thestati
onarypoi
ntwi
llbea
mini
ma.Todet erminetheconditi
onsunderwhi
ch or,weconsi
derthequadrat
icf
orm Q=au2+
2huV+bv 2subjectto substi
tuti
ngtheval
ueof fr
om t
heconst
rai
ntin weget
.
Thusthesignof wi l
lbesameast hesi
gnof.
Wenotethat
Thus ispositi
veif

Thus,taking weconcl udethatthest at

ionarypointcorr
espondst
oamaxima(ormini
ma)ifthe
si
gnoft hedeterminanti
sposi ti
ve(ornegative).
Sincet hisdet er
minantcan beobt ained bybor dering t
hedeter
minantoft
hesecond order
condi
tions,fort heunconst r
ainedextrema,byar ow andacol umn,i
tisal
socal
ledaBordered
Hessianandi sdenot edby .
TheMul ti
var
iableCase
Whent heobj ecti
vefunctiontakestheform

Thesocstil
lhi
ngesonthesi gnofd2z.
Theposit
iveornegat
ivedefini
tenessofd2zagai
ninvol
veaborder
edHessian.But
,thi
sti
met
his
condi
ti
onsmustbeexpr essedintermsofthebor
der
edprinci
pal
minor
softheHessian.
Theborder
edHessianisgivenby: -
I
tssuccessi
vebor
derpr
inci
pal
minor
scanbedef
inedas:
-

I
mpl
i
est
hesecondpr
inci
pal
minoroft
heHessi
an.

Remar
ks:
Giv
ent
heobj
ect
ivef
unct
ion andt
heconst
rai
nt t
hesecondor
dercondi
ti
onsar
e:

I
f t
henthestati
onarypointi
samini
ma.
I
f thestat
ionarypointi
samaxima,wher
e
andsoon

Condi
ti
on
Maximum
Mini
mum

Foc

Soc

….
..
..

Suffi
cientcondi t
ionf oramaxi mum wi t
honeconst raint
:-
Ifthedetermi nantoft heborderHessi anoft hatlagrangianf uncti
onev al
uatedatt hestati
onary
pointhast hesamesi gn(-1)nandt hel argestn- 1leadingprincipalminorsalter
nateinsigh,then
d2zi snegat i
vedef init
eandt hest ati
onar ypoi ntrepresentmaxi mum.
Suffi
cientcondi t
ionf orami nimum wi thoneconst r
aint:
-
Ifallofthel argestn–1l eadingpr incipalmi norsofi t
sbor deredHessianev aluat
edatt he
stati
onarypoi ntarenegat i
ve,includingt hedet ermi nantofbor deredHessiani tsel
f,thend2zi s
positi
vedef i
niteont heconst rai
ntandt hest ati
onar ypointrepresentsami ni
mum.
Exampl e1:
Findthemaxi maormi nimaoft hef unction,subj ectto24=x+2y .
Soluti
on
Whi l
eformi ngtheLagr angefunct i
on, i
tshoul dbekepti nmi ndt hattheconstrai
ntshouldbef ir
st
expressedi nimpl i
citformi.e. form.Wecanwr it
e

Fi
rstordercondi
ti
ons:
…………………………………….
.(1)
……………………………………(2)
………………………………….
.…(3)
El
iminati
ng f r
om (1)and(
2)
or or

Substi
tut
ingthi
sval
ueofyinequat
ion(
3),
weget or
Also Thusthestat
ionar
ypoi
nti
s.
Secondordercondi
ti
on:
Taki
ngdi
ff
erent
ial
oft
heconst
rai
nt,
wecanwr
it
e.

Thus, Thus, thefuncti

onhasmini
maat( 6,9).
Themi ni
mum v al
ueofthefunct
ion[cal
culatedat( 6,9)
]is612.
Example2:
Af i
rm producesr adi
osetsatt wodiffi
cul tlocati
ons.Ifx1andx2ar et
henoofr adi
oset
s
producedinl
ocationIandIIr
espect
ivel
y.Thej ointcostfunct
ionoft
hef
ir
misgi
venby

I
fthefir
m hast osuppl
yanorderof1000radi
oset s,howmanysetsshoul
ditpr
oduceateach
l
ocati
ontomi ni
mizecost
?Fi
ndtheminimum costandmar gi
nal
cost
?
Sol
uti
on:-
Theconstr
aint: 1000 = x1 + x2

…………………………….(
1)

Soc:
-

(
Young’
sTheor
em)

Thest ati
onarypoi ntisami ni
ma.
JacobianDet erminant( J)
AJacobi andet ermi nantper mit
st est i
ngf orfunct i
onaldependencef orbot
hlinearandnon- l
inear
case.Iti scomposedofal lthef i
r storderpar tialderivati
onofasy st
em equati
onar r
angedi n
orderedsequence.
Iftheequi t
ati
on’sar el i
neartheJacobi andet er
mi nanti ssameasHessi andeterminant.Howev er
,
fornon-li
nearequat ionst henormal Hessi andetermi nateisnotappl
icabl
e.
LetY1=f 1(X1,X2,X3) ,Y2=f2(X1,
X2, X3) .
Therefore,theJacobi andet erminanti sforpar t
ialder i
vati
vesisasfoll
ows
/J/=
Exampl e1:Linearcase
Y1=6X1+4X2
Y2=4X1+5X2
Soluti
on
TheJacobi andet ermi nantformi sgi veas,/J/=
Thus,/J/ =/H/=30-16=14≠0,whi chi mpliesther eisuni quesolut
ionforXandY,soy oucanf ind
thesolutionbyappl y i
ngcr amer’
sr ul e.
Exampl e2: Non-linearcase
Y1=4X1- X2
Y2=16X12+8X1X2+x22
/J/==,i
mpl ies/J/≠0.
5.1.
4.Const
rai
nedOpti
mi zat
ionandEnvel
opetheor
em
Let bet heobj
ect
ivefunctionand betheconst
rai
nt.t
heLagr
angef
unct
ioni
s
Foc

Sol
vi
ngtheequat
ions,wegett
hest
ati
onar
yval
uesof
andtheopt
imum val
ue

Di
ff
erent
iat
ingwi
thr
espectd
…………………………………( 1)

Atopt
imum : anddi
ff
erent
iat
ingwi
thr
espectd.

…………………………………………( 2).
Multi
ply
ingequat
ion(
2)byandaddt
oequat
ion(
1).
Then.
=0

=0
=0

Example:
Let beut il
it
yfuncti
onofaconsumerand behi sbudgetconst
raint
,wher
e areparamet
ers.
Deter
mi netherateofchangeofopt
imal
uti
l
itywi
thr
especteachofthethr
eepar
amet
ers.
Solut
ion:-

i
ncr easeby uni t
swi thincr
easeinM.
i
ncr easeby uni t
swi thincr
easeinP2.
5.2.Const r
ainedOpt i
mi zati
onwithInequali
tyConstraint
s
Somet i
mesopt i
mizati
onpr oblemsmayhav econstraint
sthattaketheform ofi
nequali
ti
esrather
thanequal i
ties.Constrainsmaybemor enatural
lyexpressedasinequal
i
tiesrat
herthanequali
ti
es.
Forexampl e,consumpt i
onofagoodi snon–negat ive.
Maxi mizati
onpr oblems
Letusmaxi mizesubjectt otheconstrai
nts,wheresar econst
ants.Thiscanber ewri
tt
enclearly
asbel ow.
Maximize
Subj ectto
 .

[A case wi t
h m const rai
nt s,n choi ce v ari
able non negat i
vityr estr
icti
on,Thus,we hav e3
i
ngr edients;theobject i
vef unct i
on, thesubjectfunct
ionandt henon negativi
tyconstr
aints.
Dearst udents!Whati fyouhav eaconst rai
nti n sign?
Ifweencount eropt i
mizat i
onpr oblem with≥si gn,wecansi mplymul t
ipl
ybot hsidesofi tby(-1)
andchanget hesigni nto i nequality.Wet henproceedasusual .
TheKuhn–TuckerCondi ti
on
TheKuhnTuckercondi tion, unliketheclassicalopti
mi zati
on, maynotbet henecessar ycondit
ion.
Ifwecanmakeadj ustment s,itcanbebot hnecessar yandsuf fi
cientcondi t
ions.
Thet hreecondi ti
onst osol v einequal i
ti
esusi ngKuhn- Tuckerappr oachare;
Non-negat i
vityconst r
aint
Inequal i
ti
esconst raint
Compl i
ment aryslacknesscondi ti
on
Eachst epwillbedi scussedbel ow;
EffectofNon–negat ivi
tyRest ri
cti
on
Assumet heobjectivefunct ionwi thonlyonev ar
iabl
eas
Subj ecttothenon–negat ivityrestri
cti
ons
Ther ear e3possi bil
iti
esf ormaxi mum

.
Thus`
from t
heabov
edi
scussi
on,
thecondi
ti
onf
ormaxi
mizat
ioni
s

Gener
ali
zat
ion:
-

2.EffectofInequali
tyConstrai
nts:
-
Themaxi mizationprobl
em wi t
hi nequal
i
tyconst
rai
ntsandnon-negat
ivi
tyr
est
ri
cti
onsont
he
choi
cev ariableswil
lbeclearunder.
Letusopt imize
subjecttotheconstr
aint

Step1: Changet heinequalit

yconst rai
nti ntoequalit
yconstrai
nt.Todoso,dummy( sl
ack)
variabl
esshouldbeint
roducedt ot heconstrai
ntfunct
ion.Because,t
hev aluei
nhel
efthandside
i
sl essthantheval
ueintherightside.
Theappr opr
iat
edummyv ar
iableformaxi mizati
onproblem i
sslackvari
able.

Example:Maximi
zeπ=5X1+3X2
s.
t 6X1+2X2≤36
5X1+5X2≤40
2X1+4X2≤28
X1,X2≥0.
Changeinequal
it
yt oequali
tybyaddingslackvar
iabl
es
6X1+2X2+S1=36
5X1+5X2+S2=40
2X1+4X2+S3=28
Step–2wr it
etheLagr angi
anfuncti
on:-
Step-
3setapar t
ialfir
storderder
ivateswithr
especttot
hechoi
cev
ari
abl
esl
esst
hanorequalt
o
zero( ≤0)andt hosewi t
hr espectt ot hel agrangi anmul tipli
er s≥0
Not et hatf ormi ni mi zationcount erpar tdeductt hesur plusv ariablef rom t heconst rai
ntsand
appl yal labov est eps.
Exampl e: Maxi mi zeY=10X1- X12+180X2- X22
s.t. X1+X2≤80
X1, X2≥0
Sol ution
L=10X1- X12+180X2- X22+( 80- X1- X2)
LX1=10- 2X1-=0, =10- 2X1 …………………………………………………. (1)
LX2=180- 2X2- =0, =180-2X2……………………………………………………( 2)
L=80- X1- X2=0, 80=X1+X2…………………………………………………………( 3)
From equat ions1and2=
10-2X1 =180-2X2
X1=X2- 85…………………. .………………………………………. ( 4)
Subst itut eEq4i nEq3
80=( -85+X2) +X2
80+85=2X2
165=2X2
X2=82. 5…………………………………………………………………………. (5)
Subst itut eEq5i nEq4
X1=82. 5- 85
=-2.5, whi chi snotmeani ngf ul ineconomi cs.soX1=0…………………………. (6)
Asar esul tthev alueofX2al sochanged.Hence, 80=0+X2, X2=80……………………. (7)
Tof i
ndt hev al ueof
Subst itut eEq 7i nEq2=180- 2( 80)
=20.
CHAPTERSI X
COMPARATI VESTATI CANALYSI S
6.1TheNat ur eofCompar ativeSt at ics
Compar at i
vest at icsi sconcer nedwi tht hecompar i
sonofdi fferentequi li
brium st atest hatar e
associ at edwi thdi fferentset sofv aluesofpar amet er sorexogenousv ari
abl es.Wehav et wo
classesofv ariabl es.
Choi cev ariabl es-ar eal socal ledendogenousv ari
abl esar ev ar i
ablest hatar edet er
mi nedwi t
hin
themodel .Theyar eundert hecont r olofdeci sionmaker sorf irms.
ExogenousVar iabl esorPar amet er s-ar ev ariablesout sidet hemodel .Theyar eoutoft hecont rol
oft hef irms( deci si onmaker s).Theyar edet er minedbyf actor sout si
det hemodel .
Incompar ativest at i
cs,t her ei saneedf orpoi ntofr eferenceorst arti
ngpoi ntsoast omaket he
compar i
soncl earandv isible.Forex ampl e,initi
alequi libri
um st atesar emostcommonl yusedas
star t
ingpoi nts.
Exampl e: Int het heor yofdemandandsuppl ypr icei sexogenous: Q_d=f p&Q_ s=f (p)
Inbot hcasespi sexogenousi nt hesenset hati tisnoti nfluencedbyt heact i
onsofdeci sion
maker s.I tisr at herdet ermi nedbyf act or sout sidet hecont rolofbot hconsumer sandsuppl iersin
themar ket .
Thei nitial( pre- change)equi li
br ium mayber epr esent edbyequi li
br i
um,P,andt hecor respondi ng
quant ity,Q.I fPchanges,quant itydemandedandquant itysuppl i
edcanchangeandhencet he
i
ni t
ialequi li
br i
um wi llbeupset( dist urbed) .Ifthenewequi li
br i
um st ater el
ev antt othenewpr ice
canbedef i
nedandat t
ained,t hequest iont obeposedi ncompar ativest ati
csi s‘ howwoul done
compar et henewequi l
ibrium wi t ht heol done? ’
Incompar at i
v est atics,wedi sregar dt hepr ocessofadj ust mentoft hev ariabl esi .
e.wesi mpl y
compar et hepr e- changeandt hepost -changeequi librium st at esandt hewayt hroughwhi cht he
changehascomei soutoft heconcer noft hef ield.Hence,compar at i
vest aticsisessent i
ally
concer nedwi thf indi ngt her at eofchangeoft heequi li
brium v alueoft hechoi ceorendogenous
variabl eswi thr espectt ot hechangei napar ticularpar amet erorexogenousv ar i
abl e.
From exampl e; whathappenst ot hequant i
tydemandedofagoodwhent hepr i
ceoft hegood, the
priceofr elat i
v ecommodi tiesori ncomechangesi sar epr esent ativ
equest ioni ncompar at
ive
stat i
cs.
Acompar at i
vest aticanal y siscanbeei t
herqual i
tat i
veorquant itativei nnat ur e.Ifthei nter
estisto
knownt hedi rectionofchangeoft heendogenousv ariablesasar esultoft hechangei nthe
par amet er s,t heanal y sisi squal i
tativet y pe.Howev er,i ft heconcer nisbot hondi recti
onand
magni tudeoft hechange, theanal ysiswi l
l bequant i
tative.
Ineconomi cs, theor i
esar ecommonl yt est edont hebasi sofchangesi nt hev ariableswhi chmay
ormaynotr esul tinr el ationofassumpt ions
6.2.Di ffer ent iationsandi tsAppl icationt oCompar at i
v eSt at i
cAnal y sis
Incompar at i
v est aticswemakeuseofder ivativesi nassessi ngt her ateofchangeofone
endogenousv ar i
abl easar esultoft hechangei noneormor epar amet ersorexogenousv ari
ables.
Wecanconsi derdi ffer entmodel st oi ll
ust ratet hi
spoi nt .
LetR( x)=t ot al r
ev enueasaf unct ionofoutput ,
x.
C(x)=t otal costasaf unct ionofoutput ,x.
tx=t otal t
axpai dandti saperuni tt axasapar amet erbey ondt hecont roloft hef irm.
Her e, whi lexi sanendogeneousv ariable, tisexogeneous.
Let ’
sdef inet hepr of i
tf unct ion,
π=Rx- Cx- tx
Foc: π^ '
=R^ '
x-C^ 'x- t
=0
Ifthef irmi si nper fectcompet i
tat i
v emar ket ,
P=MR; butf rom t heabov e,
MR=MC+t ⇒P=MC+t .Thi si mpl iest hatt hef i
rm choosest hel evel ofout putsucht hatMR=MC+t .
SOC: π^ ''
=R^ '
'(x)-C^ '
'(x)<0⇒R^ '
'(
x) <C^ ''
(x)i.e.,thesl opeoft hemar ginalrevenueshoul dbel essthan
thatofmar ginal cost .
Any way X i saf unct ion oft .A change i n tchange t he opt imall ev eloft he output
produced⇒X=X^ *t
orX=f (t).Ifthisi st hecase, wecani nser tthisdef init
ioni nt heFOC.
∂π^ '(x)/ ∂t=d( R^ '
(X^ *))/ dx^ *,
dx^ *
/ dt-d(C^ '
(X^ *))/dx^ *.dx^ */dt-dt/dt =0
R^''
x^ * dx^ * /dt-C^ ''
x^ *dx^ */dt =1
dx^ */dt R^ ''X^ *-C^ '
'
X^ * =1
dx^ */dt =1/ R^ '
'X^ *-C^ '
'X^ *,butl ookatt heSOCandf orπt obemi ni mi zedR^ '
'X^* -
C^ ''
X^*<0
⇒dx^ */ dt=1/ R^ '
'X^* -C^ '
'X^ *<0
dx^ */dt <0,i mpl i
esopt i
mall evelofout putandt axr atear enegat i
v elyrelat ed.i.e.Outputoft he
fir
m decr easesast het axr ateoft hef irmf acesi ncer asesandv icever sa.
Concl usi on: Apr edi ctionaboutt hesi zeanddi rect ionoft hechangeoft hechoi cev ar i
able(output
i
nt heabov ecase)canbemadebyl ooki ngatt hechangei nt hepar amet erf acingt hedeci si
on
maker -agoal incompar t
ivest atics.
Maxmi zeπx=Rx- Cx=Px- C( x)
Her et heassumpt i
oni st hatPi sanexogeneousv ariable.i.e.Pi st akenasapar amet erbey ondthe
cont rol ofdeci sionmaker .
FOC: π^ '
x=P- C^ 'x=0…. .(1)
SOC: π^ ''
x=- C^ ''
x≤0…. .(2),C^ '
'
X>0t heMCmustbeupwar dsl oppi ng.
But ,weknowt heopt i
mal l
ev elofxi saf unct i
onofPi .e.x=x^ *(P)andf rom( 1)
C^ 'x=MC=P
x=x^ * (P) Showst hesuppl yfunct ionoft hef irm st artingf rom t hepoi ntwher eMC=P.
x=x^ * (P)t el l
sushow mucht hef irm of fer stot hemar ketf orev er ymar ketdet ermi nedlevelof
price.I nser ti
ngx=x^ * (P)i nFOCP- C^ '
(x^*(P) )=0andappl y i
ngt heSOC:

dp/dpdC^ '
(x^
*)/dx^*
.dx^*/
dp=0
1-C^'
'
X^*.dx^*/dp=0⇒dx^ *
/dp1/C^'
'
(X^*)andC^'
'x^
*>0
⇒dx^ */
dp>0
Thisshowst hereisaposi t
iver
elat
ionshipbet
weentheparameterPandt
heendogenousvari
able
output
.Inot herwor dsthesupplyfuncti
onofacompeti
tiv
ef i
rmisupwardsl
opping.
i.
e,x=x^
*(p)
hasagr aphoft heslope.

Marketmodel:
letusconsi
deramar
ketmodel
forasi
ngl
ecommodi
ty.
Q_d=a-bPa,
b>0
Q_s=-
c+dPc,d>0
Wehav et woendogenousv ari
ables(P&Q)andf ourpar
amet
ers(a,b,
c&d) .Atequi
li
bri
um,
Q_d=Q_ s⇒- c+dP=a-bP⇒P=a+c/ b+d
AssumePandQar ei nit
ialequili
bri
um l
ev el
soft hePandQ,compar ati
vestati
cshelpsust o
comput et hechangei nt heinit
ialequil
i
bri
um values.
∂P/∂a=1/ b+d>0
∂P/∂b=-(a+c)/(b+d)^
2<0
∂P/∂c=1/ b+d>0
∂P/∂d=-(a+c)/(b+d)^
2<0
NB:Ther ei sani nv
er ser elati
onshipbetweent hepri
ceofthecommodi tyandthesl opesof
demandandsuppl y.
Anincreasei n‘a’i
ndicat esanupwar dshi
ftofthedemandcurve,
thati
s,asa↑⇒P↑.

Si
mil
arl
y,ani
ncr
easei
n‘c’
shi
ft
sthesuppl
ycur
vet
other
ight
.Thati
s,asc↑⇒P↑.

Ani
ncr
easei
n‘b’
incr
easest
hesl
opeoft
hedemandcur
ve.Thati
s,asb↑⇒P↓.

Ani
ncr
easei
n‘
d’i
ncr
easest
hesl
opeoft
hedemandcur
ve.Thati
s,d↑⇒P↓.

Act
ivi
ty: Expressther el
ati
onshipbetweent
hepar
ametersandtheequi
l
ibr
ium l
eveloft
he
out
putbothgraphi
cal
lyandalgebrai
cal
l
y.
Nat
ionalIncomeModel :Thediscussi
onofcompar
ati
vestat
icanal
ysi
scanalsobemadeon
nat
ionali
ncomemodel .Wehave
Y=C+I _
0+G_ 0
C=α+βY- T; α>0; 0<β<1
T=γ+δY; (γ>0; 0<δ<1)
Wher e,
α-showst heaut onomousconsumpt ionlev el.
β-measur est hemar gi nal pr oponesi tytoconsumpt ionMPC.
γ-i
ndi catest henon- i
ncomet axr ate.
δ-i
ncomet axr at e.
And, Y−Ti susual lyr ef erredt oasdi sposabl eincome( thatis,incomeaf tert ax).YandCst andfor
theendogenousv ariabl esnat ionali ncomeandconsumpt ionexpendi tur er especti
vely.I_0and
G_ 0ar eexogeneousl ydet ermi nedi nv estmentandgov ernmentexpendi tures.Thef irstequationis
anequi librium condi tion( Nat i
onalI ncome=Tot alExpendi t
ure),whi l
et hesecondandt hirdare
behav i
oralequat ions,t hati s,consumpt i
onandt axf unct i
ons.Mor eover,theequat ionsshowt hat
themodel isofcl osedt ypebecauset hetradet er
msar enoti ncor porat
ed.
NB: Theequat ionsar enei t herf unctionallydependentnori nconsistentwi theachot her.
Thus,wecandet er mi net heequi l
ibrium lev el
soft heendogeneousv ari
abl es,Y,CandTi nt er
ms
oftheexogenousv ar i
abl es.I _0andG_ 0andt hepar amet er s α, β, γand δ.
Subst itutionoft het hirdequat i
oni nt hesecondandt hent hesecondi nthef i
rst,wehav e
C=α+βY- γ+δY=α+βY- γ- δY=α- γβ+1- δY
Y=α+βY- γβ- βδY+I _ 0+G_ 0
Y+βδY- βY=α- γβ+I _ 0+G_ 0
Y1+βδ- β=α- γβ+I _ 0+G_ 0
Y=α- γβ+I _0+G_ 0/ 1+βδ- β
T=γ+δα- γβ+I _ 0+G_ 0/1+βδ- β
Thei nteresti ncompar ativ est atici stoseet heef f
ectoft hechangei noneoft heexogenous
variableont heendogenousv ariables.Todosowemakef ir
storderder i
v ati
ves.
Forexampl e:
∂Y/ ∂G=1/ 1+βδ- β>0i ndi cat et hegov ernmentexpendi turemul t
ipli
er.
∂Y/ ∂δ=- 1+βδ- β^ -2α- βγ+I _0+G_ 0β
=-βα- βγ+I _0+G_ 0/ 1+βδ- β1+βδ- β
=-βY/ 1+βδ- β<0
Thisr esul tindi cat esi ncomet axmul tipli
er.

Act
ivi
ty:
Find∂Y/
∂αand ∂Y/
∂γ,
anddet
ermi
net
hei
rsi
gns.

6.3Jacobi ansandHessi anDet er

minants
Bothf ol
low thesamepr ocedur eintheircomput ati
on.Thati s,wet akethecoef fi
cientsoff i
rst
orderder iv
ativesasel ement soft hedetermi nantofbot htheJacobianandHessi an.Or ,si
mpl ywe
evaluat ethesecondor derder i
vati
veoft heLagr angianfunctionwi t
hr espectt ot hechoi ce
variablesandt akethem asel ementsoft hedet erminants.
Thet wodi f
feri ntheirpur pose,whil
et heJacobi ansar eusedt ocheckwhet heragi vensy stem
hasasol uti
onbef oreanat t
empt ,t
heHessi ansar eusedassuf f
ici
entconditi
onst odet erminea
givenpoi ntasar el
ativemaxi mizati
onormi nimizati
onpoi nt.
Not e:Thedet ailconcept sandmat hemat icalapplicati
onofJacobi anandHessi andet erminants
addr essedi nt hepr eviouschapt er. Ther el
evanceoft hi
ssubt opichereist hatinor dert o
remi ndthest udent stor elatetheconcept swi thcompar ati
vestati
sti
csapplicati
on.

6.4Compar ati
veStat
icandGener al Functi
onModel
Inalltheabov emodelst heequili
br i
um v al
uesoft heendogenousv ari
ablescouldbeexpli
cit
ly
expressedint er
msoft heexogenousv ar
iabl
es.Accordi
ngl
yt het echniqueofsi mpl
eparti
al
dif
ferenti
ati
onwasallweneededt oobt ainthedesir
edcomparativestati
cinfor
mat i
on.
Howev er,whenamodelcont ainsf unctionsexpressedi
ngener alfor
m,expl i
citsol
uti
onsar
enot
avai l
able,i.
e.,t heequi l
ibri
um v aluesoft heendogenousv ari
ablescannotbeexpr essedexpl i
cit
ly
i
nt ermsofpar amet ersand/ orexogenousv ar
iables.Insuchcasesanew t echniquemustbe
adopt edt hatmakesuseofconcept ssuchast otaldiffer
ential
s,implicitfuncti
onr ulee.t.
c.
Let ’
st ryt oillust ratethepoi ntwi t hamar ketmodel .Consi deramar ketmodelwher eQ_ disa
funct i
onofbot hpr i
ceandanexogenousl ydeterminedi ncomeY_ 0,andQ_ sisaf unct
ionofpr i
ce
alone.
Q_ d=DP, Y_0wher e∂D/ ∂P<0and∂D/ ∂Y_0>0.
Q_ s=SP wher edS/ dP>0.
And, theequi librium posi ti
onoft hemar ketisdefinedbyQ_ d=Q_ s.Thisimpl i
es
DP, Y_0=SP
DP, Y_0-SP=0.
Event hought heabov eequat ioncannotbesol vedexpl ici
tl
yf ort heequi li
br i
um pr i
ce,P,we
assumet hatt her edoesexi stast aticequili
bri
um bef oreandaf terthechangei ntheexogeneous
variableY_ 0.
Saywehav eobt ainedP,ifincomeofconsumer schanges,t hewhol eequi li
brium willbeupset .
Thisi ndicatest hatt heopt imalv alueof Pwhi chset sthemar ketatequi l
ibr
ium i saf unct
ionof
theexogenousv ariableY_0.Thati s, P=P(Y_0).
Thechangei ni ncomeupset stheequi l
ibr
ium bycausi ngashi ftinthedemandf uncti
oni mplyi
ng
thatev eryv alueofY_ 0y i
eldsauni quev alueofP.

SinceDP, Y_ 0- SP=0i sani mplicitfunction,wecant aketitasani denti

ty:FP,
Y_0=0.
Thecompar ativest aticanalysisofsuchamodeli sconcer nedwi t
hhowachangei nYwi l
laff
ect
theequi l
ibr i
um posi tionoft hemodel .Thus, wecanr aisetwoequat i
ons:
Whati st heef fectoft hechangeof Y_ 0onPdP/ dY_ 0?
Whati st heef fectoft hechangeofY_ 0onQdQ/ dY_0?
FromDP, Y_ 0- SP=0, wecananswerquest i
onnumberone.
Appl yi
ngt hei mpl icitfunctionr ule:
dP/ dY_0=- ∂F/ ∂Y_ 0/∂F/ ∂P=-∂D/ ∂Y_ 0/
∂D/ ∂P-dS/dP=negativePosit
ive/Negat
ive-Negati
ve>0.
Hence, dP/ dY_ 0>0.Thi sindicatest hatani ncreaseinincomer esultsinanincreaseinequili
bri
um
priceandv icev ersa.
Toanswert hesecondquest i
on, thatisdQ/ dY_0:
Atequi li
br i
um, Q_ d=Q_ s=Q.Weknowt hatQ_ d=Q_s=SPbutP=P( Y_0).Thus,substi
tut
ingweget:
Q=S( P(Y_ 0))
Appl yi
ngt hechai nr ule,wehav e
dQ/ dY_0=dS/ dP. dP/ dY_ 0 wher edS/ dP>0anddP/ dY_ 0.

Theref ore,dQ/ dY_ 0>0impl yi

ngt hatani ncreasei nincomei ncreasest heequilibr
ium l
evelof
output .
Gener ally,t hecompar ativestati
cr esult
sconv eythepr oposi ti
ont hatanupwar dshif
toft he
demandcur v e(duet oar iseinY),result
sinahi gherequi l
i
brium priceandequilibri
um quant
ity
.
Theabov eder i
vati
onoft her el
ati
onshipbetweenPandQandt heexogenousv ar
iableY_0canbe
carri
edoutbyhav ingasi multaneousequat i
onappr oach.
Q_d=DP, Y_ 0;Q_s=SP
Atequi librium, Q_d=Q_ s=Q⇒Q_ d-
Q=0andQ_ s-Q=0
Writi
ngt hesei ngeneral f
orm:
DP,Y_ 0-Q=0
SP-Q=0
Conv ertingt heseequat i
onsintoani dentit
yfor m:
F^1P,Q, Y_ 0=DP, Y_0-Q=0
F^2P,Q, Y_ 0=SP- Q=0
Intheabov et woequat i
ons,wehav etwoendogenousv ari
ables,PandQ andoneexogenous
vari
abl e, Y_0.
Let’
sappl yt hepar ti
aldiff
erenti
ationrul
eonF^ 1andF^ 2,wecanf or m aJacobian,onerow the
F^1andt heot herf orF^2:
J=∂F^ 1/ ∂P&∂F^ 1/∂Q@∂F^ 2/∂P&∂F^ 2/∂Q=∂D/ ∂P&-1@dS/ dP&-1=dS/ dP-∂D/∂P>0 becausedS/ dP
 >0and∂D/ ∂P<0.
SinceJ≠0, wecanobt ainequi li
brium PandQ, thati s,P=P( Y_ 0) and
Q=Q( Y_ 0).
Thus, theequi l
ibrium condi t
ioncanbewr i
tteni nt hef or m oft hei denti
tyDP, Y_0-Q=0andSP- Q=0.
Taki ngt het otal differ entialoftheseequat ionswi thr espectt oY_ 0,wehav e
∂D/ ∂P. dP+∂D/ ∂Y_ 0. dY_ 0-dQ=0
dS/ dP. dP- dQ=0
Div i
dingbot hsi desofbot hequat ionsbydY_ 0y ields:
∂D/ ∂P. dP/ dY_ 0-dQ/ dY_ 0=-∂D/ ∂Y_ 0
dS/ dP. dP/ dY_ 0- dQ/ dY_ 0=0
Now, t
aki ngdP/ dY_ 0anddQ/ dY_0asv ariabl esandwr itingt heabov eequat ionsi nmat rixform:
∂D/ ∂P&- 1@dS/ dP&- 1dP/ dY_ 0@dQ/ dY_0=- ∂D/ ∂Y_ 0@0
Inor dert osol v efort hev ar i
ableswecanappl yt heCr amer ’sr ule,butbef oret hatl etuscheckJ≠0.
J=∂D/ ∂P&- 1@dS/ dP&- 1=dS/ dP-∂D/ ∂P>0whi chi sdi f
fer entf rom zer o.Hence, i
tispossi bet o
cont inuef i
ndt hesol ut i
onsoft hepr oblem.
dP/ dY_ 0=- ∂D/ ∂Y_ 0&- 1@0&- 1/J=∂D/ ∂Y_ 0/dS/ dP- ∂D/ ∂P>0
dQ/ dY_ 0=∂D/ ∂P&- ∂D/ ∂Y_ 0@dS/ dP&0/ J=0- ∂D/ ∂Y_ 0.dS/ dP/ dS/ dP- ∂D/∂P=∂D/ ∂Y_ 0.dS/dP/dS/dP-
∂D/ ∂P>0
Ther efor e, bothdP/ dY_ 0anddQ/ dY_ 0ar egr eat erthanzer o, i
dent i
calwithr esultsobt ainedbef ore.
6.5Li mi tationsofCompar ati
veSt ati
cAnal ysis
Compar at i
v est atici shel pfulinf indi
ngouthow adi s-equi libriati
ngchangei napar amet er
,wi ll
affectt heequi li
brium st ateofamodel .Howev er ,byitsv erynat ure, i
thast hef oll
owi ngl i
mitati
ons.
Iti
gnor est hepr ocessofadj ust mentf r
om t heol dequi l
ibr i
um t ot henewone.
Itnegl ect st het imeel ementi nv olvedinadj ust mentpr ocessf rom oneequi li
br i
um t oanot her.
Itassumes al way s a change i n a par amet erand/ orexogenous v ariabler esultsin a new
equi l
ibrium.I tdi sregar dst hepossi bil
it
yt hatnewequi li
br ium maybeat tainedev enbecauseoft he
i
nher enti nst abili
tyoft hemodel .

CHAPTERSEVEN
DYNAMI COPTI MIZATION.
7.1 Def i
nit
ionsandConcept s
Inst ati
canal ysisaneconomi cv ariabl
eisassumedt obeaf unct
ionofanot hervar
iabl
einthe
samet imeper i
od.
Exampl e:
consumpt i
onisthefunctionofincomef orthesamet i
meper i
od.
C=f(y )
2) Suppl yi
st hef uncti
onofpr i
cef orthesamet imeper iod
S=f(P)
But,dy namicanal y
sisintr
oducest heexpli
citconsiderat
ionoftimeintothepicture.
Exampl e1)Cur rentconsumpt i
onmaydependonpasti ncomes.
Ct=fYt -
1orCt =f(yt-1)
2) Quant i
tysuppl i
edinoneper iodmaydependont hepr i
ceoft hepreviousperi
od.Agood
exampl ei
sagr icultur
alfoodsupply .Theti
mepat hsofv ar i
abl
escanbest udi
edi ntwoways:
I.Discrete–timecase:-Wher et i
mecanbeconsi deredasadi scret
ev ar
iabl
e, i
nwhich(case)t
he
vari
abl eunder goesachangeonl yoncewi t
hinaper i
odoft ime.Thisut i
l
izest hemethodsof
diff
er enceequat ions.
Exampl e:
Int
erestcompoundi ngpery
ear,monthe.t
.c.Thecurveisnotsmoothher
e.
II
.Cont i
nuous-ti
mecase:- whereti
mei sconsideredasacont i
nuousv ar
iable,inwhichcase
somet hi
ng ishappening t
ot hev ar
iabl
eateach poi ntofti
me.Her e,integralcal
culusand
dif
ferenti
alequati
onsareused.Examplepopulat
iongrowth.

7.1.1Fi rst–Or derLi nearDi fferent i

al Equat ions
Differ ent i
alequat ionsar eequat i
onsi nv ol vingder ivativ
es( differ
ent ials).Theyexpr essther atesof
changeofcont inuousf unct ionsov ert ime.Theobj ectiveinwor ki
ngwi thdifferenti
alequationsis
tof indaf unct ion, wi thoutader ivative( di fferent ial)whi chsat i
sfi
est hedi f
ferent i
alequati
on.Such
af unct i
oni scal l
edt hesol utionori ntegr al oft heequat i
on.
Not e: Theor derofadi ffer ent i
alequat i
oni stheor deroft hehighestder ivativeintheequat i
on.
Thedegr eeofadi ffer ent ialequat ioni st hehi ghestpowert owhi cht heder i
v ati
veoft hehighest
orderi sr aised.
Exampl e: 1) dy /dx=2x+6: Fi rstor der, fi
rstdegr ee
2) d^ 2y /dx^ 2+dy /dx^3+x^ 2=0: Secondor der,firstdegr ee
3)d^ 2y /dx^ 2^ 7+d^ 3y /dx^3^ 5=75y : Thirdor der,fi
fthdegr ee
Gener al form: dy /dt +VY=Z wher eVandZar econst antsorf unct i
onsoft i
me.
Whenconst ant s⇒dy /dt +2Y=3
Whenf unct ionsoft ime⇒dy /dt+Vt Y=Z( t)
Not e:Foraf irstor derl ineardi fferent ialequat iondy /
dtandymustbenothi gherthanthef ir
st
degr eeandnopr oducty dy /dtmayoccur .
Gener al sol ution: y t
=Ae^ -vt+Z/ V
Wher e,A≡i sanar bi tar aryconst ant .
Ae^ -vt≡Yc⇒ t hecompl ement ar ysol ut ion.
Z/ V=Yp⇒Par ti
cul ari nt egral
Proof :Gi vent hegener al li
neardi fferent ial equat i
on:
dy/ dt+VY=Z⇒dy /
dt =Z- VY
Separ at i
ngt hev ariabl es: dy /Z-VY=dt
Integr at i
ngbot hsi des: _^ ?dy /Z-VY=_ ^?dt
-1/Vl nZ- VY=t +C_ 1
lnZ- VY=- Vt+( -
VC_ 1)
lnZ- VY=- Vt +C2 Wher e,C_2=- VC_1
Z- VY=e^ -vt+c2 Si ncei fl ny=x⇒=e^ x
Z- VY=e^ -
vt e^c2=e^ -
vt c Wher ec=e^ c2
-VY=ce^ -v t
-Z
Y=- C/ Ve^ -vt+Z/ V Wher e A=-C/V
Hence,f ort heabov egener all i
neardi fferent i
alequat i
on,t hegener alsol uti
oni sgiv
enasY=Ae^ -
vt+Z/ V .Thi sgener al sol utionwor ksf orbot hVandZasaconst ant sandf unct i
onsoft i
me.

