ForStudents

SolutionstoOddNumberedEndofChapterExercises

Chapter2
ReviewofProbability

2.1. (a) ProbabilitydistributionfunctionforY

Outcome(numberofheads) Y0 Y1 Y2
Probability 0.25 0.50 0.25

(b) CumulativeprobabilitydistributionfunctionforY

Outcome(numberofheads) Y0 0Y1 1Y2 Y2

Probability 0 0.25 0.75 1.0

(c) Y = E (Y ) (0 0.25) (1 0.50) (2 0.25) 1.00 . F Fq, .

d
 UsingKeyConcept2.3: var(Y ) E (Y ) [ E (Y )] ,
2 2

and
(ui | X i )
sothat
var(Y ) E (Y 2 ) [ E (Y )]2 1.50 (1.00) 2 0.50.

2.3. Forthetwonewrandomvariables W 3 6 X and V 20 7Y , wehave:

E (V ) E (20 7Y ) 20 7 E (Y ) 20 7 078 1454,
(a) E (W ) E (3 6 X ) 3 6 E ( X ) 3 6 070 72
W2 var (3 6 X ) 62 X2 36 0 21
7 56,

(b) V var (20 7Y ) (7) Y 49 0 1716

8 4084

2 2 2

(c) WV cov (3 6 X , 20 7Y ) 6 (7)cov (X , Y ) 42 0084 3528

3528
corr (W , V ) WV 04425
WV 756 84084

2011PearsonEducation,Inc.PublishingasAddisonWesley
2.5. LetXdenotetemperatureinFandYdenotetemperatureinC.RecallthatY0whenX32and
Y100whenX212;thisimplies Y (100/180) ( X 32) or Y 17.78 (5/9) X. UsingKey
Concept2.3,X70oFimpliesthat Y 17.78 (5/9) 70 21.11C, andX7oFimplies
Y (5/9) 7 3.89C.

Usingobviousnotation, C M F ; thus C M F and C M F 2cov( M , F ). This

2 2 2
2.7.
implies
(a) C 40 45 $85, 000 peryear.
cov( M , F )
corr ( M , F )
(b) M F ,sothat cov ( M , F ) M F corr ( M , F ). Thus cov ( M , F )
12 18 0.80 172.80, wheretheunitsaresquaredthousandsofdollarsperyear.

(c) C M F 2cov( M , F ), sothat C 12 18 2 172.80 813.60, and

2 2 2 2 2 2

C 813.60 28.524 thousanddollarsperyear.

(d) FirstyouneedtolookupthecurrentEuro/dollarexchangerateintheWallStreetJournal,the
FederalReservewebpage,orotherfinancialdataoutlet.Supposethatthisexchangerateise
(saye0.80Eurosperdollar);each1dollaristhereforewitheEuros.Themeanistherefore
eC(inunitsofthousandsofEurosperyear),andthestandarddeviationiseC(inunits
ofthousandsofEurosperyear).Thecorrelationisunitfree,andisunchanged.

ValueofY Probability
Distributionof
14 22 30 40 65 X
ValueofX 1 0.02 0.05 0.10 0.03 0.01 0.21
5 0.17 0.15 0.05 0.02 0.01 0.40
8 0.02 0.03 0.15 0.10 0.09 0.39
ProbabilitydistributionofY 0.21 0.23 0.30 0.15 0.11 1.00

(a) Theprobabilitydistributionisgiveninthetableabove.
E (Y ) 14 0.21 22 0.23 30 0.30 40 0.15 65 0.11 30.15
E (Y 2 ) 142 0.21 222 0.23 302 0.30 402 0.15 652 0.11 1127.23
var(Y ) E (Y 2 ) [ E (Y )]2 218.21
Y 14.77

2011PearsonEducation,Inc.PublishingasAddisonWesley
 (b) TheconditionalprobabilityofY|X8isgiveninthetablebelow

ValueofY
14 22 30 40 65
0.02/0.39 0.03/0.39 0.15/0.39 0.10/0.39 0.09/0.39

E (Y | X 8) 14 (0.02/0.39) 22 (0.03/0.39) 30 (0.15/0.39)

40 (0.10/0.39) 65 (0.09/0.39) 39.21
E (Y 2 | X 8) 14 2 (0.02/0.39) 22 2 (0.03/0.39) 30 2 (0.15/0.39)
402 (0.10/0.39) 652 (0.09/0.39) 1778.7
var(Y ) 1778.7 39.212 241.65
Y X 8 15.54

(c) E ( XY ) (1 14 0.02) (1 22 : 0.05) L (8 65 0.09) 171.7

cov( X , Y ) E ( XY ) E ( X ) E (Y ) 171.7 5.33 30.15 11.0
corr( X , Y ) cov( X , Y )/( X Y ) 11.0 / (2.60 14.77) 0.286

2.11. (a) 0.90

(b) 0.05
(c) 0.05

(d) When Y ~ 10 , then

2
Y /10 ~ F10, .

(e) Y Z , where Z ~ N (0,1), thus Pr (Y 1) Pr (1 Z 1) 0.32.

2

(a) E (Y ) Var (Y ) Y 1 0 1; E (W ) Var (W ) W 100 0 100.

2 2 2 2

2.13.
(b)YandWaresymmetricaround0,thusskewnessisequalto0;becausetheirmeaniszero,this
meansthatthethirdmomentiszero.

(c) Thekurtosisofthenormalis3,so 3 E (Y Y ) / Y ;solvingyields E(Y ) 3; asimilar

4 4 4

calculationyieldstheresultsforW.
X 0, sothat S W :
(d) First,conditionon
E ( S | X 0) 0; E ( S 2 | X 0) 100, E ( S 3 |X 0) 0, E ( S 4 | X 0) 3 1002.