FindingaDef initesol uti

onf oraDi ff
erent ialEquat i
on
Givenagener alsol utionofadi fferentialequat ion,ifaniniti
alcondi ti
on(boundar yvalue)isgiven,
Acanbespeci fied,inwhi chcaseadef init
esol utionispossi ble.Hence,t hear bit
raryconstanti n
thegener alsolution( i.
e.A)canbedef initi
zedbymeansofani nit
ialcondition.
GivenYt =Ae^-vt+Z/ V
Att=0,y0=Ae^ -v(
0) +Z/ V=A+Z/ V⇒A=Y0- Z/V
Thus, t
hedef ini
tesol ut i
oni sfor
medbyr eplacing‘A’ byY0-Z/Vi nt hegener alsol
utionasf oll
ows.
Yt=Y0-Z/Ve^ -
vt+Z/ V
Not e: t hepar ticulari nt
egral( Yp)r epr esentst hei nt
ertempor alequil
ibrium levelofY( t).The
compl ement arysol utionYcdenot est hedev i
ationoft hetimepat hfr
om t heequi l
ibr
ium level.For
Y(t)tobedy nami call
yst able,Ycmustappr oachzer oastappr oachesi nf
inity(∞) .
Exampl e:Fi
ndt hegener alsol
utionf oradi f
ferent i
alequati
ondy /dt+4y=12
Met hod1: si nceV=4andZ=12, thegener alsol uti
onwillbeYt =Ae^ 4t+3
Met hod 2: usi ng t he equat ion i n separ ated f orm,an expl icitsol uti
on can be f ound as
follows: dy / dt +4y =12⇒dy /dt =12- 4y
Separ atingt hev ariabl es: dy /12- 4y =dt
Integr atingbot hsi des: _^?dy /12-4y =_ ^?dt
-1/ 4ln( 12- 4y )=t+C_ 1
l
n( 12- 4y )=-4t -4C_ 1⇒) =12- 4y =e^ -
4t -4c_1
-4y=e^ -
4t e^ -4c_ 1- 12⇒y =e^ -4te^-4c_ 1/ -
4+3
Thegener al sol utioni sgi venby :
y =Ae^ -4t +3wher eA=- 1/ 4e^ -4c_ 1
7.1. 2Fi rstOr derLi nearDi ff erenceEquat ions
Firstor derl ineardi ff
erenceequat i
onsar eusedt oanal yz echangeswi threspectt ot i
mewhen
whathappensi noneper ioddependsuponwhathappenedi nt hepr eviousper i
od.Adi fference
equat ionexpr essesar el ationshi pbet weenadependentv ariabl eandal aggedi ndependent
var i
abl ewhi chchangesatdi scr et ei nter v alsoft ime. Exampl e:consumpt ioni noneper iod
dependsont hepr ev i
ousi ncome.Thati s, c_ t=fy _t-1 A_ t=f ordi scretecase,
A_ (
t)=f orcont inouscase
Not e: Theor derofadi f f
erenceequat ioni sdet er mi nedbyt hegr eatestnumberofper i
ods
l
agged.A f irstor derdi ffer enceequat i
onexpr essesat imel agofoneper i
od;secondor der
differ enceequat ionat woper i
odt i
mel agandsoon.
Exampl e: Thechangei nyastchangesf rom tt ot +1i scal ledt hef irstdi f
ferenceofy
i.
e. Δy /Δt =Δy t=y _t+1- y_t firstor der
Mostoft het imedel ta( Δ) isomi tted, andwewr itel i
ket his
I_t=ay _t-1-t_ t-2 secondor der
Q_ t=a+bp_ t-1 firstor der
y_t+3- 9y _t+2+2y _t+1+6y _t=8 thirdor der
Thesol ut ionofadi fferenceequat i
ondef inesyf orev eryv al ueoftanddoesnotcont aina
differ enceexpr essi on.
Thegener al f orm ofal i
neardi fferenceequat ioni sgi venby
y_t+1- by _t=cori nsomecasesy _t=by _t-1+c wher ebandcar econst ants.
Not e: Thedependentv ar i
abl edoesnotappearr aiset oapowerhi ghert hanoneorasacr oss
product .
Met hodsofSol v ingDi fferenceEquat i
ons
Iter at i
v emet hod
Thef irstor derdi f ferenceequat i
ondescr ibest hepat ternofchangeofybet weent woconsecut ive
per iodsonl y.Hence, oncet hedi ffer enceequat ioni sspeci fiedandani niti
alval uey _0i sgi v
en, itis
possi bl et of i
ndy ,f rom t heequat ion.Si mi lar l
y ,
oncey _1isf ound, y _2wi l
lbei mmedi atelyobt ained,
andsoon, byr epeat edappl ication( i
ter at i
on)oft hepat ter nofchangespeci fiedi nt hedi fference
equat ion.Ther esul t
sofi terationwi llenabl eust oi nt erat imepat hort hev ariableunder
consi der at ion.
Exampl es:
Fi ndt hesol utionofadi fferenceequat ionΔy _t =2assumi ngani ni t
ial v
alueof
y=15.
sol ution:
y_t=y _t+1- y _t, Δy _t =2⇒y _ t+1=y _ t+2
sinceΔThenbysuccessi v esubst i
tutionsoft =0, 1,2,3,e.t.cweobt ain
t=0: y_ 1=y _ 0+2
t=1: y_ 2=y _ 1+2=y _ 0+2+2=y _ 0+4=y _0+2( 2)
t=2: y_ 3=y _ 2+2=y _ 0+2( 2) +2=y _0+6=y _
0+3( 2)
t=3: y_ 4=y _ 3+2=y _ 0+3( 2) +2=y _0+8=y _
0+4( 2)
Foranyper iod t =y _t
=y _0+t (2)
⇒y _ t=15+2t thisi st het imepat hoft hev ar iable.
Findt hesol ut ionoft hedi fferenceequat i
ony _t+1=by _ tgiv ent hatt hei nit
ialval ueofyi sy _0.
Sol ut i
on:
Giv en y _ t+1=by _ t
Att =0: y_1=by _ 0
Att =1: y_2=by _1=bby _0=b^ 2y _0
Att =2: y _3=by _2=bb^ 2y _0=b^ 3y _0
Att =t :
y _ t=b^ ty _0
(Thesol ut ionoft hedi fferenceequat i
on)
Thesol ut ionoft hedi fferent ial equat ion:
yt=Ae^ -v t( itisnat ur al logar i
thm expr esesconst antr ateofcont inuousgr owt h)
Thesol ut ionoft hedi fference equat ion:
y_ t=Ab^ t exponent ial equat ion expr esesconst antr atesofdi scretegr owth
2. Gener al Met hod:
Givent hef i
rstor derdi ffer enceequat ion:y _t+1- by _t=ci tsgener alsol uti
onwi llconsistoft hesum
ofacompl iment ar yf unct i
on ( y_ c) andapar ticulari ntegr al( y_p) .
Def i
ni tion: y_ pequal st hei ntertempor alequi librium l ev elofyand y _ crepr esentst he
dev iationoft het imepat hf rom t hatequi l
ibrium l ev el.Fort het i
mepat ht obedy nami call
yst able,
y_cmustappr oachzer oastappr oachesi nfini ty.
y_c- ist hegener al sol utionoft her educed( homogeneous)equat ion.
y_t+1- by _t=0
y_ p- ist hegener al solut i
onoft hecompl eteornon- homogeneousequat ion.
y_t+1- by _t=c
Firstf ind y _cf ort hehomogeneousequat i
on;
Given: y _ t+1- by _ t=0⇒y _ t+1=by _ t
Thegener al sol utionwi llbeoft hef or my _t=Ab^ t(seeexampl e2abov einiterativemet hod)wher e,
Ab^ t≠0,f orot her wi se y _ tbecomes a hor i
zont alst raightl ine on t -axis.I nt hatcase t he
compl iment ar ysol ut ionwi llbe y _c=Ab^ t.
Next , findt hepar ticulari nt egr al fort henon- homogeneousequat ion:
Giveny _t+1- by _t=c,wecanchooseanysol utionpar ti
culari ntegr al( y _p).Thus,wet aket he
simpl estf or m oft het ri
alsol ut ion:y _t=k( aconst ant ).Then, sinceymai ntainst hesameconst ant
valueov ert ime,wemustal sohav e y _t+1=k.Then,subst itutet hesev aluesi ntot hecompl ete
equat ion:
y_t+1- by _t=c⇒k- bk=c⇒k1- b=c⇒k=c/ 1-b
Hence, t hepar ticul arsol ut ionbecomes y _ p=k⇒ y _p=c/ 1-b,b≠1.
Not et hatsi ncec/1- bi saconst ant, ast ationar yequi libri
um i si ndi catedi nt hiscase.
Since y _pi sundef inedatb=1,weneedt of indsomeot hersol ut i
onf ort henon- homogeneous
equat ion.So,l et ’
st ryasol ut i
onoft hef or m y _t=ktwhi chi ndi catesamov i
ngequi li
brium,
y_t=kti mpl i
es y _t +1=k( t+1) .
Subst itut ingt hesev aluesi nt hecompl eteequat ion, weobt ai n
y_t+1- by _t=c⇒kt +1- bkt =c
⇒ kt +1- kt =c, sinceb=1
⇒kt +1- t=c⇒k=ct hus, y _p=kt ⇒ y _p=ct ,forb=1
Final ly ,addi ng y _pand y _cwear ri
veatt hegener alsolut ion
y_t=y _c+ y _ p⇒ y _t=Ab^ t+c/ 1- b……. 1Gener al solut ionwhenb≠1.
y_t=Ab^ t+ct ⇒y _ t=A+ct ………. 2 Gener al sol utionwhenb=1.
Eli
mi nat ingt hear bi trar yconst ant ,t
hedef i
nitesol utionwi l
lbewr ittenas:
From ( 1)att =0, y_ 0=Ab^ 0+c/ 1-b⇒A=y _ 0-c/1-b
y_t=y _ 0-c/ 1- bb^ t+c/ 1-b 1^ 'Def init
esol ution, whenb≠1
From ( 2)att =0, y_ 0=A+c0=A
y_t=y _ 0+ct 2^ 'Def initesl ut ion, whenb=1
Exampl e: Sol vet hef i
rstor derdi ff
erenceequat iony _t+1-5y _t=1,(y _0=7/ 4).
Solut ion: Usi ngt hedef i
nitesol utionf or mul a( b≠1)
y_t=y _ 0-c/ 1- bb^ t+c/ 1-b
Sinceb=5, c=1⇒y _ t=7/ 4-1/ 1- 5( 5)^ t
-1/4
y_t=2( 5) ^t-1/ 4.
7.2.TheDy nami cSt abi li
tyofEqui li
brium
Int hecont i
nuous-t i
mecase,t hedy nami cst abi li
tyofequi l
ibrium coul dbedoneusi ngstabi l
it
y
condi tions.St abi li
t ycondi tionusi ngt hegener alf or mul af ort hedi f
ferenceequat ion y _
t=y _
0-
c/1-bb^ t+c/ 1- b , i
tcanbeexpr essedi namor egener alformul aas:
y_t=Ab^ t+a, A=y _0- c/ 1-b&a=c/ 1- b
 Thet
imepat
h y
_twi
l
lbedy
nami
cal
l
yst
abl
eonl
yift
hecompl
i
ment
aryf
unct
ionAb^
t→0as
t
→ i .e.

Assumingf ort
hemomentA=1anda=0, t
heexponenti
alexpr
essionb^
tgenerat
essev
endiff
erent
ti
mepat hsdependingontheval
ueofb( whi
chcanrangefrom - ∞t o∞).
1. Ifb>1,b^ tincr
easesatani ncreasi
ngrateasti ncr
ease.Thus,t
het i
mepat hwil
lexpl
ode
(di
verge)andmov efart
herandfart
herawayf r
om t
hehor i
zont
all
ine.
Example:Ifb=3,t
hev al
ueofb^t
atdiffer
entti
meperiodswil
lbe
t
0
1
2
3
4

b^
t
1
3
9
27
81

I
fb=1, b^t=1forall
v al
uesoft,t
hustheti
mepat hi srepr
esentedbyahor i
zont
all
i
ne.
I
f-1<b<1,(
i.
e.bisaposi t
ivef
racti
on)
, b^tdecreasesasti ncreases,
thusthet
imepat
hwoul
dbe
dampedandmov estowardstheequi
l
ibri
um li
ne( hori
zontall
ine).
Example: i fb=1/3

T
0
1
2
3
4

b^
t
1
1/
3
1/
9
1/
27
1/
81

I
fb=0,b^t
=0f oral
lval
ueoft .
I
f-1<b<0i.
e.(bisnegativefr
acti
on),b^tosci
ll
atesbetweenposi
ti
veandnegat
ivev
alues,andt
he
ti
mepat hwil
ldrawcloserandclosertothehori
zontal
line(
equi
li
bri
um).
Example:i
fb=-1/3
T
0
1
2
3
4

b^t
1
-1/
3
-1/
9
-1/
27
-1/
81

I
fb=-1,
b^t
osci
l
lat
esbet
ween+1and-
1.
Exampl
e:
T
0
1
2
3
4
5

b^t
1
-1
1
-1
1
-1

I
fb<-
1,b^
twi
l
losci
l
lat
eandmov
efar
therandf
art
herawayf
rom t
hehor
izont
all
i
neequi
l
ibr
ium.
:

.7.
3.Economi cAppl icati
onsofDy nami cOpt i
mi zation
7.3.1.UsesofDi fferentialEquati
onsi nEconomi cs
Different ialequat ionsar eused t o det erminet hecondi ti
onsf ordynami cstabil
it
yi n mi cr
o-
economi cmodel sofmar ketequi li
briaandt ot racethet i
mepat hofgr owthunderv ar
ious
condi tionsi nmacr o-model s.
Giv ent hegr owthr ateofaf unction,di f
ferentialequationsenablet heeconomi stt ofindt he
funct i
on whosegr owthi sdescr ibed.Fur t
hermor e,fr
om poi ntelasti
cit
yt heyenablehi m to
estimat et hedemandf unction.
Exampl es:
Giv en t he demand f unction Q_d=c+bP and t he supplyf uncti
on Q_s=g+hP det er
mi ne the
condi tionf orpricest abil
ityi nthemar ket( i.
e.underwhatcondi t
ionsp( t)willconverget op
(equi l
ibr i
um pr i
ce)ast i
me→ ) .
Either Q_s=g+hP, g<0andh>0Q_ s=-g+hP, g,h>0 Q_ d=c+bP,
c>0, b<0Q_ d=c-bP,c,b>0 arecor rect .
Sol uti
on:
Firstfindequi l
ibri
um pr ice( p)
:Q_d=Q_ s⇒c+bP=g+hP⇒h- bp=c-g⇒ p=c-g/h-
b
Assumet hatt her at eofchangeofpi nt hemar ketdp/ dtisaposi tivel
i
nearf uncti
onofexcess
demand( Q_ d-Q_ s):
dp/dt=m( Q_ d-Q_ s) ,wherem>0
m≡adj ust mentcoef fi
cientsubst it
utingthev aluespar amet ersofQ_ dand Q_ s;weget:
dp/dt =mc+bp- g- hp=m( c-g+(b-h) p)
Rearrangi ngt of itt hegener alformatofdi f
ferent i
al equation:dy/dt+v y
=z
dp/dt+mh- bp=mc- g
Then, si ncev =mh- bp, andz=mc- g, t
hegener al solutionbecomes
(yt
=Ae^ -
v t
+z/ v):pt =Ae^ -
m( h-
b)t+c-g/ h-
b
Thedef initesol ut ionwi l
lbe:
pt=(p0- c-g/h-b) e^ -
m( h-b)t+c-g/h-b
Usingt her elationp=c- g/h-b: pt=(p0-p)e^-m(h-b) t+p
Let k=m( h-b)⇒pt =(p0-p) e^-
kt+p
Whet herptt endst opast → dependsonwhet hert heexponent i
alexpressione^-kt→0as
t→ . sincep0andp ar econst antsandk>0k=mh- b>0,e^-kt=1/e^-kt→0as t → .
Consequent ly,t het imepat hwi llindeedl eadt hepr i
cet owar dstheequi l
ibr
ium positi
on.Inthi
s
case,t heequi l
i
br ium i ssai dtobedy namicall
yst able.
But,dependi ngont her elati
vemagni t
udeof p0 and p,t heabov esolutionyi
eldsthr
ee
possibl et imepat hs.

CobWebModel
Toi l
lustratet heuseofdi ffer enceequat ionsi neconomi canaly sis,weuseamar ketmodelofa
si
ngl ecommodi ty( cobwebmodel ).
Formanypr oduct s( suchasagr icultural commodi t
ies)whi char epl anteday earbef oremar ket
ing,
currentsuppl ydependsonl asty ear’spr ice. Q_ s,
t+1=f p_tor Q_ s,t=f(
p_t-
1) .
Whensuchasuppl yf unct ioni nter actswi thademandf unctionoft hef or m: Q_ d,t
=f(
p_t)
i
nterest i
ngdy nami cpr icepat ternswi llr
esul t.
Usingt hel inearv ersionoft hel aggedsuppl yandun- laggeddemandf unctionwegett hemar ket
model witht hef ollowi ngequat i
ons:
Q_dt =c+bp_ t(c>0, b<0)
Q_st=g+hp_ t-1(g<0, h>0)
Q_dt =Q_st equi l
ibr i
um condi ti
ons
Usingt hel astequat i
on, themodel canr educedt oasi nglef i
rstor derdiff
erenceequat i
on:
Q_dt =Q_st
c+bp_ t=g+hp_ t
-1⇒bp_ t=g- c+hp_ t-1
p_t=g-c/b+h/ bp_ t-1⇒p_ t-h/ bp_ t-1=g- c/b
Shifti
ngt heper i
odsf orwar dbyoneper i
od: p_t+1-h/bp_ t=g-c/b
Thedef i
nitesol utioni sgi v enasp_ t=p_ 0-C/ 1-bb^t+C/1-b:
p_t=p_ 0-
g-c/ b/1-h/ b( h/b) ^t+g- c/b/ 1-h/b
p_t=p_ 0-
g-c/ b-h(h/ b) ^t
+g- c/b- h l etp=g- c/b-h
Hence, p_t=p_ 0-p(h/ b) ^t+p
Twopoi ntsmaybeobser v edi nr egar dt ot histimepat h:
Thesi gnifi
canceoft heexpr essi on( p_0-p): (sameasAi nAb^ t)
It
ssi gnt ellsuswhet hert het imepat hcommences( star
ts)abov eorbel owequi librium ( mi
rror
eff
ect )
It
smagni tudehel psust odet ermi nehowf arabov eorbel owt het i
mepat his(scaleef fect )
Under‘ normal ’demandandsuppl ycondi ti
onb<0andh>0sot hath/ b<0,i
ndi cat
ingt hatt hetime
pathi sosci ll
atory .
i
fh>b, h/b>1, andp_ texpl odesdi v erges.
i
fh=b, h/b=- 1,andp_ tosci llat esuni forml y.
i
fh<b, h/b<1, andp_ tconv er gesandappr oachesp.
In shor t
,f orst abilityt hesuppl ycur vemustbef lat t
ert han t hedemand cur ve.( or ,h<b in
magni
tude).Assumingp_
0>p,
thet
imepat
hofpr
iceandquant
it
ycanbei
l
lust
rat
edasf
oll
ows.
Gr
aphi
cI l
l
ustrati
on

Fig( a) :assumet hatt hei ntersect ionofDandSy ieldst hei nter

tempor alequi l
i
br ium pr icep.Gi ven
ani ni ti
alpr i
ce( p_ 0)( wher ep_ 0>p) ,quant i
tysuppl iedi nt henextper i
od( per i
od1)wi llbeQ_ 1.But,
att hispoi nt( Q_ 1),suppl yexceedsdemand,so t hatt herei sa t endencyf orpr icet of all
.
Consequent l
y ,t
hemar ketcl ear satp_ 1.
-Assumi ng p_ 1t opr ev ail
,per i
od2’ ssuppl ywill bel imi tedt oQ_ 2(Q_2<Q_ 1).
-AtQ_ 2, demandexceedssuppl y, sot her eisat endencyf orpt or i
se.Asar esultt hemar ketclears
atp_ 2.
-Agai n,assumi ng p_ 2t opr ev ail,thesuppl yinper iod3wi llbeQ_ 3(wher eQ_ 3>Q_ 2).AtQ_ 3,
sincesuppl y>demand, t
hemar ketcl ear sat p_ 3.Repl aci ngthisreasoni ngwecant raceoutt he
pricesandquant itiesi nsubsequentper iodsbyf ol lowi ngt hear rowheads, thereby ,spinningacob
webar oundt hedemandandsuppl ycur ves.
Thet imepat hofpr icei sosci l
lat or yandexpl osivewhenSi sst eepert hanDor( h>b) .
Fig( b): p_ tosci llatesuni for ml ysot hatequi librium i sunst able. p_ tNei therdi vergesnor
conv erts.
Fig( c): p_ tisosci llatoryandconv ergentwhenSi sf l
atterthanD( h<b)sot hatequi li
brium of
ti
mepat h(p_ t
)wi llbedy nami cal l
yst able.
Exampl es:
Gi vent hedemandf unct i
on
Q_ t=20- 2p_ tandt hesuppl yf unct i
onQ_ t
=-5+3p_ t-1 ,
Find p_ tandQ_ twhen p_ 0=4
FindpandQ
Find p_ 1, Q_ 1, p_ 2,Q_ 2..e.t.c.
Usi ngcobwebscommentont hest abili
tyoft hesy st em.
Sol ution: p=g- c/ b-h=- 5-20/ -
2- 3=5
Hence, p_ t=p_ 0-p( h/ b) ^t+p⇒3/ 2>1
From Q_ dt =Q_ st⇒20- 2p_ t
==- 5+3p_ t-1⇒ p_ t=25/ 2-3/2p_ t-1
Shif ti
ngt het imeper iodsf orwar doneper i
odandr ear rangi ng:
p_ t+1+3/ 2p_ t=25/ 2
Def initesol ution:
p_ t=4- 25/ 2/ 1+3/ 2( -3/ 2)^t+25/ 2/ 1+3/ 2=( 4-5)(-3/ 2)^t+5
p_ t=-(-3/ 2)^t+5or p_ t
=5- (-3/ 2)^ t
Subst ationi nthedemandf unct i
ony i
el ds:
Q_ t=20- 25- -
3/ 2^t=10+2( -
3/ 2)^ tor Q_ t=2-3/ 2^t+10
Ift her ei sanequi l
ibr i
um ( i.e.,i ft hepr i
cei sconst antatanyper iodmeani ngP_ t=P_ t-
1=P_ t
-
2=…=P_ t-n) ,theni tf ol l
owst hat
P_ t=P_ t-1=PandQ_ t=Q
Then, from Q_ dt=Q_ st
20- 2P=- 5+3P⇒P=5
From t hedemandf unct ion, Q=20- 25⇒Q=10
UsingP_t=5-(
-3/2)^
tandQ_t
=2(-3/2)
^t+10
P_0=5--
3/2^0=4 Q_ 0=2(-
3/2)^0+10=12
P_1=5--
3/2^1=6.5 Q_1=2(-3/2)^
1+10=7
P_2=5--
3/2^2=2.75 Q_2=2(-3/2)^
2+10=14.
5
P_3=5--
3/2^3=8.38 Q_3=2(-3/2)^
3+10=3.
25
Weuset hesepoi nt
stopl
otthediagram.
UsingtheCobwebs

Theequi
l
ibr
ium i
sunst
abl
ebecauset
het
imepat
hisosci
l
lat
oryanddi
ver
gent
.

PARTI
ECONOMETRI CSONE
Chapter1
Intr
oduct i
on
1.1Def i
niti
onandscopeofeconomet r
ics
WhatI sEconomet ri
cs?
Lit
erall
y,economet rics means “ economi c measur ement ”.I tis basi cally concer ned wi th
measur i
ngeconomi cr elati
onships.
Butt he scope i s much br oader .Var i
ous economet ri
cians used di ff
er entway st o define
economet r
ics.
“Economet ricsi sthesci encewhi chi ntegrat esandappl ies:economi ctheory, economi cst ati
sti
cs,
andmat hemat i
cal economi cstoi nv estigatet heempi rical suppor tofthegener all
awsest abli
shed
byeconomi ct heor y.
Iti
saspeci al t
ypeofeconomi canal ysi
sandr esear ch.
Start
ingf rom economi ctheory,wecanexpr esst her elati
onshi pinmat hemat i
calter mssot hat
theycanbemeasur edbyt hemet hodscal ledeconomet r i
cmet hodsinor dert oobt ainnumer i
cal
estimatesoft hecoef f i
cientsoft heeconomi cr elat
ionshi ps.
WhyaSepar ateDi scipline?
Economet r
ics i ntegrate and appl i
es:-economi ct heor y,mat hematicaleconomi cs,economi c
stati
sti
cs and Mat hemat i
calst at i
st i
cs t o pr ov i
de numer icalvalues fort he par amet er
s of
economi crel ati
onshi psandv er
if
y ingeconomi ctheor ies.
Economi ct heor y-makesst at
ement sorhy pot hesest hatar emost l
yqualit
at i
vei nnature.But
Itdoesnotpr ov i
denumer icalmeasur eoft her elati
onshi pbet weeneconomi cv ari
ables.
Itdoesnott ellbyhowmucht hequant ityofonev ariablewi l
l goupordownasar esultofacer tain
changei ntheot herv ariable.
Therefore,itist hej oboft heeconomet ri
ciant opr ov i
desuchnumer i
calest i
mat es.Economet rics
givesempi rical contentt omosteconomi ct heor y.
Mat hemat ical economi cs-ex pr esseseconomi ct heor yi nmat hemat icalform( equations)wi t
hout
regar dt omeasur abi li
t yorempi rical verificat i
onoft het heory.
No essent ialdi fference bet ween mat hemat i
caleconomi csand economi ct heor y,economi c
theor yusev erbal exposi tionwhi lemat hemat ical usessy mbol s.
Bot hexpr esseconomi cr elat ionshi psi nanexactordet er mi nisti
cf orm.They( mat hemat ical
economi csandeconomi ct heor y)doesnotal low f orr andom el ement swhi chmi ghtaf fectt he
relati
onshi pandmakei tst ochast i
c.
Fur t
her mor e, theydonotpr ov idenumer ical valuesf ort hecoef fi
ci entsofeconomi cr elationshi ps.
Howev er,economet ri
cs met hod does notassume exactordet erministi
cr elationshi p. I t
assumesr andom di stur bancev ariablei nr elationshi psamongeconomi cv ar i
ableswhi chr el
at e
dev i
at i
onsf rom exactbehav ior alpat ter nssuggest edbyeconomi ct heor yandmat hemat ical
economi cs.Fur ther mor e,economet ricmet hodspr ov i
denumer icalv al
uesoft hecoef ficient sof
economi cr el ationshi ps.
Economi cst atistics-i smai nlyconcer nedwi thcol l
ect i
ng,pr ocessi ng,andpr esentingeconomi c
dat aint hef or m ofchar tsandt abl esanddescr ibingt hepat ternofeconomi cdat aov ert ime.
The dat at huscol lected const it
ut et he r aw dat af oreconomet ricwor k.Butt he economi c
statisticiandoesnotgoanyf ur thert ot esteconomi ct heor ies, howev er;t
heeconomet riciandoes.
Mat hemat ical( ori nferent i
al )st at i
st i
cs-deal s wi tht he met hod ofmeasur ementwhi ch i s
dev elopedont hebasi sofcont rolledexper i
ment s. Butst at isticalmet hodsofmeasur ementar e
not appr opr i
at ef or a number of economi cr elat ionshi ps because f or most economi c
relati
onshi pscont rolledorcar ef ullypl annedexper iment scannotbedesi gnedduet ot hef actt hat
the nat ure ofr elat ionshi ps among economi cv ariabl es ar e st ochastic orr andom.Yett he
fundament ali deasofi nf erent ialst atisticsar eappl icabl ei neconomet r
ics,butt heymustbe
adapt edt ot hepr obl em ofeconomi cl ife. Economet ricmet hodsar eadjustedsot hatt heymay
becomeappr opr iatef ort hemeasur ementofeconomi cr el ationshi pswhi char est ochast ic.The
adjust mentconsi st spr imar ilyi nspeci fyingt hest ochast i
c( random)el ement st hatar esupposed
tooper at eint her eal wor ldandent eri nt ot hedet ermi nat ionoft heobser veddat a.
1.4Economi cmodel sv seconomet ricmodel s
Economi cmodel s:
Thesi mpl ifiedanal yti
calf ramewor keconomi ct heor yi scal l
edaneconomi cmodel .I ti san
organi zedsetofr el ationshi pst hatdescr i
best hef unct ioni ngofaneconomi cent it
yunderasetof
simpl ifyingassumpt ions.Al l economi cr easoni ngi sul ti
mat elybasedonmodel s.
Economi cmodel sconsi stoft hef ollowi ngt hreebasi cst ruct ur alel ements.
Asetofv ar i
abl es
Al i
stoff undament al r
elationshi psand
Anumberofst rategi ccoef f
icient s
Economet ricmodel s:
Themosti mpor tantchar act er isti
cofeconomet r
icr el ationshi psi sthatt heycont ainar andom
elementwhi chi si gnor edbymat hemat ical economi cmodel s.
Exampl e: Mi croeconomi ct heor ypost ulatest hatt hedemandf oracommodi tydependsoni ts
price, ont hepr i
cesofot herr el atedcommodi ties,onconsumer s’ incomeandont astes. Thi si s
anexactr el ationshi pwhi chcanbewr i
ttenmat hemat ical lyas:
→ exactr elationshi p.
Howev er,manymor ef act orsmayaf fectdemandf ort hecommodi ty
. I neconomet ricst he
i
nf l
uenceoft hese‘ ot her’ factor si st akeni ntoaccountbyt hei ntroduct i
onofr andom v ar i
abl einto
theeconomi cr elat i
onshi ps.
Inourexampl e,thedemandf unct ionst udiedwi tht het ool sofeconomet ricswoul dbeoft he
stochast icf or m:
,wher e st andsf ort her andom f act orswhi chaf f
ect
t
he quant i
ty
demanded.

1.
5Met hodol
ogyofeconometr
ics
Econometri
cresear
ch isconcerned wi
ththemeasurementofthepar ameter
sofeconomi
c
rel
ati
onshi
psandwi t
hthepredicati
onoftheval
uesofeconomi
cv ar
iabl
es.Somevar
iabl
esar
e
post ulat edascausesoft hev ar iationofot herv ariabl es.
Based on t he post ul ated t heor eticalr elat ionshi ps among economi cv ari
abl es,economet ri
c
resear chori nqui rypr oceedsal ongt hef ol l
owi ngst ages:
Speci ficat iont hemodel
Estimat ionoft hemodel
Evaluat ionoft heest imat es
Evaluat ionoft hef or ecast ingpoweroft heest imat edmodel
Speci ficat ionoft hemodel
Ther elat ionshi psbet weeneconomi cv ar iabl esmustbeexpr essedi nmat hemat i
cal form.
Thisst epi nv olv est hedet er mi nat ion:
thedependentandi ndependent( expl anat or y )v ariabl esi nt hemodel .
theapr ior itheor et i
cal expect at i
onsaboutt hesi zeandsi gnoft hepar amet ers.
themat hemat ical for m oft hemodel
Itshoul dbebasedoneconomi ct heor yandi nf ormat ionr elat edt ot hephenomena.
Thus,pr esupposes knowl edge of economi ct heor y and f ami liari
t y wi th t he par ti
cular
phenomenonbei ngst udi ed.
Themosti mpor tantandt hemostdi f fi
cul tst ageofanyeconomet ricresear ch.
Int hisst aget her eexi st senor mousdegr eeofl ikel ihoodofcommi tt
inger rorsori ncor rectl
y
speci fyingt hemodel .
Someoft hecommonr easonsf ori ncor r
ectspeci ficat ionoft heeconomet r
icmodel sare:
Imper fect ions, loosenessofst at ement si neconomi ct heor ies.
Limit ationofknowl edgeoft hef act or swhi char eoper at i
vei nanypar t
icularcase.
Formi dabl eobst acl espr esent edbydat ar equi r
ement si nt heest i
mat i
onofl argemodel s.
Themostcommoner rorsofspeci ficat i
onar e:
Omi ssi onsofsomei mpor tantv ar iabl esf rom t hef unct ion.
Theomi ssi onsofsomeequat ions( forexampl e, i
nsi mul taneousequat ionsmodel ).
Incor rectmat hemat i
cal for m oft hef unct i
ons.
Estimat ionoft hemodel
Thisi spur elyat echni calst agewhi chr equi resknowl edgeoft hev ari
ouseconomet r
icmet hods,
theirassumpt i
onsandt heeconomi ci mpl icat i
onsf ort heest imat esoft hepar amet ers. Thi s
stagei ncl udest hef ol l
owi ngact iv i
ties.
Gather ingoft hedat aont hev ar iabl esi ncl udedi nt hemodel .
Exami nat i
on oft he i dent ificat ion condi tions oft he f unct ion ( especi all
yf orsi mul taneous
equat i
onsmodel s) .
Exami nat i
onoft heaggr egat ionspr obl emsi nv olvedi nt hev ar i
abl esoft hef unct i
on.
Exami nat i
onoft hedegr eeofcor rel ationbet weent heex pl anat or yv ariables( i
.e.exami nationof
thepr obl em ofmul ti
col li
near ity ).
Choi ce of appr opr iate economet rict echni ques f orest i
mat ion,i .e.t o deci de a speci fi
c
economet r
icmet hodt obeappl iedi nest imat ion; suchas, OLS, MLM, Logi t,andPr obit.
Evaluat ionoft heest imat es
Thisst ageconsi st sofdeci di ng whet hert heest imat esoft hepar amet ersar et heor eticall
y
meani ngf ul,st atist icallysat isf act or yandr eliabi l
ityoft her esul t
s.
Fort hispur poseweuset hr eegr oupsofcr iteriawhi chmaybecl assi fi
edi nt o:
Economi capr ior icr iter i
a:-det ermi nedbyeconomi ct heor yandr efert ot hesi zeandsi gnoft he
paramet er sofeconomi cr elat ionshi ps.
Statist i
calcr iter i
a( fi
rst -or dert est s) :-det er mi nedbyst atisticalt heor ywhi chf ocusesont he
evaluat i
on oft he st at i
st icalr eliabi li
ty oft he est imat es oft he par amet ers oft he model .
Correl at ioncoef ficientt est ,st andar der rort est ,t-test ,F- test ,andR2- testar esomeoft hemost
commonl yusedst atistical test s.
Economet ri
ccr iter i
a( second- or dert ests) :-
Setbyt het heor yofeconomet ricsi nv est igat ingwhet hert heassumpt ionsoft heeconomet ri
c
met hodempl oy edar esat isf iedornot .
Itser v easasecondor dert est( asst at isticalt est s)i .e.t heydet ermi net her eliabil
ityoft he
statistical criter i
a;
Help us t o est abl ish whet hert he est imat es hav et he desi rabl e pr oper ties ( unbiasedness,
eff
iciencyandconsi stency ).
Economet riccr iteriaai m att hedet ect ionoft hev iolationorv alidit
yoft heassumpt ionsoft he
vari
ouseconomet rict echni ques.
4)Ev al uat ionoft hef orecast ingpoweroft hemodel :
Forecast ingi soneoft heai msofeconomet ricr esear ch. I ti spossi blet hatt hemodelmaybe
economi cal lymeani ngf ulandst atisticallyandeconomet ri
cal lycor rectf ort hesampl eper iodf or
whicht hemodelhasbeenest imat ed;howev er,itmaynotbesui t
abl ef orf orecast ingduet o
vari
ousf act or s( reasons) .
Theref or e,t his st age i nv olves t he i nvest igat ion oft he st abi li
tyoft he est i
mat es and t heir
sensitiv i
t yt ochangesi nt hesi zeoft hesampl e.
Desirabl epr oper tiesofaneconomet ricmodel
An economet ri
c modeli s a modelwhose par amet ers est imat ed wi th some appr opriate
economet rict echni que.
The‘goodness’ oft hemodel isjudgedaccor di ngt ot hef ol l
owi ngdesi rabl epr operties.
Theor et ical pl ausi bility:-compat iblewi tht hepost ulatesofeconomi ct heor y.
descr ibet heeconomi cphenomenat owhi chi trelates.
Explanat or yabi li
ty :
- shoul dbeabl et oexpl aint heobser vat ionsoft heact ualwor l
d.
mustbeconsi stentwi tht heobser v edbehav ioroft heeconomi c v ariableswhoser elationshi pit
determi nes.
Accur acyoft heest imat es:- coef fi
ci entsshoul dbeaccur at e, appr oximat etothet r
ue
par amet ersoft hest ructur almodel .
shoul dpossesst hedesi rabl epr oper ti
esofunbi asedness, consi stencyandef fi
ciency .
Forecast ingabi li
ty :
-t hemodel shoul dbeabl et opr edi ctf utur ev aluesoft hedependent
(endogenous)v ariabl es.
Simpl icity :
- Themodel shoul dr epr esentt heeconomi cr elationshi ps, fewert heequat ions
andt hesi mpl ermat hemat i
cal f
or m wi thoutaf fect ingdesi rablepr oper ti
es.
GoalsofEconomet ri
cs
Threemai ngoal sofEconomet ricsar e:
Analy sis-t est ingeconomi ct heor y
Poli
cymaki ng-obt ai ningnumer i
calest imat esoft hecoef ficient sofeconomi cr elati
onshi psf or
poli
cysi mul at ions.
Forecast ing-usi ngt henumer i
calest i
mat esoft hecoef fi
cient si nor dert of orecastt hef uture
val
uesofeconomi cmagni tudes.
TheSour ces, Ty pesandNat ureofDat a
Thesuccessofanyeconomet r i
canal ysisul timat elydependsont heav ailabil
ityoft heappr opriate
data.I ti st her efor eessent ialt hatwespendsomet imedi scussi ngt henat ure,sour ces,and
l
imitat ionsoft hedat at hatonemayencount eri nempi r
ical anal y sis.
TypesofDat a
Threet ypesofdat amaybeav ailablef orempi ricalanal ysi s.
TimeSer i
esDat a
i
sasetofobser vat i
onsorv aluest hatav ariabl et akesatdi ffer entt i
mes.
maybecol lect edatr egul art imei nter vals( dai ly,weekl y, mont hly, quar terl
y ,annual l
y).
mostempi rical wor kbasedont imeser i
esdat aassumesst ationar y( i
tsmeanandv ar i
ancedonot
varysy stemat ical l
yov ert ime) .
Cross- Sect ionDat a
dataononeormor ev ariabl escol lectedatt hesamepoi nti nt ime,suchast hecensusof
popul at i
onconduct edbyt heCensusBur eauev ery10y ear s.
Justast imeser iesdat acr eat et heirownspeci alpr obl ems( stationar it
yi ssue) ,cr oss-sect ional
datahav et hei rownpr oblems( t
hepr obl em ofhet erogenei ty).
PooledDat a: -
Inpool ed, orcombi ned, dat aar eel ement sofbot ht imeser iesandcr oss- sectiondat a.
Panel ,Longi tudi nal, orMi cr opanel Dat a-i saspeci al typeofpool eddat ai nwhi cht hesamecr oss-
sectional uni t(say ,af ami lyoraf i
rm)i ssur v ey edov ertime.
Chapt erTwo
2.1THECLASSI CALREGRESSI ONANALYSI S
[TheSi mpl eLi nearRegr essi onModel ]
Economi ct heor iesmai nlyconcer nedont her elati
onshi psamongeconomi cvari
ables.
Wheni tphr asedi nmat hemat i
cal t
erms, canpr edictt heef fectofonev ari
ableonanot her.
Thef unct ionalr elationshi psoft hesev ariablesdef inet hedependenceofonev ari
ableupont he
otherv ariable( s)i nt hespeci ficf orm.
Thespeci ficf unct ionalfor msmaybel inear,quadr atic,logarit
hmi c,exponent i
al,hyperbolic,orany
otherf orm.
Inthischapt erweshal lconsi derasi mplel i
nearr egr essi onmodel ,i
.e.ar el
ati
onshi pbetweent wo
variabl esr elated i nal inearf orm.We shal lf i
r stdi scusst wo i mpor tantfor msofr el
ation:
stochast i
cand non- st
ochast ic,among whi ch weshal lbeusi ng t hef ormeri n economet ri
c
anal ysis.
2.1.St ochast i
candNon- st ochast icRel ati
onshi ps
Ar elationshi pbet weenXandY, asY=f (X)issai dt obe:
Det ermi nistic-i ff oreachv alueoft hei ndependentv ari
able( X)t herei soneandonl yone
correspondi ngv alueofdependentv ari
able( Y).
Stochast ic-i ff orapar t
icul arv alueofXt herei sawhol epr obabili
sticdistri
buti
onofv aluesofY.
Insuchacase,f oranygi v env alueofX,t hedependentv ari
ableYassumessomespeci f i
cv al
ue
onlywi thsomepr obabi l
it
y .
Let’si ll
ust ratet hedi sti
nct ionbet weenst ochast icandnonst ochast icrelati
onshipswi tht hehelp
ofasuppl yf unct ion.
Assumi ngt hatt hesuppl yf oracer taincommodi tydependsoni tspr i
ce( otherdeterminant staken
tobeconst ant )andt hef unct ionbei ngl inear,ther elat i
onshi pcanbeputas:

From t heabov er elati

onshi pf orapar ticularv alueofP, t
her eisonl yonecor respondi ngv alueofQ.
Ther efore,iti sadet erminist i
c( non- stochast i
c)r elati
onshi psi ncef oreachpr icet herei sal ways
onlyonecor respondi ngquant it
ysuppl i
ed.
Thisi mpl iest hatal lthev ariati
oni nYi sduesol elytochangesi nX,andt hatt her earenoot her
factorsaf f
ect i
ngt hedependentv ariable.
Ifpl ottedonat wo- dimensi onalpl ane,woul df allonast raightl ine. Howev er,ifwegat her
obser vat i
onsont hequant i
tyact uallysuppl i
edi nt hemar ketatv ariouspr icesandwepl ott hem
onadi agr am weseet hattheydonotf allonast raightl i
ne.
Theder ivationoft heobser vationf rom t helinemaybeat tributedbyt hef actor s:
Omi ssi onofi mpor t
antv ariablesf rom t hef unct ion
Random behav iorofhumanbei ngs
Imper fectspeci ficationoft hemodel
Errorofaggr egat ion
Errorofmeasur ement
Tot akei ntoaccountt heabov esour cesofer ror swei ntroducear andom v ariabl e(err
ort erm or
random di sturbanceorst ochast icter m)whi chisdenot edbyt hel etter‘
u’ or‘ ’int heeconomet ri
c
funct i
on.Socal l
edbecauseui ssupposedt o‘ disturb’theexactl inearrel ationshi pwhi chex i
st
betweenXandY.
Byi ntroduci ngt hisr andom v ari
abl ei nt hef unctiont hemodel isr ender edst ochast i
coft hef orm:
………………………………………………………. (2.2)
Thusast ochast i
cmodeli samodeli nwhi cht hedependentv ar i
ablei snotonl ydet erminedbyt he
explanat or yv ariable(s)incl udedi nt hemodelbutal sobyot her swhi char enoti ncludedi nthe
model .
2.2.Si mpl eLi nearRegr essi onmodel .
Theabov est ochast icrelationshi p( 2.2)wi thoneexpl anat or yv ariabl
ei scal ledsi mpl el i
near
regressi onmodel .
Thet r
uer elat i
onshi pwhichconnect st hev ariablesi nvolvedi sspl itintotwopar ts:
apartrepresent
edbyal ineand
apartrepresent
edbyt her andom t erm‘u’.
Thescatterofobservationsr epr esent
st hetr
uer el
ati
onshipbet weenYandX.Thelinerepr
esents
theexactpartoftherelationshi pandt hedeviat
ionoft heobservati
onfrom t
heli
nerepr
esentsthe
random componentoft her el
ationship.
Wereitnotf ortheerror sint hemodel ,wewoul dobserveal lthepoint
sontheli
ne. Howev er
becauseoft herandom di sturbance,weobser v epointsdeviati
ngf r
om theli
ne. Thesepoints
diver
gefrom theregr
essi onl ineby.

Thef irstcomponenti nt hebr acketi st hepar tofYexpl ai

nedbyt hechangesi nXandt hesecond
i
st hepar tofYnotexpl ainedbyX, t
hati stosayt hechangei nYi sduet other andom i nf l
uenceof .
2.2.1Assumpt ionsoft heCl assi cal LinearSt ochast icRegr essi onModel
Thecl assi calsmadei mpor tantassumpt i
oni nt hei ranal ysisofr egr ession. Themosti mpor t
antof
theseassumpt i
onsar edi scussedbel ow.
Themodel i
slineari npar amet er s.
Theyassumedt hatt hemodelshoul dbel ineari nt hepar amet er sr egar dlessofwhet herthe
expl anat or yandt hedependentv ar i
abl esar el inearornot .
becausei fthepar amet er sar enon- l
ineari ti sdi f
ficultt oest imat et hem si ncet heirv al
uei snot
knownbuty ouar egi venwi tht hedat aoft hedependentandi ndependentv ariable.
isar andom r eal variable
thev aluewhi chumayassumei nanyoneper ioddependsonchance; itmaybeposi t
iv e,negative
orzer o.
Themeanv al
ueoft her andom v ariable( U)iszer o
foreachv alueofx, ther andom v ariable( u) mayassumev ariousv alues, somegr eat ert hanzer o
andsomesmal lert hanzer o,
i
fweconsi der eral lt heposi tiveandnegat i
v ev aluesofu,f oranygi venv alueofX,av eragev al
ue
equal tozer o. I not herwor dst heposi ti
v eandnegat ivev aluesofucancel eachot her .
Mat hemat ically, ………………………………. .….(2.3)
Thev ar ianceoft her andom v ariabl e(U)i sconst anti neachper i
od
(Theassumpt ionofhomoscedast i
city)
Mat hemat ical ly;
; (since) .
Thisconst antv ar i
ancei scal ledhomoscedast icityassumpt ionandt heconst antvar i
ancei t
selfis
calledhomoscedast icv ariance.
Ther andom v ariabl e( U)hasanor mal distribut i
on
thev aluesofu( foreachx)hav eabel lshapedsy mmet ri
cal distri
but ionaboutt hei
rzer omeanand
const antv ari
ance, i
.e.
(………………………………………. .……2. 4
Ther andom t er msofdi fferentobser vat i
onsar ei ndependent .
(Theassumpt i
onofnoaut ocor relation)
Thismeanst hev aluewhi cht her andom t erm assumedi noneper ioddoesnotdependont he
valuewhi chi tassumedi nanyot herper i
od.
Algebr aical l
y ,
…………………………. .….(2.5)
The ar easetoff ixedv aluesi nt hehy pot het i
calpr ocessofr epeat edsampl i
ngwhi chunder l
ies
thel inearr egressi onmodel .
Thismeanst hat ,i nt akingl ar genumberofsampl esonYandX,t he v aluesar ethesamei nall
sampl es, butt hev aluesdodi ff
erf rom sampl et osampl e,andsoofcour sedot hev aluesof.
Ther andom v ariabl e( U)i sindependentoft heexpl anat oryv ariables.
Thismeanst her ei snocor rel ationbet weent her andom v ar i
ableandt heexpl anatoryv ar i
able.
Iftwov ariablesar eunr elatedt hei rcov ariancei szer o.
Hence………………………………………. .…. (2.6)
Theexpl anat or yv ariablesar emeasur edwi thouter ror
Random v ariabl e( u)absor bs t he i nfluence ofomi tted v ariables and possi bly er r
ors of
measur ementi nt hey ’
s. i .e.,wewi llassumet hatt her egressor sar eer rorf ree,whi l
eyv al
ues
mayormaynoti ncludeer rorsofmeasur ement .
Dearst udents! Wecannowuset heabov eassumpt i
onstoderi
vet
hefol
l
owi ngbasi
cconcept
s.
A.Thedependentv ar i
abl e isnormallydistr
ibuted.
i.
e………………………………( 2.7)
Theshapeoft hedi stributi
onof i sdet er
mi nedbyt heshapeofthedistri
but
ionofwhichis
normal byassumpt ion4.
Since,beingconstant ,theydon’taff
ectthedi str
ibut
ionof.
Furthermor e,t
hev aluesoft heexplanatoryv ar
iabl
e,,areasetoffi
xedvaluesbyassumpti
on5
andt hereforedon’
taf fecttheshapeoft hedi str
ibut
ionof.

B.successi vev al uesoft hedependentv ariablear ei ndependent ,i .e

2.2.2Met hodsofest i
mat ion
Speci fyi
ngt hemodel basedont heassumpt ionsi st hefirstst ageofanyeconomet ri
cappl i
cat
ion.
The nextst ep i st he est imat i
on oft he numer i
calv alues oft he par amet ers ofeconomi c
relationships.
Thepar amet er soft her egressionmodel canbeest imat edbyv ariousmet hods.
Thr eeoft hemostcommonl yusedmet hodsar e:
Ordi naryleastsquar emet hod( OLS)
Maxi mum l i
kel i
hoodmet hod( MLM)
Met hodofmoment s( MM)
2.2.2.1Theor dinaryl eastsquar e(OLS)met hod
Themodeli scal l
edt het ruer elationshi pbet weenYandXbecauseYandXr epresentthei
r
respect i
v epopul ationv alue, and
aret het ruepar amet er ssincet heyar eest imat edf rom t hepopul ationv alueofYandX.
Buti tisdi ff
icultt oobt aint hepopul at i
onv alueofYandXbecauseoft echnicaloreconomi c
reasons.
Sowet aket hesampl ev alueofYandX.Thepar amet ersest i
mat edf r
om t hesampl ev al
ueofY
andXar ecal l
edt heest imat orsoft het ruepar amet ers andar esy mbol i
zedas.
Themodel ,iscal l
edest imat edr elationshi pbet weenYandXsi nce ar eest i
mat edf r
om the
sampl eofYandXand
repr esentst hesampl ecount erpar toft hepopul at i
onr andom di sturbance.
Est i
mat ionof byl eastsquar emet hod( OLS)i nv olvesf indingv aluesfort heest imat or
s whi ch
willmi nimizet hesum ofsquar eoft hesquar edr esiduals( ).
From t heest imat edr elat i
onshi p: ,
weobt ain:
……………………………( 2.6)
………………………. (2.7)
Tof i
ndt hev aluesof t hatmi nimi zet hissum,wehav et opar ti
allydifferentiat
ewi threspectto
andsetequal tozer o.
1.
Rear rangingt hisexpr essi onwewi llget : ………. (2.9)
Ifyoudi v i
de( 2.9)by‘ n’andr earrange, weget

2.
Note:thetermintheparenthesi
sinequati
on2.8and2.11i
stheresi
dual
,.Hencei
tispossi
bleto
rewri
te(2.
8)and(2.11)as and.
I
tfoll
owst hat
;
and
I
fwer earr
angeequation(2.
11)weobt ai
n;
………………………………………. ( 2.
13)
Equati
on(2.9)and(2.13)ar
ecal l
edt
heNor mal
Equati
ons. Subst
it
uti
ngtheval
uesoffr
om (2.
10)
to(2.
13),weget:
 =(
2)
…………………. (
2.14)
Equat
ion(
2.14)canberewr
itt
eninsomewhatdi
ff
erentwayasf
oll
ows;

Subst
it
uti
ng(
2.15)and(
2.16)i
n(2.
14)
,weget

Now, denot i
ngas, andasweget ;
………………………………………( 2.17)
Theexpr essioni n(2.17)istermedast hedevi
ationfor
m.
2.2.
2.2Est i
mat i
onofaf uncti
onwi thzeroint
ercept
Suppose: , subjecttot herestr
icti
on
Toest imat e,puti naf orm ofr estri
ctedminimizati
onpr
obl
em andt
henLagr
angemet
hodi
s
appli
ed.
Wemi nimi ze:
Subjectto:
Thecomposi tefunctionthenbecomes
wher eisaLagr angemul tipl
ier.
Wemi ni
mi zet hefunctionwi t
hr espectto

Subst
it
uti
ng(
ii
i)i
n(i
i
)andr
ear
rangi
ngweobt
ain:

……………………………………..
(2.
18)
Thi
sfor
mulai
sinact
ual
observat
ionnoti
ndev
iat
ionf
orms.

2.2.2.3.St atisti
cal Proper tiesofLeastSquar eEst imat ors
Anest i
mat emustbecl oset ot hev alueoft het ruepopul ati
onpar amet er s( smallrangeof
variationar oundt het ruepar amet er )
.How ar ewet ochooseamongt hedi fferenteconomet ric
met hods,t heonet hatgi ves‘ good’est imat es?Weneedsomecr i
teri
af orj udgi ngt he‘goodness’
ofanest imat or.‘Cl oseness’oft heest i
mat etot hepopul at
ionpar amet eri smeasur edbyt he
meanandv ari
anceorst andar ddev i
ati
onoft hesampl i
ngdi stri
butionoft heest i
mat esoft he
differenteconomet ricmet hods.Weassumet heusualpr ocessofr epeat edsampl ingi .
e.we
assumet hatwegetav erylargenumberofsampl eseachofsi ze‘n’;wecomput et heestimat es’ s
fr
om eachsampl e,andf oreacheconomet ri
cmet hodandwef ormt hei rdi stributi
on.Wenext
compar et hemean( expect edv alue)andt hev ariancesoft hesedi st
r i
but i
onsandwechoose
amongt heal t
ernat i
veest imat est heonewhosedi stri
but i
oni sconcent ratedascl oseaspossi bl
e
aroundt hepopul ationpar amet er.
Pr oper ti
esofOLSEst i
mat or s
Opt imum pr opertiesofOLSest imat orsmaybesummar izedbywel lknownt heor em knownast he
Gauss- Mar kovTheor em.
Stat est hat : “ t
hel eastsquar esest imat orsar el i
near,unbi asedandhav emi nimum v ari
ance,
BLUEpr oper ty
.
Anest imat oriscal ledBLUEi f:
Linear :
-al inearfunct ionoft her andom v ari
abl e,suchas, thedependentv ariableY.
Unbi ased: -itsaver ageorexpect edv al
uei sequal tot het ruepopul ati
onpar amet er.
Mi nimum v ar i
ance: -Ithasami nimum v ari
ancei nt hecl assofl i
nearandunbi asedest i
mat ors.An
unbi asedest imatorwi ththel eastv ar i
ancei sknownasanef fi
cientestimat or.
Accor dingt ot heGauss- Mar kovt heor em,t heOLSest i
mat orspossessal lt heBLUEpr oper t
ies.
Thedet ail
edpr oofoft hesepr opertiesar epr esent edbel ow
Linear i
ty : f or( )
Pr oposi tion: ar el i
neari nY.
Unbi asedness:
Proposi tion: ar et heunbi asedest imat orsoft het ruepar amet ers
Recal lthati fi sanest imat orof t henand
i
fi stheunbi asedest imat orof t henbi as=0i .
e.
Inourcase, ar eest imat orsoft het r
uepar amet ers. Toshow t hatt heyar etheunbi ased
estimat or soft heirr espect i
v epar amet ersmeanst hat :
Mi nimum v arianceof
Now, wehav etoest ablisht hatoutoft hecl assofl inearandunbi asedest imat orsof, possesst he
smal lestsampl i
ngv ariances. Fort his,weshal lf i
rstobt ainv arianceof andt henest abli
sht hat
eachhast hemi ni mum v ar i
ancei ncompar isonoft hev ariancesofot herl i
nearandunbi ased
estimat or sobt ai
nedbyanyot hereconomet ri
cmet hodst hanOLS.
Wehav ecomput edt hev ar i
ancesOLSest imat or s.
Now,checkwhet hert hesev ariancesofOLSest imat or sdopossessmi nimum v ari
ancepr operty
compar edt othev ar i
ancesofot herest i
mat orsoft het rue, otherthan.
Toest abl isht hat possessmi ni
mum v ariancepr oper ty,wecompar et hei rv ari
anceswi tht hatof
thev ariancesofsomeot heral ternat i
velinearandunbi asedest imat orsof ,say and.Now,we
wantt opr ovet hatanyot herl inearandunbi asedest imat oroft het ruepopul ati
onpar amet er
obt ainedf rom anyot hereconomet ricmet hodhasl argerv ariancet hant hatofOLSest imator s.
Let sf i
rstshowmi nimum v arianceof andt hent hatof.
Thev ar ianceoft her andom v ariable( Ui)
TheOLSest imat or si nv olv
e(pop/ nv ari
anceoft her andom di sturbancet erm.
Butdi fficultt oobt aint hepopul ationdat aoft hedi sturbancet erm becauseoft echnicaland
economi cr easons.
Hence,di f
ficulttocomput e;t hisi mpl i
est hatv ar i
ancesofOLSest imat esar eal sodi f
ficultt o
comput e.
Butwecancomput et hesev ariancesi fwet aket heunbi asedest i
mat orofwhi chis f rom t he
sampl ev alueoft hedi sturbancet erm e_ if
rom t heexpr ession:
…………………………………. .2. 30
Touse i nt heexpr essionsf ort hev ariancesof,wehav etopr ovewhet her i sunbi ased
estimat orof, i.
e.,

2.2.2.
4.St atisti
cal testofSi gnif
icanceoft heOLSEst imat ors(FirstOr dert ests)
Afterest imat i
onoft hepar amet ers,weneedt oknowhow‘ good’ isthef itoft her egressionl inet o
thesampl eobser v ationofYandX.Thati s,weneedt omeasur ethedi spersionofobser vat i
ons
aroundt her egr essi onl i
ne.
Thecl osert heobser vationt ot hel ine,t hebet tert hegoodnessoff i
t,i.e.thebet teri st he
explanat ionoft hev ar iationsofYbyt hechangesi nt heexpl anat oryv ar
iabl es.
Wedi v idet heav ailabl ecr i
teriaintot hreegr oups:t het heoreticalapr ioricr it
eria,thest at i
stical
crit
eri
a,andt heeconomet ri
ccr it
er i
a.Undert hi
ssect ion,ourf ocusi sonst ati
sticalcr i
teria( f
irst
ordertest s).Thet womostcommonl yusedf i
rstordert est sineconomet r i
canal ysisar e:
The coef fi
cientofdet ermi nati
on ( R2):-i s used f orj udging t he expl anatory poweroft he
i
ndependentv ar i
abl e( s).
Thest andar der rort est soft heest imat ors:isusedf orjudgi ngt hest atisti
calr eli
abili
tyoft he
estimat esoft her egr essi oncoef fi
cient s.
1.TheCoef f i
cientofDet er minationr 2: AMeasur eof“ GoodnessofFi t”
Aftertheest imat ionoft hepar amet ers,weneedt oknowhow‘ good’i st hef i
toft her egr ession
l
inet ot hesampl eobser vati
onofY andX.Thati s,weneedt omeasur et hedi sper sionof
observ ationsar oundt her egressionl ine.
Thecl osert heobser v ationt othel ine,t hebet tert hegoodnessoff i
t,i .
e.itbet t
erexpl ainst he
variat
ionsofYbyt hechangesi nt heexpl anatoryv ariables.
Weuset hreecr i
ter i
a: -thet heoreticalapr i
oricri
teri
a,t hest atisticalcriteria,andt heeconomet ric
crit
eri
a.
Thet womostcommonl yusedSt atistical cri
teri
a( f
irstor dertest s)ineconomet r
icanal ysisar e:
The coef fi
cientofdet erminat
ion (R2) :
-is used t o deci de the explanatory poweroft he
i
ndependentv ariable(s).
Thest andarder rortestsoft heest i
mat ors(s.e):i
susedf orjudgingthestatist
icalreli
abi
li
tyofthe
esti
mat esoft her egressi oncoeffi
ci
ents.
1.TheCoef f
icientofDet erminat
ionr2:AMeasur eof“ GoodnessofFi t”
r2isameasur et hattellshowwel lt
hesampl eregressionl i
nef it
sthedata.
Itmeasur est hepr opor t
ionorper centageoft het otalv ari
ationi nYexplainedbyt heregressi
on
model .
Ifal
ltheobser vati
onsl ieont heregressionline,a“ perfect”fit,butthi
sisrarelythecase, t
herewill
besomeposi t
ive u_ iandsomenegat ive u_ i
.
Byf i
tt
ingt hel i
ne wet rytoobtaintheexpl anat i
onoft hev ari
ati
onoft hedependentv ar
iableY
producedbyt hechangesi ntheexplanatoryv ar
iableX.

………………………………. ..(2.47)
_ ^??y _i
^2=( ^2_ ^??x_ i
^2+_ ^??u_i^2
OR,
i.
e.
…t hisshows t het otalv ariat i
oni nt heobser vedYv aluesaboutt hei rmeanv al
uecanbe
partiti
onedi nt ot wopar ts:
ESS=at tri
but ablet ot her egressi onl ineand
RSS=r andom f or cesbecausenotal lactual Yobser vati
onsl i
eont hef i
tt
edl ine.
Nowdi vidingt heabov eEquat ionbyTSSonbot hsides, weobt ai
n:
1=ESS/ TSS+RSS/ TSS
1=_ ^?(Y_i
-Y)^2/_ ^?(
Y_i-
Y)^2+_ ^?? u_ i
^ 2/_^ ?(Y_ i-Y) ^2
Wenowdef i
ner 2as:
r^2=_ ^?(Y_ i-
Y) ^2/_^?(Y_i-Y)^2=ESS/ TSS or r
^2=1
-_^??u_i^2/ _^? (Y_i-Y) ^2 r^2=1- RSS/ TSS
Proper ti
esofr 2
Itisanonnegat ivequant i
ty
Itslimitsar e0≤r2≤1.
r^2=1meansaper fectf it,thatis, i =Yi foreachi .Ont heot herhand,
r^2=0means nor elationshi pbet weent her egr essandandt her egressor.
2.6Conf idenceI nt erv alsandHy pot hesi sTest i
ng
Tot estt hesi gni fi
canceoft heOLSest i
mat orsweneed:
Var i
anceoft hepar amet erest i
mat ors
Unbi asedest imat orof
Theassumpt ionofnor mal di str
ibut ionofer rorterm:
TheOLSest imat or s ar eobt ai
nedf rom asampl eofobser vationsonYandX. Si ncesampl ing
errorsar ei nev i
tabl ei nal lest i
mat es,i tisnecessar yt oappl yt estofsi gnificancei nordert o
measur et hesi zeoft heer roranddet ermi net hedegr eeofconf i
dencei nor dert omeasur et he
validit
yoft heseest i
mat es. Thi scanbedonebyusi ngv arioust ests.Themostcommonones
are:
Standar der rortest
Student ’st -test
Conf idencei nt erv al
Alloft heset est ingpr ocedur esr eachont hesameconcl usion.Letusnow seet hesetesting
met hodsonebyone.
Standar der ror( s.et est
Todeci dewhet hert heest imat ors ar esi gnifi
cant l
ydi f
ferentf rom zero,i.e.whet hert hesampl e
from whi cht heyhav ebeenest i
mat edmi ghthav ecomef rom apopul ati
onwhoset ruepar ameters
arezer o, .
For mal l
ywet estt henul lhy pothesis
Nul l hypot hesi s, against
Alternativehy pot hesis,
Thest
andar
derr
ortestmaybeout
li
nedasfoll
ows.
1st: Comput
estandar
derr
orofthepar
ameter
s.

2nd: compar etheSEwi tht henumer icalvaluesof.

Deci sionr ule:
If, acceptH_ 0andr ejectH_ 1.
Weconcl udet hatisst ati
sticall
yinsi gnificant.
If,rejectH_ 0andacceptH_ 1.
Weconcl udet hatisst ati
sticall
ysi gnificant.
Theaccept anceorr ej
ectionoft henullhypot
hesishasdefini
teeconomicmeaning.
Accept anceoft heH_ 0: (
thesl opepar amet eriszero)impli
esthat
:
theexpl anat or
yv ariabletowhi cht hisest imaterelat
esdoesnotinf l
uencethedependentv
ari
abl
e
Y.
evidencedt hatchangesi nXl eaveYunaf f
ected.
,i.
e.nor elationshipbet weenXandY.
Numer i
cal exampl e:
Supposet hatf r
om asampl eofsi zen=30, weest i
matethefol
lowingsupplyfunct
ion.

Testt
hesi
gni
fi
canceoft
hesl
opepar
amet
erat5%l
evel
ofsi
gni
fi
canceusi
ngt
heSEt
est
.

,
,
Thei mpl icationi s isstati
st i
callysignifi
cantat5%l ev elofsignifi
cance.
Note:TheSEt estisanappr oximat edt est(whi chisappr oxi
mat edfrom thez-testandt -t
est)and
i
mpl iesat wot ail
testconduct edat5%l evel ofsignifi
cance.
Student ’st -
test
I
mpor tantt ot estthesignifi
canceoft hepar amet ers.
atestofsi gni fi
cancei sapr ocedur ebywhi chsampl er esult
sareusedt ov eri
fythet r
uthorf al
sity
ofanul l hy pothesis.
ast atistici ssai dt obest atisti
call
ysi gnif
icanti fthev al
ueoft het eststat
ist
icliesint hecrit
ical
regi
on.I nt hiscaset henullhy pothesisisrejected.
Byt hesamet oken,at estissai dtobest at i
sti
cal
lyi nsignifi
cantifthev al
ueoft heteststati
sti
c
l
iesi nt heaccept anceregion.
From st atistics,anyv ari
ableXcanbet ransf ormedi ntotusi ngthegener alf
ormul a:
,wi thn- 1degr eeoffreedom.
Wher e ( =v alueoft hepopul ationmean
sampl eest i
mat eofthepopul ati
onst andar ddev iati
on

sampl esize
Recallthatundert henor mal i
tyassumptionwecander i
vet het-v
alueoft heOLSest i
mates
withn- kdegr eeoff r
eedom.
Wher e: SE=i sst andarderror
k=numberofpar amet ersinthemodel .
foll
owst hetdistr
ibuti
onwi thn-2df.
Andsi ncet histestst atisticfol
lowsthetdi stri
buti
on,theconf i
dence-interv
alst at
ementscanbe
constructedast hef ollowi ng:
Pr-t
_(/2≤(_i-?(
_i^*/se( _i
≤t_ (
/2=1-(
Wher e? (
_ i
^*i st hev alue( _i
under H_ 0andwher e-t_
(/2and t _
(/2ar et hevaluesoft( t
he
cri
ti
cal tvalues)obt ainedf rom thettabl
e.
Rearrangingtheabov eequationgives:-
Pr?(_
i^* -
t_(
/ 2se(_i≤(_ i≤? (_i
^*+t_(/
2se( _i
=1-(
whichgivest heinter
vali nwhi ch(_iwi
llf
allwit
h
1-αprobabi l
ity
.
100(1-α)%CIest abli
shedi nabov eequat
ionisknownast her egi
onofaccept anceofH_ 0and
theregi
on(s)outsi
det heCIis(are)call
edther egion(s)ofrejecti
on(ofH0)orthecri
ti
calregi
on(s).
theconfi
dencelimits,theendpointsoftheconf i
dencei nter
val,arealsocall
edcr
it
icalval
ues.
Si
ncewehav et wopar ametersi nsimplelinearr egr
essionwi thinter
ceptdi
ff
erentfrom zero,
ourdegreeoffreedom i sn-
2.
Formal testi
ngprocedureoft hehy pothesis:
Step1:Constructnullandalter
nat i
vehypothesis
fortheslopepar ameter;

and
f
orthei
nter
cept
.

Step2: Comput e t estst atist ic( t^ *),byt akingt hev alueof i nthenul l
hy pot hesis.
Step3: Choosel ev el ofsi gnif icance.
i
st hepr obabi lityofmaki ng‘ wr ong’ deci sion
thepr obabi l
ityofr ej ect ingt hehy pot hesi swheni tisact ual l
yt rueort hepr obabi li
tyofcommi tt
ing
at ypeIer r
or .
Step3: Checkwhet heri tisonet ail testort wot ail test.
Ifthei nequal itysi gni nt heal ter nat iv ehy pot hesisi s,t heni ti mpl iesat wot ailtestanddi vi
det he
chosenl ev el ofsi gni ficancebyt wo; deci det hecr iti
cal r
egi onorcr iticalval ueoftcal l
ed t _c.
i
ft hesi gni sei ther>or<, onet ailt estandnoneedt odi videt hechosenl ev elofsignificanceby
twot oobt aint hecr itical v aluef rom t het -tabl e.
Exampl e:
Ifwehav e
agai nst:
Thent hisi sat wot ailt est . I ft hel ev elofsi gni f
icancei s5%,di v i
dei tbyt wot oobt aincr i
tical
valueoftf rom t het -tabl e.
Step4: Obt ain t _cat wi thn- 2dff ort wot ail t
est .
Step5: Compar et ^*( thecomput edv alueoft )wi tht _c(cr i
tical valueoft )
Ift^*>t _c, rejectH_ 0andaccept H_ 1.
Concl usi on: i sst at i
st ical lysi gni fi
cant .
Ift^*<t _c, acceptH_ 0andr eject H_ 1.
Concl usi on: i sst at i
st ical lyinsi gni ficant .
Ast atistici ssai dt obest at istical lysi gnificanti fthev alueoft het estst atisticliesint hecr i
tical
region.I nt hiscaset henul lhy pot hesi si sr ejected.
Byt hesamet oken,at esti ssai dt obest atisti
callyi nsignificanti fthev alueoft het estst at
isti
c
l
iesi nt heaccept ancer egi on.
TABLE2. 1: ThetTestofSi gni ficance: Deci sion
Numer ical Exampl e:
Supposet hatf rom asampl esi zen=20weest i
mat et hef ollowi ngconsumpt i
onf unct i
on:
C=100 + 0. 7Y +e
(75.5) (0.21)
Thev aluesi nt hebr acket sar est andar der rors. Wewantt ot estt henul l hypot hesis:
agai nst
,uset het -
testat5%l ev el ofsi gni ficance.
thet -val uef ort het estst atist i
ci s:
=
Sincet heal ter nat ivehy pot hesi s( H1)i sst at edbyi nequal i
tysi gn( ),iti
sat wot ailtest,hencewe
divide t oobt ai nt hecr iticalv al ueof‘ t’ at=0. 025and18 degr eeoff reedom ( df)i .
e.(n-2=20-
2). Fr om t he
t-table‘ tc’ at0. 025l ev el ofsi gni fi
canceand18df i s2. 10.
Sincet *=3. 3andt c=2. 1, t* >tc. I ti mpl i
est hat i sst at i
sticallysigni ficant.
Conf i
dencei nt erv al
Conf i
denceI nt erv alsf orRegr essi onCoef fi
cient s
Reject ionoft henul lhy pot hesi sdoesn’ tmeant hatourest imat es ar ecor rectest imat eoft he
truepopul at i
onpar amet er .I tsi mpl ymeanst hatourest i
mat ecomesf rom asampl edr awnf rom
apopul ationwhosepar amet er i sdifferentf rom zero.
Inordert odef i
nehow cl oset heest i
mat et ot hetruepar ameter,wemustconst ructconf idence
i
ntervalf ort het ruepar amet er ,inot herwor dswemustest abl i
shlimit i
ngv aluesar oundt he
esti
mat e wi t
hi n which t he true par amet eri s expected tol ie within a cer t
ain“ degr ee of
confidence” .I nthi
srespectwesayt hatwi thagi venprobabi
lit
yt hepopul ati
onpar amet erwi llbe
wit
hint hedef i
nedconf idencei nt erval(confidencel imits).
Wechooseapr obabil
ityinadv anceandr efert oitasconf i
dencel evel(int
er valcoeffi
cient).I tis
customar ilyineconomet ricst ochooset he95%conf i
dencelev el. Thi smeanst hati
nr epeat ed
sampl ingt heconf i
dencel i
mits,comput edf rom thesampl e,woul dincludet het ruepopul ation
paramet eri n95%oft hecases. I nt heot her5%oft hecasest hepopul at i
onpar amet erwi llfal
l
outsidet heconf idenceint erv
al.
Inat wo- t
ai ltestat(l evelofsi gnifi
cance,t hepr obabili
tyofobt aini
ngt hespeci fict-v
al ueei ther
–tcort ci satn- 2degreeoff reedom. Thepr obabil
it
yofobt ai
ni nganyv alueoftwhi chi sequal
to atn- 2degr eeoff r
eedom i s.
i.e. …………………………………………( 2.57)
but ……………………………………………………. (2.58)
Subst it
ut e( 2.
58)i n(2.
57)weobt ainthef ol
lowi ngexpression.
………………………………………. .
(2.59)

Theli
mitwi
thi
nwhi chthetrue li
esatdegreeofconfi
dencei
s:
;wher
e isthecri
ti
calv al
ueoftatconfi
denceint
ervalandn-
2degr
eeoff
reedom.
Thetestprocedur
eisoutl
inedasfoll
ows.

Decisionr ul
e: Ifthehy pothesizedv alueof i nthenullhypot
hesi
siswithi
ntheconfi
dence
i
nterval,acceptH0andr ejectH1. Thei mpl
icat
ionisthat isst
ati
sti
cal
l
yinsi
gni
fi
cant
;whil
eif
thehy pothesi
zedvalueof i nt henul lhypot
hesisisoutsi
detheli
mit
,rej
ectH0andacceptH1.
Thisindicates i
sstati
sti
callysignif
icant.
*A100( 1−α)%conf i
denceinter
valf
orβ2.

Numerical
Exampl
e:
Supposewehaveest
imat
edt
hef
oll
owi
ngr
egr
essi
onl
i
nef
rom asampl
eof20obser
vat
ions.

Thevaluesinthebracketarestandarderr
ors.
Construct95%confidenceint
ervalfort
heslopeofparameter
Testthesigni
fi
canceoft hesl
opepar ameterusi
ngconstr
uctedconf
idencei
nter
val
.
Solut
ion:
Theli
mi twit
hinwhichthetrue liesat95%confidencei
nterv
ali
s:

,
,at0.
025l
evel
ofsi
gni
fi
canceand18degr
eeoff
reedom i
s2.
10.

Theconf i
dencei ntervalis:
(1.09, 4.67)
Thev al
ueofi nt henul lhypot hesisiszer
owhi chimpliesitisoutsi
detheconf i
denceinterval
.
Hence i sst at i
sticallysignif
icant.
ReviewQuest ions
Economet r
icsdeal swi t
ht hemeasur ementofeconomi crelati
onshi
pswhichar estochasti
cor
random.Thesi mpl estform ofeconomi cr elat
ionshi
psbetweent wov ar
iabl
esXandYcanbe
representedby :
; wher ear er egressionpar amet ersand t hestochasti
cdistur
banceter
m
Whatar et her easonsf orthei nser
ti
onofU- terminthemodel ?
Thef oll
owingdat ar eferstot hedemandf ormoney( M)andt herateofint
erest(R)inforeight
diff
erenteconomi cs:
M(Inbi
l
li
ons)
56
50
46
30
20
35
37
61

R%
6.
3
4.
6
5.
1
7.
3
8.
9
5.
3
6.
7
3.
5

Assumi ngar el
ati
onship,obt
aintheOLSest i
matorsof
Calculatethecoeff
ici
entofdeter
mi nati
onfort
hedataandi
nterpr
etit
sv al
ue
Ifina9t heconomyt herateofi nter
estisR=8.1,pr
edi
ctthedemandf ormoney(M)int
his
economy .
Thef oll
owingdatareferst
othepriceofagood‘ P’
andthequanti
tyofthegoodsuppl
i
ed,‘
S’
.
P
2
7
5
1
4
8
2
8

S
15
41
32
9
28
43
17
40

Esti
matetheli
nearregressi
online
Esti
matethestandarderror
sof
Testthehypot
hesisthatpri
ceinfluencessuppl
y
Obtai
na95%conf i
denceinter
val f
or
Thefoll
owingresul
tshav ebeenobt ai
nedfrom asi
mpl
eof11obser
vat
ionsont
hev
aluesofsal
es
(Y)ofafi
rm andthecorrespondingprices(X)
.

Est
imat
etheregressionli
neofsal
eonpr
iceandint
erpr
ettheresul
ts
Whati
sthepartofthev ar
iati
oni
nsal
eswhichi
snotexplai
nedbytheregr
essi
onl
i
ne?
Est
imat
ethepriceelasti
cit
yofsal
es.
Thefol
lowi
ngt
abl
eincl
udest
heGNP(
X)andt
hedemandf
orf
ood(
Y)f
oracount
ryov
ert
eny
ear
s
peri
od.
year
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989

Y
6
7
8
10
8
9
10
9
11
10

X
50
52
55
59
57
58
62
65
68
70

Esti
matethef oodf uncti
on
Computet hecoef f
icientofdeterminat
ionandfindtheexplai
nedandunexpl
ainedv ari
ati
oninthe
foodexpenditur
e.
Computethest andar derroroftheregressi
oncoeff
ici
entsandconductt
estofsignifi
canceatthe
5%levelofsignif
icance.
Asampl eof20obser vati
oncorrespondingtot
heregressi
onmodel gavethef ol
lowingdata.

a.Est
imate
b.Cal
culatethevari
anceofourest
imates
c.
Est
imatet
hecondit
ionalmeanofYcor r
espondi
ngtoav al
ueofXfixedatX=10.
Supposethatar
esear
cheresti
mat esaconsumpti
onsfuncti
onandobt
ainsthefol
lowi
ngr
esul
ts:

whereC=Consumpt i
on, Yd=di sposablei ncome,andnumbersinthepar
ent
hesi
sar
ethe‘
t-
rati
os’
Testthesigni
fi
cantofYdstat
ist
ical
lyusingt -
rat
ios
Determinetheest
imatedst
andarddev i
ationsoft hepar
amet
erest
imat
es
StateandproveGuass-Mar
kovtheorem
Giv
enthemodel
:
wit
husual
OLSassumpt
ions. Der
ivet
heexpr
essi
onf
ort
heer
rorv
ari
ance.

CHAPTERTHREE
TheMult
ipl
eLi
nearRegr
essi
onAnal
ysi
s

3.1I ntroduct ion

Adependentv ariabl eYcandependonmanyf act or s( expl anat oryv ar iables)orr egressors. For
i
nst ance, indemandst udieswest udyt her elationshi pbet weenquant it
ydemandedofagoodand
priceoft hegood,pr iceofsubst i
tutegoodsandt heconsumer ’si ncome. Themodelwe
assumei s:
---
-----
-----
-----
--( 3.1)
Wher equant itydemanded,P1i spr iceoft hegoodi t
sel f,P2i spr iceofsubst i
tutegoods,Xii s
consumer ’si ncome, andar eunknownpar amet ersand i st hedi st ur bance.
Equat i
on ( 3. 1)i s a mul tipler egr essi on wi tht hr ee expl anat oryv ar i
abl es.I n generalf orK-
explanat or yv ar i
abl ewecanwr itet hemodel asf ollows:
---
---
-(3. 2)
Wher ear eexpl anat or yv ariables,Yii st hedependentv ariabl eandar eunknownpar amet ersand
i
st hedi st urbancet erm.
Thedi st urbancet erm hassi mi l
arnat uret ot hati nsi mpl er egr ession, r
ef l
ect ing:
thebasi cr andom nat ur eofhumanr esponses
errorsofaggr egat i
on
errorsofmeasur ement
errorsi nspeci ficat ionandanyot herf act or s, othert hant hatmi ghti nf l
uenceY.
Let ’
sst ar tourdi scussi onwi t
ht heassumpt ionsoft hemul tipleregr essi onsandwewi l
lpr oceed
ouranal ysi swi tht hecaseoft woexpl anat or yv ariabl esandt henwewi l
lgener ali
zet hemul ti
ple
regressi onmodel int hecaseofk- expl anat oryv ar iabl esusi ngmat rixal gebr a.
3.2Assumpt ionsofMul t
ipleRegr essi onModel s
Mul tiplel inearr egr essi onmodelassumpt ionsar et hesamewi tht hesi ngl eexpl anatoryv ar i
able
model dev elopedear lierexceptt heassumpt ionofnoper f
ectmul ticol li
near i
ty.
Theseassumpt ionsar e:
Randomnessoft heer rort erm: Thev ariabl eui sar eal random v ar iabl e.
Zer omeanoft heer r ort erm:
Homoscedast icity :Thev arianceofeach i st hesamef oral lthe v al ues.
i.
e. ( const antv ar iance)
Nor mal ityofu: Thev aluesofeach ar enor mal lydi stribut ed.
i
.e.
Noaut ocor r
el ationorser i
al corr elationbet weent hedi sturbancet er ms:
The v al ues of ( cor respondi ng t o )ar ei ndependentf rom t he v alues ofany ot heru_ j
(correspondi ngt oXj )f ori (j.
i.
e.cov (u_ i,
u_ j,
)=0 i
.e. f or
Independenceof andXi :
Ever ydi st urbancet er m i sindependentoft heexpl anat oryv ariables,
cov( u_ i,X_ 1i)=cov( u_ i,
X_ 2i)=0. i .e.
Nospeci f
icat ionbi as: Themodel mustbecor rect lyspeci fied.
Noper f ectmul ticol l
inear ity:-
Theexpl anat oryv ar iabl esar enotper fect l
yl i
near l
ycor rel
ated.
Informal ly,no col linear it
y means none oft he r egr essor s can be wr itt
en as exactl inear
combi nat ionsoft her emai ningr egr essor si nt hemodel .
For mal ly,nocol linear itymeanst hatt her eexi stsnosetofnumber s, λ2andλ3, notbothzer osuch
that: λ2X2i +λ3X3i =0.
Ifsuchanexactl inearr elationshi pexi sts,t henX2andX3ar esai dt obecol l
inearorl inearly
dependent .Ont heot herhand, ifλ2=λ3=0, thenX2andX3ar esai dt obel inear l
yindependent .
3.3.AModel wi thTwoExpl anat or yVar iabl es
Inor dert ounder standt henat ur eofmul t i
pler egr essi onmodeleasi ly,west artouranal y
si swi th
twoexpl anat oryv ariabl es, thenext endt hi st ot hecaseofk- expl anat or yv ariables.
3.3.1 Est i
mat i
onofpar amet er soft wo-explanator
yvar i
ablesmodel
Themodel , PRF: ……………………………………( 3.3)
Letussupposet hatt hesampl edat ahasbeenusedt oest i
mat ethepopul at
ionregressi
on
equati
onandequat i
on( 3.4)hasbeenest i
mat edbysampl eregressionequation,whichwewr it
e
as:
………………………………………………. (3.5)
Wher e ar eestimat orsoft he andi sknownast hepredictedvalueofY.
Givensampl eobser vati
onson, weest i
mat e(3.3)usi
ngt hemet hodofl eastsquare(OLS).
………………………………………. (3.
6)
issampl erelati
onbet ween.
…………………………………. .
(3.7)
Toobt ai
nexpr essi onsf ort hel eastsquar eest i
mators,wepar t
iall
ydiffer
enti
atewi t
hrespectto
andsett hepart i
alder i
vativesequal tozero.
……………………….( 3.8)
…………………….( 3.9)
…………………. .(
3.10)
Summi ngf rom 1t on, themul t
ipleregressionequat i
onpr oducest hreeNormal Equati
ons:
……………………………………. (3.11)
…………………………( 3.12)
………………………. ..(3.
13)
From (3.
11)weobt ain
--
--
--
----
--
--
--
---
--
----
---
--
---
--
---
---
----
--
--(3.14)
Substit
uting(3.14)i n(3.12), weget :

-
--
--
--(
3.15)
Weknowthat

Substi
tutingt heaboveequat i
onsi nequati
on(3.
14),thenormalequat
ion(3.
12)canbewr
it
teni
n
devi
ationf orm asfoll
ows:
…………………………………………( 3.16)
Usi
ngt heabov eprocedureifwesubst i
tut
e(3.
14)in(3.13)
,weget
………………………………………. .
(3.
17)
Let
’sbr i
ng( 2.17)and(2.18)together
………………………………………. (
3.18)
………………………………………. (
3.19)
and caneasi lybesolvedusingmat r
ix
Wecanr ewr i
tetheabovet woequat i
onsinmatri
xform asfol
lows.
………….(
3.20)
=

I
fweuseCr amer ’srul
etosol vetheabovematr
ixweobtai
n
…………………………. .……………. .(3.21)
…………………. ………………………( 3.22)
3.3.
2Coeffi
cientofMul t
ipleDet erminat
ion(
R2)
I
nmul ti
pleli
nearr egr
essionmodel sR2measuresthepropor
ti
onofv
ari
ati
oni
nthedependent
vari
abl
eexplainedbyal l
expl anatoryvar
iabl
esi
ncl
udedint
hemodel.
Thecoeffi
cientofdeterminat i
onR2hasbeendef i
nedas:

I
nthr
eev
ari
abl
ecase

R^
2measur
est
hepr
edi
cti
onabi
l
ityoft
hemodel
overt
hesampl
eper
iod,
or
Itmeasur eshowwel l t
heest i
matedr egressionl i
nefitsthedat a.
Thev alueofR^ 2i sequalt ot hesquar edsampl ecor r
elationcoeffi
ci entbet ween. Si ncethe
samplecor relati
oncoef ficientmeasur est helinearassociat i
onbetweent wov ar
iables,
IfR^2ishi gh,thereisacl oseassoci ati
onbet weent hev aluesofandt hev al
uespr edictedbyt he
model, .I nt hi
scase, themodel issaidt o“fit”thedatawel l.
IfR^2i slow,t hereisnoassoci ati
onbet weent hev al
uesof andt hev aluespredict edbyt he
model, andt hemodel doesnotf i
tthedat awel l
.
AdjustedCoef f
ici
entofMul ti
pleDet er
mi nati
on( )
Onedi ffi
cultywithi sthati tcanbemadel ar gebyaddi ngmor eandmor ev ar
iables,ev enifthe
vari
ablesaddedhav enoeconomi cjusti
fication. Algebr ai
cal l
y,i
tisthef actthatast hev ar
iabl
es
areaddedt hesum ofsquar ederrors(RSS)goesdown( itcanr emainunchanged, butt hisisrar
e)
andthusgoesup.
Adjust
edi sanal ternativewayofmeasur inggoodnessoff i
t ,
andsy mbol izedas.
Iti
scomput edas:

Thismeasur edoesnotal way sgoesupwhenav ari

ableisaddedbecauseoft hedegreeof
freedom term n-kinthedenomi nator.
Ast henumberofv ar
iablesincreases,RSSgoesdown,butsodoesn-k.Theeff
ectondependson
theamountbywhi ch f all
s.
Whi l
esolv i
ngonepr oblem,t hiscorrect
edmeasureofgoodnessoff i
tunfort
unatelyi
ntr
oduces
anotherone. I tlosesitsinterpret
ati
on;isnol
ongert
hepercentofvari
ati
onexplai
ned.
Var i
ancesandSt andar dErrorsofOLSEst imat
ors

v ar
(_0=1/ n+?X_1^2_ ^?
?x_2i^ 2+?X_ 2^2_^??x_1i
^2-
2X_ 1X_ 2_ ^? x_1ix_2i/_ ^??x_ 1i^2_^? ?x_2i^ 2-(_^? x_1ix_2i)^2.
(^2
s.e(_0=&var(_0
Wher e( _0i sint erceptcoef f
icient
var (
_1= _^??
x_2i
^2/_ ^??x_ 1i^2._^??x_ 2i^2–( _^?x_ 1ix_2i)^2.(^2
var(
_1=
(^2/ _
^ ??x_ 1i^2.(1–? r_12^ 2 ,wher er _12ist hesampl ecoeffi
cientofcorr el
ati
onbet ween X_ 1
and X_ 2.
s.
e(_1=&var(_1
var(_2= _^??x_ 1i ^2/_^??x_ 1i^
2. _^??x_2i
^2–( _
^?x_1i
x_2i)^2.(
^2
var(_2=( ^2/_ ^??x_2i ^2.(1–? r_12^ 2
s.e(_2=&v ar (_2
Proper tiesofOLSEst imat orsandGauss- Mar kovTheor em
Stat i
stical Pr oper t
iesoft hePar amet ers( Matr i
xAppr oach)
(Gauss- Mar kovTheor em)
Inmul ti
pler egressi ons, theOLSest imat orsmustsat isfythesmal lsamplepr opert
yofest imators
(BLUEpr oper ty)
.Nowwepr oceedt oexami net hedesi redproperti
esoft heest i
mat orsinmat r
ix
not ati
onsasf oll
ows:
Linear i
ty
Unbi asedness.
3.Mi nimum v ariance
3.5.Hy pothesi sTest ingi nMul tipleRegr essi onModel
Inmul ti
pler egressi onmodel swewi llunder t
aket wot est sofsignifi
cance.
Significanceofi ndi vi
dual par amet ersoft hemodel and
Ov erallsigni ficanceoft hemodel .
1.Test sofi ndividual signi ficance
Assumi ngt hat , wecanuseei thert het -
testorst andarderrortesttotestahy pothesi
sabout
i
ndi vidual par t
ialregr essi oncoef ficient.
Toi ll
ust rat econsi dert hef ol l
owi ngexampl e.
Let…………………………………( 3.
51)
i)
 i
i
)

Thenul lhypot hesi sin( i

) st at
est hat,holdingX2const antX1hasnoi nfl
uenceonY.
Simi l
arlyhypot hesi s( i
i)st ates thathol ding X_ 1 constant,X_ 2 has no i nf
luence on the
dependentv ar iabl eY_ i
.
Tot estthesenul l hypothesiswewi lluset hef oll
owi ngtests:
Standar derrort est :Undert histestingmet hodl et
’st estonlyfor.Thetestfor wi l
lbedoneint he
sameway .
;wher e
If,weacceptt henul lhypot hesisthati s,wecanconcl udet hattheestimatei snotstat
ist
icall
y
signifi
cant.
If SE( _1<1/2( _ 1,wer ejectt henul lhy pothesis,andwecanconcl udet hattheest i
matei s
statist
ical
lysigni ficant.
Not e:The smal l
ert he st andar d errors,t he st r
ongert he evidence thatt he est
imates are
statist
ical
lyreliabl e.
Thest udent’st -test: Wecomput et het-ratioforeach
,wher eni snumberofobser vationand ki snumberofpar ameters.Ifwehav e3
paramet ers,thedegr eeoff r
eedom wi l
l be( n-3)
.So;
;wit
h(n- 3)degr eeoff reedom
Inournul lhypot hesi s t het *becomes:

Ift^
* (comput ed)<t( t
abulat ed),
weacceptt henul lhypot hesis(H_0),wecanconcl udet hati snotsignificant.Thisimpl i
est he
regressordoesnotcont r
ibut etotheexplanati
onoft hevariati
onsinthedependentv ari
able(
Y) .
Ift^
* comput ed>t( tabulated) ,
wer ejectthenul lhy pothesi s(H_0)andweacceptt heal t
ernative(H_1);i
sst at i
sti
call
ysignif
icant.
Thisi mpl i
est her egressordoescont ri
butet otheexplanationoft hevariationsint hedependent
vari
abl e(Y).
Thus, t
hegr eaterthev alueoft *thest r
ongertheevidencet hatisstat
ist
icallysignifi
cant.
TestofOv erallSignifi
canceoft heModel
Int his sect i
on we ext end si gnif
icance testoft he i
ndividualestimat ed par t
ialregression
coeff i
cienttoj oi
ntt estoft herelevanceofal lt
heincludedexpl anator
yv ariabl
es.Now consi der
thefol l
owing:

atl eastoneoft he( _iisnon- zero

Thi snul lhypot hesi sisaj oi
nthy pothesi sthatar ej ointlyorsi multaneousl yequaltozer o. At est
ofsuchahy pot hesi si scal led at estofov erallsi gnificanceoft heobser ved orest imated
regr essi onl i
ne, thati s,whet herYi slinearl
yr elatedt o.
Thet estpr ocedur ef oranysetofhy pothesi scanbebasedonacompar i
sonoft hesum of
squar eder rorsf rom t heor iginal,theunr est r
ictedmul tipleregressi onmodelt ot hesum of
squar eder rorsf rom ar egressi onmodeli nwhi cht henul lhypot hesisi sassumedt obet r
ue.
Whenanul lhy pothesi si sassumedt obet rue, wei nef f
ectpl acecondi t
ionsorconst raints, onthe
valuest hatt hepar amet erscant ake,andt hesum ofsquar ederrorsi ncreases. Thei deaoft he
testi st hati fthesesum ofsquar eder rorsar esubst anti
al l
ydifferent,thent heassumpt ionthat
thej oi ntnul lhy pot hesisi struehassi gnificant lyreducedt heabi li
tyoft hemodelt of itthedat a,
andt hedat adonotsuppor tthenul lhypothesi s.
Ifthenul lhypot hesi sist rue, weexpectt hatt hedat aar ecompat iblewi tht heconditi
onspl acedon
thepar amet ers. Thus,t herewoul dbel i
ttlechangei nt hesum ofsquar ederrorswhent henul l
hy pot hesi sisassumedt obet rue.
Lett heRest ri
ctedResi dualSum ofSquar e( RRSS) bet hesum ofsquar ederror
si nt hemodel
obt ainedbyassumi ngt hatt henul lhy pothesi si str ueandURSSbet hesum oft hesquar eder r
or
oft heor i
ginalunr estrictedmodeli .
e.unr est r
ictedr esidualsum ofsquar e(URSS). I ti sal ways
truet hatRRSS-URSS0.
 Consi
der.
Thi
smodeliscal
l
edunr
est
ri
cted. Thet
estofj
ointhy
pot
hesi
sist
hat
:

atleastoneoftheisdiff
erentfr
om zero.
Thesum ofsquareder rorwhent henullhypot
hesi
sisassumedt obet r
uei scall
edRest
ri
cted
Resi
dualSum ofSquare(RRSS)andt hi
sisequaltot
hetotal
sum ofsquare(TSS).
Ther at
io: (hasanF- dit
ri
buti
onwi t
hk-1andn-kdegr
eesoff r
eedom forthenumerat
or
anddenominatorr
espectivel
y)

……………………………………………….(3.54)
I
fwedi
vi
det
heabovenumer
atoranddenomi
nat
orby t
hen:

…………………………………………. .(3.55)
Thisi mpl iest hecomput edv alueofFcanbecal culatedei therasar atioofESS&TSSorR2&1-
R2. I ft henul lhy pot hesisi snott rue,t hent hedifferencebet weenRRSSandURSS( TSS&RSS)
becomesl ar ge,i mpl yingt hatt heconst raintspl acedont hemodelbyt henul lhypot hesi
shav e
l
argeef fectont heabi lit
yoft hemodelt of itt hedat a,andt hev al
ueofFt endst obel arge. Thus,
wer ejectt henul lhy pot hesisi ftheFt estst at icbecomest ool arge. Thi sv alueiscompar edwith
thecr i
ticalv alueofFwhi chl eav est hepr obabi li
tyof i ntheuppert ai
loft heF-distri
but i
onwithk-
1andn- kdegr eeoff reedom.
I
ft hecomput edv alueofFi sgr eat ert hant hecr it
icalv al
ueofF( k-1,n- k),thenthepar ametersof
themodelar ej oint l
ysi gnificantort hedependentv ariabl
eYi sl i
near l
yr elatedt othei ndependent
variabl esi ncl udedi nt hemodel .
Appl icat ionofMul ti
pleLi nearRegr essi onModel s
Exampl e1: Consi dert hedat agi v enbel owf ital i
nearf unction:
Y
Σy i
2=619; Σx1x2=240; Σx2x3=- 420; Σx1x3=-
330;
Σx12=270
Σx22=630; Σx32=750; Σx1yi=319; Σx2y i
=492; Σx3yi
=-
625;
N=10; =52; =42 ; =62; =200
Smal llet tersar ei ndev i
ationf or m.Basedont hei nformat iongi venabov eandmodelanswert he
foll
owi ngquest ion.
Estimat et hepar amet erest i
mat or susi ngt hemat ri
xappr oach
Comput et hev ar i
anceoft hepar amet er s.
Comput et hecoef f
icientofdet ermi nat i
on( R2) andAdj ustedcoef fici
entofdet er
mi nati
on
()andi nter prett her esul t
Repor tther egressi onr esul t
.
Testsi gni ficanceofeachpar ti
al slopecoef ficientsat5%si gnif i
cancel evel
[
uset 0. 025( 6)=2. 447]
Testt heov erall signi f
icanceoft hemodel .[
useF0. 05( 3,6)
=4. 76]

Chapt erFour
Violati
onsofbasi cCl assical Assumpt ions
4.Introduct i
on
Thecl assical
istsseti mpor t antassumpt i
onsaboutt hedi str
ibut
ionof andt her andom er
ror
term‘ut ’intheregressi onmodel s.
Theyassumedt hatt heer rort erm( ui)f ollow normaldistri
butionwithmeanzer oandconstant
variat
ion, ,
andt hatthe
Errorscor respondingt odi f
fer entobser vationar euncorr
elated, (fort(s)and
I
nmul tipleregression,noper fectcorrelationbet weentheindependentv ari
abl
es.
Now, weaddr esst hef oll
owi ng‘ whatif’quest ionsinthi
schapt er.
Whati ft heerrorvar i
ancei snotconst antov erallobser
vations?
Whati fthedi fferenterrorsarecor r
el ated?
Whati ftheexpl anatoryvariabl
esar ecor related?
4.1Het eroscedast i
cit
y
4.1.
1Thenat ureofHet er
oscedast i
ct y
I
nl i
nearr egr essi onmodel
Thedi stri
but ionoft hedisturbancet ermr emai nssameov eral lobservat
ionsofX;i.
e.thevari
ance
ofeachi st hesamef orallthevaluesoft heexpl anat oryv ar i
able. Sy mbolical
ly
,
const antv alue.
Thisfeat ureofhomogenei tyofvariance( orconst antv ariance)i sknownashomoscedast ici
ty.
I
ft he di sturbance t erms do nothav et he same v ariance ( non-constantvar
iance)ornon-
homogenei ty ofv ar
iance is known as het er oscedast ici
ty. Thus,we say t hatU’s are
heteroscedast icwhen:
(
aconst ant)but
(av aluet hatv ari
es)
Theassumpt ionofhomoscedast i
cit ystatest hatt hev ar i
ationofeach ar ounditszeromean
doesnotdependont hev al
ueofexpl anator yvariable.
Mathemat i
cal l
y,i
snotaf unct ionofX; i
.e.
I
f dependsont hev al
ueofX( i
.e )t hedi spersionofhet roskedast i
cerrorter
m(vari
ance)wil
l
i
ncreasewi thX
greaterint hemi ddle rangeofX’ s, taperingof ftowar dt heext remes.
greaterforl owv al
uesofX.

I
tspat
ter
ndependont
hesi
gnsandv
aluesoft
hecoef
fi
cient
soft
her
elat
ionshi
pandt
aket
he
f
orms:

etc.
4..
1. 4Ex ampl esofHet eroscedast icf unct i
ons
Consumpt ionFunct ion:
Suppose: ;
wher e:Consumpt i
onexpendi tureofi thhousehol d; Disposabl ei ncomeofi thhousehold
Atlowl evelsofi ncome, theav erageconsumpt i
oni sl ow, thev ariati
oni slow.
Athi ghi ncomest he wi llbehi gh,whi leatl ow i ncomest hewi l
lbesmal l.Ther efore,t
he
assumpt ionofconst antv ari
anceofi sdoesnothol dwhenest i
mat ingt heconsumpt i
onf uncti
on
fr
om acr osssect ionoff ami lybudget s.
Product ionFunct i
on:
Supposet hepr oduct i
onf unct i
on.
Disturbance t er ms st and f ormanyf act ors;like ent repreneur shi p,t echnol
ogicaldi ff
erences,
sell
ingandpur chasi ngpr ocedur es,di ffer encesi nor ganizations,et c.otherthaninput s,l
abor( L)
and capi tal( K)consi der ed int he pr oduct i
on f unct i
on. These f actorsshow consi derable
variancei nl argef i
rmst hani nsmal lones. Thi sl eadst obr eakdownofourassumpt ionon
homogenei tyofv ariancet erms.
4.1.5Reasonsf orHet r
oscedast i
ci ty
Ther ear esev eralreasonswhyt hev ariancesofmaybev ar i
abl e. Someoft heseare:
Errorlear ningmodel :i tstatest hataspeopl el earnt heirer rorofbehav i
orbecomesmal lerover
ti
me.I nthiscase i sexpect edt odecr ease.
Exampl e: ast henumberofhour soft ypingpr act i
cei ncr eases,t
heav eragenumber
of
typi nger rorsandaswel l
ast hei rvariancedecr eases.
Asdat acol lectiont echniquei mpr ov es, i
sl ikelytodecr ease.
Het eroscedast i
ci t
ycanal soar iseasar esul tofthepr esenceofout liers.
Obser vationt hati smuchdi fferent( eitherv erysmal lorv erylar ge)int hesampl e.
Het eroscedast i
ci t
ycan ar i
sef rom v i
ol ating theassumpt ion t hatl inearregression modeli s
correctlyspeci fi
ed.
Het eroscedast i
ci t
ycanar iseduet oskewnessi noneormor er egressor sinthemodel
unev endi st ri
but ionofdat aoft her egr essor s
Exampl es: dist ributionofi ncomeandweal thinmostsoci etiesisunev en,bei ngownedbyaf ewat
thet op.
Het eroscedast icitycan al so ar i
se because of :incorrectdat at ransformat ion and incorr
ect
funct i
onal for m( e.
g., l
inearv ersusl og–l i
nearmodel s).
4.1.6ConsequencesofHet roscedast i
cityfortheLeastSquar eestimat ors
Het eroscedast icityhast hef ol l
owi ngconsequences:
TheOLSest imat orswi l
l hav enobi as( unbi ased)
Var i
anceofOLSest i
mat orswi l
l bei ncor rect
OLSest imat or sshal l bei nef ficient:
theOLSest imat orsdonothav et hesmal lestv ar
ianceint heclassofunbi asedest i
mat or
sbot hin
smal l andl argesampl es.
i
.e.,theOLSest imat ori sl i
near ,unbi asedandconsi stent,butitisineffi
cient..
Undert hehet er oscedast icassumpt ion:
Asar esul tt het rues. e( )shal l beunder estimated.
thet -testandF- testassoci at edwi thi twi l
lbeov eresti
mat edwhi chmi ghtl eadt otheconclusion
thatinaspeci ficcaseathand i sst atisti
cal l
ysignifi
cant(whichi nfactmaynotbet rue).
Ouri nf erenceandpr edi cti
onaboutt hepopul ati
oncoef fi
cientswoul dbei ncor r
ect.
Int hepr esenceofhet eroscedast i
city ,theBLUE est i
mat orsar epr ovidedbyt hemet hodof
wei ght edl eastsquar es( WLS) .
4.1.7Det ect ingHet eroscedast i
ci t
y
Met hodsoft est i
nghet eroscedast i
ci ty. Thesear e:
Informal met hod
For mal met hod
Informal met hod
Iti
sat estbasedont henat ureoft hegr aph.
Whent her eexi st sasy stemat icr elationbet weenr esidualsquaredandt hemeanv alueofYi .
e.or
with, het er oscedast icityi nt hedat a.
When t her e i s no sy stemat ic pat t
ern bet ween t he t wo v ariables, may suggest no
hetroscedast icityispr esenti nt hedat a.

Formalmet hods
Thef oll
owingareformalmet hodsoft est
ingheter
oscedastict
y.
Parktest
Suggeststhatt hevari
anceofu_ i(
)isaf uncti
onoftheexplanatoryvar
iabl
e.
Thefunctional f
orm hesuggestedwas:
Or 3. 14
wherev ii
st hestochasti
cdist
urbancet erm.
Using ? û_i^2asapr oxyfor
,andr unni
ngt hefoll
owingregressi
on.
l
nûi2
l
nûi23.15 si
nce isconst ant.
Testi
ngwhet herther
eisasignifi
cantrelat
ionbetweenûi 2.
Let v s

I
fi sstat
isti
call
ysi gnif
icant,itwouldsuggesthet r
oscedast
ici
tyinthedat
a.
I
fi sinsi
gnifi
cant,wemayacceptt heassumpt i
onofhomoscedast i
cit
y.
Thepar ktestisthusat wo- stagetestprocedure;
runOLSr egressiondi sr
egar dingtheheteroscedasti
ci
tytoobtainûi
2.and
thenrunther egressioninequat ion(3.
15)abov e.
Exampl e: Supposet hatf rom asampleofsi zen=100weest
imat
ether
elat
ion
between
compensat i
onandpr oducti
vi
ty.

Ther
esi
dual
obt
ainedf
rom (
3.16)wer
eregr
essedongi
vest
hef
oll
owi
ngr
esul
t.
Theabov eresultrev ealedthat :
thesl opecoef f
icienti sst atisti
callyinsigni fi
canti mplyingt hereisnost atist
ical
lysigni
fi
cant
rel
ati
onshi pbet weent hetwov ari
abl es.
wemayconcl udet hatt hereisnohet er
oscedast icit
yi ntheer rorvari
ance.
GoldfeldandQuandthav earguedt hepar kt esti nt hattheer rorterm ent eri
ngintotheequation
mayi tselfbehet eroscedast ic.
Glej
sert est:
Glej
sersuggest edr egr essi
ngt heabsol utev al
ueofont hev ari
ablethatisthoughttobeclosely
associatedwi th.
Inhisexper i
ment ,Glejseruset hef ollowingf unct ionalforms:
;
wher eiser rorterm.
GoldfeldandQuandtpoi ntoutpr oblemswi t
hGl ej sertestist hat:-
Expect edvalueofer rortermi snon- z er
o, i
.e.iti
sser iall
ycor r
el at
edandhet er
oscedastic.
Model ssuchas:and ar enon- l
i
neari npar amet ersandt hereforecannotbeest imatedwit
ht he
usualOLSpr ocedur e.
Glej
serhas f ound t hatf orl ar
ge sampl es t he f ir
stf ourpr eceding models gi
ve generall
y
sati
sfactoryresul tsindet ectinghet eroscedast icity.

Goldfi
eld-Quandttest
Thismet hodisappl icabl
eifthehet er
oscedast
icvar
iancei
sposi
ti
vel
yrel
atedt
ooneoft
he
explanatoryvari
ablesintheregressi
onmodel.
Forsimpl i
cit
y,considertheusualtwovari
abl
emodels:

Suppose ? (_
i^ 2isposi tivel yrelatedt o X_ ias:
3.
18; wher eisconst ant.
I
nt heabov eequat ionwoul dbel arger
,f orlargerv al
uesof .Ift hatturnsouttobet hecase,
heteroscedast icityismostl ikelytobepr esenti nt hemodel .
Tot estt hisexpl icit
ly,Gol df eldandQuandtsuggestt hef oll
owingst eps:
Step1: Or dert heobser v at i
onsaccor di
ngt othev al
uesof begi nningwiththelowestvalue.
Step2: Omi tCcent ralobser vationswher eCi sspeci fi
edapr iori,anddivi
detheremaining(n-c)
observ ationsi ntot wogr oupseachof obser vations
Step3: Fi tsepar ateOLSr egressiont ot hef i
rstobser vat
ionsandt helastobservati
ons,and
obtaint her espect iver esi dualsumsofsquar esRSS,RSS1andRSS2,RSS1r epresenti
ngRSS
from ther egressi oncor respondi ngt ot
hesmal l
erv alues( t
hesmal lvari
ancegroup)andRSS2
thatfrom t hel argerv alues( t
hel argev ar
iancegr oup).
TheseRSSeachhav e or, wher e: Ki st henumberofpar ameter
st obeest imated,
i
ncludingt heint erceptt erm; and dfi
sthedegr eeoff reedom.
Step4: comput e
foll
owsFdi stributionwi thnumer atoranddenomi natordfeach( n-c-2k)
/2.

Ifthecomput ed >Fcr it
ical atthechosenlevel ofsi
gnif
icance,
thenrejectthehy pothesi
sofhomoscedast i
cit
y,(heter
oscedast i
cit
yisv eryl
ikely
).
Exampl e: From adat aonconsumpt ionexpendi tur
ei nr
elati
ont oincomef oracr oss-sect
ionof
30f ami li
es. Supposet hatconsumpt ionexpendi t
ureisl i
near l
yr elat
edt oi ncomebutt hat
heteroscedasti
cityispr esentint hedata.Bydr oppingthemi ddl
ef ourobser vations,theOLS
regressionbasedont hef ir
st13andt helast13obser vati
onsandt hei
rassoci atedsumsof
squaresofr esidualsareshownnext( standarderrorsinparentheses).
Regr essi
onbasedont hefi
rst13observati
ons

Regr
essi
onbasedont
hel
ast13obser
vat
ions

Fr
om t
heser esult
sweobt ai
n:
F_
Cwithdfof( 11,11)atthe5%lev
el i
s2.82.
Si
nceF_cal
(=()>F_0.05(11,
11),
wemayconcl udet hatt
hereishet
eroscedast
ici
tyi
ntheer
rorv
ari
ance.
Spear
man’s rank cor r
elat
ion test
, Breusch-Pagan-
Godfrey test and Whi
te’
s Gener
al
Heter
oscedast
ici
tyt
estarealsoanot
hert
estmet hodsheter
oscedasti
cit
y.

4.1.8Remedi al measur esf orthepr obl emsofhet eroscedast i

cit
y
Knowi ng t he consequences ofhet eroscedast i
city,itmay be necessar yt o seek r emedi al
measur es.Thepr obl em her ei st hatwedonotknowt het ruehet er oscedast i
cv ar i
ances, ?(_i^
2, for
theyar er arelyobser v
ed.I fwecoul dobser v
et hem,t henwecoul dobt ainBLUEest imatorsby
dividingeachobser vationbyt he( het eroscedast ic)(_iandest imat ethet ransf ormedmodelby
OLS.Thi smet hod ofest i
mat i
on i sknown ast hemet hod ofWei ghted ( Gener ali
zed)l east
squar es( WLS) .
Unf ortunat el
y, t
het rue? (_i^2i srarelyknown.Thenwhati st hesol uti
on?
Inpr act i
ce,wemakeeducat edguessesaboutwhat? (_i^
2 mi ghtbeandt ransf ormt heor iginal
regressi onmodeli nsuchawayt hati nt het ransformedmodelt heer r
orv ariancemi ghtbe
homoscedast i
c.I nshor tGLSi sOLSont het ransf ormedv ari
ablest hatsat i
sfyt hest andar dleast
squar esassumpt ions. Theest i
mat or sthusobt ainedar eknownasGLSest imat or s,andi ti s
theseest i
mat orst hatar eBLUE.
Someoft het ransf or mat ionsusedi npr acticear easf ollows:
Ifthet r
ueer rorv ariancei spr opor ti
onal t
ot hesquar eofoneoft her egressors, wecandi vi
debot h
sidesofEquat ionbyt hatv ari
abl eandr unt het r
ansf ormedr egr ession.Wet hensubj ectt his
regressi ont ohet er oscedast icit
yt ests,suchast heBPandWhi t
et est s.Iftheset est si ndicatet hat
therei snoev i
denceofhet eroscedast icity,wemayt henassumet hatt het r
ansf or meder r
ort erm
i
shomoscedast ic.
Exampl e:Supposet hehet eroscedast icityisoft hef orm
,thet r
ansf ormi ngv ar i
ableisi f
wher e .
Thet r ansformedmodel i
s:
Ifthet r
ueer rorv ariancei spr opor ti
onalt ooneoft her egressor s, wecanuset heso- calledsquar e
transf ormat i
on,t hati s,wedi videbot hsi desofequat ionbyt hesquar er ootoft hechosen
regressor .Wet henest imat et her egr essiont hust r
ansf or medandsubj ectt hatr egr essiont o
heteroscedast i
cityt ests.I ftheset estsar esatisf actory,wemayr elyont hi
sr egressi on.
Exampl e:Supposet hehet eroscedast icityisoft hef orm :
Thet ransf ormingv ar i
abl eis
Thet r ansformedmodel i
s:

Thelogar i
thmict r
ansformat
ion:sometimes,i
nst
eadofest imatingregression,wecanr egressthe
l
ogar i
thm oft hedependentv ari
ableontheregressors,whichmaybel inearorinlogf orm.The
reasonf orthisisthatthelogt r
ansfor
mationcompr essesthescal esinwhi chthev ari
ablesare
measur ed,therebyreduci
ngat enfol
ddif
fer
encebet weentwov aluestoat wofolddi
fference.

4.
2Aut ocor
relati
on
4.2.
1Thenat ureofAutocor r
elat i
on
Theassumpt i
onsoft hecl assicali
stsay st hatt hewhi chi mpli
est hatsuccessi v
ev aluesof
dist
urbancet er
m U ar et empor ar
il
yi ndependent ,i
.e.di sturbanceoccur ri
ng atonepoi ntof
observationisnotr el
at edtoanyot herdisturbance. Thi smeanst hatwhenobser vati
onsar e
madeov erti
me, theeff
ectofdi stur
banceoccur r
ingatoneper ioddoesnotcar ryoverint oanother
peri
od.
Iftheabov eassumpt i
oni snotsat i
sfi
ed,t hatis,ifthev alueofU i nanypar t
icularper i
odi s
correl
ated withi tsown pr ecedi ng value(s),wesayt her ei saut ocorr
elat
ion oft her andom
vari
ables. Hence,aut ocorrelati
oni sdef i
nedasa‘ correlation’betweenmember sofser i
esof
observationsorderedint i
meorspace.

Thereisadi f
fer
encebetween‘cor
rel
ati
on’andautocor
rel
ati
on.Autocorr
elat
ioni
saspeci alcase
ofcorrel
ationwhi chref
erstotherel
ati
onshipbetweensuccessi
vev al
uesofthesamev ari
able,
whil
ecor relati
onmayal sorefertotherelat
ionshi
pbetweent woormor edif
ferentvari
ables.
Autocorr
elationisalsosometi
mescalledasser i
alcor
rel
ati
onbutsomeeconomi stsdi
sti
nguish
bet weent heset wot er ms. Accor di ngt oG. Tinner ,aut ocorrelationi sthel agcor relat i
onofa
giv enser ieswi thi tsel fbutl aggedbyanumberoft imeuni ts. Thet erm ser i
alcor relat ioni s
def inedbyhi m as“ lagcor relat i
onbet weent wodi ffer entser i
es.”Thus,cor relationsbet weent wo
timeser iessuchas and,wher et hef or meri st hel att
erser i
esl aggedbyonet imeper iod,i s
aut ocor relat ion.Wher eascor r
el ationbet weent i
meser i
essuchasand wher eUandVar et wo
differ entt imeser ies, iscal ledser ial cor relat ion.
Gr aphi cal repr esent at i
onofAut ocor relat i
on
Sinceaut ocor r elat i
oni scor relationbet weenmember sofser i
esofobser vationsor der edi nt i
me,
wewi llseegr aphi cal lyt het r
endoft her andom v ar iablebypl ottingt i
mehor izont allyandt he
random v ar iabl e v ertical l
y .
Consi dert hef ollowi ngf igur es
Thef i
gur es( a)–( d)abov e,showacy cl i
calpat t
er namongt heU’ sindicatingaut ocor relat i
oni .e.
figur es( b)and( c)suggestanupwar danddownwar dl i
neart rendand( d)i ndi catesquadr atic
trend i nt hedi st urbancet erms.Fi gur e( e)i ndicat esno sy stemat icpat tern suppor ting non-
aut ocor relat ionassumpt ionoft hecl assi cal linearr egr essionmodel .
Wecanal so show aut ocor relationgr aphi cal lybypl ottingsuccessi vev aluesoft her andom
dist urbancet er mv ertical ly( ui)andhor izont ally(uj).
The abov ef igur es fand g si mi l
ar lyi ndicat es us posi ti
v e and negat ive aut o-cor relati
on
respect ivel ywhi lehi ndi cat esnoaut ocor relat i
on.
In gener al,i ft he di stur bance t er ms f ollow sy st emat ic pat t
er n as i n( f)and ( g)t her ei s
aut ocor relat ionorser ialcor relationandi ft her ei sno sy stemat i
cpat tern,t hi si ndicat esno
cor relat ion.
42. 3Reasonsf orAut ocor relation
Ther ear esev er al reasonswhyser ial oraut ocor relationar ises. Someoft hesear e:
Cy clical fluct uat ions
Timeser i
esdat asuchasGNP,PI ,pr oduct i
on,empl oy mentandunempl oymentexhi bitbusi ness
cy cle.Wheneconomi cr ecov eryst ar ts, mostoft heseser i
esmov eupwar d,thev alueofaser i
esat
onepoi nti nt imei sgr eat ert hani tspr ev iousv alue.
Thus,t her ei sa‘ moment um’bui lti nt ot hem,andi tcont i
nuesunt ilsomet hinghappens( e.g.
i
ncr easei ni nter estr ateort axesorbot h)t osl owdownt hem. Ther efore,r egressi oni nv olvi
ng
timeser i
esdat a, successi v eobser vat ionsar el i
kelyt obei nterdependent .
Speci ficat ionbi as
Thi sar isesbecauseoft hef oll
owi ng.
Excl usi onofv ar iabl esf rom t her egr essi onmodel
Incor rectf unct i
onal form oft hemodel
Negl ect ingl aggedt ermsf rom t her egr essi onmodel
Let ’sseeonebyonehowt heabov especi ficat i
onbi asescausesaut ocorrelati
on.
Excl usi onofv ar iabl es: Theer rort er m wi llshowasy stemat icchangewheni mpor t
antv ariablei s
excl udedf rom t hemodel .
Forexampl e, supposet hecor r
ectdemandmodel isgi venby : 3.21
wher e quant ityofbeefdemanded,pr iceofbeef ,consumeri ncome,pr i
ceofpor kandt i
me.
Now, supposewer unt hef ol l
owi ngr egr essi on: -
---
--3.22
runni ngequat i
on3. 22whenequat ion3. 21i st he‘ cor rect’modelort ruer elation,l etting. The
er r
orordi st urbancet erm Vwi llref l
ectasy stemat icpat tern,thuscr eatingaut ocor r el
at i
on.
Incor rectf unct i
onal form oft hemodel :
Itisal soonesour ceoft heaut ocor rel ationofer r
ort er m.
Suppose t he cor rect model i n a cost -
out put st udy i s as f ol l
ows.
Mar ginal cost =3. 23
Howev er, wei ncor r
ect lyf itt hef ollowi ngmodel . ---
--
----
--
---
--
----
---
--
---
--
-3.24
TheMCcur vecor respondi ngt ot he‘ true’modeli sshowni nt hef i
gur ebel ow al ongwi ththe
‘
incor rect ’ li
nearcostcur v e.

Bet
weenpointsA andB thel i
nearmarginalcostcur
veconsi
stentl
yov eresti
matethetrue
margi
nal
cost;
whereas,
Out
sidet
hesepoi
ntsitwil
lconsi
stent
lyunderest
imat
ethet
ruemar gi
nalcost. Thisr
esul
tisto
beexpect edbecauset hedi sturbancet erm i s,inf act,equalt o( output )2+ui ,andhencewi l
l
catcht hesy stemat icef fectoft he( output )
2t erm ont hemar ginalcost . I nt hi
scase,Viwi l
l
refl
ectaut ocor relationbecauseoft heuseofani ncorr ectf unct ional for m.
Negl ect ingl aggedt ermf rom t hemodel :-
Ifthedependentv ar i
abl ei st obeaf fect edbyt helaggedv alueofi tselfort heexpl anat oryvari
able
andi snoti ncl udedi nt hemodel ,t heer rorterm oft hei ncor rectmodelwi llrefl
ectasy st
emat i
c
patter nwhi chi ndi cat esaut ocor r
el ationi nthemodel .
Supposet hecor rectmodel forconsumpt i
onexpendi turei s:
--
------
--
------
----
---
---
---
--
----
3.25
(thi si sknownasaut oregr essi
onbecauseoneoft heex planat oryv ariablesi sthel aggedv alue
oft hedependentv ariabl e.
(Ther ational ef orsuchamodeli sconsumer sdonotchanget hei rconsumpt ionhabi ts r eadily
forpsy chol ogi cal,technol ogi cal,orinst i
tuti
onal reasons.
(Ifwenegl ectt hel aggedt er mfr om t heEqu/ n,ther esult i
nger rort erm wi llrefl
ectasy st
emat i
c
patter nduet ot hei nf luenceofl aggedconsumpt i
ononcur rentconsumpt ion.
4.2.5Thecoef fi
cientofaut ocor r
elat i
on
Aut ocor r
elat ioni saki ndofl agcor r
el ati
onbet weensuccessi vev aluesofsamev ariabl es. Thus,
wet reataut ocor relat i
oni nt hesamewayascor rel
at i
oni ngener al.
Ifthev al ueofUi nanypar ti
cularper ioddependsoni tsownv aluei nt hepr ecedingper i
odalone,
wesayt hatU’ sf ollowaf i
rstor deraut oregr essiveschemeAR( 1)i .e.
.--
-----
------
----
---
----
--
--
-----
-----
--
----
---
--
---
--
---
------
--
----
--3.28
If u_ tdependsont hev aluesoft het wopr ev i
ousper iods, then:
u_ t
=f (u_ t
-1,u_ t -
2)t hi sf orm ofaut ocor rel
at i
oni scalledasecondor derARschemeandsoon.
Gener allywhenaut ocor relationispr esent ,weassumesi mpl efirstf or m ofaut ocor r
elat i
on:
u_ t=f(u_ t-1,u_ t-
2) ,
---
----
-----
-------
----
---
-----
------
--
-----
3.30
wher e t hecoef fi
ci entofaut ocor r elati
onandVi sar andom v ariablesat i
sfyingal lt hebasic
assumpt ionofOLS, Theabov er elationshipst atest hesi mplest
possi bl ef or m ofaut ocor relati
on; i fweappl yOLSont hemodelgi veni n(3. 30)weobt ain:

-
---
--
----
--
----
--
-----
-----
---
-3.31
Givent hatf orl argesampl es:, weobser vet hatcoef f
icientofaut ocor r
elationrepresent sasimpl e
correlationcoef ficientr . (Why ?)-
---
--
----
---
--
--
---
-3.32
si
nce - -
-----
----
----
----
---
---
--
---
---
----
-----
---3.33
Thispr ov est hest atement“ wecant reataut ocorrelati
oni nt hesamewayascor r
elat
ioni n
gener al
” . Fr om st atistics
i
fr=1wecal l i
tper fectposi t
ivecor relat ion,
i
fr=- 1, per fectnegat ivecor r
elationand
i
fr=0, ther ei snocor relati
on.
Byt hesameanal ogy :
i
f =1, per fectposi ti
v eaut ocor r
elat i
on,
i
f =- 1, per fectnegat iveaut ocor relationand
i
f( =0, , noaut ocor relation.
4.2.7Ef fectofAut ocor rel
ationonOLSEst imat ors.
Ift
heer rort er msar ecor r
elated,thef ollowi ngconsequencesf oll
ow:
TheOLSest imat or sar est il
lunbi asedandconsi st
ent.
Theyar est il
l nor mal l
ydi stri
but edi nlar gesampl es.
Butt heyar enol ongeref fi
cient.Thati s, theyar enol ongerBLUE( bestlinearunbi asedest i
mat or)
.
InmostcasesOLSst andar der r
or sar eunder esti
mat ed,whi chmeanst heestimat edt-val
uesar e
i
nflated, giv i
ngt heappear ancet hatacoef fi
cientismor esignificantt hani tact
uallymaybe.
Asar esul t,asi nt hecaseofhet eroscedast icit
y,t hehy pot hesi s-
testingpr ocedurebecomes
suspect , sincet heest imat edst andar der rorsmaynotber el
iable, evenasy mptoti
cal l
y(i.
e.inlarge
sampl es) .Inconsequence, theusual t-test andF-t est smaynotbev alid.
i.
e. ,I fisunder estimat ed, i salsounder esti
mat ed, thismakest -ratiolarge.
Thi sl arget -
rati
omaymake st ati
stical
lysi gni
ficantwhi l
ei tmaynot .Wrong
 predi ctionsandi nfer encesaboutt hechar acteristicsoft hepopul ation.
4.2.8Det ect ion( Test i
ng)ofAut ocor r
elat ion
Ther e ar et wo met hods t hatar e commonl yused t o det ectt he exi stence orabsence of
autocor rel
ationi nt hedi stur bancet erms. Thesear e:
Graphi cmet hod
Detect ionofaut ocor relationusi nggr aphswi llbebasedont woway s.
Appl yOLSt ot hegi v endat awhet heri ti saut ocor rel
atedornotandobt aintheer rort erms.Pl ot
horizont allyand v er ti
cal l
y .i.
e.pl otthef ollowi ngobser vations.
Ifi
ti sf oundt hatmostoft hepoi nt sfall inquadr antIandI I
I,asshowni nf ig(a)bel ow, wesayt hat
thegi vendat ai saut ocor relatedandt het y peofaut ocorrelati
oni sposi t
iveaut ocor relati
on.
Ifmostoft hepoi nt sf all inquadr antI IandI V, asshowni nf ig( b)bel owtheaut ocor relationissai d
tobenegat ive.Buti ft hepoi ntsar escat teredequal l
yi nal lt hequadr antsasshowni nf i
g( c)
below, thenwesayt her ei snoaut ocor relat i
oni nt hegi vendat a.
Formal t
est i
ngmet hod
Itisbasedonei thert hez- test,t-test,F- testorX^ 2-t
est.I fat estappl i
esanyoft heabov e,itis
call
ed f ormalt est ing met hod.The mostf r equent l
y and wi dely used t est i
ng met hods by
resear cher sar et hef ollowi ngs:
Runt est :
Runi st henumberofposi ti
veand negat ivesi gnsoft heer rort erm ar r
anged i nsequence
accor dingt ot hev aluesoft heexpl anat oryv ar i
ables,like“ ++++++++- -----
--
---
--++++++++- ---
--
---
--
-
++++++”
Iftherear et oomanyr uns,i twoul dmeant hechangesi gnf requent l
y,t husi ndicat i
ngnegat ive
serialcor r
elation. Si mi larly,ift
her earet oof ewr uns, t
heymaysuggestposi t
iveaut ocor relat
ion.
Now let : tot al number of obser vati
ons
numberof+sy mbol s; numberof–sy mbol s; andnumberofr uns
Undert henul lhy pot hesi st hatsuccessi v eout comes( here,r esidual
s)ar ei ndependent ,and
assumi ngt hat, t henumberofr unsi sdi stributed( asympt otical l
y)nor mallywi th:
Mean:
Var iance:

Deci
sionrule:
Donotrejectthenullhypothesi
sofrandomnessori
ndependencewi
th95%conf
idence.
I
f;
Rej
ectthenullhypothesis
I
ftheesti
mat edkliesout si
dethesel
imits.
I
nahy potheti
calexampl eof weobt ai
n:

Hencet
he95%conf
idencei
nter
val
is:

sincek=5,i tcl earlyfall

sout si
det hisint erv al. Ther eforewecanr ejectt hehy pot
hesist hatthe
obser v edsequenceofr esidualsi srandom ( areofi ndependent )wi th95%conf i
dence.
TheDur bin- Wat sondt est:
Themostcel ebr atedtestf ordet ectingser ialcor relat ioni sonet hatisdev elopedbyst at
isti
cians
Dur binandWast on. Itispopul ar l
yknownast heDur bin-Wast ondst atist
ic,whi chisdef i
nedas:
---
----
-----
-----
----
--
--
---
--
--
----
3.47
Not et hat ,i nt he numer atorofd st at i
st ict he numberofobser vati
ons i s because one
obser v ationi sl osti ntakingsuccessi vedi fferences.
Itisi mpor tantt onot etheassumpt i
onsunder l
yingt hed- stati
stics
Ther egr essi onmodel includesani nterceptt erm.Ifsucht ermi snotpr esent, asinthecaseoft he
regressi ont hr ought heor igin,i
tisessent ialt orerunt her egressioni ncludingt hei nt
ercepttermt o
obtai nt heRSS.
Theexpl anat or yv ari
ables, t
heX’ s,arenon- st ochast ic, orf i
xedinr epeat edsampl i
ng.
Thedi stur bancesar egener atedbyt hef i
rstor deraut oregr essivescheme:
Ther egr essi onmodeldoesnoti ncludel aggedv alueoft hedependentv ariableasoneoft he
expl anat oryv ariables. Thus, t
het esti sinappl i
cabl et omodel soft hef oll
owi ngt ype
Wheret heoneper iodl aggedv alueofyissuchmodel sareknownasaut oregressi
v emodel s.Ifd-
testi
sappl i
edmi stakenl y,thev al
ueofdi nsuchcaseswi l
loft
enbear ound2,whi chist hev alue
ofdint heabsenceoff ir
stor derautocorrel
ation.Durbindevel
opedt heso-call
edh- stat
isti
ctot est
seri
alcor r
elati
oni nsuchaut oregressi
ve.
Therear enomi ssingobser vationsinthedat a.
Inusingt heDur bin–Wat sont est,iti
s,therefore,t
onot ethatitcannotbeappl i
edi nviol
ationof
anyoft heabov ef i
veassumpt i
ons.
From equat i
on3. 47t hev alueof
Squaringt henumer atoroft heabov eequation,weobt ain
---
---
--
--
----
----
3.48
Howev er,forlar
gesampl esbecausei nbot hcasesoneobser v
ationislost.Thus,

Fr
om t
heabov
erel
ati
on,
ther
efor
e

Thusweobt aintwoi mpor tantconclusions

Valuesofdl iesbet ween0and4
Ift
her ei
snoaut ocorrelati
on
Whenev er,therefore,thecalculatedv al
ueofdt urnsoutt obesuffi
cientl
ycl oseto2,weaccept
nullhypothesis,andi fi ti
scl osetozer oorf our,wer ejectthenul
lhy pothesist
hatthereisno
autocorrel
ation.
Howev er,becauset heexactv alueofdi snev erknown,t her
eexistrangesofv al
ueswi t
hin
whichwecanei theracceptorr ej
ectnullhypothesis. Wedonotal sohav euniquecri
ti
calval
ue
ofd-stasti
cs.
Wehav e-lowerboundand upperboundoft heinit
ialval
uesofdt oacceptorr ej
ectthenull
hypothesis.
Forthet wo-tail
edDur binWat sont est,wehav esetf i
ver egi
onstothev aluesofdasdepi ct
edin
thefigurebelow.