Similarly,

E ( S | X 1) 0; E ( S | X 1) 1, E ( S | X 1) 0, E ( S | X 1) 3.
2 3 4

2011PearsonEducation,Inc.PublishingasAddisonWesley
 Fromthelawofiteratedexpectations
E ( S ) E ( S | X 0) Pr (X 0) E ( S | X 1) Pr( X 1) 0
E ( S 2 ) E ( S 2 | X 0) Pr (X 0) E ( S 2 | X 1) Pr( X 1) 100 0.01 1 0.99 1.99
E (S 3 ) E (S 3 | X 0) Pr (X 0) E ( S 3 | X 1) Pr( X 1) 0
E ( S 4 ) E ( S 4 | X 0) Pr (X 0) E ( S 4 | X 1) Pr( X 1)
3 1002 0.01 3 1 0.99 302.97

(e) S E ( S ) 0, thus E ( S S ) E ( S ) 0 frompart(d).Thusskewness0.Similarly,

3 3

S2 E ( S S ) 2 E ( S 2 ) 1.99, and E ( S S ) 4 E ( S 4 ) 302.97. Thus,

kurtosis 302.97 / (1.992 ) 76.5

9.6 10 Y 10 10.4 10
Pr (9.6 Y 10.4) Pr
4/n 4/n 4/n
9.6 10 10.4 10
Pr Z
2.15. (a) 4/n 4/n
whereZ~N(0,1).Thus,
9.6 10 10.4 10
Pr Z Pr (0.89 Z 0.89) 0.63
(i) n20; 4/n 4/n
9.6 10 10.4 10
Pr Z Pr( 2.00 Z 2.00) 0.954
(ii) n100; 4/n 4/n
9.6 10 10.4 10
Pr Z Pr(6.32 Z 6.32) 1.000
(iii) n 1000; 4/n 4/n

c Y 10 c
Pr (10 c Y 10 c) Pr
4/n 4/n 4/n
c c
Pr Z .
(b) 4/n 4/n
c
Asngetlarge 4 / n getslarge,andtheprobabilityconvergesto1.
(c) Thisfollowsfrom(b)andthedefinitionofconvergenceinprobabilitygiveninKeyConcept2.6.

Y=0.4and Y 0.4 0.6 0.24

2

2.17.
Y 0.4 0.43 0.4 Y 0.4
Pr Pr 0.6124 0.27
(a) (i) P( Y 0.43) 0.24/n 0.24/n 0.24/n

2011PearsonEducation,Inc.PublishingasAddisonWesley
 Y 0.4 0.37 0.4 Y 0.4
Pr Pr 1.22 0.11
(ii) P( Y 0.37) 0.24/n 0.24/n 0.24/n
0.41 0.40
0.41 1.96
(b) WeknowPr(1.96Z1.96)0.95,thuswewantntosatisfy 24 / n
0.39 0.40
1.96.
and 24 / n Solvingtheseinequalitiesyieldsn9220.

l
Pr (Y y j ) Pr ( X xi , Y y j )
i 1
l
Pr (Y y j | X xi )Pr ( X xi )
2.19. (a) i 1

k k l
E (Y ) y j Pr (Y yj ) yj Pr (Y yj |X xi ) Pr ( X xi )
j 1 j 1 i 1

l k



yj Pr (Y yj |X xi ) Pr ( X xi )

i 1 j 1

l
E (Y | X xi )Pr ( X xi )
(b) i 1

(c) When X and Y areindependent,

Pr (X xi , Y y j ) Pr (X xi )Pr (Y yj )

so
XY E[( X X )(Y Y )]
l k
( xi X )( y j Y ) Pr ( X xi , Y y j )
i 1 j 1
l k
( xi X )( y j Y ) Pr ( X xi ) Pr (Y y j )
i 1 j 1

l
k

i 1
( xi X ) Pr ( X xi ( y j Y ) Pr (Y y j
)
j 1
E ( X X ) E (Y Y ) 0 0 0,
XY 0
corr(X , Y ) 0
XY XY

E ( X )3 E[( X ) 2 ( X )] E[ X 3 2 X 2 X 2 X 2 2 X 2 3 ]
E ( X 3 ) 3E ( X 2 ) 3 E ( X ) 2 3 E ( X 3 ) 3E ( X 2 ) E ( X )
3E ( X )[ E ( X )]2 [ E ( X )]3

2.21. (a) E ( X 3 ) 3E ( X 2 ) E ( X ) 2 E ( X ) 3

2011PearsonEducation,Inc.PublishingasAddisonWesley
 E ( X ) 4 E[( X 3 3 X 2 3 X 2 3 )( X )]
E[ X 4 3 X 3 3 X 2 2 X 3 X 3 3 X 2 2 3 X 3 4 ]
E ( X 4 ) 4 E ( X 3 ) E ( X ) 6 E ( X 2 ) E ( X ) 2 4 E ( X ) E ( X )3 E ( X ) 4

(b) E ( X 4 ) 4[ E ( X )][ E ( X 3 )] 6[ E ( X )]2 [ E ( X 2 )] 3[ E ( X )]4

2011PearsonEducation,Inc.PublishingasAddisonWesley
2.23. XandZaretwoindependentlydistributedstandardnormalrandomvariables,so
X Z 0, X2 Z2 1, XZ 0.

(a) Becauseoftheindependencebetween X and Z , Pr ( Z z| X x) Pr ( Z z ), and

E ( Z |X ) E ( Z ) 0. Thus E (Y | X ) E ( X 2 Z | X ) E ( X 2| X ) E ( Z |X ) X 2 0 X 2

(b) E ( X ) X X 1, and Y E ( X Z ) E ( X ) Z 1 0 1
2 2 2 2 2

(c) E ( XY ) E ( X ZX ) E ( X ) E ( ZX ). Usingthefactthattheoddmomentsofastandard
3 3

normalrandomvariableareallzero,wehave E ( X ) 0. Usingtheindependencebetween
3

X and Z , wehave E ( ZX ) Z X 0. Thus E ( XY ) E ( X ) E ( ZX ) 0.