Themechani smsoft heD- Wt estareasf oll

ows,assumi ngt hatt heassumpt i
onsunder ly
ingthe
testsaref ul fi
ll
ed.
Runt heOLSr egr essi onandobt ainther esiduals
Obt ai
nt hecomput edv alueofdusi ngthef ormulagiveninequat ion3. 47
Fort hegi vensampl esi zeandgi vennumberofexpl anat oryv ariables,fi
ndoutcr i
tical and
values.
Nowf oll
owt hedeci sionr ulesgivenbel ow.
Ifdi slesst hat orgr eat erthan wer ejectthenullhypot hesisofnoaut ocorrel
ati
oni nfavorof
thealternat ivewhi chi mpl iesexistenceofaut ocorrel
ati
on.
If,dli
esbet ween and, acceptt henull hypothesisofnoaut ocorrelati
on
Ifhowev ert hev alueofdl iesbet weenandorbet ween and, theD- Wt estisinconcl
usive.
Exampl e1. Supposef orahy potheti
cal model :,
i
fwef ound
0.1380; 1. 37; 1.50
Basedont heabov ev al
uest estforautocor rel
ati
on
Solution: Firstcomput e and andcompar ethecomput edv alue
ofdwi t
h,,and
=4- 1.37=2. 63
=4-1.5=2. 50
since d i s l ess t han we r ej
ect t he nul l hy pothesi s of no aut ocor rel
ati
on
Exampl e2. Consi dert hemodel wit
ht hef oll
owingobserv ationonXandY
X
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

Y
2
2
2
1
3
5
6
6
10
10
10
12
15
10
11

Testf orautocor
rel
ati
onusi
ngDur
bin-Watsonmethod
Soluti
on:
regressYonX: i
.e. :
From t heabovetabl
ewecancomputethefol
lowi
ngvalues.

Val
uesofandon5%l
evel
ofsi
gni
fi
cance,
wit
hn=15andoneexpl
anat
oryv
ari
abl
ear
e:=1.
08 and
=1.
36.

Sinced*l iesbet ween,acceptH0.Thi simpl iest hedat ai snotaut ocorrelated.

AlthoughD- Wt estisextremelypopul ar ,t
hed- t
esthasonegr eatdr awbacki nthatifi
tf all
sint he
i
nconcl usivezoneorr egion,onecannotconcl udewhet heraut ocor rel
at i
ondoesordoesnotex ist
.
Sev eralauthorshav eproposedmodi ficati
onsoft heD- Wt est.
Inmanysi t
uations,howev er,ithasbeenf oundt hattheupperl i
mi t i sappr oximatelythet rue
signifi
cancel i
mi t. Thus, t
hemodi fi
edDW t estisbasedon i ncaset heest i
mat eddv alueliesin
thei nconclusi
v ezone,onecanuset hef ollowi ngmodi fieddt estpr ocedur e.Gi v
ent helev elof
signifi
cance; i
f
iftheest i
mat ed, rej
ectH0atl evel,thati st hereisst atistical
l
ysi gnif
icantposi t
ivecorrelat
ion.
iftheest imatedorr ejectH0atl evelst ati
st icall
yt herei ssignifi
cantev idenceofaut ocor r
elati
on,
posi t
iveornegat ive.
4.
2.9Remedi alMeasur esf ort hepr oblemsofAut ocorrelat
ion
Sincei nthepresenceofser i
al cor
relationt heOLSest i
mat orsarei nef
ficient,i
tisessent i
altoseek
remedi almeasur es.Ther emedyhowev erdependsonwhatknowl edgeonehasaboutt henature
ofinterdependenceamongt hedi sturbances. Thi smeanst heremedydependsonwhet herthe
coeffi
cientofaut ocorrelationi sknownornotknown.
A. when i sknown- Whent hest ruct
ur eofaut ocorrel
ati
oni sknown,i .eisknown,t he
appropr iatecor rectivepr ocedur eist ot ransformt heoriginalmodelordat asot haterrorterm of
thetransf ormedmodelordat ai snonaut ocor related. Whenwet r
ansform,wear ewi ppyoft he
eff
ectof .
Supposet hatourmodel is
3.
49
and3. 50
Equation( 3.50)i ndi
catest heexi stenceofaut ocor rel
ati
on.If isknown, thentransform Equation
(3.
49)i nt oonet hatisnotaut ocor r
elated.Thepr ocedureoft r
ansformati
onwi l
lbegiv enbelow.
Taket hel aggedf or
m ofequat i
on( 1)andmul tiplythroughby. .
3.51
Subtract ing( 3)from (1), wehav e:
3.52
Byrear rangi ngt hetermsi n(3. 50),wehav e
whi chonsubst itutingt helastt er
m of( 3. 52)gives

Let
:

Equat
ion(
3.53)maybewr
it
tenas:

I
tmaybenot edt hati
ntransfor
mi ngEquati
on(
3.49)i
nto(3.
54)oneobservati
onshallbelost
becauseoflaggingandsubtracti
ngin(3.52)
.WecanapplyOLStothetransformedrelat
ionin
(3.
54)toobtai
nforourtwoparamet er
s.
anditcanbeshownt hat

Becauseisperf
ectl
yandl
i
nearl
yrel
atedto. Agai
nsincesati
sfi
esal
lst
andar
dsassumpt
ions,
thevari
anceofwoul
dbegi
venbyourstandar
dOLSf
ormulae.

Estimat orsobt ainedi nequat ion6ar eef fi

cient ,onlyifoursampl esizeislargesothatl ossofone
obser v ationbecomesnegl igi
ble.
B.When i snotknown
When i s notknown,we wi lldescr ibe t he met hods through whi ch the coef f
icientof
autocor relationcanbeest i
mat ed.
Met hodI : Apr iori inf
or mat i
onon
Manyt imesani nv estigatormakessomer easonabl eguessaboutt hev alueofautor egressi
v e
coef ficientbyusi nghi sknowl edgeori nstitutionaboutt herelati
onshipunderst udy . Many
resear cher susual lyassumet hat=1or- 1.
Undert hismet hod, thepr ocessoft ransf ormat i
onist hesameaswhen i sknown.
When=1, t hetransf ormedmodel becomes;
; wher e
Not et hatt heconst antt ermi ssuppr essedi nt heabov e. i sobtainedbyt akingmer elythef i
rst
differencesoft hev ariableandobt ai ningl inet hatpassest hr
ought heor i
gin. Supposet hatone
assumes=- 1inst eadof=1,i .ethecaseofper f
ectnegat iveautocorrel
ati
on. I nsuchacase,t he
transf ormedmodel becomes:
Or
Thismodeli sthencal ledt woper i
odmov ingav er
ager egressi
onmodelbecauseact uallywear e
regr essingt hev al ueofonemov ingav erageonanot her
Thismet hodoff ir
stdi f
ferencei nqui t
epopul arinappl iedresearchf oritssimpli
city.Butt he
met hod r ests on t he assumpt ion t hatei thert her eis perfectposi t
ive orperfectnegat i
v e
autocor relationi nt hedat a.
Met hodI I: Est i
mat i
onof f rom d- statistic:
From equat
ion(3.47),weobt
ained. Supposewecal
cul
atecer
tai
nval
ueofd-
stat
ist
ici
nthe
caseofcert
aindata.
Giventhed-
valuewecanest
imat
e f r
om t
his.

Asal readypoi ntedout , wi llnotbeaccur ateifthesampl esi zei ssmal l. Theabov erelati
onshi p
i
st rueonl yf orlargesampl es.Forsmal lsampl es,TheilandNagarhav esuggest edt hef ol
lowi ng
relation:
………………………………………………. .3.55
wher en=t ot alnumberofobser vation; d=Dur binWat sonst atist
ic; k=number of coef f
icients
(i
ncl udi ngi nterceptt erm) .Usingt hi sv alueof wecanper formt heabov et r
ansfor mat i
ont oav oid
aut ocor relationf rom t hemodel .
Met hodI I
I: TheCochr ane-Or cutti terativeprocedur e:Inthismet hod, wer emov eaut ocor r
elat i
on
gradual lyst arti
ngf r
om t hesi mpl estf orm ofaf irstorderscheme.Fi rstweobt aint her esiduals
and appl y OLS to t
hem;
……………………………………………………. 3.56 Weest i
mat e
from t heabov er elation.Wi tht heest imat ed,wet ransformt heor i
ginaldat aandt henappl yOLS
to t
he model .
……………. ..
..
.…3.57
we once agai n appl y OLS to the newly obt ained
residual s
……………………………………………………………3. 58 Weuset hi
s
secondest imat e t ot r
ansf or mt heor iginalobser vati
onsandsoonwekeeppr oceedi ngunt ilthe
valueoft heest imat eof conv er ges. I tcanbeshownt hatt hepr ocedur eisconv ergent.When
thedat aist ransf ormedonl ybyusi ngt hissecondst ageest imat eof ,itisthencal l
edt wost ages
Cochr ane- Or cuttmet hod. Howev eronecanf oll
owanal ternativeappr oacht ouseateachst ep
ofi nter acti
on, t
heDur binWat sond- statisti
ct oresidualsforaut ocor r
elationortil
lt heest i
mat esof
donotdi ffersubst antiallyfrom oneanot her.
Met hodI V: Dur bin’
st wo- stagemet hod:Assumi ngthef i
rstor deraut oregressi
vescheme, Dur bi
n
suggest sat wo-stagepr ocedur ef orr esol vi
ngt heser i
alcorrelati
onpr obl em.Thest epsundert hi
s
met hodar e:
Giv en- ---
--
-----
---
--
----
--
-----
----
---
-(3.59)
Taket helaggedt er m oft heabov eandmul t
iplyby
--
-----
-----
--
---
-----
--
--(
3.60)
Subt ract(3. 60)from ( 3. 59)
--
----(3.61)
Rewr ite(3.61)i nt hef ollowi ngf orm

Thisequat i
oni snow treatedasr egressi
onequat ionwi ththreeexplanat
oryv ari
abl
es Thi
s
providesesti
mat eof whi chisusedt oconstructnewv ari
ables I nthesecondst ep,
esti
mat or
s
ofareobt ai
nedfrom theregressionequat i
on:
; wher e
4.3 Mul ti
coll
inear
it
y
4.3.
1Thenat ureofMul ti
coll
inearit
y
Multicoll
i
neari
tymeantt heexistenceofa‘ perfect’
or‘exact’
,l
inearrel
ati
onshipamongsomeoral l
explanatoryv ar
iables of a r egression model .For r egr
ession modeli nvolvi
ng k-
variabl
e
explanatoryvari
ables,anexactl inearr el
ati
onshipissai dtoexi sti
ft hefoll
owingcondit
ionis
sati
sf i
ed.

wherear econst antssucht hatnotal lofthem aresimul

taneouslyzer
o.
Notet hat:-multicol
li
nearityref ersonlytolinearrel
ati
onshi
psamongt heexpl
anat
oryv
ari
ables.
I
tdoesnotr ul
eoutnon- li
nearr elati
onshi
psamongt hex-v
ari
ables.
Forexampl e:
Where: Y- Totalcostand X- output.
Thev ariablesar eobv i
ouslyf unct i
onal
lyrelatedtoxibutther el
ati
onshi
pisnon-l
i
near
.Str
ictl
y,
therefore,model ssuch( 3.31)donotv iolat etheassumpt ionofnomul ticolli
neai t
y. Howev er,in
concr eteappl icat i
ons,t heconv ent i
onal lymeasur edcor r
elationcoef fi
cientwi llshowt obehi ghly
cor r
elated,whi chasweshal lshow,wi l
lmakei tdiff
icul
ttoest i
mat ethepar amet erswi t
hgr eater
precision( i.e.wi thsmal l
erst andar der ror s).
4.3.2 Reasonsf orMul ticollinear ity
Thedat acol lectionmet hodempl oy ed: Exampl e:Ifwer egr essonsmal lsampl evaluesoft he
popul ation;t heremaybemul t
icol li
near i
tybuti fwet akeal lthepossi blev alues,itmaynotshow
mul ti
col l
inear it
y.
Const raintov ert hemodel ori nt hepopul at ionbei ngsampl ed.
Forexampl e: i nt her egr essionofel ect ricityconsumpt iononi ncome( x1)andhousesi ze( x2),
therei saphy sicalconst raintint hepopul ationint hatfami li
eswi thhigheri ncomegener al l
yhav e
l
ar gerhomest hanwi thloweri ncomes.
Ov erdeter mi nedmodel :Thi shappenswhent hemodelhasmor eexplanat oryv ari
ablest hant he
numberofobser vati
ons.Thi scoul dhappeni nmedi calr esearchwher et her emaybeasmal l
numberofpat i
ent saboutwhom i nfor mat ioni scol l
ectedonal argenumberofv ari
ables.
4.3.3ConsequencesofMul t
icolli
near i
ty
Whydoest hecl assicallinearr egr essionmodelputt heassumpt ionofnomul ticoll
i
nearityamong
theX’ s? I ti sbecauseoft hef ollowi ngconsequencesofmul ti
colli
neari
tyonOLSest i
mat or s.
Ifmul ticollinear i
t yisper fect,t her egr essi oncoef fi
cientsar ei ndeterminat eandt heirst andar d
errorsar ei nfinit
e.
Proof :-Consi deramul t
ipler egr essionmodel asf oll
ows:

Recal
lthef
ormul
asof
and f
rom ourdi
scussi
onofmul
ti
pler
egr
essi
ons

Assume-
--
--
--
--
--
--
--
--
--
--
--
-3.32
Wherei
snon-zer
oconst ant
s. Subst
it
ute3.
32i
ntheabov
eandf
ormul
a:

(i
ndeter
minate.
Apply
ingthesamepr ocedure,weobt
ainsimilarresul
t(i
ndet
ermi
nateval
ue)f
or.
Li
kewise,f
rom ourdiscussi
onofmulti
pleregressionmodel
,var
ianceof i
sgiv
enby:

Subst
it
uti
ngi
ntheabov
evar
iancef
ormul
a,weget
:

(i
nfi
nit
e.
Thesearetheconsequencesofper
fectmul
ti
col
l
inear
it
y.Onemayr
aiset
hequest
ion
on
consequences of l ess t han per fect correl
ation.In cases of near or hi gh
mult
icolli
near i
ty,
onei sli
kelytoencount erthef ol
lowingconsequences.
Ifmul t
icoll
inearit
yi sl ess than per fect( i
.e nearorhi gh mul t
icol
li
near
ity)
,the regr
ession
coef
fici
entsar edeterminate
Proof : Considerthetwoexpl anator
yv ariabl
esmodel aboveindevi
ati
onform.
I
fweassume i ti
ndi catesusper fectcorrel
ati
onbet ween because
t
he change i n x2 is compl etel
y because of t he change in x1.Inst
ead of exact
mult
icolli
near i
ty,
wemayhav e: Wher eisst ochasticerrorterm suchthat.In
t
hiscasex2i snotonlydeter minedbyx1, but al soaffectedbysomeot hervar
iabl
esgiven
by
v_i(
st ochasticerrort
erm).
Substi
tuteintheformul aof abov e

(det
ermi
nate.
Thi
spr
ovest
hati
fwehavelessthanper
fectmul
ti
col
l
inear
it
ytheOLScoef
fi
cient
sar
e
 deter
minate.
Theimplicati
onofi
ndeter
minati
onofregressi
oncoeffici
ent
sinthecaseofperf
ect
multi
coli
neari
tyi
sthati
tisnotpossi
bletoobser
v ethesepar
ateinf
luenceof. But
suchextremecaseisnotveryfr
equentinpract
icalappli
cat
ions. Mostdataexhi
bitl
ess
t
han
perf
ectmul ticol
li
nearit
y.
I
fmul ti
coll
ineait
yi slessthanperfect(i
.e.nearorhighmulti
col
l
inear
it
y) ,OLSesti
mator
sretai
n
t
hepr opertyofBLUE
Hence, i
fthebasicassumpt ionswhichareimportanttoprovet
heBLUEpr oper
tyar
enotvi
olat
ed
whethermul ti
coll
inearit
yexistornot,t
heOLSest i
mat or
sareBLUE.
Al
thoughBLUE, theOLSest i
matorshavelargevari
ancesandcovari
ances.

Mul
ti
plyt
henumer
atorandt
hedenomi
nat
orby

Where i st
hesquareofcorr
elat
ioncoeff
ici
entbet
ween,
I
f,whathappentothevar
ianceofas islineri
ses.
As t endsto1orascol li
near
ityi
ncreases,t
hev ar
ianceoft
heest
imat
ori
ncr
ease
andi
n
t
hel imitwhen v ar
ianceofbecomesi nfi
nit
e.
Si
mil
arl
y. (why?)
Asi ncreasestowar done,t
hecov ari
anceofthetwoestimatorsi
ncr
easeinabsol
ute
v
al ue.Thespeedwi thwhi
chv ar
iancesandcov ar
iancei
ncreasecanbeseenwitht
he
variance-i
nfl
ati
ngf act
or(VI
F)whi chisdefi
nedas:

VI
Fshowshowt hev
ari
anceofanesti
matorisi
nfl
atedbythepr
esenceof
mult
icol
l
inear
it
y. As approaches1,theVIFapproachesi
nfi
nit
y. Thati
s,ast
he
ext
ent
ofcol
l
inear
it
yincr
ease,t
hev
ari
anceofanest
imat
ori
ncr
easesandi
nthel
i
mitt
he
v
ari
ance
becomesi nfinite. Ascanbeseen, ift herei snomul t i
colli
near it
ybet ween,
VIFwi llbe1.
Usingt hisdef initi
onwecanexpr essandi nt ermsofVI F
VIFandVI F
Whi chshowst hatv ari
ancesofar edi r
ect l
ypr opor tional tot heVI F.
Because oft he large v ar i
ance oft he est imat ors,whi ch means l arge st andar d errors,t he
conf idencei ntervaltendt obemuchwi der,leadi ngt otheaccept anceof“ zeronul lhy pothesi s”(i
.e.
thet ruepopul at i
oncoef f i
cienti szer o)mor er eadi ly
.
Becauseofl argest andar der r
oroft heest i
mat or s,thecomput edt -rati
owi l
lbev erysmal lleading
oneormor eoft hecoef ficientst endt obest atisticallyinsigni f
icantwhent estedi ndiv i
dually.
Althought het -r
atioofoneormor eoft hecoef ficient sisv er ysmal l(whi chmakest hecoef fici
ents
stat i
stical l
yinsignificanti ndi vi
dual ly
) ,
R2, t
heov er allmeasur eofgoodnessoff i
t, canbev eryhigh.
Exampl e:if
Int hecasesofhi ghcol li
near i
t y
,iti spossi blet of indt hatoneormor eoft hepar tialsl ope
coef ficientsar eindi vi
dual l
yst at i
sti
callyinsi gnificantont hebasi soft -test.Butt heR2i nsuch
situat i
onsmaybesohi gh,i nsuchacaseont hebasi sofFt estonecanconv i
nci nglyrejectt he
hypot hesi st hat I ndeed, thisisoneoft hesi gnal sofmul ticolli
near ity-insignificantt -
valuesbuta
highov erallR^2( i
.easi gni fi
cantF- value).
TheOLSest i
mat orsandt heirstandar der r
orscanbesensi tivet osmal lchangei nt hedat a.
4. 3.4Det ectionofMul ti
collinearit
y
Ar ecognizabl esetofsy mpt omsf ortheexi st enceofmul ticolli
near ityare:
Highcoef fici
entofdet ermi nation(R2)
Highcor relati
oncoef ficient samongt heexpl anat oryv ariables
Lar gest andarder ror sandsmal lert-
ratiooft her egr essionpar amet ers
Not e:Noneoft hesy mpt omsbyi tsel
far easat i
sfact oryi ndicatorofmul ticol l
ineari
t y
.
 Because:
Largest andar der r
orsmayar isef orv ari
ousr easonsandnotonl ybecauseoft hepresenceso
l
inearr elationshipsamongex planat or yvari
ables.
Ahi ghi sonl ysuf fi
cientbutnotanecessar ycondi tion(adequat econditi
on)fort heexistenceof
mul ti
collinearit
ybecausemul ti
col l
inear it
ycanal soexi steveni fthecorr
elat
ioncoef f
ici
entislow.
Howev er,t he combi nation of al lt hese cr it
eria shoul d help the det ection of
mul ti
collinearit
y.
4.3.4.
1TestBasedonAuxi l
i
ar yRegr essions:
Sincemul ti
coll
ineari
tyarisesbecauseoneormor eoft her egr essor
sar eexactorappr oximatel
y
l
inearcombi nati
onsoft heot herr egr essors,onemayoff i
ndi ngoutwhi chXv ariabl
eisr el
atedto
otherXv ariablestor egresseachXiont heremai ningXv ari
abl esandcomput ethecorresponding
R2,whi chwedesi gnateas,eachoneoft heser egr
essi onsi scalledauxiliarytot hemai n
regressionofYont heX’ s. Then,f oll
owingt her elati
onshi pbet weenFandR2est abl i
shedin
chapt erthreeunderov eral lsi
gni fi
cance, t
hev ariable,

wher e: - nisnumberofobser vation

- ki snumberofpar amet ersi ncl
udingt heintercept
I
ft hecomput edFexceedst hecr iti
calFatt hechosenl evelofsi gnifi
cance,itist akentomean
thatt hepar t
icularXicol l
inearwi thot herX’s;ifitdoesnotexceedt hecrit
icalF,wesayt hatitis
notcol l
i
nearwi thotherX’ sinwhi chcasewemayr etainthev ar iableinthemodel .
I
fFii sst at
isticall
ysignifi
cant , wewi llhavet odecidewhet hert hepar t
icul
arXishoul dbedropped
from t hemodel .
Not et hataccor di
ngt oKl i
er i
’sr uleoft humb,whi chsuggestt hatmul ti
colli
nearitymaybea
troublesomepr oblem onl yi fR2obt ainedfrom anauxi l
iaryr egr essionisgr eaterthant heoveral
l
R2, thatisobt ainedfrom ther egressionofYonal lr
egressor s.
4.3.4.2TheFar rar-
Glaubert est-Theyuset hreest ati
sti
csf ort est i
ngmut l
i
col l
i
near i
tytherear
echi
-squar e,F-
rati
oandt -r
atio. Thi stestmaybeout li
nedi nthreest eps.
A.Comput ationof ( 2 t ot estor t
hogonal i
tly: twov ari
abl esar ecall
edor thogonali f i .
e.if
therei snoanycol inearit
ybet weent hem.Inourdi scussionofi nmul ti
pleregressionmodel s,we
hav eseent hemat ri
xr epresent ati
onofat hreeexpl anatoryvar iablemodel whichi sgivenby

Di
vi
deeachel
ement
sofx’
xby andcomput
ethedet
ermi
nant

Thev alueoft hedet ermi nantisequaltozer ointhecaseofper f

ectmulti
colli
neari
ty.( since
On theot herhand,i tt hecaseofor t
hogonalit
yoft hex’ s,theref
oreand t hev alueoft he
determinanti sunity.I tf ol
lows,t
herefore,thati
fthevalueoft hisdeterminantli
esbet weenzer o
and uni ty,t here exi st some degr ee of mul ti
col
li
near i
ty.For det ecti
ng the degr ee of
mul t
icoll
i
near i
tyov ert hewhol esetofexpl anatoryvari
ables,GlauberandFar r
arsuggest s( 2to
testinthef oll
owingway .
areor t
hogonal (i
.e.
arenotor t
hogonal (i
.e.
GlauberandFar rarhav ef oundthatthequant it
y
(2=- [n–1–1/ 6(2k+5) ].loge{valueoft hestandardi
zeddet erminant}
;hasa( 2dist
ribut i
onwith
1/2k(k-1)df .Ifthecomput ed( 2isgreat erthanthecr it
icalvalueof( 2,rej
ectH0i nf avourof
mul t
icoll
i
near i
ty.Butifi tisless,t
henacceptH0.

B.Comput ati
onoft
-r
atiototestthepatternofmul t
icoll
i
near
it
y
Thet-testhel
pstodetectthosev ar
iableswhichar ethecauseofmul ti
col
li
near
it
y. Thistesti
s
perf
ormed based on thepar t
ialcorrelat
ion coeffi
ci
entsthrough t
hef ol
lowi
ng pr
ocedureof
hypothesi
s.

I
nthet
hreev
ariabl
emodel
(How?)
 (
How? )
i
ft*>t(t
abulat
e) ,
H0i srej
ect ed
t*
<t(tabulat
ed),H0i saccepted,weacceptXiandXjar enotthecauseofmul t
icol
l
inear
it
ysi
nce
(i
snotsignif
icant)
4.
3.4.3Testofmul t
icoll
ineari
tyusingEi
genval
uesandconditi
onindex:
UsingEigenvalueswecandr i
v eanumbercall
edcondi
ti
onnumberKasf ol
lows:
I
nadditionusingthesev aluewecandri
vethecondit
ioni
ndex(CI)def
inedas

Decisi
onrul
e: i fKisbetween100and1000t her
eismoder atet ostr
ongmul t
icoll
inear
it
yandifi
t
exceeds1000t her
eissev ermul ti
coll
i
nearity
. Al t
ernati
velyifCI()i
sbetween10and30, t
her
eis
moderatetostrongmul t
icoll
ineari
tyandifitexceeds30t hereissevermulti
coll
ineari
ty.
Example.Ifk=123,
864andCI =352–Thi ssuggestexi st
enceofmul t
icol
li
neari
ty
4.3.
4.4 Testofmul ti
coll
inearit
yusingToleranceandv ari
ancei nfl
ati
onfactor

wher eist heintheauxi l

iaryr
egressionofXjont heremai ni
ng(k-2)regressorsandVI Fisvar
iance
i
nf lati
onfactor.
Someaut horstherefor
euset heVI Fasani ndi
catorofmul ti
coll
ineari
ty: Thel argeri
sthev al
ue
ofVI Fj,t
hemor e“tr
oublesome”orcol l
inearisthev ari
ableXj. Howev er
, howhighshouldVIFbe
bef orear egressorbecomest roublesome?Asar uleoft humb,i fVIFofav ari
ableexceeds10
(thiswillhappensi f exceeds(0.9)t hev ar
iabl
eissaidtobehi ghlycoll
inear.
Ot herauthorsuset hemeasur eoft olerancetodetectmul ti
col
li
near i
ty.I tisdefi
nedas
Clear l
y,TOLj=1i fXjisnotcorrelat
edwi t
ht heotherregressor
s,wher easi tiszeroifi
tisper
fectl
y
relatedtoot herregressors.

VIF( ort olerance)asameasur eofcol l

inear i
ty,isnotf reeofcr i
ti
cism.Aswehav eseenear l
ier;
dependsont hr eef actor s Ahi ghVI Fcanbecount erbal ancedbyl ow Toputdi fferently,a
highVI Fi snei thernecessar ynorsuf f
ici
entt ogethi ghv ari
ancesandhi ghst andarder r
or s.
Ther ef
or e,hi ghmul ticollinear ity,asmeasur edbyahi ghVI Fmaynotnecessar ycausehi gh
standar der ror s.
4.3.5.Remedi al measur es
The f ollowi ngcor rect i
v epr ocedur eshav ebeensuggest edi ftheprobl em ofmul ti
colli
near i
tyi s
foundt obeser ious.
Increaset hesi zeoft hesampl e:
Mul ticollinear itymaybeav oidedorr educedi fthesi zeoft hesampl eisi ncr eased.
Ifthev ariabl esar ecol lineari nt hepopul ation, thepr ocedureofi ncreasi ngt hesizeoft hesampl e
will nothel pt or educemul ticollinearit
y .
Introduceaddi tional equat ioni nt hemodel :
Theaddi t i
onofnew equat i
ont ransformsoursi ngleequat i
on( ori
ginal )modelt osimul taneous
equat i
onmodel .
Ther educedf or m met hod( whi chi susual l
yappl i
edf orestimat i
ngsimul t aneousequat i
onmodel s)
cant henbeappl iedtoav oi dmul t
icoll
ineari
ty .
Useext raneousi nformat ion: –
Ext raneousi nf ormat ioni st hei nformat i
onobt ai
nedf r
om anyot hersour ceoutsidethesampl e
whi chi sbei ngusedf ort heest imat ion.
Ext raneousi nfor mat ionmaybeav ail
ablef rom economi ctheoryorf rom someempi ri
calst udies
alreadyconduct ed.
Thr eemet hods, t
hroughwhi chext raneousi nf ormat i
oni sutil
izedinordert odeal withthepr oblem
ofmul ticol l
inear i
ty.
Met hodsoft ransf ormi ngv ariables:Thi smet hodi susedwhent herel at i
onshipbetweencer tain
par amet ersi sknownasapr iori
PARTTWO
ECONOMETRI
CSTWO

CHAPTERONE
1.Regr essi onAnal ysiswi thQual itativ eInformat i
on: Binar y(orDummyVar iables)
1.1.Descr i
bingQual i
tat i
veI nf ormation
Qual i
tat ive f act ors of ten come i n t he f orm of bi nar y i nformat ion: aper son i s
femal e ormal e; aper sondoesordoesnotownaper sonal comput er ; af irm of fer s a
cert ai
n ki ndofempl oyee pensi onpl an ori t doesnot ;ast ateadmi ni sters capi tal
puni shment ori tdoesnot .Inal l oft heseexampl es,t her elevant i nfor mat i
on canbe
capt ur ed bydef i
ning a bi nar y v ariable orazer o-one v ariable. I neconomet ri
cs,
binar y v ariables ar emostcommonl y cal led dummy v ar iables, al though t hisname i s
notespeci ally descr i
ptive.
Indef ini ng adummy v ariable, wemust deci dewhi ch ev ent i sassi gned t hev alue one
andwhi ch i sassi gned t hev alue zer o.Forexampl e, i nast udyofi ndi vi
dual wage
det ermi nat i
on, wemi ght def inef emal e t obeabi nary v ariable t aki ng ont he v al
ue
onef or f emal es andt he v alue zer ofor mal es.
Thenamei nt hiscasei ndi cat est heev entwi tht hev alueone.Thesamei nfor mat ion i scapt ured
bydef ini ng mal e t obeonei ft he per son i smal e andzer o i ft he per soni s f emal e.
Either oft hese i sbet ter t han usi ng gender because t his name doesnot make i t
clear when t he dummy v ariabl e i sone: doesgender =1cor respondt omal e or
femal e? Whatwecal l ourv ariabl es i suni mpor tant f orget ti
ng r egr essi on r esult s, but
i
t al way s hel ps t o choose names t hat cl arify equat ions andexposi tions.Var i
ables
thatassumesuch0and1v aluesar ecal leddummyv ari
ables. Such v ariabl es ar et hus
essent ial l
y adev i
ce t ocl assi fy dat ainto mut ual l
y excl usi ve cat egor ies such asmal e
orf emal e.DummyVar i
abl esusual lyi ndicatest hedi chot omi zed“ presence”or“ absence” ,“y es”or
“no” , et c.Var iables i ndi cat es a“ quali
ty ” oranat tri
bute, such as“ mal e” or“ femal e”,
“black” or“ whi te”,“ ur ban” ornon-ur ban” , “ bef ore” or“ after”,“ Nor th” or“ sout h”,
“East ” or“ West ” , mar it
al st at us, j
obcat egor y,r egion, season, etc. Wequant if
y such
variabl es byar ti
fi
ciall
y assi gni ng v alues t ot hem ( for exampl e, assi gni ng 0and1t o
sex,wher e0i ndi cates mal e and1i ndicat es f emal e), anduset hem i nt her egr ession
equat i
on t oget her wi th t heot heri ndependent v ariables. Suchv ariablesar ecal leddummy
variabl es. Al ternativenamesar ei ndi catorv ari
ables,bi naryv ar iables,cat egor i
calv ariabl es,and
dichot omousv ariables.
Dummyv ariables can be i ncor por ated inr egressi on model sj ustas easi lyas quant i
tati
ve
variabl es.Asamat ter off act ,ar egr ession model may cont ai
n r egr essor s t hatar eall
excl usiv el y dummy , orqual itativ e, i nnat ure.Suchmodel sar ecal l
edAnal y sisofVar iance
(ANOVA)model s. Regr essi on model s i n mosteconomi c r esear ch i nvol ve quant i
tati
ve
explanat ory v ar iabl es i naddi tion t odummy v ariabl es. Suchmodel sar eknownas
Anal ysis ofCov ar iance ( ANCOVA)model s.
1.2.DummyasI ndependentVar iabl es
1.2.1. ASi ngl eDummy I ndependent Var iabl e
Howdowei ncor por atebi nar yi nf ormat ion i nt or egr essi on model s? I nt hesi mpl est case,
with onl yasi ngl e dummy expl anat ory v ariabl e, we j ust addi t asani ndependent
variabl e i n t heequat ion.
Forexampl e, consi der t hef ol l
owi ng si mpl e model ofhour ly wagedet ermi nat i
on:
wage=( _ 0+( _0D+( _ 1educ+u_ i,
wher ewage=wager at eofcer taini ndiv i
dual ,educ=l ev el ofeducat i
on
D=1, iff emal e0, ot her wise
Inmodel ( 1. 1),onl y t woobser vedf act ors af fect wage: gender andeducat ion. Si nce
D=1whent he per son i s f emal e, andD=0when t he per son i s mal e, t he par amet er
(0has t he f ollowi ngi nt er pretat ion:( 0<0 i s t he di f f
er ence i n hour ly wage bet ween
femal es and mal es, gi v en t he sameamount ofeducat i
on ( andt hesameer r
or t erm
u).Thus,t hecoef ficient( 0 det er mines whet her t her ei sdi scrimi nat i
on agai nst women:
i
f( 0<0 , t hen, f ort hesame l evel ofot her f act or s, women ear nlesst han men on
aver age.
Inter msofexpect ations, i f weassume t hezer ocondi ti
onal mean assumpt i
onEu=0,
then
(_0=Ewage/ D=1, educ- E( wage/ D=0, educ)
Thekeyher ei st hatt hel ev elofeducat i
oni st hesamei nbot hexpect ations;t hedi fference, (_0i s
duet ogenderonl y .Thesi tuat i
oncanbedepi ctedgr aphi cal lyasani nter ceptshi ftbet weenmal es
andf emal es. I nFi gur e1. 1, thecase ( _0<0i s shown, sot hat men ear naf i
xed amount
mor e perhour t han women. The di ffer ence doesnotdepend ont heamount of
educat i
on, andt his expl ai ns why t hewage-educat ionpr of i
l
es f orwomen andmenar e
paral l
el.
Att his poi nt ,y oumaywonder whywedonotal soi ncl ude i n( 1.1)adummy v ariable, say
D,whi chi s onef or mal es and zer o f or f emal es. Ther easoni s t hat t hi s woul d
ber edundant .I n ( 1.1) ,thei nt ercept f ormal es i s( 0, andt hei nt er
cept f orfemal es i s
(_0+( _0. Si ncet her ear ej ustt wogr oups,weonl yneedt wodi f
ferenti nt
er cepts. Thi smeans
that,i naddi ti
ont o( _ 0,weneedt ouseonl y onedummy v ari
able; wehav e chosen t o
i
nclude t hedummy v ar i
abl e f orf emal es. Usi ng t wodummyv ar iables woul d i nt r
oduce
perfectcol l
inear it
y becausef emal e+ mal e=1,whi ch meanst hatmal ei saper fect l inear
funct ion off emal e. I ncl uding dummy v ar i
abl es f orbot hgender s i st he si mpl est
exampl e oft he so-cal led dummy v ar iabl e t rap, whi ch ar ises when t oomany
dummy v ar i
abl esdescr i
beagi v en number ofgr oups.
Fig.1. 1
Model( 1.1)cont ains onequant itati
v e v ariabl e ( lev el feducat i
on) andonequal i
tative
variabl e ( sex)t hat has t wo cl asses ( or cat egor i
es) , namel y, mal e and f emal e.
What i s t he meani ng of t hi sequat i
on? Assumi ng, asusual ,t hatEu_ i
=0, weseet hat :
Meansal aryoff emal ecol legepr of essor :
EY/ D=1, educ=( _ 0+( _ 0+( _1educ……………………. 1. 2
MeanSal aryofmal ecol l
egepr ofessor :
E( Y/D=0, educ) =( _0+( _ 1educ………………………………. .1.3
Iftheassumpt ion ofcommon sl opesi sv alid, at estoft he hy pothesi s t hatt het wo
regressi ons ( 1. 2)and ( 1.3) hav e t he same i nter cept ( i.e., t her e i s no sex
discr i
mi nat ion) can be made easi l
y byr unni ng t her egressi on ( 1.2)andnot i
ng t he
statist i
cal si gni ficance oft heest i
mat ed ( _0ont hebasi sof t he t raditi
onal t t est.I f
the t t est shows t hat i t i s st at isti
cal ly si gni f
icant, we r ej
ect t he nul l
hypot hesi s t hatt hemal e andf emal e col lege pr of essor s’ l
evels ofmean annual sal ar y
aret hesame.
Befor e pr oceedi ng f ur ther , not e t he f ol l
owi ng f eat ures of t hedummy v ari
abl e
regressi on model consi der ed pr eviousl y.
Todi stingui sht het wocat egor ies, mal eandf emal e, wehav ei nt r
oducedonl yonedummyv ari
abl e
D. For exampl e, i fD_ i=1al way s denot es amal e, when D_ i=0weknowt hati tis a
femal e si nce t her e ar e onl y t wo possi ble out comes. Hence, one dummy
variabl esuf f i
ces t odi st ingui sh t wocat egor ies. Thegener al r ule i st his: I faqual itative
variabl e has‘ m’ cat egor ies, i nt roduce onl y ‘ m -1’ dummy v ar iables. I n our
exampl e, sex has t wocat egor ies, and hence we i ntroduced onl y a si ngl e
dummy v ar iabl e. I ft his r ule i s notf ol lowed, we shal lf al li nto what mi ght be
called t he dummy v ar i
abl e t rap, t hat i s, t hesi tuation ofper fect mul t
icol l
inear ity.
Theassi gnment of1and0v al ues t ot wocat egor i
es, suchasmal e andf emal e, i s
arbitrar yi nt hesenset hati nourexampl e wecoul d hav eassi gned D=1f orf emal e andD=0
formal e.
Thegr oup, cat egor y , or cl assi f i
cat ion t hat i sassi gned t hev alue of0i sof ten
refer r
ed t oast hebase, benchmar k, cont rol , compar i
son, r ef erence, or omi tted cat egor y
.Iti st hebasei n t hesenset hatcompar isons ar emadewi th t hatcat egor y .
Thecoef f i
ci ent at tachedt ot hedummy v ar iable Dcanbecal led t hedi fferent i
al i nt ercept
coef ficient becausei tt el ls byhowmuch t hev alue oft hei nter cept t er m oft hecat egor y
thatr ecei v es t hev alue of1di ffer s f rom t hei ntercept coef ficient oft hebasecat egor y .
Exampl e1. 1( Ef fect sofComput erOwner shiponCol legeGPA)
Inor dert odet ermi net heef fect sofcomput erowner shi poncol l
egegr adepoi ntav er age,we
estimat et hemodel : col GPA=( _0+( _0D+( _
1hsGPA+( _2ACT+u,wher e t he
dummyv ar iabl e D equal s one i fa st udentowns a per sonal comput erand zer o
other wi se.Ther ear ev ar i
ousr easonsPCowner shipmi ghthav eanef f ectoncol GPA.Ast udent ’
s
wor kmi ghtbeofhi gherqual ityi fi tisdoneonacomput er,andt imecanbesav edbynothav ing
towai tatacomput erl ab. Ofcour se,ast udent mi ghtbemor ei ncl i
nedt opl aycomput er
gamesorsur ft heI nt er neti fheorsheownsaPC,soi ti snotobv i
oust hat( _0i sposi ti
v e.The
variabl eshsGPA( highschoolGPA)andACT( achi evementt estscor e)ar eusedascont rols:i t
coul dbet hatst rongerst udent s, asmeasur edbyhi ghschool GPAandACTscor es, aremor el ikely
toown comput ers.Wecont rolf ort hesef act or sbecausewewoul dl iket oknow t heav erage
effectoncol GPAi fast udenti spi ckedatr andom andgi v enaper sonal comput er.
Iftheest imat edmodel i
s
colGPA=1. 26+0. 157D+0. 447hsGPA+0. 0087ACT, R^ 2=0. 219
SE .33 (.057) (.049) (.0105)
Thi sequat ioni mpl iest hatast udentwhoownsaPChasapr edictedGPAabout. 16poi nthi gher
thanacompar abl est udentwi thoutaPC( remember , bot hcol GPAandhsGPAar eonaf ourpoi nt
scal e).Theef fecti sal sov er yst atist i
cal lysi gni ficant ,wi tht_PC=0. 157/ 0.057≈2. 75.
1.2.2Usi ngDummyVar iabl esf orMul tipl eCat egor ies
Supposet hat ,ont hebasi soft hecr oss-sect ional dat a,wewantt or egr esst heannual
expendi tur eon heal th car e by an i ndi vidual on t he i ncome and educat ion of
the i ndi vidual . Si nce t hev ariabl e educat ion i squal it
ativ e i nnat ur e, supposewe
consi dert hr eemut ual l
y excl usi ve l ev elsofeducat ion: l ess t han hi gh school , hi gh
school , andcol lege. Now,unl i
ket hepr ev iouscase,wehav emor et hant wocat egor i
esoft he
qual itativ ev ar iabl eeducat ion. Ther ef ore, f ol l
owi ngt her ule t hatt henumber ofdummi es
beonel esst han t henumber ofcat egor ies oft hev ariable,weshoul d i nt roduce t wo
dummi es t ot akecar eoft het hr eel ev els ofeducat ion. Assumi ng t hatt het hr ee
educat ional gr oups hav eacommon sl opebutdi f
fer ent i ntercept s i n t her egr essi on of
annual expendi t
ur e onheal th car eonannual i ncome, wecanuset hef ollowi ng model :
Y_ i
=( _1+( _2D_ 2i+( _3D_ 3i +( X_ i+u_ i
…………………………………………………1. 4
Wher e Yi =annual expendi tur e onheal th car e
Xi =annual expendi tur e
D_ 2=1, ifhi ghschool educat ion0, ot her wi se
D_ 3=1, ifcol legeeducat i
on0, ot her wi se
Not et hati n t hepr ecedi ng assi gnment oft hedummy v ariabl es wear ear bi traril
y
treating t he“ lesst hanhi gh schooleducat ion” cat egor y ast hebasecat egor y. Ther efor e,
thei nt er cept( _ 0wi llref lectt he i nt er cept f or t his cat egor y.The di ffer ential i nter cept s
(_1and( _2t ell by how much t hei nter cept s oft heot hert wocat egor ies di ffer f rom
thei nt er cept oft hebasecat egor y, whi ch canber eadi l
y checkedasf ollows: Assumi ng
E(ui)=0, weobt ain f rom ( 1.4)
EY_ i/D_ 2=0, D_ 3=0, X_ i
=( _ 1+( X_ i
EY_i/
D_2=1,
D_ 3=0,
X_i
=((
_ 1+(
_2)+(
X_i
EY_i/
D_2=0,
D_ 3=1,
X_i
=((
_ 1+(
_3)+(
X_i
which are,respecti
vel
y t hemeanhealth careexpendi
tur
e f unct
ions forthethr
eelevel
s
ofeducati
on,namely,l essthan high school, hi
gh school, andcoll
ege. Geometr
icall
y,
thesit
uat
ion i sshown i nfi
g 1.2(fori
l
lust
rat
ive pur
posesiti
sassumed t hat
((_
2>(
_1)
.