3

cov (XY ) E[( X X )(Y Y )] E[( X 0)(Y 1)]

E ( XY X ) E ( XY ) E ( X )
0 0 0
0
corr (X , Y ) XY 0
(d) XY XY
n n

axi (ax1 ax2 ax3 L axn ) a ( x1 x2 x3 L xn ) a xi

2.25. (a) i 1 i 1

(x
i 1
i yi ) ( x1 y1 x2 y2 L xn yn )

( x1 x2 L xn ) ( y1 y2 L yn )
n n
xi yi
(b) i 1 i 1

a (a a a L a ) na
(c) i 1

n n

(a bx
i 1
i cyi )2 (a 2 b 2 xi2 c 2 yi2 2abxi 2acyi 2bcxi yi )
i 1
n n n n n
na 2 b 2 xi2 c 2 yi2 2ab xi 2ac yi 2bc xi yi
(d) i 1 i 1 i 1 i 1 i 1

2.27 (a) E(W)E[E(W|Z)]E[E(X X%)|Z]E[E(X|Z)E(X|Z)]0.

2011PearsonEducation,Inc.PublishingasAddisonWesley
 (b) E(WZ)E[E(WZ|Z)]E[ZE(W)|Z]E[Z0]0
(c) Usingthehint:VWh(Z),sothatE(V2)E(W2)E[h(Z)2]2E[Wh(Z)].Usingan
argumentlikethatin(b),E[Wh(Z)]0.Thus,E(V2)E(W2)E[h(Z)2],andtheresult
followsbyrecognizingthatE[h(Z)2]0becauseh(z)20foranyvalueofz.

Chapter3
ReviewofStatistics

3.1. Thecentrallimittheoremsuggeststhatwhenthesamplesize( n )islarge,thedistributionofthe

2
N Y , Y2 Y2 Y .
sampleaverage( Y )isapproximately with n Givenapopulation Y 100,
Y2 430, wehave
Y2 43
Y2 043,
(a) n 100, n 100 and
Y 100 101 100
Pr (Y 101) Pr (1.525) 09364
043 043

Y2 43
Y2 06719,
(b) n 64, n 64 and
101 100 Y 100 103 100
Pr(101 Y 103) Pr
06719 06719 06719
(36599) (12200) 09999 08888 01111

Y2 43
Y2 02606,
(c) n 165, n 165 and
Y 100 98 100
Pr (Y 98) 1 Pr (Y 98) 1 Pr
02606 02606
1 ( 39178) (39178) 10000 (rounded to four decimal places)

Denoteeachvoterspreferenceby Y . Y 1 ifthevoterpreferstheincumbentand Y 0 ifthe

3.3.
voterprefersthechallenger. Y isaBernoullirandomvariablewithprobabilityPr (Y 1) p and
Pr (Y 0) 1 p. FromthesolutiontoExercise3.2, Y hasmean p andvariance p(1 p).

2011PearsonEducation,Inc.PublishingasAddisonWesley
 215
p 05375.
400
(a)

p ) p (1 p ) 0.5375 (1 0.5375) 62148 104.

var(
(b)Theestimatedvarianceof p is n 400 The
1

standarderrorisSE ( p ) (var( p )) 00249.

2

(c) Thecomputedtstatisticis
p p0 05375 05
t act 1506
SE( p ) 00249

Becauseofthelargesamplesize (n 400), wecanuseEquation(3.14)inthetexttogetthe

pvalueforthetest H 0 p 05 vs. H1 p 05 :
p -value 2 (|t act |) 2 (1506) 2 0066 0132

(d) UsingEquation(3.17)inthetext,thepvalueforthetest H 0 p 05 vs. H1 p 05 is

p -value 1 (t act ) 1 (1506) 1 0934 0066
(e) Part(c)isatwosidedtestandthepvalueistheareainthetailsofthestandardnormal
distributionoutside(calculatedtstatistic).Part(d)isaonesidedtestandthepvalueisthe
areaunderthestandardnormaldistributiontotherightofthecalculatedtstatistic.
(f) Forthetest H 0 p 05 vs. H1 p 05, wecannotrejectthenullhypothesisatthe5%
significancelevel.Thepvalue0.066islargerthan0.05.Equivalentlythecalculatedtstatistic
1506 islessthanthecriticalvalue1.64foraonesidedtestwitha5%significancelevel.The
testsuggeststhatthesurveydidnotcontainstatisticallysignificantevidencethatthe
incumbentwasaheadofthechallengeratthetimeofthesurvey.

3.5. (a) (i) Thesizeisgivenby Pr(| p 0.5| .02), wheretheprobabilityiscomputedassuming

that p 0.5.

Pr(|p 0.5| 0.02) 1 Pr( 0.02 p 0.5 .02)

0.02 p 0.5 0.02
1 Pr
0.5 0.5/1055 0.5 0.5/1055 0.5 0.5/1055
p 0.5
1 Pr 1.30 1.30
0.5 0.5/1055
0.19
wherethefinalequalityusingthecentrallimittheoremapproximation.
(ii) Thepowerisgivenby Pr(| p 0.5| 0.02), wheretheprobabilityiscomputedassuming
thatp0.53.

2011PearsonEducation,Inc.PublishingasAddisonWesley
 Pr(|p 0.5| 0.02) 1 Pr( 0.02 p 0.5 .02)
0.02 p 0.5 0.02
1 Pr
0.53 0.47/1055 0.53 0.47/1055 0.53 0.47/1055
0.05 p 0.53 0.01
1 Pr
0.53 0.47/1055 0.53 0.47/1055 0.53 0.47/1055
p 0.53
1 Pr 3.25 0.65
.53 0.47/1055
0.74
wherethefinalequalityusingthecentrallimittheoremapproximation.
0.54 0.50
t 2.61, and Pr(|t | 2.61) 0.01,
(0.54 0.46) / 1055
(b) (i) sothatthenullisrejectedatthe
5%level.
(ii) Pr(t 2.61) .004, sothatthenullisrejectedatthe5%level.

(iii) 0.54 1.96 (0.54 0.46) / 1055 0.54 0.03, or 0.51 to 0.57.

(iv) 0.54 2.58 (0.54 0.46) / 1055 0.54 0.04, or 0.50 to 0.58.