Figur e1. 2.Expendi t

ur eonheal thcar ei nr el ationt oi ncomef ort hreel ev elsofeducat ion
1.2. 3.Regr essi ononOneQuant itativeVar i
abl eandTwoQual itat i
v eVar iabl es
Thet echni queofdummyv ariabl ecanbeeasi lyext endedt ohandl emor et hanonequal itati
ve
var iabl e. Letusr ev ertt ot hecol legepr of essor s’sal ar yr egr essi on( 1. 1) ,butnowassumet hatin
addi t
ion t oy ear s of t eachi ng exper ience andsex, t he ski n col or of t he t eacher
i
s al so ani mpor tant det er minant ofsal ary . Forsi mpl icity,assumet hatcol orhast wo
cat egor i
es: bl ackandwhi te. Wecannowwr i
te( 1. 1)as:
Y_ i=(_ 1+( _2D_ 2i+( _3D_ 3i +( X_ i+u_ i…………………………………………………1. 5
Wher e Yi =annual sal ar y
Xi =y ear soft eachi ngexper i
ence
D_ 2=1, i
ff emal e0, other wi se
D_ 3=1, i
fwhi te0, other wi se
Not i
cet hateachoft het woqual it
at i
v ev ar i
abl es,sexandcol or ,hast wocat egor i
esandhence
needsonedummyv ar iabl ef oreach.Not eal sot hatt heomi tted,orbase,cat egorynowi s“ black
femal epr of essor .”
Assumi ngE( u) =0, wecanobt ain t hef ollowi ng r egr essi on f r
om ( 1. 5)
Meansal aryf orbl ackf emal epr of essor :EY_ i
/ D_ 2=0, D_ 3=0, X_ i=( _1+( X_ i
Meansal aryf orbl ackmal epr ofessor :EY_ i
/D_ 2=1, D_ 3=0, X_ i=((_ 1+( _2)+( X_ i
Meansal aryf orwhi tef emal epr of essor: EY_ i
/ D_ 2=0, D_ 3=1, X_ i
=( (
_ 1+( _3) +( X_ i
Meansal aryf orwhi temal epr ofessor :EY_ i
/ D_ 2=1, D_ 3=1, X_i=( (
_ 1+( _2+( _ 3)+( X_ i
Once agai n, i t i s assumed t hat t he pr ecedi ng r egr essi ons di ffer onl y i n t he
i
nt er ceptcoef ficientbutnoti nt hesl opecoef ficient( .An OLS est i
mat ionof( 1. 5)wi l
lenabl e
ust ot esta v ar i
et y ofhy pot heses. Thus, i f ( _3i sst at i
st i
cal lysi gni ficant ,i
twi l
lmeant hat
col ordoesaf fectapr of essor ’ssal ary . Si mi larly,i f( _2i sst atistical lysi gni fi
cant ,i twillmean
thatsexal soaf fect sapr of essor ’ssal ar y. I fbot ht hesedi ffer ent i
ali nt ercept sar est at
isticall
y
signi ficant ,i twoul d mean sexaswel lascol ori sani mpor tantdet er mi nantofpr ofessor s’
sal aries.
Fr om t hepr ecedi ng di scussi on i tf ollows t hatwecanext end ourmodel t oi nclude mor e
than onequant i
tat ive v ariabl e andmor e t han t woqual i
tativ e v ar i
ables. The onl y
pr ecaut i
on t obet aken i st hat t henumber ofdummi es f or eachqual itative v ar i
able
shoul d beonel ess t han t henumber ofcat egor ies oft hatv ariabl e.
Exampl e 1. 2
(LogHour ly Wage)
Let us est imat e amodel t hat al lows f or wage di ffer ences among f our gr oups:
mar ried men,mar ried women, si ngl e men, and si ngl e women. Todot hi
s,wemust
sel ectabasegr oup; wechoosesi ngl e men. Then, wemust def ine dummy v ariables
foreachoft he r emai ni ng gr oups. Cal l t
heseD1( mar r i
edmal e) ,D2( mar r
iedf emal e),andD3
(singl ef emal e). Put tingt heset hreev ariabl esi nt o( 1. 1)andi ft heest i
mat edequat iongi vest he
followi ngr esul t,
l
ogwage=0. 321+. 213D_ 1-.198D_ 2- .
110D_ 2+. 079educ+. 027exp
s.e .100 .055 .058 .
056 .007
(.005)
Toi nt erpr ett hecoef ficient sont hedummyv ariables,wemustr emembert hatt hebasegr oupi s
singl e mal es. Thus, t he est i
mat es on t he t hree dummy v ariables measur e t he
pr opor ti
onat edi f
f erence i n wage r el ative t osi ngl e mal es. Forexampl e, mar r
ied men
ar e est imat ed t oear nabout21. 3%mor et han si ngl e men, hol ding l ev els ofeducat i
on
andexper ience f ixed. Amar riedwoman, ont heot herhand,ear nsapr edicted 19. 8%l ess
thanasi ngl e man wi th t hesamel ev elsoft heot herv ariabl es.
1.2. 4.I nter act i
onsAmongDummyVar iabl es
Consi der t hef ollowi ng model :
Y_ i
=( _1+( _2D_ 2i+( _3D_ 3i+( X_ i
+u_ i……………………………………………………. .
1. 6
Wher eYi =annual expendi ture D2=1i ff emal e0, ot her wi se
Xi= annuali ncome D_ 3=1, i fcol legegr aduat e0,
other wi se
Impl iciti nt hi s modeli st heassumpt i
on t hatt hedi ffer ent ial ef f
ectoft hesexdummyD2i s
const antacr osst he t wol ev els of educat ion andt he di ffer ent ial ef fect of t he
educat i
on dummyD3i s al soconst ant acr oss t he t wo sexes. That i s, i f , say ,t he
mean expendi tur e oncl ot hi ng i s hi gher f orf emal es t han mal es t his i ssowhet her
they ar ecol lege gr aduat es ornot .Li kewi se, i f, say ,col legegr aduat es ont he av erage
spendmor e oncl ot hi ng t han non-col l
ege gr aduat es, t hi s i ssowhet hert hey ar e
femal e ormal es.
Inmanyappl icat i
onssuchanassumpt ionmaybeunt enabl e. Af emal ecol l
egegr aduat emay
spendmor eoncl ot hingt hanamal ecol legegr aduat e. I not herwor ds,t heremaybei nteract ion
betweent he t wo qual i
tativ e v ariabl esD2andD3. Ther ef ore t hei r ef fect on mean Y
may not besi mpl yaddi tiv e asi n( 1. 6)butmul ti
pl i
cat i
v e aswel l, asi nt hef ollowi ng
model :
Y_ i
=( _1+( _2D_ 2i+( _3D_ 3i+( _4D_ 2i D_3i +( X_ i+u_ i…………………………………………. .
1. 7
From ( 1. 7)weobt ain:
EY_ i/D_ 2=1, D_ 3=1, X_ i
=( _1+( _ 2+( _ 3+( _4+( X_ i…………………. .1.8
whi ch i st hemeancl ot hingexpendi tur eofgr aduat e f emal es.Not icet hat
(_2=di ffer ent i
al eff ectof bei ngaf emal e, ( _3=di ff
erent ial eff ectofbei ngacol l
egegr aduat e
(_4=di f ferent ial ef f ectof bei ng a f emal e gr aduat e whi ch shows t hat t he mean
clothi ng expendi tur e of gr aduat e f emal es i s di ffer ent ( by( _4f rom t he mean
clothi ngexpendi tur eoff emal esorcol l
egegr aduat es. I f( _ 2,(_3and( _4andar eal lposi t
ive,t he
aver agecl ot hi ngexpendi ture off emal esi shi gher ( than t hebasecat egor y ,whi ch her ei s
mal enogr aduat e) , buti ti smuchmor esoi ft hef emal esal sohappen t obegr aduat es.
Simi lar l
y ,t heav er age expendi tur e on cl ot hing by a col legegr aduat e t ends t obe
higher t han t hebasecat egor ybutmuch mor esoi ft hegr aduat e happens t obeaf emal e.
Thisshowshowt hei nter act ion dummy modi fi
es t he ef fect of t he t wo at tr
ibut es
consi der ed i ndi vidual ly.
Whet hert hecoef fici entoft hei nteract iondummyi sst at i
st ical lysi gni ficantcanbet est edbyt he
usualt -test .Ifi tt ur nsoutt obesi gni fi
cant ,t hesi mul t aneouspr esenceoft het woat tri
but eswi l
l
attenuat eorr ei nf or cet hei ndi vidualef fect soft heseat tribut es.Needl esst osay ,omi tti
nga
signif i
canti nt er act iont er mi ncor rect lywi l
l l
eadt oaspeci fi
cat i
onbi as.
1.2.5.Test i
ngf orSt ruct ur alSt abi lityofRegr essi onModel s
Unt il now,i nt hemodel s consi der edi nt hi s chapt erweassumed t hatt hequal i
tative
variabl es af fectt he i nter cept but not t he sl ope coef ficient of t he v ar i
ous
subgr oup r egr essi ons. Butwhat i f t hesl opesar eal sodi fferent ? I fthesl opesar ei nf act
different , t est ing f ordi ffer ences i n t he int er cept s may be ofl ittle pr actical
signif i
cance. Ther ef ore, weneedt odev elop agener al met hodol ogy t of ind out
whet her t wo ( or mor e) r egr essi ons ar e di ffer ent , wher e t he di ffer ence may be
i
n t hei nter cept s ort hesl opesorbot h.
Supposewear ei nt erest ed i n est i
mat ing asi mpl e sav ing f unct ion t hat r el
at es
domest i
chousehol d sav ings ( S)wi th gr oss domest ic pr oduct ( Y)f orEt hiopi a.
Supposef ur thert hat ,atacer tain poi nt oft i
me, aser i
es ofeconomi c r ef orms hav e
beeni nt roduced. The hy pot hesi s her ei st hatsuchr efor ms mi ght hav econsi der abl y
i
nf l
uenced t hesav ings-i ncome r elat ionshi p,that i s,t her elat i
onshi p bet ween sav i
ngs
andi ncome mi ght bedi ffer ent i n t he postr ef orm per iodascompar ed t ot hat i n
thepr e-r ef or m per iod. I ft hishy pot hesi si st rue,t henwesayast ruct ur alchange has
happened. Howdowechecki ft hisi sso?
Wr itet hesav ingsf unct ionas:
S_ t
=( _0+( _1D_ t+( _2X_ t+( _ 3Y_ tD_ t+u_ i…………………………. ……………. .1.8
wher eSi shousehol dsav ingatt imet ,Yi sGDPatt i
metand:
D_ t
=0, i
fpr e-ref or m( <19911, ifpost -
r efor m( ≥1992
Her e(_3i sthedi f
ferenti
al sl ope coef f
ici
ent indi
cating how much t heslopecoef fi
cient
oft he pr er
eform per iodsav i
ngs f uncti
on di f
fers f r
om the sl opecoef ficient oft he
sav i
ngs f uncti
on i n t hepostr eform per i
od. I f(_1 and (_
3 ar ebot h statisti
cal l
y
signif
icant asj udged byt het -t
est,then t he pr e-
refor
m and post -reform r egressions
diff
er i nbot hthei nterceptandt hesl ope.Howev er,ifonly(_1is st atisti
cal
ly si gni fi
cant ,
then t hepr e-reform andpost -reform r egr
essions di ffer onl y i nthe i nt er
cept
(meani ng t hemar gi
nal pr opensit
y t osave(MPS)i sthe same f orpre-refor
m and post -
reform per iods).Simi l
arl
y,ifonly(_3i sstati
sti
cal
lysignif
icant
,thenthet wor egressionsdi ffer
onlyint heslope( MPS) .

Exampl
e: S=- 20. 76005+5. 99916D_ t+2. 616285Yt -0. 5298177( YtDt)
SE ( 6.
04) (6.4) (.57) (.6035149)
Sincebot h ( _ 1and( _3ar ebot hst atisti
cal ly i nsi gnifi
cant ,t herei s nodi ffer ence bet ween
pre-r efor m andaf terr efor m r egr essi on f ort hesav ing model .
1.3. DummyasDependentVar iabl e
Int hel ast sev er al sect ions, west udi ed how,t hr ough t heuseofbi nar y i ndependent
variabl es, wecani ncor porate qual it
at ive i nfor mat i
on asexpl anator y v ar i
abl es i na
mul tiple r egr essi on model . Inal l of t he model s up unt i
l now, t he dependent
variabl e Y has had quant i
tat i
v e meani ng ( forex ampl e, Y i s adol lar amount ,a
testscor e,aper cent , ort he l ogs oft hese) . Whathappensi fwewantt ousemul tiple
regressi ont oexpl ainaqual itativ
eev ent ?
Int hesi mpl est case,andonet hat of ten ar i
ses i n pr act i
ce, t heev ent we woul d
l
ike t oexpl ai n i sabi nar y out come. I not herwor ds, ourdependentv ariabl e, Y, takesononl y
twov alues: zer oandone.Forexampl e, Ycanbedef ined t oi ndicat e whet her anadul th
asahi gh schooleducat ion; orYcan i ndi cat e whet her acol lege st udent used i ll
egal
drugs dur ing agi ven school y ear; orY cani ndi cat e whet her af i
rm was t aken
over by anot her f ir
m dur i
ng agi ven y ear .I neachoft heseexampl es,wecanl etY=1
denot eoneoft heout comesandY=0t heot herout come.
Ther ear esev er almet hodst oanal yzer egressi onmodel swher et hedependentv ariabl ei sbinar y.
Thesi mpl estpr ocedur eist oj ustuset heusual OLSmet hod. I nthiscaset hemodel i
scal l
edt he
l
inearpr obabi litymodel ( LPM) .Theot heral ternativei st osayt hatt her ei sanunder l
y i
ngor
l
at entv ariabl eY^ *whi chwedonotobser v e.Whatweobser v ei s:
Y=1, i
fY^ *>00, Other wi se. Thi sist hei deabehi ndt hel ogi tandpr obi tmodel s.
1.3.1.TheLi nearPr obabi l
ityModel s( LPM)
When weuse al inear r egressi on model t oest imat e pr obabi li
ties, wecal l t he
model t he l inearpr obabi lity model . Consi dert hef ol l
owi ngmodelwi thwhenYi sabi nar y
variabl e.
Y=( _0+( _1X_ 1+( _2X_ 2+…( _kX_ K+u…………………………………………………………. 1.9
SinceYcant akeononl y t wov alues,( _jcannotbei nt erpr eted ast hechange i nYgi ven a
one-uni ti ncr ease i nX_ j
, hol ding al l ot herf actor s f ixed: Y ei ther changes f r
om
zerot ooneorf rom onet ozer o. Nev er thel ess, t he( _ jst i
ll hav e usef ul i nt erpr etat
ions.
Ifweassume t hatt hezer ocondi tional mean assumpt i
on hol ds, thatis,Eu=0, then wehav e,
asal way s,
EY=( _0+( _1X_ 1+( _ 2X_ 2+…( _
kX_ K=X(
The keypoi nt i st hat when Y i sabi nar y v ar i
abl e t aki ng ont hev alues zer oandone,
i
tis al way st ruet hatP( Y=1/ X)=E( Y/ X):t hepr obabi li
ty of“ success”—t hat i s, thepr obabi li
ty
thatY=1-i st he sameast heexpect edv alue ofY.Thus, wehav et hei mpor tantequat ionist he
sameast heexpect edv alue ofY.Thus, wehav et heimpor tantequat i
on
P(Y=1/ X) =( _0+( _ 1X_ 1+( _2X_ 2+…( _kX_ K=X( ………………………………. 1.10
whi ch say st hat t hepr obabi li
ty ofsuccess, sayP( X) =P( Y=1/ X)=E( Y/X) ,isal inear f unction
oft heXj .
Equat i
on ( 1.10) i s anexampl e of a bi nar y r esponsemodel , andP( Y=1/ X)i s al so
called t her esponse pr obabi li
ty. Becausepr obabi l
itiesmustsum t oone, PY=0/ X=1- P( Y=1/ X),
(whi ch i scal l
ed t henon-r esponse pr obabi lity
)i sal soal inear f unct i
on oft heXj .
The mul tiple l i
near r egressi onmodel wi th a bi nar y dependent v ariabl e i s cal l
ed
the l inearpr obabi li
ty model ( LPM)because t her esponsepr obabi li
ty i sl i
near i n t he
par amet er s ofX_ j
.I nt heLPM,( _j
measur es t hechange i n t hepr obabi lity ofsuccess
whenX_ jchanges, hol di ng ot herf act or sf ixed. Thi s i st heusual l inear r egr essi on
model . Thi s makesl inear pr obabi lit
y model s easyt oest imat e andi nter pret, buti tal so
highl i
ght s someshor tcomi ngs oft heLPM.
Thedr awbacksoft hi smodel ar e:
Ther ighthandsi deofequat i
on( 1.9)i sacombi nat ionofdi scret eandcont inuousv ariabl eswhi l
e
thel eft hand si dev ar i
abl e i sdi scret e.
Usual l
ywear bitrar il
y( orf orconv enience)use0and1f orY.I fweuseot herv al uesf orY,say3
and4, βwi ll al sochange ev eni ft hev ect oroff act ors X r emai ns unchanged.
Uassumes onl y t wov al ues:i fY=1t henu=1- X( , i
fY=0t heu=-X( Consequent ly,u
i
s not nor mal ly di st ribut ed but r ather hasadi scr ete ( binar y) pr obabi li
tydi st ri
but ion
def i
ned.
Iti s easy t o see t hat ,i f we pl ug i n cer t
ain combi nations of v alues f or t he
i
ndependentv ar iabl es i nt o ( 1.9) ,wecangetpr edict i
ons ei ther l ess t hanzer oorgr eater
thanone.Si ncet hesear epr edi ctedpr obabi lities, andpr obabi lit
iesmustbebet weenzer oandone,
thiscanbeal ittleembar rassi ng.
Duet opr obl em 3, thev ar iance ui s het er oscedast ic.
1.3.2.TheLogi tandPr obi tModel s
The l inear pr obabi li
ty model i s si mpl e t oest imat e and use, but i t has some
drawbacks t hat we di scussed i n Sect i
on 1. 3.1.The t wo most i mpor tant
disadv ant ages ar et hatt hef itted pr obabi li
tiescan be l ess t han zer o or gr eater
than one and t he par ti
al ef f
ect of any expl anat ory v ari
abl e( appear ing i nl ev el
form)i sconst ant . Thesel imi tat i
onsoft heLPM canbeov ercomebyusi ngmor esophi sticat ed
binar yr esponsemodel s.
Inabi nar yr esponsemodel ,inter estl iespr imar i
lyi nt her esponsepr obabi li
ty
PY=1/ X) =P( Y=1/ X_ 1,X_ 2…. X_ k……………………………………. .1.11
wher e we use X t o denot e t he f ull set of expl anat ory v ariabl es. For exampl e,
when Y i s an empl oy ment i ndi cat or, X mi ght cont ain v ar i
ous i ndi vidual
char act erist i
cs such as educat ion, age,mar ital st at us,andot herf act orst hataf fect
empl oy ment st at us,i ncl udi ng abi nar y i ndicat or v ariablef orpar tici
pat ion i nar ecentj ob
training pr ogr am.
Speci fy ingLogi tandPr obi t Model s
Int heLPM, weassumet hatt her esponsepr obabi lit
yi slinear i nasetof par amet ers, (_j. To
av oidt heLPM l imi tat ions, consi deracl assofbi nar yr esponsemodel soft hef orm
P( Y=1/ X) =G( ( _0+( _ 1X_ 1+( _2X_ 2+…( _kX_ K) =G( X()………………………………. 1.10
wher e G i s af unct ion t aki ng on v alues st ri
ct l
y bet ween zer o andone: 0<G( z)<1,
for al l r ealnumber s z.Thi sensur est hatt heest i
mat edr esponsepr obabi l
iti
esar est rictly
bet weenzer oandone.
Asi near l
ierEconomet ricsI ,wewr ite X( =(_ 0+( _1X_ 1+( _2X_ 2+…( _kX_ K
Var ious nonl i
near f unct ions hav ebeensuggest ed f or t hef unction G i nor dert omake
sur et hatt hepr obabi lities ar ebet ween zer oandone.Thet wowewi llcov erher ear eusedi n
thev astmaj or ityofappl icat ions( alongwi th t heLPM) .I nt hel ogitmodel ,G i st hel ogi sti
c
funct ion:
Gz=exp ( z)/1+expz=( z=e^ X( /1+e^ X( …………………………………………1. 13
whi ch i sbet weenzer oandonef oral l r ealnumber s z.Thi si st hecumul at i
v edi st ri
but ion
funct ion( cdf )f orast andar dl ogi sticr andom v ariabl e.
Her et her esponsepr obabi li
tyP( Y=1/ X)i sev aluat edas
P=P( Y=1/ X)=e^ X( /1+e^ X(
Simi larly ,thenon-r esponsepr obabi li
tyi sev aluat edas:
1-P=PY=0/ X=1- e^ X( /1+e^ X( =1/ 1+e^ X(
Not et hatt her esponseandnon- r esponsepr obabi l
iti
es bot hl i
e i n t hei nterval [ 0,1],
andhence, arei nter pr etabl e.
Fort hel ogi tmodel , ther at i
o:
P/ 1-P=PY=1/ X=e^ X( / 1+e^ X( /PY=0/ X=1/ 1+e^ X( =e^ X( =e^ (
_ 0+(_1X_ 1+( _2X_ 2+…+( _ kX_ k
i
st her ati
o oft heoddsofY=1agai nstY=o.Thenat urall
ogari
thm oftheodds(
log-odds)i
s:
l
nP/ 1-P=( _0+(_1X_ 1+(_2X_2+…( _
kX_k
Thus, thelog-oddsi sal inearfuncti
onoft heexplanatoryvar
iables.
I
n t he pr obit model, G i s t he st andard nor mal cumul ati
ve dist
ri
but
ion funct
ion
(cdf),which i sexpr essedasani nt
egral:
Gx=(x=_ -∞^ ∞? (xdx…………………………………. .
1.14
Wher e( zist hest andard nor mal densi ty
(x=1/&2( ^2(e^-(
x- (
)^2/2(^2
Gx=_ -
∞^ x_0? 1/&2(^2(e^-x-
(2/2(^
2
Thest andar dnor mal cdfhasashapev erysimil
artot hatoft
hel ogi
sti
ccdf.

Theest imat ing model t hatemer ges f rom t henor mal CDFi spopul ar l
y knownast he
probi t model , although somet imes i tisal soknownast he nor mi t model .
Not et hatbot ht hepr obi tandt hel ogi tmodel sar eest i
mat edbyMaxi mum Li keli
hoodEst i
mat ion.
1.3.3.I nt erpr etingt hePr obitandLogi tModel Est i
mat es
Giv en moder n comput ers, f rom apr act ical per spect ive, t he mostdi ffi
cul t aspectof
l
ogi t orpr obi tmodel s i spr esent ing andi nt er preting t her esul t
s. The coef fi
cient
est i
mat es, t hei r st andar d er rors,andt hev alue oft hel og-l ikeli
hood f unct i
on ar e
repor tedbyal l sof twar epackagest hatdol ogi t andpr obi t, andt heseshoul d ber epor t
edi n
anyappl icat ion.
Thecoef f
icient s gi ve t hesi gns oft hepar t i
al ef fect s ofeachXjont her esponse
probabi lity, andt hest atisti
cal si gni ficance of Xj i sdet ermi ned bywhet her wecan
reject H0:( j=0ata suf fi
cientlysmal l si gnificance l ev el. Howev er,themagni tudeoft he
est i
mat edpar amet er s( dZ/ dX)hasnopar ti
cul ari nter pr etation. Wecar eaboutt hemagni tudeof
dPr ob( Y) /dX. Fr om t hecomput erout putf orapr obi torl ogi testimat i
on,y oucani nterpr ett he
stat i
stical signi ficanceandsi gn ofeachcoef f
icient di rect l
y .Assessi ngmagni t
udei strickier .
GoodnessofFi tSt atistics
The conv ent ional measur e of goodness of f i
t,R2, isnot par t
icularl
y meani ngf ul i n
binar yr egr essandmodel s. Measur essi mi lart oR2,cal ledpseudoR2,ar eav ailabl e,andt her e
areav ariet yoft hem.
Measur esbasedonl ikel i
hood r at i
os
LetLURbet hemaxi mum l i
kelihood f unct ion when maxi mized wi th r espectt oal lt he
par amet er s and LR be t he maxi mum likelihood f unct ion when maxi mi zed wi th
rest ri
ct i
ons( j
=0.
R^2=1- (L_ R/ L_ UR) ^2/ n
CraggandUhl er ( 1970)suggest ed apseudoR2t hatl i
es bet ween 0and1.
R^2=? L_ UR^ 2/ n- ?L_ R^ 2/n/ (1-?L_R^ 2/n) ?L_ UR^ 2/n
McFadden( 1974)def inedR2as:
R^2=1- logL_ UR/ logL_ R
Anot hergoodness-of-f i
t measur e t hati susual ly r epor tedi st heso-cal led per cent
cor rect l
ypr edi ct ed,whi ch i scomput ed asf ol l
ows. Foreachi ,wecomput et heest i
mat ed
probabi lityt hatYit akesont hev alue one, Y_ i
.IfY_ i≥0. 5t hepr ediction ofYi i suni ty
, andi f
Y_i<0. 5,Yii si s pr edict ed t obezer o. The per cent age of t imes t he pr edict edY_ i
mat ches t he act ual Yi( whi ch weknow t obezer oorone) i stheper cent cor r
ect ly
predi cted.
CountR^ 2=No. ofcor rectpr edictions/ Tot al no. ofobser vations
Numer ical Exampl e:
Thi s sessi on showsanexampl e ofpr obi t andl ogi t r egr ession anal y si
s wi th St ata.
Thedat ai nt hisexampl e wer egat her edonunder gr aduat es appl y i
ng t ogr aduat eschool
andi ncl udes under gr aduat eGPAs,t her eput at i
on oft heschool oft heunder graduat e ( a
topnot ch i ndi cat or ), t hest udent s' GREscor e,andwhet her ornott hest udent was
admi tted t ogr aduat e school . Usi ng t his dat aset ,wecanpr edi ctadmi ssion t ogr aduat e
school usi ng under gr aduat e GPA,GREscor es,andt her eputation oft heschool oft he
under graduat e. Ourout comev ariablei sbi nar y,andwewi l
luseei t
herapr obi toral ogi t
model . Thus,ourmodelwi llcal cul ateapr edi ct edpr obabi l
ityofadmi ssionbasedonour
pr
edi
ctor
s.

Iteration Hi stor y - Thi s i s al isting of t he l og l ikel ihoods ateach i teration f or

the pr obi t
/logi tmodel .Remembert hatpr obit/logi t r egr essi onusesmaxi mum l ikelihood
est i
mat ion, whi chi sani t
er ativepr ocedur e. The f i
rst i t erat ion ( cal l
ed I teration 0)i s t he
l
og l i
kel i
hood of t he " null" or" empt y" model ;t hati s,amodel wi t
h nopr edi ctor s.
Att henext i ter ati
on ( called I teration 1) ,the speci fied pr edi ctor s ar eincluded i nt he
model . I nthi s exampl e, t hepr edict or s ar eGRE, t opnot chand GPA. Ateachi terat i
on,
thel og l i
kelihood i ncreases becauset hegoal i stomaxi mi ze t hel ogl ikeli
hood.Whent he
differencebet weensuccessi vei ter ations i sv ery smal l, t hemodel i ssai d t ohav e
"conv er ged" andt hei terat i
ng st ops.
Logl ikelihood - Thi s i sthel og l ikeli
hood oft hef itted model . I tis usedi n t he
Likelihood Rat ioChi-Squar e t estofwhet her al l pr edi ct or s' r egr essi on coef f
icient s i n
themodel ar esi mul t
aneousl yzer o.
LRchi 2(3) - Thi s i st heLi kel ihood Rat i
o ( LR) Chi-Squar e t estt hatatl east oneof
thepr edi ctors' regressi on coef ficient i snotequalt ozer o.Thenumber i n t hepar ent heses
i
ndi cat es t hedegr eesoff reedom oft heChi-Squar e di stribut i
on usedt ot esttheLRChi-
Squar e st atistic andi sdef i
ned byt henumber ofpr edi ct or s i nt hemodel ( 3).
Pr ob >chi 2 - Thi s i sthepr obabi li
ty ofget t i
ng aLRt estst at i
st i
c asext reme as, or
mor eso, thant he obser ved st at i
st i
c under t he nul l hy pot hesi s; t he nul l hy pot hesi s
i
s t hat al l oft he r egr essi oncoef ficients ar esi mul taneousl y equal t ozer o. I not her
wor ds, t
hi s i st he pr obabi li
ty ofobt ainingt his chi -squar e st atist ic oronemor eext reme
i
f t her ei si nf act noef fect oft hepr edictor v ari
abl es.Thi s p-v alue i scompar edt oa
speci fied al pha l ev el, ourwi ll
ingness t oacceptat y peIer ror ,whi chi s t ypicall
y set at
0.05or 0. 01.The smal l p-v alue f r
om t he LR t est , 0. 0001,woul d l ead us t o
concl ude t hat at l east one of t he r egr essi on coef f i
ci ent s i n t he model i s not
equal t ozer o. Thepar amet er oft hechi-squar e di stribut ion usedt ot estt henul l
hy pothesi s i sdef ined byt hedegr eesoff reedom i nt hepr i
orl ine, chi 2(3) .
PseudoR2 - Thi s i sMcFadden' spseudoR-squar ed.Because t his st at i
stic doesnot
meanwhatRsquar emeansi n OLSr egr ession ( t hepr opor t
ion ofv ar i
ance oft her esponse
var i
abl e expl ained byt hepr edi ctors) ,i tshoul d be i nt er pret ed wi th gr eatcaut ion.The
i
nt erpr etation of t he coef f i
cient s canbeawkwar d. Forexampl e, f or aone uni t
i
ncr ease i nGPA, t
hel ogoddsofbei ng admi tted t ogr aduat eschool (vs.notbei ng admi tt
ed)
i
ncr eases by. 667.Fort his r eason, many r esear cher s pr ef er t oexponent iate t he
coef ficients and i nterpr et t hem asodds-r atios. Lookatt hef ol lowi ngr esult.

Now wecansayt hatforaoneuni t increase i n GPA,t heoddsofbei ng admi t

ted t o
graduat
e school( vs.notbei
ng admi t
ted) i ncreased byaf actor of1.94.Si
nce GRE
scoresi
ncrease only in units of10,wecant aketheoddsr at
io andr ai
se itt
ot he10th
power,e.
g.1.00248^10=1. 0250786,andsay f ora10uni t increase in GRE scor e,the
oddsofadmi ssi
on tograduat
e school i ncreased byaf actorof1.025.
CHAPTERTWO
2. I nt roduct i
ont oBasi cRegr essi onAnal ysiswi t
hTi meSer iesDat a
2.1.TheNat ureofTi meSer iesDat a
Time ser i
es dat a ar e dat a col lect ed f or a si ngle ent ity ( person, f i
rm, and
count ry ) col lect ed( obser ved)atmul ti
pl e t i
me per iods. Exampl es:
Aggr egat e consumpt ion and GDP f or a count ry ( for exampl e, 20 y ear s of
quar terlyobser vat i
ons= 80obser vat i
ons)
Birr/$, pound/ $ andEur o/ $ exchange r at es( dai l
y dat afor2y ear s=730obser vat i
ons)
Inflat i
on r at ef orEt hiopi a ( quar terly dat af or30y ear s=120obser v at i
ons )
Gr ossdomest i
c i nv est ment f orEt hiopi a ( annual dat afor 40y ear s=40obser vat i
ons )
An obv ious char act er ist i
c of t ime ser ies dat awhi ch di stingui shes i tf rom cr oss-
sect ional dat ai sthat at i
me ser ies dat aset comes wi th at empor al or der i
ng. For
exampl e, i n t he abov e dat a setment ioned i nt heexampl es, wemust knowt hat t he
dat af or1970i mmedi atel y pr ecedet hedat af or1971. Foranal yzing t ime ser i
es dat a i n
the soci al sci ences, wemust r ecogni ze t hat t he pastcanaf f ect t hef utur e, butnot
vice v er sa.Toemphasi zet hepr operor der ing oft ime ser i
es dat a,Tabl e2. 1gi ves a
par tial l isting oft hedat aon Et hiopi a gr osscapi t
al f or mat i
on ( GCF)andgr ossdomest i
c
sav ings ( GDS)bot hi n mi lli
on ETBsi nce 1969-2010.
Anot herdi fferencebet weencr oss-sect ionalandt imeser iesdat ai smor esubt l
e. I nChapt ers2
and3ofEconomet ricsI ,west udi edst at i
sticalpr oper tiesof t he OLSest i
mat ors basedon
the not iont hat sampl es wer er andoml y dr awn f r
om t he appr opriat e popul at i
on.
Under standi ng why cr osssect ional dat ashoul d bev i
ewed asr andom out comes i s
fairly st rai ght forwar d: adi ffer ent sampl edr awnf rom t hepopul ati
on wi ll gener all
y
yield di f fer ent v alues oft hei ndependent anddependentv ariables ( such as educat i
on,
exper ience, wage, and so on) . Ther ef or e,theOLSest imat escomput ed f rom di fferent
random sampl es wi ll gener ally di ffer , andt hi s i s why weconsi der t heOLS
est i
mat or s t ober andom v ariabl es.
How shoul dwet hink aboutr andomness i nt imeser iesdat a?Cer t ainly, economi c t ime
ser i
essat isf yt hei ntuitiv e r equi rement s f orbei ng out comes ofr andom v ariables. For
exampl e, t oday wedonotknowwhatt hest ockpr i
ce wi l
l beat i ts cl oseatt heendof
thenext t radi ng day .Wedonotknow whatt heannualgr owt hi nout putwi llbei nEt hiopia
dur ing t hecomi ng y ear .Si ncet heout comesoft hesev ariablesar enotf oreknown, theyshoul d
clear lybev iewed asr andom v ariabl es.For mal ly, asequence ofr andom v ariables
i
ndexed byt ime i scal led ast ochast ic pr ocess orat imeser i
espr ocess.( “St ochast ic” i s
asy nony m f orr andom. )Whenwecol lect at ime ser iesdat aset ,weobt ain one possi ble
out come, or r ealizat ion, oft he st ochast ic pr ocess. Wecanonl yseeasi ngl er eal i
zat i
on,
becausewecannotgobacki nt imeandst ar tt hepr ocessov eragai n. ( This i sanal ogoust o
cross-sect ional anal y sis wher e we can col l
ect onl y one r andom sampl e.
)
Howev er,i f cer taincondi tions i n hi stor y hadbeen di fferent , wewoul d gener all
y
obt ain adi fferent r eal izat ion f ort hest ochast i
c pr ocess, andt his i swhywet hi nk of
time ser iesdat aast heout come ofr andom v ar i
ables.
Thesetofal lpossi bler eal izat i
onsofat imeser iespr ocesspl ayst her oleoft hepopul ationi n
crosssect ional anal ysis.
2.2.St at ionar yandnon- st at i
onar ySt ochast icPr ocesses
2.2.1.St ochast icPr ocesses
Ar andom orst ochast icpr ocessi s acol l
ect ion of r andom v ari
abl es or der edi n t i
me. I f
wel et Ydenot e ar andom v ariabl e, andi f i ti s cont inuous, we denot e i t asY( t),
but i f i t i s di scr ete, wedenot ed i tasYt.Anexampl eoft hef orm er i san
elect rocar diogr am, andanexampl e oft hel atter i sGDP, PDI, GDI ,GDS, et c.Si ncemost
economi cdat aar ecol lect ed atdi scr et e poi nt s i n t ime, f orourpur posewewi ll uset he
not at i
on Yt r ather t han Y( t).Ifwel etY r epr esent GDS,f or ourdat awehav eY1, Y2, Y
3,.
..,Y39, Y40, Y41, wher et hesubscr i
pt 1denot est hef ir
st obser v ation ( i.e.,GDSof1969)and
the subscr ipt 41 denot est hel ast obser vation ( i.
e., GDS of 2010) .Keepi n mi nd
thateachoft heseY’ sisar andom v ar iable.
Stat ionar ySt ochast icPr ocesses
At ypeofst ochast i
cpr ocesst hathasr ecei v edagr eatdealofat t
ent ionandanal ysi sbyt ime
ser iesanal yst s i s t heso-cal led st at i
onar yst ochast i
cpr ocess. Br oadl y speaki ng,a
stochast icpr ocessi ssai d t o be st at i
onar y i fi ts mean and v ariance ar e const ant
ov er t i
me and t he v al ue of t hecov ariancebet weent het wot i
meper i
odsdependsonl y
ont hedi st anceorgaporl agbet weent het wot imeper iodsandnott heact ualt i
meatwhi cht he
cov ar iancei scomput ed. I nt het imeser iesl iteratur e,suchast ochast icpr ocessi sknownasa
weakl yst ationar y, orcov ar iancest ati
onar y, orsecond- or derstationar y ,orwi desense, stochast i
c
process. Fort he pur pose oft hischapt er ,and i n mostpr act i
calsi t
uat ions,t hist ype of
stat ionar ityof tensuf fices.
Toexpl ainweakst ationar ity,letYt beast ochast ic t ime ser i
eswi th t hese pr oper ti
es:
Mean: EY_ t=( ……………………………………………………………………1
Var iance: varY_ t=E( Y_ t-
()^2=( ^2…………………………………………. .2
Cov ariance: (_ k=EY_ t
-(Y_ t+k- (…………………………………………. .3
Wher e ( _k, thecov ariance ( oraut ocov ariance) atl ag k, isthecov ar i
ance bet ween t he
values ofYt and Yt +k,t hati s,bet ween t woY v alues kper i
ods apar t.Ifk=0,weobt ain
(_0,whi ch i ssi mpl y t hev ariance ofY( =σ2) ;if k=1,( _1i st hecov ariance bet weent wo
adj acent v alues ofY.
Supposeweshi fttheor i
gi nofYf rom Yt t o Yt +m ( say,from 1969t o1974f orourGDS
dat a) .Nowi fYt i st obe st ationar y ,t hemean, v ariance, and aut ocov ari
ances ofYt +m
mustbet he sameast hoseofYt.I nshor t,ifat i
meser iesis st ationar y ,i t
smean, v ari
ance,
andaut ocov ar iance ( atv ar i
ous l ags)r emai n t hesamenomat ter atwhatpoi nt we
measur e t hem; t hati s, theyar e t i
mei nv ariant. Suchat i
me ser ies wi llt end t or eturn
toi ts mean ( called mean r ev ersi on) and f l
uct uat i
ons ar ound t hismean ( measur ed
byi tsv ariance) wi l
l hav eabr oadl y const ant ampl it
ude.I f at ime ser i
es i s not
stat ionar y i n t hesense j ust def ined, i t i scal led anonst ationary t ime ser i
es( keep
i
n mi nd wear et alking onl y about weak st at i
onar it
y).I not herwor ds,anonst ationar y
ti
meser ieswi l
l hav e at i
me-v ary i
ng mean orat i
me-v aryi
ng v ar i
ance orbot h.
Whyar e st at i
onar yt i
me ser ies so i mpor tant ? Because i f at ime ser ies i s
nonst at i
onar y, we can st udy i ts behav i
or onl y f or t he t ime per iod under
consi der ation. Eachsetoft imeser iesdat awi llther ef ore bef orapar t
icul ar epi sode. Asa
consequence,i ti snotpossi blet ogener ali
zei t t oot hert i
me per iods. Ther efor e, f or t he
pur pose of f orecast i
ng,such ( nonst at i
onar y) t i
me ser ies may beofl ittl
e pr actical
value.
How doweknow t hatapar ticulart imeser iesi sst at i
onar y? I npar ti
cul ar,ist het imeser ies
showni nFi gur e 2. 1 st ationar y? We wi l
lt ake t hi s impor tant t opic up i n Sect i
ons
2.5, wher e we wi llconsi der sev eral t est sofst at ionar i
ty. But i f wedependoncommon
sense, i twoul d seem t hat t he t ime ser ies depi cted i n Fi gures 2. 1 i s
nonst at i
onar y, at l east i n t he mean v alues

Bef
ore wemov e on,wement i
on aspeci al t ypeof st ochasti
c pr ocess(or t i
me
ser
ies), namel y, apur el
yr andom,or whi t
enoi se,process.Wecal lastochast i
cpr ocess
pur
elyr andom ifi
thaszer omean,constantv ar
ianceσ2,and i s ser i
all
y uncor rel
ated.You
mayr ecallt hat t
he er r
or t erm ut, enteri
ng t he classical nor mal linear regression
model t hat we di scussed i n Economet ri
cs I wasassumed t o be a whi te
noi
se pr ocess, which we denot ed as ut ∼I ID N(0,σ2); that i s, ut i s
i
ndependent ly and identi
cally dist
ri
buted asa nor mal di str
ibuti
on wi th zer o mean
and const antv ariance.
Nonst at ionar ySt ochast icPr ocesses
Although ouri nt erest i si n st ationary t i
me ser ies, one of ten encount er s
nonst at ionar y t ime ser ies, thecl assic exampl e bei ng t he r andom wal kmodel ( RWM) .
Itisof ten sai d t hatassetpr ices, suchas st ock pr i
ces or exchange r ates, f ollow a
random wal k; t hat i s, t hey ar e nonst ationar y. Wedi stinguisht wot y pesofr andom
wal ks: ( 1)r andom wal kwi thoutdr if
t( i
.e.,noconst ant ori nt erceptt er m) and( 2)r andom
wal kwi th dr i
ft( i
.e., aconst ant t ermi spr esent ).
Random Wal kwi thout Dr ift: Supposeuti sawhi te noi se er rort erm wi th mean 0and
var i
ance( 2. Then t heser iesYti ssai dto bear andom wal kif:
Y_ t=Y_ t-1+u_ t……………………………………………………………………………. .2.4
In t her andom wal kmodel , as( 2.4)shows, thev al
ue ofYatt imeti sequal t oi tsv al ue at
time ( t−1)pl us ar andom shock;t hus i tisan AR( 1)model i n t hel anguage of
Chapt er4ofEconomet rics I .Wecant hinkof( 2.4)asar egr essionof Yatt ime toni tsv alue
l
agged oneper iod.Bel iev ers i nt heef fici
ent capi tal mar ket hy pot hesi s ar gue t hat
stockpr ices ar eessent ially r andom and t herefor et her e i s noscopef or pr ofit
abl e
specul at i
on i n t hest ockmar ket:I f onecoul d pr edict t omor row’ spr i
ceont hebasi sof
today ’spr ice, wewoul d al l bemi llionaires.
Nowf r om ( 2.4)wecanwr ite
Y_ 1=Y_ 0+u_ 1
Y_ 2=Y_ 1+u_ 2=Y_ 0+u_ 1+u_ 2
Y_ 3=Y_ 2+u_ 3=Y_ 0+u_ 1+u_ 2+u_ 3
Ingener al ,ifthepr ocessst artedatsomet ime0wi th av alue ofY0, wehav e
Y_ t=Y_ 0+_ ^? u_ t………………………………………………………. 2.5
Ther efor e,
E(Y_ t)=EY_ 0+_ ^?u_ t=Y_ 0………………………………………………………. 2.6
Inlikef ashi on, i tcanbeshownt hat
var Y_ t=t (^ 2………………………………………………………………………. 2.7
Ast hepr ecedi ngexpr essi onshows, themean ofYi sequal t oi tsi niti
al, or st arti
ng, v alue,
whi chi s const ant , but ast i ncreases, i ts v ariance i ncr eases i ndef i
nitely, t hus
violat ing acondi ti
on ofst ationar it
y . I nshor t
,t heRWM wi thoutdr ifti sanonst at i
onar y
stochast icpr ocess.I npr act i
ceY0i sof tensetatzer o,inwhi chcaseE( Yt)=0.
Ani nter est ingf eatur e ofRWM i st heper sistence ofr andom shocks( i.
e.,r andom er rors),
whi ch i scl earf rom ( 2.5):Yt i sthesum ofi niti
al Y0pl us t hesum of r andom
shocks.Asar esul t,t hei mpactofapar t
icularshockdoesnotdi eaway .Forexampl e,i fu2=2
rathert hanu2=0, t hen al l Yt’ sf rom Y2onwar dwi l
l be2uni ts hi gher andt heef f
ect
oft hi s shocknev er di esout .Thati swhyr andom wal ki ssai dt ohav eani nfinit
ememor y.
Thei mpl icat ioni st hat ,random wal kr emember st heshockf or ever ;thati s,ithasi nf i
ni t
ememor y .
Inter est ingl y ,ify ouwr i
te ( 2.4)as
Y_ t-Y_ t-1=ΔY_ t=u_ t………………………………………………………………2. 8
Wher eΔi st hef irstdi ffer enceoper at or?Itiseasyt oshowt hat ,whi leYti snonst at i
onar y ,itsfirst
differ encei s st at i
onar y .I n ot her wor ds, t he f ir
st di f
fer ences of a r andom wal k
time ser i
es ar est at i
onar y.
Random Wal kwi thDr ift.Letusmodi fy(2.4)asf ollows:
Y_ t=θ+Y_ t-1+u_ t………………………………………………………………2. 9
Wher eθi sknownast hedr iftpar amet er. Thenamedr iftcomesf r
om t hef actt hati fwewr ite
thepr ecedi ng equat ion as
Y_ t-Y_ t-1=ΔY_ t=θ+u_ t………………………………………………………………2. 10
i
tshowst hatYt dr ifts upwar dordownwar d,dependi ng onδ being posi tive or
negat iv e. Not et hatmodel( 2.9)i sal soanAR( 1)model .Fol lowi ngt hepr ocedur edi scussed f or
random wal k wi t
houtdr i
f t,i tcan beshown t hatf ort her andom wal kwi th dr i
ft model
(2.9) ,
E(Y_ t)=Y_ 0+t .
θ+u_ t………………………………………………………………2. 11
v ar Y_t=t (^2……………………………………………………………………………2. 12
Asy ou cansee, for RWM wi th dr i
ftt hemean aswel l ast hev ari
ance i ncr eases ov er
ti
me, agai nv iolat i
ng t he condi tions of ( weak) st ationar ity. I n shor t, RWM, wi th
or wi thout dr ift,i s anonst at ionar y st ochast ic pr ocess. Ther andom wal k model
i
sanex ampl e of what i sknown i n t hel iter atur e asauni tr ootpr ocess.
Uni tRootSt ochast icPr ocess
Letuswr itet heRWM ( 2. 4)as:
Y_ t=(Y_ t-1+u_ t;
-1≤( ≤1……………………………………………………………. 2.13
Thi s model r esembl es t he Mar kov f irst-or der aut oregr essi v e model t hat we
discussed onaut ocor relation. I f ρ=1, (2. 4.1)becomes a RWM ( wi thout dr i
ft).I f ρi s
i
n f act 1, wef ace what i s
known ast heuni t r oot pr obl em, t hat i s,asi t
uat ion of nonst ationar y; weal ready
knowt hati n t hiscaset hev ariance ofYt i snotst ati
onar y . Thenameuni tr ooti sduet ot he
factt hat ρ=1.Thust het ermsnonst ationar y ,r andom wal k,anduni trootcanbet reatedas
sy nony mous.I f,howev er,| ρ| <1, thati si ft heabsol ute v alue ofρi slesst han one, then i t
canbeshown t hatt het i
me ser i
esYt i s st at i
onar y i n t hesensewehav e def i
ned i t
.
2.3.Tr endSt ati
onar yandDi fferenceSt at i
onar ySt ochast i
cPr ocesses
Ift het rendi nat imeser i
esi scompl etelypr edi ctabl eandnotv ariabl e,wecal li tadet ermi nisti
c
trend,wher easi fi tisnotpr edi ctabl e,wecal litast ochast i
ct rend.Tomaket hedef i
niti
onmor e
formal , consi dert hef ollowi ngmodel ofthet i
meser i
esYt .
Y_ t=(_ 1+( _ 2t+( _3Y_ t
-1+u_ t…………………………………………………………………. 14
Wher euti sawhi tenoi seer rort er m andwher eti st i
memeasur edchr onol ogical ly.Nowwehav e
thef ollowi ngpossi bili
ties:
Pur er andom wal k: I fin ( 2.14)β1=0, β2=0, β3=1, weget
Y_ t=Y_ t -
1+u_ t………………………………………………………………………………. 2.15
whi chi snot hing butaRWM wi thout dr ift andi st heref ore nonst ationar y. Butnot ethat ,i
f
wewr ite ( 2.15)as
ΔY_ t=Y_ t-Y_ t-1+u_ t
……………………………………………………………………. 2.16
i
tbecomesst ationar y , asnot edbef ore. Hence,aRWM wi thout dr ift i sadi ff
erence
stat i
onar ypr ocess( DSP) .
Random wal kwi thdr ift:I fi n( 2. 14)( _1≠0, (_ 2=0, (_3=1, we get
Y_ t=(_ 1+Y_ t
-1+u_ t……………………………………………………………………. 2.17
whi ch i sar andom wal kwi th dr ift andi sther ef ore nonst ationar y.I fwewr i
tei tas
Y_ t-Y_t -1=ΔY_ t=(_ 1+u_ t…………………………………………………………. 2.17a
thismeans Yt wi ll exhi bit aposi ti
v e ( β1>0)ornegat i
v e ( β1<0)t rend.Suchat rendi s
cal l
edast ochast ict rend. Equat ion( 2.16a)i saDSPpr ocessbecauset henonst ationarityi nYt
canbeel imi nated byt aking f i
rst di fferences oft het ime ser ies.
Det er mi ni stict rend: I fi n( 2.14) ,(_ 1≠0, (_2≠0, (_
3=0weget
Y_ t=(_ 1+( _ 2t+u_ t
……………………………………………………………………. 2. 17b
whi ch i scal led at rendst ationar y pr ocess ( TSP) .
Alt hough t he mean ofYt i s
β1+β2t ,whi ch i s notconst ant ,i tsv ar iance ( =σ2)i s. Oncet hev aluesof β1 and β2
areknown, themean canbef or ecast edper fect ly. Ther efore, i f we subt ract t hemeanof
Yt f rom Yt,t her esul ting ser ies wi l
l best at i
onar y,hence t he name t rend
stat i
onar y. Thi s pr ocedur e ofr emov ing t he ( deter mi nist ic) trend i s cal l
ed
det rendi ng.
Random wal kwi thdr iftanddet er mi nisti
ct r end: Ifi n( 2.14) (_1≠0, ( _
2≠0, (_3=1, weobt ai
n:
Y_ t=(_ 1+( _ 2t+Y_ t
-1+u_ t……………………………………………………………………. 2.18 we hav ear andom walk
wi th dr ift andadet ermi ni stic t rend, whi ch canbe seeni fwewr it
e t hisequat ion as:
Y_ t-Y_t -1=ΔY_ t=(_ 1+(_ 2t+u_ t …………………………………………………………. 2.18a
whi chmeans t hatYt i snonst at ionar y.
Det er mi ni stict rendwi thst at i
onar yAR( 1)component :I fi n(2. 14)( _1≠0, (_2≠0, (_ 3<1, weobt ain:
Y_ t=(_ 1+( _ 2t+( _3Y_ t
-1+u_ t……………………………………………………………………. 2.19
whi chi sst ationar y ar oundt hedet ermi nist i
c t rend.
2.4.I nt egr at edSt ochast icPr ocess
The r andom wal k model i s but aspeci fic case of amor e gener al cl ass of
stochast ic pr ocessesknown asi ntegr at ed pr ocesses.Recal l t hat t he RWM wi thout
drift i s nonst at i
onar y, but i t s f irstdi ff erence, asshown i n ( 2.8) ,is st ationar y.
Ther ef or e, wecal ltheRWM wi thoutdr i
fti nt egr at edofor der1, denot edasI (1).Si milarly,ifat i
me
ser i
eshast obedi ffer encedt wi ce( i.
e. ,taket hef irstdi fference oft he f i
rst di fferences) t o
makei t st at ionar y, wecal l such at i
me ser i
es i ntegr at ed ofor der 2. I n gener al,
i
f a( nonst at ionar y) t i
me ser i
es has t o be di f
fer enced dt i
mes t o make i t
stationar y ,t hatt ime ser i
es i ssai dt obei ntegr at ed ofor derd.At imeser i
esYti ntegr ated
ofor derdi sdenot edasYt∼I (d) .Ifat i
meser ies Yt i sst ationar y t obegi n wi t
h ( i
.e.,it
doesnotr equi re anydi ffer enci ng) ,i t i ssai d t obei ntegr at ed ofor derzer o, denot ed by
Yt ∼I (0) .
Thus, wewi l
l uset het er ms “ stat i
onar y t ime ser i
es” and “ time ser i
es
i
nt egrat ed ofor derzer o” t omean t hesamet hi ng.
Mosteconomi ct ime ser iesar egener ally I (1);t hati s,they gener al l
y becomest ati
onar y
only af tert aki ngt hei rf irst di fferences.
Proper tiesofI nt egr at edSer i
es
The f ol lowi ng pr oper ties ofi nt egr ated t ime ser i
es may benot ed: LetXt, Yt, andZt
bet hree t imeser ies.
If Xt∼I (0) and Yt∼I (1),then Zt=( Xt +Yt)=I (1) ;t hat i s, al i
near combi nat i
on or
sum ofst at i
onar y andnonst at i
onar y t ime ser iesi s nonst at i
onar y.
If Xt ∼I (d) ,t hen Zt=( a+bXt)=I (d), wher e a and b ar e const ants. That i s, a
l
inearcombi nat ion ofanI (d)ser i
es i sal soI ( d).
Thus, if Xt ∼I( 0) ,then Zt =( a+bXt) ∼I (0) .
IfXt∼I ( d1)andYt∼I (d2) ,then Zt =( aXt+bYt) ∼I(d2) ,wher ed1<d2.
IfXt ∼I (d) and Yt ∼I (d) ,thenZt =( aXt +bYt) ∼I(d* )
;d*i sgener ally equalt od,buti n
somecasesd* <d.
2. 5.Test sofSt ationar ity: TheUni tRootTest
At est of st at ionar i
ty ( ornonst at ionar i
ty) t hat has become wi dely popul ar ov er
the pastsev er aly ear si st heuni t r oott est .Thest ar t
ingpoi nti st heuni tr oot( stochast i
c)
processt hatwedi scussedi nSect ion 2. 2.West artwi th ( 2.4. 1)
Y_t=( Y_ t-1+u_ t;- 1≤( ≤1………………………………………………………………. 2.20
Wher eut i sawhi t
e noi se er rort er m.
Weknow t hat i f ρ =1,t hat i s, i n t he caseoft heuni t r oot ,(2.20)becomes a
random wal k modelwi thout dr ift, whi ch weknow i s anonst ationary st ochast i
c
process.Ther ef ore, why not si mpl yr egr essYt oni ts( one per iod)l agged v alue Yt −1
andf i
nd outi ft heest imat ed ρi sst at istical l
y equalt o1?I fitis, t
henYt i snonst ationar y.
Thi sist hegener al i
deabehi ndt heuni tr oott estofst at ionar i
ty.
Fort heor eticalr easons,wemani pul at e( 2. 20)asf ol l
ows: Subt ract Yt −1 f rom bot hsi des
of( 2.20)t oobt ai n:
Y_t-Y_ t-1=( Y_ t-1-Y_ t
-1+u_ t; …………………………………………………. 2.21
whi chcanbeal ter nat ivelywr itten as:
ΔY_ t=( Y_ t-1+u_ t;…………………………………………………. 2. 22
Wher e δ=( ρ−1)and( ,
ist hef irst -differ ence oper at or.
In pr act ice,t her ef or e,i nst ead ofest i
mat ing ( 2.20) ,we est imat e( 2.22)and t estt he ( null)
hypot hesi st hat δ = 0.I f δ = 0, t hen ρ = 1, t hat i s we hav e a uni t
root , meani ng t he t ime ser i
es underconsi der ati
on i snonst at i
onar y. Unf ortunat ely,
undert henul lhy pot hesi st hatδ =0( i.e., ρ=1) ,thetv alue oft heest imat ed coef fi
cient
ofYt −1 doesnotf ollow t hetdi str i
but ion ev eni n l argesampl es;t hat i s, i t doesnot
hav e an asy mpt ot i
c nor mal di stribut ion. Di ckeyandFul ler hav e shown t hatunder
thenul l hy pot hesi s t hatδ = 0,t heest imat ed tv alue oft hecoef ficient ofYt −1 i n
(2.22)f ol lowst heτ( tau)st at istic. I nt hel iter atur et het aust atisticort esti sknownast he
Dickey –Ful ler( DF)t est ,inhonorofi tsdi scov er ers.
Toal low f ort hev ar iouspossi bi l
ities, t heDF t esti sest imat ed i nt hreedi ff
er ent f or ms,
thatis, undert hr eedi fferent nul l hy pot heses.
Y_tisr andom wal k: ΔY_ t
=( Y_ t-1+u_ t;…………………………………………………. 2. 23a
Y_tisr andom wal kwi thdr ift:ΔY_ t=(_1+( Y_ t-1+u_ t;…………………………………. 2. 23b
Y_tisr andom wal kwi thdr iftandt rend: ΔY_ t=(_ 1+( _2t +(Y_ t
-1+u_ t;
………………. 2.23c
Wher et i st het imeort r
endv ar i
abl e.I neachcase,t henul lhy pot hesi si st hatδ=0;t hati s,
therei sauni tr oot —t het imeser iesi snonst at i
onar y . Theal ter nat ivehy pot hesi sist hatδi sl ess
thanzer o;t hati s, t het i
meser iesi sst ationar y.I fthenul lhy pot hesisi srejected, i
tmeanst hat
Yt i sast ationar ytimeser ies.
TheAugment edDi ckey–Ful ler(ADF)Test
Inconduct ingt heDFt estasi n( 2.23a-c) ,i twasassumedt hatt heer rort erm ut was
uncor related.Buti ncaset heutar ecor r
elated, Di ckeyandFul l
erhav edev elopedat est,knownas
theAugment edDi ckey –Ful ler( ADF)t est. Thi st estisconduct edby“ augment i
ng”t hepr ecedi ng
threeequat ionsby addi ng t he l agged v alues of t he dependent v ariable ΔYt- i
.To
bespeci f i
c,supposeweuse( 2.23c) .TheADFt esther econsi st sofest imat i
ngt hef oll
owi ng
regression:
ΔY_ t=(_ 1+( _2t +(Y_ t-
1+_ ^? (_i ΔY_ t-1+( _i
;………………. 2.
24
Wher eεt i sa pur ewhi tenoi seer rort erm and wher e ΔYt −1= ( Yt −1−Yt −2),ΔYt −2 =
(Yt−2−Yt −3) ,etc.I nADFwest i
lltestwhet herδ=0andt heADFt estf oll
owst hesameasy mpt ot i
c
distr
ibut i
onast heDFst atist ic,sot hesamecr it
ical valuescanbeused.
Togi v e a gl impse of t hi
s pr ocedur e, we est i
mat ed ( 2.24)f or t he GDP ser ies
using one l aggeddi ffer ence of nat ural l ogof GDP ofEt hiopia; t her esults wer eas
foll
ows:
ΔGDP_ t=- 0.2095+0. 0016t +0. 0197157GDP_ t
-1+0. 0269ΔGDP_ t-1
t=τt au -
0.28 0.67 0.27 0.15
1%CV=- 4.242 5%CV=- 3.540
The t (=τ)v al ue oft heGDPt −1 coef ficient (=δ)i s0. 27,butt his v alue i nabsol ute
termsi smuch l esst hanev ent he1%and5% crit
ical τv al ue of −4. 242and -3. 540
respect i
v ely, agai n suggest ing t hatev enaf t ertaking car eofpossi bl e aut ocorrelation i n
theer rort erm, theGDPser i
esi snonst ati
onar y .
ThePhi ll
ips–Per ron( PP)Uni tRootTest s
An impor tantassumpt i
on oft heDF t esti st hatt heer rort er msutar eindependent l
yand
i
dent icallydi str i
but ed. The ADF t estadj ust st he DF t estt ot ake car e ofpossi ble ser ial
correlationi nt heer rorter msbyaddi ngt hel aggeddi fferencet ermsoft her egressand. Phi l
lips
andPer ronusenonpar amet ri
c st atisti
cal met hods t ot akecar eoft he ser i
al cor r
el ation
i
n t heer ror t erms wi thoutaddi ng l agged di f
ference t erms.
ThePhi l
lips-Per ront estinv olvesf it
tingt hef ollowi ngregr essi on:
Y_t=( _1+( _ 2t+(Y_ t-
1+u_ t
Under t he nul l hy pot hesi s t hat ρ=0,t hePPZ( t) and Z( ρ) st atistics hav e t he
same asy mpt ot i
cdi stribut ions ast he ADFt -statistic andnor mal i
zed bi as st at i
stics.
Oneadv ant ageoft hePPt est sov ert heADFt estsi sthatt hePPt est sarer obustt ogener alf orms
ofhet eroscedast icityint heer r
ort er m ut.Anot her adv ant age i s t hat t he user does
not hav e t o speci f y a l ag l engt h f or t he t estr egr ession. Now l et’
st estf or
whet herl nGDPi sst ati
onar yornotusi ngPPt est.

Theabov
eresul
tshowst
hatl
nRGDPi
snonst
ati
onar
yatl
evel
.

2.6ThePhenomenonofSpur iousRegr essi on

Considerthefollowingt worandom wal kmodel s:
Y_t=Y_t-1+u_ t
X_t=X_t-1+v _t
Wher eut∼N( 0,1)andv t∼N(0,1).Weal soassumedt hatutandv tareserial
l
yuncor r
elatedaswel l
asmut uallyuncorrel
ated.Asy ouknowbynow,bot htheset imeser iesarenonstati
onar y;thatis,
theyareI (1)orexhibitstochastictrends.
Supposewer egr
essYtonXt .SinceYtandXtar euncor rel
atedI (1)processes,t
heR2f rom the
regressionofYonXshoul dt endt ozero; thatis,t
her eshouldnotbeanyr el
ati
onshipbet weent he
twov ariables.
Y_t=13.26+0. 337X_t, R^2=0. 104, d=0.0121
t
= 0. 62 (
0.044)
Thecoef fici
entofXi shi ghlyst ati
sti
cal l
ysignificant,and,al t
hought heR2v aluei slow,i tis
stati
sti
cal l
ysignifi
cantlydif
ferentf r
om zer o.From t heseresults,youmaybet emptedt oconcl ude
thatther eisasi gnif
icantstatisti
calrelationshipbet weenYandX,wher easapr i
orithereshoul d
be none.Thi si si n a nut
shellthe phenomenon ofspur i
ous ornonsense r egression,fir
st
discov eredbyYul e.Thatthereissomet hingwr ongint heprecedingregressi
oni ssuggest edby
theext remelylowDur bin–Watsondv al
ue,whichsuggest sverystrongfi
rst-or
deraut ocorrel
ati
on.
Accor dingt oGrangerandNewbol d,anR2>dt heestimatedregressi
oni sspurious.TheR2and
thetst atist
icfrom suchaspur iousr egressionaremi sleadi
ng,andt hetst atisti
csar enot
distr
ibut edas( Student’
s)tdistri
butionand,t heref
ore,cannotbeusedf ortesti
nghy potheses
aboutt hepar ameters.

2.7Coi nt egr at ion: Regr essi onofaUni tRootTi meSer i

es
onAnot herUni tRootTi meSer ies
Wehav ewar nedt hatt her egr essi onofanonst ati
onar yt imeser iesonanot hernonst ationar yt ime
ser i
esmaypr oduceaspur i
ousr egr essi on.Supposet hatY_ tandX_ taret imeser iesdat aand
i
ndi v idual lyt heybot har eI (
1) ;thati s,t heycont ainast ochast i
ct rend.I tisqui tepossi blet hatt he
twoser iesshar et hesamecommont rendsot hatt her egressi onofoneont heot herwi l
lnotbe
necessar il
yspur i
ous.
Y_t =(_ 1+( _ 2X_ t+u_ t
u_t=Y_ t-(_ 1-(_ 2X_ t
Supposewenowsubj ectu_ tt ouni tr ootanal ysisandf indthati ti sst at i
onar y;thatis, i
tisI (0) .The
variabl esY_ tand X_ tar ei ndiv i
dual lyI (1),t hati s,t heyhav est ochast ict rends,t heirl i
near
combi nat ioni sI (0) .So, thel inearcombi nat ioncancel soutt hest ochast ict rendsi nthet woser ies.
Int hiscasewesayt hatt het wov ariabl esar ecoi ntegr ated.Economi cal l
yspeaki ng, t
wov ar i
abl es
will becoi nt egr atedi ftheyhav eal ong- ter m, orequi librium, relat i
onshi pbet weent hem.
ErrorCor rect ionMechani sm ( ECM)
Ift he v ariabl es Y_ tand X_ tar e coi nt egr ated,t hati s,t herei sal ong- term,orequi libr i
um,
relationshi pbet weent het wo.Ofcour se,i nt heshor t-
runt her emaybedi sequi libri
um.Ther efor e,
wecant reatt heer rort er mi nt hef ol l
owi ngequat i
onast he“ equi l
ibr i
um er ror.”Andwecanuse
thiser rort ermt ot i
et heshor t-runbehav i
orofY_ ttoi tslong- runv alue:
u_t=Y_ t-(_ 1-(_ 2X_ t-(_3t…………………………………………………………………2. 25
Iftwov ar i
abl esYandXar ecoi ntegr ated,t her elati
onshi pbet weent het wocanbeexpr essedas
ECM.
Nowconsi dert hef oll
owi ngmodel :
ΔY_ t=(_ 0+( _1ΔX_ t +(_2u_ t-
1+( _i………………………………………………………………. .2.26
ECM equat i
on( 2.25)st atest hatΔY_ tdependsonΔX_ tandal soont heequi li
br i
um er rort erm.I f
theequi l
ibr ium er r ort ermi snonzer o,t hent hemodeli soutofequi l
ibri
um.Suppose,i fΔX_ tis
zero and u_ t-1isposi ti
v e,Y_ t-1i st oo hi ght o bei nequi l
ibrium,t hati s,Y_ t
-1i sabov ei ts
equi librium v alueof( ( _0+( _1X_ t-1).Si nce( _2i sexpect edt obenegat i
v e,t het erm( _2u_ t-1i s
negat iveand,t heref ore,ΔY_ twi llbenegat ivet or estor etheequi li
br i
um.Thati s,if Y_ tisabov e
i
tsequi libr ium v alue,i twi llstar tfallingi nt henextper i
odt ocor rectt heequi l
ibr i
um er ror ;hence
thenameECM.Byt hesamet oken,i fu_ t-1i snegat i
v e( i
.e.,Y_ ti sbel ow i tsequi l
ibrium v alue) ,
(_2u_ t-
1 wi ll beposi t
ive, whi chwi llcauseΔY_ ttobeposi t
ive, leadi ngY_ ttor i
sei nper iodt .Thus,
theabsol ut ev alueof ( _ 2deci deshowqui cklytheequi li
brium i sr est ored.
Inpr act ice, weest imat e u_ t-1by :
u_t-1=Y_ t-(_1- (_2X_ t
-(_3t ,Not et hatt heer r
orcor recti
oncoef fi
cient( _2i sexpect edt obenegat ive.
Exampl e: ΔY_ t
=0. 061+0. 29ΔX_ t-
0. 122u_ t-1, R^ 2=0.1658, d=2. 15
t=9. 6753 6. 2282 (-
3.8461)
Stat istical ly, theECM t er mi ssi gnificant , suggest i
ngt hat Y_ t adj ust st o X_ t wi t
hal ag; only
about12per centoft hedi screpancybet weenl ong- t
er m andshor t-term.
CHAPTERTHREE
3.INTRODUCTI ONTOSI MULTANEOUSEQUATI ONMODELS
3.1.TheNat ur eofSi mul t
aneousEquat ionModel s
In many si tuat i
ons, such a one-way or uni directional cause-and-ef fect
relationshi p i s notmeani ngf ul. Thi soccur si fYi sdet er minedbyt heX’ s,andsomeoft heX’ s
are, int ur n, det ermi nedbyY.I nshor t, t
herei sat woway ,orsimul t
aneous, relat i
onshi pbet weenY
and( someof )t heX’ s, whi chmakest hedi sti
nctionbet weendependentandexpl anat or yv ariables
ofdubi ousv alue. I ti sbet tert ol umpt oget herasetofv ari
abl est hatcanbedet ermi ned
simul taneousl ybyt he r emai ning setofv ariables—pr eciselywhati sdone i n si mul taneous-
equat i
onmodel s. I nsuchmodel s t hereismor e t han one equat ion—one f or each of
the mut ual ly, or j oint l
y , dependent or endogenousv ariables. Andunl iket hesi ngl e-
equat i
onmodel s,i nt hesi mul taneous-equat i
onmodel sonemaynotest imat e t hepar amet er s
ofasi ngl e equat i
on wi t
hout t aking i nt o account i nfor mat ion pr ov i
dedbyot her
equat i
ons i n t hesy stem.
Exampl e:
Thecl assi cexampl eofsi mul taneouscausal i
tyineconomi csissuppl yanddemand. Bot hpr ices
andquant i
ties adj ustunt ilsuppl yanddemandar einequi l
ibrium. Ashockt odemandor
suppl ycausesbot h pr i
cesandquant iti
est omov e. Asi swel l known,t he pr icePofa
commodi ty andt hequant it
y Q sol d ar e det ermined by t he i nt ersect ion of t he
demand-and-suppl y cur ves f or t hatcommodi ty.
Figur e3. 1. I nter dependenceofPr i
ceandQuant ity
Thus, assumi ng f or si mpl ici ty t hat t he demand-and-suppl y cur v es ar e l inear
andaddi ng t hest ochast ic di stur bance t ermsu1andu2,wemaywr it
e t heempi rical
demand-and-suppl yf unct ions as:
Demandf unct i
on: Q^ d=( _0+( _1P_ it+( _2Y_ t+u_it
, (_
1<0…………………. 3.1
Suppl yf unct i
on: Q^ s=(_ 0+( _1P_ it+u_ 2t , (_1>0………………………. .
……. 3.2
Equi li
br ium condi tion:Q^ d=Q^ s
Wher e? Q_ t^ d=quant it
ydemanded,? Q_ t^
s=quant i
tysuppl iedt=t imeandt he( ’sandβ’ sar et he
par amet er s.
Now i ti snott oodi f
ficultt oseet hatPandQar ejoi ntl
ydependentv ariables. I f,forexampl e,
u1ti n ( 3.1)changes becauseofchanges i n ot herv ari
ables af fect i
ng ? Q_ t^d( such as
i
ncome, weal t h, andt ast es) ,
t hedemandcur vewi ll shi ft upwar d i fu1t i sposi tive and
downwar d i fu_ 1t i snegat ive. Theseshi f
tsar eshowni nFi gur e3. 1. Ast he f igur e
shows, ashi ft i n t he demand cur ve changes bot hPand Q.Si mi l
ar l
y, a change
i
n u2t ( becauseofst ri
kes, weat her ,i mpor t orexpor trestricti
ons, et c.)wi ll shi ft t he
suppl y cur ve,agai n af fecting bot hP and Q.Because oft hi s si mul taneous
dependencebet ween Q and P,u1t and Pt i n ( 3.1)andu2t and Pt i n ( 3. 2)
cannot bei ndependent . Ther ef or e, ar egressi on of Q on P asi n ( 3.1)woul d
violate an i mpor tant assumpt i
on of t he cl assi cal l i
nearr egr ession model , namel y,
the assumpt ion ofno cor relat i
on bet ween t he expl anat or y v ariable(s)and t he
distur bance t er m.
Def initi
onsofSomeConcept s
Thev ariabl esPandQar ecal l
edendogenousv ariabl es becauset hei rv aluesar edet ermi ned
withi nt hesy stem wehav ecr eat ed.
Thei ncomev ar i
abl eYhasav al uet hati sgi vent ous,andwhi chi sdet ermi nedout sidet his
syst em. I ti scal ledanexogenousv ariable.
Predet ermi nedv ar iables areexogenousv ar i
abl es,laggedexogenousv ar i
ablesandl agged
endogenousv ar i
abl es. Pr edet er mi nedv ari
ablesar enon-st ochast icandhencei ndependentof
thedi st ur bancet er ms.
Struct ur almodel s:A st r
uct ur almodeldescr ibest hecompl etest ruct ureoft her elationshi ps
amongt heeconomi cv ari
ables. Structural equationsoft hemodel maybeexpr essedinter
msof
endogenousv ariabl es, exogenousv ariablesanddi sturbances( random v ari
ables).
Reducedf orm oft hemodel : The r educed f orm of ast ructural model i s t he
model i n whi ch t heendogenous v ari
abl es ar eexpr essed a f uncti
on oft he
predet ermi ned v ar i
abl es andthe er r
ort erm only.
Exampl e:Thef ollowi ngsi mpl eKeynesi anmodelofi ncomedet erminationcanbeconsi der
edasa
structur almodel .
C=( +(Y+U, for(>0, 0<( >1………………………………………………. .3.3
Y=C+Z…………………………………………………………………………………3. 4
wher e: C=Consumpt ion expendi t
ur e
Z=non-consumpt i
onexpendi ture
Y=nat ionali
ncome
CandYar eendogenousv ar
iabl
eswhi leZi sexogenousv ari
able.
Reducedf or m oft hemodel :
The r educed f orm of a st ructural model i s t he model i n whi ch t he
endogenous v ar iables ar eexpressedaf unction oft hepredetermi ned v ariabl
es andt he
errorter m onl y
.
Il
l
ust ration: Findt her educedf or
m oft heabov est ructuralmodel .
SinceC andYar eendogenousv ariablesandonl yZi stheexogenousv ari
ables,wehav eto
expr essCandYi nt er msofZ.Todot hissubst it
ut eY=C+Zi ntoequat i
on( 3.4).
C=( +((C+Z) +U
C=( +(C+( Z+U
C-(C=( +(Z+U
C(1- (
)=(+( Z+U
C=( /1-(+(/1- (
Z+1/ 1- (U…………………………………………………………. .3.5

Substi
tutingagain(3.5)i
nto (3.
4)weget;
Y=(/1-
(+1/1-(Z+1/1-
(U…………………………………………………………. .
3.6
Equati
on( 3.5)and(3.6)arecal
ledther
educedf
orm oft
hestr
uct
uralmodeloft
heabov
e.Wecan
writ
ethismor efor
mallyas:
Str
ucturalform ofEquation
Reducedf orm ofEquation

C=(
+(Y+U
C=(
/1-(
+(/
1-(
Z+1/1-
(U
C=(
_01+(_
11Z+v_
11

Y=C+Z
Y=(/
1-(
+1/1-
(Z+1/1-
(U
Y=(_
02+(_
12Z+v_12

Parameters of t he r educed f orm measur e t he t otal ef fect(direct and i ndi

rect)
of a changei nexogenous v ariables ont heendogenous v ari
abl e
3.2.Si
mul tanei
tyBias
Ifoneapplies OLSt o est imat e t he par amet ers of each equat ion di sregar ding ot her
equati
ons of t he model ,t heest i
mat es so obt ained ar e notonl y bi ased butal so
i
nconsistent ;i.e.evenif t he sampl e si zeincreases i ndefinitely,t heest i
mat or
s do
notconverge t othei
rt ruev alues.
The bi as ar ising f r
om appl ication of such pr ocedure ofest imation whi ch t r
eat s
eachequat ion oft hesimul taneous equat ions model ast hough i t wer easi ngl e model
i
sknown as si multaneitybiasorsi multaneousequat i
on bi as. I tisuseful tosee, i
nasi mpl e
model,t hatan expl anatoryv ari
ablet hati s determined simul t
aneousl ywi t
ht he dependent
vari
abl
ei sgener all
ycorrelatedwi tht heer r
ort erm,whi chleadst obiasandi nconsi stencyi nOLS.
The t wo-way causat i
on i n ar elationship l eadst ov i
olation oft he i mportant
assumption oflinearregr
essi
on model,i .
e.onevar
iable canbedependent v ar
iabl
e in
oneoftheequat i
on butbecomesal soexplanat
ory variabl
e i ntheotherequat
ions ofthe
si
multaneous-equation model. Inthi
scaseE[ Xi
Ui] may bedi ff
erent from zer o. To
showsimultanei
tybias,l
et’
sconsi
dert
hefol
lowingsi
mplesi mul
taneousequati
onmodel.