(v) 0.54 0.67 (0.54 0.46) / 1055 0.54 0.01, or 0.53 to 0.55.
(c) (i) Theprobabilityis0.95isanysinglesurvey,thereare20independentsurveys,sothe
probabilityif 0.95 0.36
20

(ii) 95%ofthe20confidenceintervalsor19.
1.96 SE( p ) .01 or 1.96 p(1 p) / n .01.
(d) Therelevantequationis Thusnmustbe
1.96 p(1 p )
2
n ,
chosensothat 0.012 sothattheanswerdependsonthevalueofp.Notethatthe
largestvaluethatp(1p)cantakeonis0.25(thatis,p0.5makesp(1p)aslargeas
1.962 0.25
n 9604,
possible).Thusif 0.012 thenthemarginoferrorislessthan0.01forall
valuesofp.

3.7. Thenullhypothesisisthatthesurveyisarandomdrawfromapopulationwithp=0.11.Thet
p 0.11
t ,
statisticis SE( p ) where SE( p ) p (1 p )/n. (AnalternativeformulaforSE( p )is
0.11 (1 0.11) / n, whichisvalidunderthenullhypothesisthat p 0.11). Thevalueofthet

statisticis2.71,whichhasapvalueofthatislessthan0.01.Thusthenullhypothesis p 0.11
(thesurveyisunbiased)canberejectedatthe1%level.

3.9. Denotethelifeofalightbulbfromthenewprocessby Y . Themeanof Y is andthestandard

deviationof Y is Y 200 hours. Y isthesamplemeanwithasamplesize n 100. The
Y 200
Y 20
standarddeviationofthesamplingdistributionof Y is n 100 hours.The

2011PearsonEducation,Inc.PublishingasAddisonWesley
 hypothesistestis H 0 : 2000 vs. H1 2000 . Themanagerwillacceptthealternative
hypothesisif Y 2100 hours.

(a) Thesizeofatestistheprobabilityoferroneouslyrejectinganullhypothesiswhenitisvalid.

Thesizeofthemanagerstestis

size Pr(Y 2100| 2000) 1 Pr(Y 2100| 2000)

Y 2000 2100 2000
1 Pr | 2000
20 20
1 (5) 1 0999999713 287 10 7 ,

where Pr(Y 2100| 2000) meanstheprobabilitythatthesamplemeanisgreaterthan2100

hourswhenthenewprocesshasameanof 2000 hours.

(b) Thepowerofatestistheprobabilityofcorrectlyrejectinganullhypothesiswhenitisinvalid.We
calculatefirsttheprobabilityofthemanagererroneouslyacceptingthenullhypothesiswhenitis
invalid:

Y 2150 2100 2150

Pr(Y 2100| 2150) Pr | 2150
20 20
( 25) 1 (25) 1 09938 00062

Thepowerofthemanagerstestingis 1 1 00062 09938.

(c) Foratestwith5%,therejectionregionforthenullhypothesiscontainsthosevaluesofthe
tstatisticexceeding1.645.
act
2000
t act Y 1645 Y act 2000 1645 20 20329
20
Themanagershouldbelievetheinventorsclaimifthesamplemeanlifeofthenewproductis
greaterthan2032.9hoursifshewantsthesizeofthetesttobe5%.

3.11. Assumethat n isanevennumber.Then Y%isconstructedbyapplyingaweightof1/2tothen/2

oddobservationsandaweightof3/2totheremainingn/2observations.
1 1 3 1 3
E (Y%
) E (Y1 ) E (Y2 ) L E (Yn 1 ) E (Yn )
n 2 2 2 2

1 1 n 3 n
Y Y Y
n 2 2 2 2
1 1 9 1 9
var(Y%
) 2 var(Y1 ) var(Y2 ) L var(Yn 1 ) var(Yn )
n 4 4 4 4
1 1 n 2 9 n 2 2
Y Y 1 25 Y
n2 4 2 4 2 n

2011PearsonEducation,Inc.PublishingasAddisonWesley
3.13 (a) Samplesize n 420, sampleaverage Y 646.2samplestandarddeviation sY 195.
s 19.5
(Y ) Y 09515.
Thestandarderrorof Y isSE n 420 The95%confidenceintervalfor
themeantestscoreinthepopulationis
Y 196SE(Y ) 6462 196 09515 (64434 64806)

(b) Thedataare:samplesizeforsmallclasses n1 238, sampleaverage Y 1 6574, sample

s 194;
standarddeviation 1 samplesizeforlargeclasses n2 182, sampleaverage
Y 2 6500, samplestandarddeviation s2 179. Thestandarderrorof Y1 Y2 is
s12 s22 19.42 17.92
SE (Y1 Y2 ) 18281.
n1 n2 238 182
Thehypothesistestsforhigheraverage
scoresinsmallerclassesis
H 0 1 2 0 vs H1 1 2 0
Thetstatisticis

t act Y 1 Y 2 6574 6500 40479

SE(Y 1 Y 2) 18281
Thepvaluefortheonesidedtestis:
p-value 1 (t act ) 1 (40479) 1 0999974147 25853 10 5
Withthesmallpvalue,thenullhypothesiscanberejectedwithahighdegreeofconfidence.
Thereisstatisticallysignificantevidencethatthedistrictswithsmallerclasseshavehigher
averagetestscores.

3.15. Fromthetextbookequation(2.46),weknowthatE( Y )Yandfrom(2.47)weknowthat

Y2
var( Y ) n .Inthisproblem,becauseYaandYbareBernoullirandomvariables, p a Ya , p b
Yb , Ya2 p (1p )and Yb2 p (1p ).Theanswersto(a)followfromthis.Forpart(b),notethat
a a b b

var( p a p b )var( p a )var( p b )2cov( p a , p b ).But,theyareindependent(andthushave

cov(p a ,p b ) 0because p a and p b areindependent(theydependondatachosenfromindependent

samples).Thusvar( p a p b )var( p a )var( p b ).Forpart(c),useequation3.21fromthetext

(replacing Y with p andusingtheresultin(b)tocomputetheSE).For(d),applytheformulain
(c)toobtain

2011PearsonEducation,Inc.PublishingasAddisonWesley
 0.859(1 0.859) 0.374(1 0.374)

95%CIis(.859.374)1.96 5801 4249 or0.4850.017.

Ym , 2008 Ym , 1992 1.96 SE(Ym , 2008 Ym , 1992 )

3.17. (a) The95%confidenceintervalis where
2 2
s s 11.782 10.17 2
SE(Ym , 2008 Ym , 1992 ) m ,2008
m ,1992
0.37;
nm ,2008 nm ,1992 1838 1594
the95%confidence
intervalis(24.9823.27)0.73or1.710.73.
Yw, 2008 Yw, 1992 1.96 SE(Yw, 2008 Yw, 1992 )
(b) The95%confidenceintervalis where
2 2
s s 9.662 7.782
SE(Yw, 2008 Yw, 1992 ) w,2008
w ,1992
0.31;
nw,2008 nw,1992 1871 1368
the95%confidenceinterval
is (20.87 20.05) 0.60 or0.820.60.
(Y Y ) (Yw, 2004 Yw, 1992 ) 1.96 SE[(Ym, 2008 Ym,1992 )
(c) The95%confidenceintervalis m , 2004 m, 1992
(Yw, 2008 Yw, 1992 )], SE[(Ym , 2008 Ym , 1992 ) (Yw, 2008 Yw, 1992 )]
where
sm2 ,2008 sm2 ,1992 sw2 ,2008 sw2 ,1992
11.782 10.17 2 9.66 2 7.782

nm,2008 nm,1993 nw,2008 nw,1992 1594 1871 1368 0.48. The95%
1838
confidenceintervalis(24.9823.27)(20.8720.05)1.960.48or0.890.95.

E (Yi 2 ) Y2 Y2 and E (YiY j ) Y2 for i j.

3.19. (a) No. Thus
2
1 n 1 n
1 n
1
E (Y ) E Yi 2
2
E (Y i
2
) E (Y Y )
i j
2
Y Y2
n i 1 n i 1 n2 i 1 j i n
2
(b) Yes.If Y getsarbitrarilyclosetoYwithprobabilityapproaching1asngetslarge,then Y
getsarbitrarilycloseto Y withprobabilityapproaching1asngetslarge.(Asitturnsout,
2

thisisanexampleofthecontinuousmappingtheoremdiscussedinChapter17.)

3.21. Setnmnwn,anduseequation(3.19)writethesquaredSEof Ym Yw as

2011PearsonEducation,Inc.PublishingasAddisonWesley
 1 1
in1 (Ymi Ym ) 2 in1 (Ywi Yw )2
(n 1) (n 1)
[ SE (Ym Yw )]
2

n n

in1 (Ymi Ym ) 2 in1 (Ywi Yw )2

.
n( n 1)
Similarly,usingequation(3.23)
1 n 1
i 1 (Ymi Ym ) 2 in1 (Ywi Yw ) 2
2( n 1) ( n 1)
[ SE pooled (Ym Yw )]2
2n

in1 (Ymi Ym ) 2 in1 (Ywi Yw ) 2

.
n(n 1)

Chapter4
LinearRegressionwithOneRegressor

4.1. (a) Thepredictedaveragetestscoreis

2011PearsonEducation,Inc.PublishingasAddisonWesley

TestScore 5204 582 22 39236
(b) Thepredictedchangeintheclassroomaveragetestscoreis
TestScore (582 19) (582 23) 2328


(c) Usingtheformulafor 0 inEquation(4.8),weknowthesampleaverageofthetestscores
acrossthe100classroomsis
TestScore 0 1 CS 5204 582 214 39585

(d) Usetheformulaforthestandarderroroftheregression(SER)inEquation(4.19)togetthe
sumofsquaredresiduals:
SSR (n 2)SER 2 (100 2) 1152 12961
2
Usetheformulafor R inEquation(4.16)togetthetotalsumofsquares:
SSR 12961
TSS 13044
1 R 1 0082
2

13044 1318.
Thesamplevarianceis sY n1
2 TSS
s sY2 115.
99 Thus,standarddeviationis Y

4.3. (a) Thecoefficient9.6showsthemarginaleffectofAgeonAWE;thatis,AWEisexpectedto

increaseby$9.6foreachadditionalyearofage.696.7istheinterceptoftheregressionline.It
determinestheoverallleveloftheline.
(b) SERisinthesameunitsasthedependentvariable(Y,orAWEinthisexample).ThusSERis
measuredindollarsperweek.
(c) R2isunitfree.
(d) (i) 696.7 9.6 25 $936.7;
(ii) 696.7 9.6 45 $1,128.7
(e) No.Theoldestworkerinthesampleis65yearsold.99yearsisfaroutsidetherangeofthe
sampledata.
(f) No.Thedistributionofearningispositivelyskewedandhaskurtosislargerthanthenormal.

(g) 0 Y 1 X , sothat Y 0 1 X . ThusthesamplemeanofAWEis696.79.641.6
$1,096.06.

2011PearsonEducation,Inc.PublishingasAddisonWesley
4.5. (a) uirepresentsfactorsotherthantimethatinfluencethestudentsperformanceontheexam
includingamountoftimestudying,aptitudeforthematerial,andsoforth.Somestudentswill
havestudiedmorethanaverage,otherless;somestudentswillhavehigherthanaverage
aptitudeforthesubject,otherslower,andsoforth.
(b) BecauseofrandomassignmentuiisindependentofXi.Sinceuirepresentsdeviationsfrom
averageE(ui)0.BecauseuandXareindependentE(ui|Xi)E(ui)0.
(c) (2)issatisfiedifthisyearsclassistypicalofotherclasses,thatis,studentsinthisyears
classcanbeviewedasrandomdrawsfromthepopulationofstudentsthatenrollintheclass.
(3)issatisfiedbecause0Yi100andXicantakeononlytwovalues(90and120).
(d) (i) 49 0.24 90 70.6; 49 0.24 120 77.8; 49 0.24 150 85.0
(ii) 0.24 10 2.4.

4.7. Theexpectationof 0 isobtainedbytakingexpectationsofbothsidesofEquation(4.8):
1 n
E ( 0 ) E (Y 1 X ) E 0 1 X ui 1 X
n i 1
n
1
0 E ( 1 1 ) X E (ui )
n i 1
0

wherethethirdequalityintheaboveequationhasusedthefactsthatE(ui)0andE[( 1 1) X ]

E[(E( 1 )| X ) X ]because E[( 1 1 ) | X ] 0 (seetextequation(4.31).)
1

4.9. (a) With 1 0, 0 Y , and Yi 0 Y . ThusESS0andR20.