Y=(_ 0+( _ 1X+U …………………. .

3.7
X=(_ 0+( _ 1Y+( _2Z+V
Supposet hatt hef ollowi ngassumpt i
onshol dt r
ue.
EU=0 EV=0
EU^ 2=? (_ u^ 2
EV^ 2=? (_v^ 2
EU_ iU_ j=0 EV_iV_j
=0,
and EU_ iV_ i=0
wher eXandYar eendogenous v ariables andZi sanexogenous v ariable.Thereducedf orm
ofXoft heabov emodel i sobt ai ned bysubst ituti
ng Yi ntheequat i
on ofX.
X=( _
0+( _ 1( (_0+(_ 1X+U) +(_2Z+V.
X=( _0+( _0(_ 1/ 1-(_1(_1+( _2/1-(_1( _1Z+( _
1U+V/ 1-
(_1(_1…………………………………………………. 3.8
Appl ying OLSt ot hef irst equat ion oft he abov e st ructuralmodel wi l
l r esult i n
biased est imat orbecausecovX_ i
U_ i
=E(X_ iU_ j)
≠0.
Now, let ’
spr oofwhet hert hisexpr ession.
Cov XU=E[ {X- EX} {U- EU]
EX-EXU……………………………………………………………………. .
3. 9
E(_0+( _ 0( _1/ 1-(_ 1(_1+( _ 2/ 1-(
_1( _1Z+( _1U+V/ 1- (
_1( _1-(_0+(_0(_1/1-(_
1( _1-(_2/1-
(_1(_1ZU………………. .3.10
Subst itut ingt hev alueofXi nequat ion( 3.8)i ntoequat i
on( 3.9)
EU/ 1-(_ 1( _1( (_0-(_0( _1+( _ 2Z+V- (
_ 0+( _0(_1-(_2Z
EU/ 1-(_ 1( _1( (_1U+V
1/1-(_ 1( _1E( (_1U^ 2+UV)
(_1/1- (_1( _1EU^ 2=(_ 1? (_u^ 2/1- (
_ 1( _
1≠0, since E( UV)=0
Thati s,cov ar i
ancebet weenXandUi snotzer o.Asaconsequence,i fOLSi sappliedt oeach
equat ionoft hemodelsepar at elyt hecoef ficient swi llturnoutt obebi ased. Now, l
et’sexami ne
how t henon-zer oco-v arianceoft heer rort erm andt heexpl anat oryv ar
iabl
ewi l
lleadt o
biasnessi nOLSest i
mat esoft hepar amet ers.
Ifweappl yOLSt ot hef irstequat ionoft heabov est r
uct uralequation( 3.3)
Y=(_ 0+( _ 1X+U, weobt ain
(_i
=_ ^? xy /_^ ?x^ 2=_ ^?xY- Y/ _ ^?
x^ 2=_ ^?xY/_^? x^2, sinceY_ ^
?x/_^?x^2iszero
=_^?x( _ 0+( _ 1X+U/ _^ ?x^ 2=( _0_^ ?x/ _^?x^2+( _1_ ^?xX/ _^? x^2+_^?xU/_^?x^2
But, weknowt hat_ ^?x=0and_ ^? xX/ _^?x^2=1, hence,
(_i
=( _1+_ ^? xU/ _ ^?x^2……………………………………………………………………………………11
Taking t he expect edv alues onbot hsi des;
E((_i)=( _1+E_ ^?xU/ _^? x^ 2. Si nce,wehav eal r
eadypr ovedt hatE(_^?xU) ≠0whi chisthesameas
E(XU) ≠0.Consequent l
y ,whenEXU≠0; E((_i
)≠( _1, that( _1willbebiasedbyt heamountequi valentto
_^?xU/ _^? x^ 2.

3.3.TheIdentificat
ionProbl em
Insimultaneousequat ionmodel s,theProblem ofidentif
icationi
saproblem ofmodel f
ormulat
ion;
i
tdoesnotconcer nwi t
ht heestimationoft hemodel. Theest imati
onofthemodeldependsup
ont heempi ricaldataandt heform oft hemodel . I ft hemodel i s noti n t heproper
stati
sti
cal f or m, itmay t ur
n outt hatthe par amet ers may notuni quely estimated
even t hough adequat e andr elevantdataareav ail
abl e. Inalanguageofeconomet ri
cs,a
modeli ssaidt obei dentifi
edonlywheni tisinuniquest ati
sti
calf
ormt oenabl eust oobt ai
n
uniqueestimat esofitspar ametersfrom thesampl edata.
Byt heidentificati
onpr oblem wemeanwhet hernumer i
calestimatesoft hepar ametersofa
struct
uralequat ioncanbeobt ai
nedf rom theesti
mat edr educed-f
orm coef
fi
cients.Ifthi
scanbe
done, wesayt hatt hepar t i
cul arequat ioni sident if
ied.I ft hiscannotbedone, thenwesayt hatt he
equat ionunderconsi der ationi suni dent ifi
ed,orunderi dent i
fi
ed.Ani dent ifi
edequat ionmaybe
eitherexact ly( orf ull
yorj ust )i dent if
iedorov eri dent i
fied.
Itissai dt obeexact lyi dent i
fiedi funi quenumer icalv aluesoft hest r
uctur alpar amet er scanbe
obt ained.I ti ssai dt obeov eri dent ifiedi fmor et hanonenumer icalv al
uecanbeobt ainedf or
someoft hepar amet er soft hest ruct ur alequat ions.Theci r
cumst ancesunderwhi cheachof
thesecasesoccur swi l
lbeshowni nt hef ollowi ngdi scussi on.Thei dentifi
cat i
onpr obl em ar ises
becausedi ff er entset sofst ruct ur alcoef f i
cient smaybecompat iblewi ththesamesetofdat a.To
putt hemat terdi fferent ly,agi v enr educed- form equat ionmaybecompat i
blewi thdi fferent
struct uralequat ionsordi ffer enthy pot heses( model s),and i tmaybedi ffi
cultt ot ellwhi ch
par ticul arhy pot hesi s( model )wear ei nv est igat ing.I nt her emai nderoft hissect ionweconsi der
sev er al exampl est oshowt henat ur eoft hei dent if
icat ionpr oblem.
Consi dert he demand- and- suppl y model ,t oget herwi tht he mar ket-clear i
ng,orequi l
ibrium,
condi tiont hatdemandi sequal tosuppl y.
Q^ d=α0+α1Pt+u1t ……………………………. .……………………………………………………………………………………3. 12
Qs=β0+β1Pt+u2t……………………………………………………………………………. 3.13
Byt heequi l
ibr ium condi tion, weobt ain:
α0+α1Pt+u1t=β0+β1Pt+u2t
Sol vingt hisequat i
on, weobt aint heequi librium pr ice
P_ t=( _0+v _t………………………………………………………………………………. 3. 14
wher e ( _
0=( _ 0-( _0/(_ 1-(_1, v _t=u_ 2t-u_ 1t /(_1- (
_ 1
Subst itutingPtf rom Eq.( 3. 14)i nt oEq.( 3. 13)or( 3. 12) ,weobt aint hef oll
owi ngequi li
br i
um
quant ity:
Q_ t=( _1+w_ t…………………………………………………………………………………3. 15
Wher e( _1=( _ 1( _0- (_0(_ 1/(_1- (_1, w_ 1=( _1u_ 2t-(_1u_ 1t /(_1-(_1
Equat ions( 3.14)and( 3.15)ar er educed- for m equat i
ons.Now ourdemandandsuppl ymodel
cont ai nsf ourst ruct ur alcoef fi
cient s( _ 0,( _1,(_ 0and( _ 1.Theser educed- form coef ficient scont ain
allf ourst ruct ur alpar amet er s,butt her ei snowayi nwhi cht hef ourst r
uct uralunknownscanbe
est i
mat edf rom onl yt wor educed- form coef ficient s.
Ther ei sanal ter nat ivewayofl ooki ngatt hei dent i
ficat i
onpr obl
em.Supposewemul tiplyEq.( 3. 12)
byλ( 0≤λ≤1)andEq.( 3. 13)by1−λt oobt aint hef ollowi ngequat ions:
(Q^ d=( α0+( α1Pt+( u1t
(1-()Qs=( 1- ()β0+( 1-()β1Pt+( 1-()u2t
Addi ngt heset woequat i
onsgi vest hef ollowi ngl inearcombi nat ionoft heor i
ginaldemandand
suppl yequat ions:
Q_ t=( _0+( _1P_ t+w_ t,…. ..……………………………………………………………………. 3.15
Wher e, (_ 0=( (_0+( 1-( )(_0
(_1=( ( _1+(1- ()( _1
w_ t=( u_ 1t+( 1- ( )u_ 2t
The“ mongr el,”equat ion( 3.15)i sobser vational l
yi ndi stingui shabl ef rom eit hersuppl yordemand
equat ionbecauset heyi nv olv et her egr essi onofQandP.Foranequat iont obei dent ified, thati s,
fori tspar amet er st obeest i
mat ed, itmustbeshownt hatt hegi vensetofdat awi llnotpr oducea
struct uralequat iont hatl ookssi mi l
ari nappear ancet ot heonei nwhi chwear ei nter est ed.I fwe
setoutt oest imat et hedemandf unct ion,wemustshow t hatt hegi vendat aar enotconsi stent
witht hesuppl yf unct ionorsomemongr el equat ion.
A f unct i
on ( an equat ion) bel ongi ng t o a sy stem ofsi mul taneous equat ionsi s
i
dent ifiedi fi thasauni que st at i
st i
calf or m, i.
e. i fther ei sno ot herequat ionint hesy st em, or
formedbyal gebr aicmani pul ationsoft heot herequat ionsoft hesy stem,cont ainst hesame
variabl esast hef unct i
on( equat ion)i nquest ion.
Ident ificationpr obl emsdonotj ustar iseonl yont woequat ion-model s. Usi ng t heabov e
procedur e, we can check i dent ificat ion pr obl ems easi l
y i f we hav e t wo or t hree
equat ions i n a gi vensi mul taneous equat i
on model . Howev er,for‘ n’equat ions
simul taneousequat i
onmodel , suchapr ocedur ei sv er ycumber some. I ngener al foranynumber
ofequat i
onsi nagi ven si mul taneousequat ion, we hav e t wocondi tions t hat need t o
be sat isfi
ed t o say t hat t he model i s i n gener alident i
f i
ed ornot .Int hef ollowi ng
sect
ionwewi
l
lseet
hef
ormal condi
ti
onsf
ori
dent
if
icat
ion.

3.4.OrderandRankCondi ti
onsofI dent i
fi
cation(wi t
houtpr oof )
Theor dercondi t
ionfori dentifi
cati
on
Thiscondi ti
oni sbasedonacount i
ngr uleoft hev ariablesi ncludedandexcl udedf rom t he
parti
cularequat ion.Iti sanecessar ybutnotsuf fi
cientcondi ti
onf ort hei denti
fi
cat
ionofan
equati
on.Theor dercondi ti
onmaybest atedasf ol l
ows.Foranequat i
ont obei denti
fi
edt hetot al
numberofv ariables(endogenousandexogenous)excl udedf rom i tmustbeequalt oorgr eater
thanthenumberofendogenousv ar i
ablesint hemodell essone.Gi vent hatinacompl etemodel
thenumberofendogenousv ari
ablesisequalt othenumberofequat ionsoft hemodel ,theor der
conditi
onfori dentif
icationi ssomet i
messt atedint hef oll
owi ngequi valentform.Foranequat ion
tobei dentifi
edt het otalnumberofv ari
ablesexcl udedf r
om i tbuti ncludedi notherequat i
ons
mustbeatl eastasgr eatast henumberofequat i
onsoft hesy stem lessone.
Let:G=t otalnumberofequat i
ons( =totalnumberofendogenousv ariables)
K=numberoft ot alvari
ablesi nthemodel (endogenousandpr edetermined)
M=numberofv ari
ables,endogenousandexogenous, i
ncludedi napar ti
cul
arequation.
Thent heordercondi tionf oridenti
ficati
onmaybesy mbolicallyexpr essedas:

Forexampl e,i fasy st em cont ains10equat i

onswi th15v ar
iables,t enendogenousandf ive
exogenous,anequat ioncont aining11v ariablesi snoti dentifi
ed,whi leanot hercont aini
ng5
variabl esi sident i
fied.
Fort hef i
rstequat i
onwehav e
G=10, K=15, M=11
Ordercondi tion:
(K-M)( G- 1(( 15-11=4)<( 10- 1=9) ;thatis, bytheor dercondi t
ioniti snotsat i
sfi
ed.
Theor dercondi ti
onf ori dent if
icat i
oni snecessar yf orar elat
iont obei dentified,buti tisnot
suf fi
ci ent ,thati s,itmaybef ulfi
l
ledi nanypar ti
cularequat i
onandy ett her elat i
onmaynotbe
i
dent ified.
Ther ankcondi t
ionf ori dent ifi
cation
Ther ankcondi ti
onst at est hat:inasy stem ofGequat i
onsanypar ti
cularequat ioni sidentif
iedif
andonl yi fitispossi blet oconst ructatl eastonenon- zerodet erminantofor der( G-1)from t he
coef f i
ci entsoft hev ariabl esexcl udedf rom t hatpar t
icularequat ionbutcont ainedi nt heot her
equat ionsoft hemodel .Thepr act i
calst epsf ort r
acingt hei denti
fiabl yofanequat i
onofa
struct ur al model maybeout li
nedasf ol l
ows.
Firstly .Wr i
tet hepar amet ersofal lt heequat ionsoft hemodeli nasepar atetabl e,notingthatt he
par amet erofav ariabl eexcl udedf rom anequat ioni sequal tozer o.
Forexampl eletast ruct ur al model be:
y_1=3y _2-2x_ 1+x_ 2+u_ 1
y_2=y _3+x_ 3+u_ 2
y_3=y _1- y_2+2x_ 3+u_ 3
wher et hey ’
sar et heendogenousv ari
abl esandt hex’ sar ethepr edet erminedv ari
ables.Thi s
model mayber ewr itteni nt hef orm
y_1- 3y _2- 0y_3+2x_ 1- x_ 2+0x_ 3=u_ 1
0y_ 1+y _2- y_
3+0x_ 1+0x_ 2-x_ 3=u_ 2
-y_1+y _2+y _
30x_ 1+0x_ 2- 2x_ 3=u_ 3
Ignor ingt her andom di st urbancet het ableoft hepar amet ersoft hemodel i
sasf ol l
ows:

Secondl
y. St r
ikeoutther ow ofcoef
fi
cientsoftheequat i
onwhichi
sbei ngexaminedf
or
i
denti
fi
cati
on.Forexampl
e,ifwewanttoexami net
heidenti
fi
abil
it
yoft
hesecondequat
ionoft
he
modelwestri
keoutthesecondrowofthetableofcoef
fi
cient
s.
Thirdl
y. St rike out t he columns i n which a non- zero coef f
ici
entof the equation
beingexami ned appear s. By del eti
ng the r elev
ant r ow and col umns we ar e
l
eft wi t
h t hecoeffi
cient
sofv ari
abl
es noti ncluded i nt hepar t
icul
ar equation,but
containedi nt heotherequati
ons oft he model .For exampl e, i fwe ar eexami
ning f or
i
dent i
ficat
ion t hesecondequati
on oft hesystem,we wi llstr
ikeoutthesecond,t
hir
d and
thesixthcolumnsoft heabove table,
thusobt
aining t hefol
lowingtabl
es.

Fourt
hly.Formthedeter
minant
(s)oforder(G-
1)andexami net
hei
rval
ue.I
fatleastoneoft hese
deter
minantsisnon-
zero,t
heequati
onisident
if
ied.I
fallt
hedet
ermi
nantsoforder(G-1)ar
ezer o,
theequationi
sunderi
denti
fi
ed.
I
n t heabov e exampleofexplor
ati
onoftheidenti
fi
abil
i
tyoft
hesecondstr
ucturalequat
ionwe
have threedeter
minantsofor
der(G-1)=3-
1=2.Theyare:

(thesy mbol Δst ands f or ‘ det erminant ’) Weseet hatwecanf or mt wonon- zer odet ermi nants
ofor derG- 1=3- 1=2; hencet hesecondequat i
onofoursy stem i si dent if
ied.
Fifthl y
.Toseewhet hert heequat i
oni sexact lyi dent
ifi
edorov er i
dent ifi
ed weuset heor der
condi tion( K-M) ((G-1) .
Wi tht hi scr i
terion,iftheequal it
ysi gni ssat isfied,thatisi f(K- M)=( G-1),the
equat ioni sexact l
yi dent i
fied.I ft heinequal i
tysi gnhol ds, thati s, i f
(K- M) >( G-1) ,t
heequat i
oni sov er i
dent if
ied.
Int hecaseoft he secondequat i
onwehav e:
G=3 K=6 M=3
And t hecount ingr ule
(K- M) ((G- 1)gives( 6-3)>( 3-1) .Ther eforet hesecondequat ionoft hemodel i
sov eridentified.
Est imat ionofSi mul taneousEquat ionsModel s
1. I ndi rectLeastSquar es( ILS)Met hod
Int hi s met hod, wef irst obt ain t heest imat es oft her educed f orm par amet ers by
appl yi
ng OLSt ot her educedf orm equat i
ons andt heni ndirect ly gett heest i
mat es oft he
par amet ers oft hest ruct ur almodel . Thi smet hodisappl iedt oexact l
yident i
fiedequat ions.
Steps:
Obt aint her educedf orm equat i
ons( thati s,expr esst heendogenousv ari
abl esi nt ermsof
predet er mi nedv ariables) .
Appl yOLSt ot her educedf or m equat i
onsi ndi viduall
y. OLSwi ll y i
eld consi stentest imat es
oft her educedf orm par amet er s ( since each equat ion i nv olves onl y nonst ochast i
c
(pr edet er mi ned) v ari
abl es t hatappearas‘ independent ’v ariabl es).
Obt ain ( orr ecov erback)t heest i
mat es oft heor i
ginal st ruct ural coef f
icient susi ng t he
est imat es i nst ep( b).
2. Two- StageLeastSquar es( 2SLS)Met hod
The2SLSpr ocedur ei sgener allyappl icablef orest imat ionofov er-identif
iedequat i
onsasi t
prov i
desuni que est i
mat ors.
Steps:
Est imat e t her educedf or m equat ions byOLSandobt ai n t hepr edictedY_ i.
Repl ace t her i
ght handsi deendogenous v ariabl
es i nt hest ruct ural equat i
ons by
thecor respondi ng Y_ iandest imat e t hem byOLS.
Consi dert hef ollowi ngsi mul taneousequat i
onsmodel :
Y_ 1=a_ 1+b_ 1Y_ 2+c_ 1z_1+c_ 2z_ 2+u_ 1………………………………. …………. .(a)
Y_ 2=a_ 2+b_ 2Y_ 1+c_ 3z_3+u_ 2……………………………………………………. (b)
Wher e Y_ 1andY_ 2ar eendogenouswhi lez_ 1, z_2andz_ 3ar epr edet ermined.
The2-SLSpr ocedur eofest imat ionofequat ion( b)(whichi sov er-ident ifi
ed)is:
Wef irst est imat e t her educedf or m equat i
ons byOLS;t hati s,wer egress Y1onz_ 1,
z_ 2andz_ 3usi ng OLSandobt ainY_ i.Wet hen r eplaceY1byY_ iandest i
mat e equat i
on( b)
by OLS, thati s, weappl yOLSt o:Y_ 2=a_ 2+b_ 2Y_ 1+c_3z_ 3+u_ 2

CHAPTERFOUR
4. I
NTRODUCTI
ONTOPANELDATAREGRESSI
ONMODELS
4.1.I nt roduct ion
Inpaneldat at hesamecr oss- sect i
onaluni t( sayaf ami l
yoraf i
rm orast ate)i ssur vey edov er
ti
me.I nshor t,panel dat ahav espaceaswel last imedi mensi ons.
Hy pot het i
cal exampl es:
Dat aon200Et hiopi anSomal i r egionalst at eschooli n2004andagai n i n2005,f or400
obser vationst ot al.
Dat aon9r egi onalst atesofEt hiopi a,eachst atei sobser vedi n5y ear s,f orat ot alof45
obser vations.
Dat aon1000i ndi v idual s, i
nf ourdi fferentmont hs, for4000obser vat i
onst ot al.
Ther e ar e ot her names f or panel dat a, such as pool ed dat a ( pool ing of t ime
ser i
es and cr osssect i
onal obser v ations) , combi nat ion of t i
me ser i
esand cr oss-
sect ion dat a ( cr oss- sect ionalt ime- ser i
esdat a), andmi cropanel dat a, l ongi tudi nal
dat a ( ast udy ov ert ime ofav ar iable orgr oupofsubj ect s) .
WhyShoul dWeUsePanel Dat a?Thei rBenef it
sandLi mitat ions
Bal tagi ( 2005) l ist sev er al benef its f rom usi ng panel dat a.Thesei ncl udet he
followi ng.
Cont rol l
ingf ori ndi v idualhet er ogenei ty. Paneldat aal l
ows y out ocont rol f orv ari
abl es y ou
cannot obser veormeasur e l ike cul tural f act or s ordi fference i n busi ness pr act i
ces
acr osscompani es; orv ariabl es t hatchange ov ert ime butnotacr ossent i
ties ( i.e.
nat ional pol i
cies,f eder al r egul at i
ons, i nt er national agr eement s, et c.). Thi si s,i t
account sf ori ndi v idualhet erogenei ty. Ti me-ser iesandcr oss-sect i
onst udiesnotcont rolling
thishet erogenei tyr unt her iskofobt ainingbi asedr esul ts.
Panel dat a gi v e mor e i nfor mat ive dat a, mor e v ariability,l ess col l
inear it
y among
thev ariables, mor e degr ees off r
eedom andmor eef ficiency . Ti me-ser iesst udi esar e
plaguedwi thmul ticol linear ity.
Paneldat aar ebet terabl et ost udyt hedy nami csofadj ust ment . Cr oss-sect i
onaldi st ri
but i
ons
thatl ookr elat i
vel yst abl ehi deamul t i
tudeofchanges.
Paneldat aar ebet terabl et oi dent ifyandmeasur eef fectst hatar esi mpl ynotdet ect abl ei npur e
cross-sect ionorpur et ime-ser i
esdat a.
Paneldat amodel s al low ust oconst ruct andt estmor ecompl i
cat ed behav ioral model s
than pur ely cr oss-sect ion ort ime-ser i
es dat a.Forexampl e,techni calef ficiencyi sbet ter
studi edandmodel edwi thpanel s.
Mi cr o panel dat a gat her ed on i ndividual s,f irms and househol ds may be mor e
accur at elymeasur ed t han si mi lar v ariables measur ed at t he macr o l evel. Bi ases
resul t i
ngf rom aggr egat ionov erf i
rmsori ndi vi
dual smayber educedorel i
mi nated.
Limi t ationsofpanel dat ai nclude:
Desi gnanddat acol lect ionpr obl ems. These i ncl ude pr obl ems of cov erage ( i
ncompl ete
account of t he popul at i
on of inter est ), nonr esponse ( due t o l ack of
cooper ation of t her espondentorbecauseofi nter v
iewer er ror ),recal l ( r
espondent not
remember ing cor r ect ly),frequency of i nterv iewi ng, i nt erview spaci ng, r ef erence per i
od,
the use ofboundi ng andt ime-i n-sampl e bi as.
Di st ortionsofmeasur ementer rors. Measur ementer ror s may ar i
se because of f aul ty
responsesduet ouncl ear quest ions, memor y er ror s,del i
ber at e di st orti
on ofr esponses
(e.g.pr est ige bi as) ,inappr opr iate i nfor mant s, mi srecor ding ofr esponsesandi nter viewer
effect s.
Select ivitypr obl ems. Thesei ncl ude:
Self-sel ect ivity. Peopl echoosenott owor kbecauset her eser v ationwagei shi ghert hant he
offer edwage. I nt hi scaseweobser v et hechar act eristicsoft hesei ndiv i
dual sbutnott hei rwage.
Sinceonl yt hei rwagei smi ssi ng,t hesampl ei scensor ed.Howev er ,i
fwedonotobser veal ldat a
ont hesepeopl et hi swoul dbeat runcat edsampl e.
Nonr esponse.Thi scanoccuratt hei niti
alwav eoft hepanelduet or ef usalt opar ticipat e,nobody
athome,unt racedsampl euni t,andot herr easons.I tem ( orpar tial) nonr esponseoccur swhen
oneormor e quest ions ar el ef t unanswer ed orar ef ound not t opr ov ide ausef ul
response.
Attrition. Whi le nonr esponse occur sal so i n cr oss-sect ion st udies, i t i samor e
seri
ouspr oblem i n panel s because subsequent wav es of t he panel ar e sti
ll
subject to nonr esponse. Respondent smaydi e,ormov e,orfindthatt hecostofr espondi
ng
i
shi gh.
Shortt i
me-ser iesdi mensi on. Ty picalmi cropanel sinvolveannualdat acov eringashor tt
ime
spanf oreachi ndividual. Thi s means t hatasy mpt oti
c ar gument s r el
y cr ucially onthe
number of i ndividuals t ending t oinfi
nity.I ncreasing t he t ime span of t he panel
i
s not wi t
hout cost ei t
her.I n f act,t his i ncreases t he chances of at triti
on and
i
ncreases t he comput ati
onal di ffi
culty f orl i
mi t
ed dependent v ari
abl e paneldat a
model s.
Cross-sectiondependence.Macr opanel soncount ri
es orr egions wi th l ong t i
me ser i
es
thatdonotaccount f orcross-country dependence mayl eadtomi sleading i nference.
Notationforpanel dat a
Adoubl esubscr i
pti susedt odisti
nguishent i
ti
es( states,family,countr
y ,indi
v i
dual s,et c.
) and
ti
me per i
ods.
Considerthef oll
owi ngsi mplepanel dataregressionmodel :
Y_it
=(_0+(_1X_ 1it
+(_2Z_ t+u_it
…………………………………………4. 1
i =1, …,n,T=1, …,T
Wher e i =ent i
ty ( state),n=number ofent i
ti
es, soi =1,…,n,t=t i
me per iod( year ,month,
quar ter,
etc.),
T=number oft ime per i
ods, sot hat t=1, …,T
Panel dat awith k r egressors:
Y_it
=(_0+(_1X_ 1it
+(_2X_ 2it+…(_kX_ki
t+u_it
…………………………………………4. 2

Estimat ion ofPanel Dat aRegr essi onModel

TheFi xedEf fect s( Entity/Time Fi xed) Appr oach
Youmayappl yent i
tyf i
xedef fectsr egr essionwheny ouwantt ocont r
olforomi t
tedv ariablest hat
differamongpanel sbutar econst antov ert i
me.Ont heot herhand,i ft herear eunobser v ed
effect st hatv aryacr osst imer at hert hanacr osspanel s,weappl yt imef ixedef fectsr egressi on
model. Usef ixed-ef fects ( FE)whenev er y ou ar eonl y i nt
erest ed i nanal yzing t heimpact
of v ar i
abl es t hatv ar y ov er t i
me. FEexpl orest her elationshi pbet weenpr edictorand
outcomev ar i
abl eswi t
hinanent ity( count ry,per son,company ,et c.)
.Eachent i
t yhasi t
sown
i
ndi vidualchar act eristi
cs t hatmayormaynot i nfluence t hepr edict or v ariables ( for
exampl e bei ng amal e orf emal e coul d i nf luencet he opi nion t owar dcer tain i ssue
ort he pol itical sy stem ofapar ti
cul ar count ry coul d hav e someef fect on t r
adeor
GDPort hebusi ness pr act i
ces ofacompany may i nfluence i t
sst ock pr i
ce).Whenusi ng
FEweassume t hat somet hing wi thin t hei ndi v i
dual mayi mpact or bi ast hepr edictor
orout come v ar i
abl es andweneedt ocont rol f ort his. Thi sist her ati
onal ebehi ndt he
assumpt ionoft hecor relationbet weenent it
y ’
ser rort erm andpr edictorv ar i
ables. FEr emov e
theef f
ect oft hoset ime-i nv ariant char act eristi
cs f rom t he pr edictor v ar i
ables sowe
can assess t he pr edictor s’ netef fect.
Anot her i mpor tant assumpt i
on of t he FE model i s t hat t hose t ime-i nvariant
char act eristics ar eunique t ot hei ndi vi
dual and shoul d not becor related wi th ot her
i
ndi vidual char act eri
stics. Eachent i
ty i s di ffer ent t herefor e t he ent it
y ’
s er ror t erm
and t he const ant ( whi ch capt ures i ndiv i
dualchar acteristics) shoul d notbecor rel
at ed
with t heot her s.I ftheer rort ermsar ecor r
elat edt henFEi snosui tablesincei nferencesmaynot
becor rectandy ouneedt omodel thatr elationshi p( pr obabl yusingr andom -ef f
ect s.
Entity-demeanedOLSRegr essi on
Thinkoft hef ollowi ngt wov ar i
abl espanel regr essionmodel i nf ixedef fectf orm:
Y_it=(_i+( _1X_ 1it+u_ i
t……………………………………………………. …………4. 3
(_iiscal led an“ entit
y f ixed ef fect” or“ ent ity ef fect” –i tist heconst ant ( fi
xed)
effect ofbei ng i nent ity i .
Thest ateav er
agessat isfy :
1/T_ t=1^ T? Y_ it=(_i+( _11/T_ t=1^ T? X_ i
t+1/ T_^ ?u_ i
t
Dev iation f rom ent it
y av erages:
Y_it-1/T_ t=1^ T? Y_it=( (
_i-
(_i)+(_1X_ 1it-1/T_ t
=1^ T? X_it+( u_ i
t-1/T_^? u_ i
t)
(Y_i
t=(
_1X_i
t+u_
it, wher
e: X_
it=(
X_1i
t-1/
T_t=1^T?
X_it),
u_i
t=(
u_i
t-
1/T_
^?u_
it)
Thenweappl
yOLS Y_ i
t=(_
1X_i
t+u_itt
oestimate(_1.

3.4.2.TheRandom Ef fects( RE)Appr oach

Ifyoubel i
ev et hatsomeomi tted v ariables may beconst ant ov ert i
me butv ary among
panel s, andot hersmaybef ixed amongpanel sbutv aryov ertime, theny oucanappl yrandom
effects r egr essi onmodel .Random ef fectsassumet hatt heent i
ty’
ser rortermi snotcor r
elated
witht hepr edict orswhi chal lowsf ort i
me-i nvar i
antv ariablest o playar oleasexpl anator y
variables.I nr andom -ef f
ect s y ouneed t o speci f
y t hose i ndiv i
dual char acteri
stics t hat
may or may not i nfl
uence t he pr edictorv ar i
ables.
Thebasi ci deaofr andom ef fect smodel istost ar twi th( 4.3):
Y_it=(_i+(_1X_ i
t+u_ it……………………………………………………. …………4. 3a
Insteadoft reating ( _i asf i
xed,weassumet hati tisar andom v ari
ablewi thameanv alueof
(( nosubscr ipti her e).Andt hei nterceptv aluef orani ndi vi
dual ent i
tycanbeexpr essedas:
(_i
=(+( _i
t,i=1……n…………………………………………………………………4. 4
Wher e ( _i i
sar andom er rort er m wi thamean v alueofzer oandv ari
anceof ?(_(^
2.
What wear eessent i
all
y say ing i s t hat t he ent i
ti
es i ncluded i n oursampl e ar ea
drawi ng f rom amuch l arger uni verse ofsuchpopul ation andt hat t heyhav ea
common meanv alue f ort hei ntercept( =()andt hei ndiv i
dual di ff
erences i n t hei ntercept
values ofeachent ity ar er efl
ect ed i n t heer rorter m( _i
.
Subst ituti
ng ( 4.4)i nto ( 4.3a) ,weget :
Y_it=(+( _
1X_ it+( _i
t+u_ it……………………………………………………. …………4. 5
Y_it=(+( _
1X_ it+w_ i
t, wher e:w_ it
=(_ i
t+u_ it
In r andom ef f
ect s model ( REM)orer rorcomponent model ( ECM) i tisassumedt hat
thei nterceptof ani ndiv i
dual uni t i s ar andom dr awi ng f rom amuch l arger
popul ation wi th aconst ant meanv al
ue. Thei ndiv i
duali nterceptist henexpr essedasa
dev i
at i
onf rom t hisconst antmeanv alue. Oneadv ant ageofECM ov erFEM i sthati ti s
economi cali ndegr eesoff reedom,aswedonothav et oest imat eNcr oss-sect ionalintercept s.
Weneedonl yt oest imat ethemeanv alueoft heinter ceptandi tsvariance. ECM i s appr opr i
ate
i
n si tuations wher e t he ( r
andom) i ntercept of each cr osssect ional uni t i s
uncor related wi th t heregr essor s.

3.
5.ChoosingBetweenFi
xedandRandom Effect
s
Ifyouaren'
texact
lysurewhichmodels,f
ixedef
fect
sorrandom ef
fect
s,youshoul
duse,youcan
do at estcall
ed Hausmant est
.To runaHausmant estin St ata,youneed t
o savethe
coeff
ici
entsfrom eachoft hemodels anduset hestor
edresul
ts i nthet
est
.Tost or
ethe
coeff
ici
ents,
youcan
use"esti
mates store" command.

Thehausmant esttestst henul

l hy pot
hesis t hat t hecoeff
ici
ents est i
mat ed byt he
eff
ici
ent r andom ef f
ect s esti
mat or ar ethe same ast he ones est imated byt he
consistent f i
xed ef fects esti
mat or
.I ftheyare,t
henitissaf
etouser andom eff
ects. I
fyou
getast ati
stical
lysigni
ficantP-value,however,youshoul
dusef i
xedeffects.I
nthisexample,t
he
P-v alueisst atist
ical
l
ysi gni
fi
cant. Therefore,fi
xedeff
ectswouldbemor eappropri
ateinthi
s
case.

Fi
xedEff
ect Model
Random Eff
ect Model

Funct
ional form
Y_i
t=(
(+(_i
t)
+(_1X_i
t+u_
it
Y_i
t=(
+(_1X_i
t+((
_it
+u_i
t)

I
nter
cept
s
Vary
ing acr
ossgr
oups and/
ort
imes
Const
ant

Err
orv ar
iances
Constant
Varyi
ng acr ossgr
oupsand/
ort
imes

Slopes
Constant
Constant

Est
imati
on
LSDV,wi
thi
n ef
fect met
hod
GLS,
FGLS

Hypothesi
s test
I
ncremental Ftest
Br
eusch-Pagan LM t est

OtherTests/Diagnosti
cs
Testi
ngf orTime- fi
xedEff
ects
I
tisajoint t esttoseeift hedummi
es f
oral
lyear
sar
eequal t
o0,
ift
hey ar
ethen no
ti
me f ixed ef fectsar
eneeded.

Wef ail
edt orej
ectthenullthatallyears’coeffi
cientsarejoi
ntlyequalt
ozer other
eforenot i
me
fi
xedef fectsareneeded.
Testi
ngf orRandom Eff
ects:
Breusch-PaganLagr angeMul t
ipl
ier(LM)
TheLM t esthelpsyoudecidebet weenar andom ef fect
sregressionandasimpleOLSr egression.
Thenul lhy pot
hesi
sintheLM t estisthatvariancesacrossentiti
esarezero.Thi
sisnosi gnif
icant
di
fferenceacr ossunit
s(i.
e.nopanel effect)
.

Her ewer ej
ectt henul landconcl udet hat r andom ef fectsisappr opriate. Thisi s,thereis
evidenceofsi gni fi
cantdif f
erencesacr osscount ri
es,thereforeyoushoul dr unarandom ef f
ects
regr ession.
Test ingf orCr oss- Secti
onal Dependence/ Cont empor aneousCor relati
on:
Usi ngBr eusch- PaganLM TestofI ndependence
Accor dingt oBal tagi,cross-sect i
onaldependencei sapr obl em inmacr opanel swithl ongt i
me
series( ov er20-30y ears).Thisi snotmuch of apr oblem i nmi cr
o panel s ( f
ew y earsand
l
ar genumber ofcases) .The nul l hy pothesis i ntheB-P/ LM testofi ndependence i sthat
residual s acr ossent it
ies ar enot cor rel
ated.
Thecommandt or unthis t esti sxttest2 ( runitaf t
er xt reg, fe):
xtregf atali
ty beer tax,fe
xt
test2
Test ingf orCr oss- Secti
onal Dependence/ Cont empor aneousCor relati
on:
Usi ngPesar an CDTest
Pasar an CD ( cross-sect i
onaldependence)t esti sused t otestwhet hert he residuals are
cor relatedacr ossent i
ti
es* . Cr oss-sect i
onal dependence can l ead t o bias i n t ests
resul ts ( also cal l
edcont empor aneous cor r
elati
on). Thenul l hypothesisisthatresidualsare
notcor rel ated.
Thecommandf ort hetestisxt csd, youhav et oinstal
li tt
y pi
ng ssci nst
al l xt
csd
xtreg f atalit
y beer tax,f e
xtcsd, pesar an/ xtcsd, f r
ees/ xtcsd, f rees

Si
nce t
her
eis cr
oss-sect
ional dependence i
nourmodel
,iti
ssuggest
ed t
ouseDr
iscol
l
andKr
aayst
andar
d er
ror
s.

Atestforhet
eroscedast
ici
tyi
sav
ail
abl
efort
hef
ixed- ef
fect
smodelusi
ngt
hecommand
xt
test
3.

Thenul lishomoscedast i
city(orconst antv ar
iance). Abov ewer ejectthenul landconcl ude
heteroscedast
ici
ty.Seri
alcorrelati
ont estsapplyt omacr opanelswithlongt i
meser i
es(ov er20-
30y ears).Notapr obl
em inmi cr
opanel s(withv eryfew year
s).Serialcorrel
ati
oncausest he
standarderror
soft hecoeff
icientstobesmal l
erthantheyactual
lyareandhi gherR-squared.
ALagr ange-Mult
ipliert
estf
orser i
alcorrelat
ionisav ai
labl
eusingthecommandxt seri
al.

Thenullisnoser i
alcorrelation.Abov ewer ej
ectt henull
andconcl udethedat adohav efirstorder
autocorr
elati
on.
Testi
ngf orUnitRoots/Stationari
ty
TheLev i
nLin-Chu ( 2002) ,Harris-Tzavali
s ( 1999),Brei
tung ( 2000;Breitung andDas2005) ,
Im -PesaranShin ( 2003) ,and Fi sher-type ( Choi 2001)t estshave ast henul l hypothesis
thatallt hepanelscont ain auni tr oot.TheHadr i(2000)Lagr angemul ti
plier(LM)testhasas
thenullhypothesisthatal lthepanel sare( t
rend)stati
onary.Thet opoft heout putforeacht est
makesexpl ici
tthenullandal ter
nat i
vehy potheses.Optionsallowy outoi ncludepanel-speci f
ic
means( fi
xedef f
ects)andt i
met rendsinthemodel ofthedat a-generat
ingpr ocess.