(b) IfR20,thenESS0,sothat Yi Y foralli.But Yi 0 1 X i , sothat Yi Y foralli,

whichimpliesthat 1 0, orthatX isconstantforalli.IfX isconstantforalli,then
i i

n
(Xi X ) 0
2

i 1 and 1 isundefined(seeequation(4.7)).

(a) Theleastsquaresobjectivefunctionis i 1
n
(Yi b1 X i ) 2.
4.11. Differentiatingwithrespecttob1
n
(Yi b1 X i ) 2
2 i 1 X i (Yi b1 X i ).
i 1 n

yields b1 Settingthiszero,andsolvingfortheleast
n

X Y i i
1 i 1
n
.
X i
2

squaresestimatoryields i 1

2011PearsonEducation,Inc.PublishingasAddisonWesley
 n

X (Y 4)
i i
1 i 1
n
.
X i2
(b) Followingthesamestepsin(a)yields i 1

4.13. TheanswerfollowsthederivationsinAppendix4.3inLargeSampleNormalDistributionofthe
OLSEstimator.Inparticular,theexpressionforiisnowi(XiX)ui,sothatvar(i)
3var[(XiX)ui],andtheterm2carrythroughtherestofthecalculations.

Chapter5
RegressionwithaSingleRegressor:Hypothesis
TestsandConfidenceIntervals

5.1 (a) The95%confidenceintervalfor 1 is {582 196 221}, thatis

10152 1 14884.
(b) Calculatethetstatistic:
1 0 582
t act 26335
SE( 1) 221

Thepvalueforthetest H 0 1 0 vs. H1 1 0 is
p -value 2 ( |t act |) 2 (26335) 2 00042 00084
Thepvalueislessthan0.01,sowecanrejectthenullhypothesisatthe5%significancelevel,
andalsoatthe1%significancelevel.
(c) Thetstatisticis
1 (5.6) 022
t act 0.10
SE ( 1) 221

Thepvalueforthetest H 0 : 1 5.6 vs. H1 : 1 5.6 is

p -value 2 (|t act |) 2 (0.10) 0.92
Thepvalueislargerthan0.10,sowecannotrejectthenullhypothesisatthe10%,5%or1%
significancelevel.Because 1 5.6 isnotrejectedatthe5%level,thisvalueiscontainedin
the95%confidenceinterval.
(d) The99%confidenceintervalfor 0 is {520.4 2.58 20.4}, thatis, 467.7 0 573.0.

5.3. The99%confidenceintervalis1.5{3.942.580.31)or4.71lbsWeightGain7.11lbs.

2011PearsonEducation,Inc.PublishingasAddisonWesley
5.5 (a) Theestimatedgainfrombeinginasmallclassis13.9points.Thisisequalto
approximately1/5ofthestandarddeviationintestscores,amoderateincrease.
13.9
t act
5.56,
(b) Thetstatisticis 2.5 whichhasapvalueof0.00.Thusthenullhypothesisis
rejectedatthe5%(and1%)level.
(c) 13.92.582.513.96.45.

3.2
2.13
5.7. (a) Thetstatisticis 1.5 withapvalueof0.03;sincethepvalueislessthan0.05,the
nullhypothesisisrejectedatthe5%level.
(b) 3.21.961.53.22.94

(c) Yes.IfYandXareindependent,then 1 0; butthisnullhypothesiswasrejectedatthe

5%levelinpart(a).
(d) 1 wouldberejectedatthe5%levelin5%ofthesamples;95%oftheconfidenceintervals
wouldcontainthevalue 1 0.
1
(Y1 Y2 L Yn )
n

5.9. (a) X sothatitislinearfunctionofY1,Y2,,Yn.

(b) E(Yi|X1,,Xn)Xi,thus
1
n (Y1 Y2 L Yn )
E ( |X 1 , K , X n ) E | ( X 1 , K , X n )
X

1
( X1 L X n )
n .
X
sm
SE (Ym )

Usingtheresultsfrom5.10, 0 Ym and 1 Yw Ym . FromChapter3,
nm
5.11. and
2 2
s s
SE (Yw Ym ) m
. w

Plugginginthenumbers 0 523.1 and SE ( 0 ) 6.22; 1 38.0 and
nm nw

SE ( 1 ) 7.65.

5.13. (a) Yes,thisfollowsfromtheassumptionsinKC4.3.

(b) Yes,thisfollowsfromtheassumptionsinKC4.3andconditionalhomoskedasticity
(c) Theywouldbeunchangedforthereasonsspecifiedintheanswerstothosequestions.
(d) (a)isunchanged;(b)isnolongertrueastheerrorsarenotconditionallyhomosckesdastic.

var ( m,1 w,1 )

5.15. Becausethesamplesareindependent, m,1 and w,1 areindependent.Thus
var ( m,1 ) var( w,1 ). Var ( m,1 ) [SE( m ,1 )]2 Var (w,1 )
isconsistentlyestimatedas and is

2011PearsonEducation,Inc.PublishingasAddisonWesley
 [SE( w,1 )]2 , var( m ,1 w,1 )
consistentlyestimatedas sothat isconsistentlyestimatedby

[SE( m ,1 )] [SE( w,1 )] ,
2 2
andtheresultfollowsbynotingtheSEisthesquarerootofthe
estimatedvariance.

2011PearsonEducation,Inc.PublishingasAddisonWesley
Chapter6
LinearRegressionwith
MultipleRegressors

6.1. Byequation(6.15)inthetext,weknow
n 1
R2 1 (1 R 2 ).
n k 1
2
Thus,thatvaluesof R are0.175,0.189,and0.193forcolumns(1)(3).
6.3. (a) Onaverage,aworkerearns$0.29/hourmoreforeachyearheages.
(b) Sallysearningspredictionis 440 548 1 262 1 029 29 1567 dollarsperhour.
Betsysearningspredictionis 440 548 1 262 1 029 34 1712 dollarsperhour.
Thedifferenceis1.45
6.5. (a) $23,400(recallthatPriceismeasuredin$1000s).
(b) InthiscaseBDR1andHsize100.Theresultingexpectedchangeinpriceis23.4
0.15610039.0thousanddollarsor$39,000.
(c) Thelossis$48,800.
R 2 1 n n k11 (1 R 2 ), R 2 1 n n k 11 (1 R 2 ),
(d) Fromthetext so thus,R20.727.

6.7. (a) Theproposedresearchinassessingthepresenceofgenderbiasinsettingwagesistoo

limited.Theremightbesomepotentiallyimportantdeterminantsofsalaries:typeofengineer,
amountofworkexperienceoftheemployee,andeducationlevel.Thegenderwiththelower
wagescouldreflectthetypeofengineeramongthegender,theamountofworkexperienceof
theemployee,ortheeducationleveloftheemployee.Theresearchplancouldbeimproved
withthecollectionofadditionaldataasindicatedandanappropriatestatisticaltechniquefor
analyzingthedatawouldbeamultipleregressioninwhichthedependentvariableiswages
andtheindependentvariableswouldincludeadummyvariableforgender,dummyvariables
fortypeofengineer,workexperience(timeunits),andeducationlevel(highestgradelevel
completed).Thepotentialimportanceofthesuggestedomittedvariablesmakesadifference
inmeanstestinappropriateforassessingthepresenceofgenderbiasinsettingwages.
(b) Thedescriptionsuggeststhattheresearchgoesalongwaytowardscontrollingforpotential
omittedvariablebias.Yet,therestillmaybeproblems.Omittedfromtheanalysisare
characteristicsassociatedwithbehaviorthatledtoincarceration(excessivedrugoralcohol
use,gangactivity,andsoforth),thatmightbecorrelatedwithfutureearnings.Ideally,dataon
thesevariablesshouldbeincludedintheanalysisasadditionalcontrolvariables.

6.9. Foromittedvariablebiastooccur,twoconditionsmustbetrue:X1(theincludedregressor)is
correlatedwiththeomittedvariable,andtheomittedvariableisadeterminantofthedependent
 variable.SinceX1andX2areuncorrelated,theestimatorof1doesnotsufferfromomitted
variablebias.

6.11. (a) (Y b X
i 1 1i b2 X 2i )2

(Yi b1 X 1i b2 X 2i ) 2
2 X 1i (Yi b1 X 1i b2 X 2 i )
b1
(Yi b1 X 1i b2 X 2 i )2
2 X 2i (Yi b1 X 1i b2 X 2 i )
(b) b2

(c) From(b), 1 satisfies

X 1i (Yi 1 X 1i 1 X 2 i ) 0

X 1iYi 2 X 1i X 2i
1
or X 12i
andtheresultfollowsimmediately.
(d) Followinganalysisasin(c)
X 2iYi 1 X 1i X 2i
2
X 22i

andsubstitutingthisintotheexpressionfor 1 in(c)yields

X 1iY 2 i i X1
X Y X 1i X 2 i
X 1i X 2 i
1 2
2i
.
X 12i

Solvingfor 1 yields:

X 22i X 1iYi X 1i X 2 i X 2iYi

1
X 12i X 22i ( X 1i X 2 i ) 2

(e) Theleastsquaresobjectivefunctionis
(Yi b0 b1 X 1i b2 X 2 i ) 2
andthepartialderivative
withrespecttob0is
(Yi b0 b1 X 1i b2 X 2 i ) 2
2 (Yi b0 b1 X 1i b2 X 2i ).
b0

Settingthistozeroandsolvingfor 0 yields: 0 Y 1 X 1 2 X 2 .

(f) Substituting 0 Y 1 X 1 2 X 2 intotheleastsquaresobjectivefunctionyields
(Yi 0 b1 X 1i b2 X 2i )2 (Yi Y ) b1 ( X1i X1 ) b2 ( X 2i X 2 ) ,whichisidentical
2

totheleastsquaresobjectivefunctioninpart(a),exceptthatallvariableshavebeenreplaced
withdeviationsfromsamplemeans.Theresultthenfollowsasin(c).

2011PearsonEducation,Inc.PublishingasAddisonWesley
Noticethattheestimatorfor1isidenticaltotheOLSestimatorfromtheregressionofYonto

X1,omittingX2.Saiddifferently,when 1i
( X X 1 )( X 2 i X 2 ) 0
,theestimatedcoefficient
onX1intheOLSregressionofYontobothX1andX2isthesameasestimatedcoefficientin
theOLSregressionofYontoX1.
Chapter7
HypothesisTestsandConfidence
IntervalsinMultipleRegression

7.1and7.2
Regressor (1) (2) (3)
College(X1) 5.46** 5.48** 5.44**
(0.21) (0.21) (0.21)
Female(X2) 2.64** 2.62** 2.62**
(0.20) (0.20) (0.20)
Age(X3) 0.29** 0.29**
(0.04) (0.04)
Ntheast(X4) 0.69*
(0.30)
Midwest(X5) 0.60*
(0.28)
South(X6) 0.27
(0.26)
Intercept 12.69** 4.40** 3.75**
(0.14) (1.05) (1.06)

(a) Thetstatisticis5.46/0.2126.0,whichexceeds1.96inabsolutevalue.Thus,thecoefficient
isstatisticallysignificantatthe5%level.The95%confidenceintervalis5.461.960.21.
(b) tstatisticis2.64/0.2013.2,and13.21.96,sothecoefficientisstatisticallysignificant
atthe5%level.The95%confidenceintervalis2.641.960.20.

7.3. (a) Yes,ageisanimportantdeterminantofearnings.Usingattest,thetstatisticis

0.29 / 0.04 7.25, withapvalueof4.21013,implyingthatthecoefficientonageis
statisticallysignificantatthe1%level.The95%confidenceintervalis0.291.960.04.
(b) Age[0.291.960.04]5[0.291.960.04]1.451.960.20$1.06to$1.84

college,1998 college,1992
t
SE( college,1998 college, 1992 )
7.5. Thetstatisticforthedifferenceinthecollegecoefficientsis .Because
college,1998 college,1992
and arecomputedfromindependentsamples,theyareindependent,which

2011PearsonEducation,Inc.PublishingasAddisonWesley
 cov( college,1998 , college,1992 ) 0 var( college,1998 college,1992 )
meansthat Thus, =
var( college,1998 ) var( college,1998 ) 1
SE( college,1998 college,1992 ) (0.212 0.202 ) 2 .
.Thisimpliesthat
5.48 5.29
t act 0.6552.
Thus, 0.212 0.20 2 Thereisnosignificantchangesincethecalculatedtstatistic
islessthan1.96,the5%criticalvalue.

7.7. (a) Thetstatisticis 0.485 / 2.61 0.186 1.96. Therefore,thecoefficientonBDRisnot

statisticallysignificantlydifferentfromzero.
(b) ThecoefficientonBDRmeasuresthepartialeffectofthenumberofbedroomsholdinghouse
size(Hsize)constant.Yet,thetypical5bedroomhouseismuchlargerthanthetypical
2bedroomhouse.Thus,theresultsin(a)sayslittleabouttheconventionalwisdom.
(c) The99%confidenceintervalforeffectoflotsizeonpriceis2000[0.0022.580.00048]
or1.52to6.48(inthousandsofdollars).
(d) Choosingthescaleofthevariablesshouldbedonetomaketheregressionresultseasytoread
andtointerpret.Ifthelotsizeweremeasuredinthousandsofsquarefeet,theestimate
coefficientwouldbe2insteadof0.002.
F
(e) The10%criticalvaluefromthe 2, distributionis2.30.Because0.082.30,thecoefficients
arenotjointlysignificantatthe10%level.

7.9. (a) Estimate

Yi 0 X 1i 2 ( X 1i X 2i ) ui

andtestwhether0.
(b) Estimate
Yi 0 X 1i 2 ( X 2 i aX 1i ) ui

andtestwhether0.
(c) Estimate
Yi X 1i 0 X 1i 2 ( X 2i X1i ) ui

andtestwhether0.

7.11. (a) Treatment(assignmenttosmallclasses)wasnotrandomlyassignedinthepopulation(the

continuingandnewlyenrolledstudents)becauseofthedifferenceintheproportionoftreated
continuingandnewlyenrolledstudents.Thus,thetreatmentindicatorX1iscorrelatedwithX2.
Ifnewlyenrolledstudentsperformsystematicallydifferentlyonstandardizedteststhan
continuingstudents(perhapsbecauseofadjustmenttoanewschool),thenthisbecomespart
oftheerrortermuin(a).ThisleadstocorrelationbetweenX1andu,sothatE(u|Xl)0.

BecauseE(u|X )0,the 1 isbiasedandinconsistent.
l

(b)Becausetreatmentwasrandomlyassignedconditionalonenrollmentstatus(continuingor
newlyenrolled),E(u|X1,X2)willnotdependonX1.Thismeansthattheassumptionof

conditionalmeanindependenceissatisfied,and 1 isunbiasedandconsistent.However,
becauseX2wasnotrandomlyassigned(newlyenrolledstudentsmay,onaverage,have
attributesotherthanbeingnewlyenrolledthataffecttestscores),E(u|X1,X2)maydependof

X2,sothat 2 maybebiasedandinconsistent.
2011PearsonEducation,Inc.PublishingasAddisonWesley
Chapter8
NonlinearRegressionFunctions

198 196
100 1.0204%.
8.1. (a) Thepercentageincreaseinsalesis 196 Theapproximation
is100[ln(198)ln(196)]1.0152%.
205 196
100 4.5918%
(b) WhenSales2010205,thepercentageincreaseis 196 andthe
approximationis100[ln(205)ln(196)]4.4895%.WhenSales2010250,thepercentage
250 196
100 27.551%
increaseis 196 andtheapproximationis100[ln(250)ln(196)]
500 196
100 155.1%
24.335%.WhenSales2010500,thepercentageincreaseis 196 andthe
approximationis100[ln(500)ln(196)]93.649%.
(c) Theapproximationworkswellwhenthechangeissmall.Thequalityoftheapproximation
deterioratesasthepercentagechangeincreases.

8.3. (a) Theregressionfunctionsforhypotheticalvaluesoftheregressioncoefficientsthatare

consistentwiththeeducatorsstatementare: 1 0 and 2 0. When TestScore isplotted
against STR theregressionwillshowthreehorizontalsegments.Thefirstsegmentwillbefor
valuesof STR 20; thenextsegmentfor 20 STR 25; thefinalsegmentfor STR 25. The
firstsegmentwillbehigherthanthesecond,andthesecondsegmentwillbehigherthanthe
third.
(b) Ithappensbecauseofperfectmulticollinearity.Withallthreeclasssizebinaryvariables
includedintheregression,itisimpossibletocomputetheOLSestimatesbecausetheintercept
isaperfectlinearfunctionofthethreeclasssizeregressors.

8.5. (a) (1) Thedemandforolderjournalsislesselasticthanforyoungerjournalsbecausethe

interactiontermbetweenthelogofjournalageandpricepercitationispositive.(2)There
isalinearrelationshipbetweenlogpriceandlogofquantityfollowsbecausetheestimated
coefficientsonlogpricesquaredandlogpricecubedarebothinsignificant.(3)Thedemand
isgreaterforjournalswithmorecharactersfollowsfromthepositiveandstatistically
significantcoefficientestimateonthelogofcharacters.
 (b) (i) Theeffectofln(Pricepercitation)isgivenby[0.8990.141ln(Age)]ln(Priceper
citation).UsingAge80,theelasticityis[0.8990.141ln(80)]0.28.

(ii) Asdescribedinequation(8.8)andthefootnoteonpage261,thestandarderrorcanbefound
bydividing0.28,theabsolutevalueoftheestimate,bythesquarerootofthe
Fstatistictestingln(Pricepercitation)ln(80)ln(Age)ln(Pricepercitation)0.
Characters
ln ln(Characters ) ln(a)
(c) a foranyconstanta.Thus,estimatedparameteron
Characterswillnotchangeandtheconstant(intercept)willchange.

8.7. (a) (i) ln(Earnings)forfemalesare,onaverage,0.44lowerformenthanforwomen.

(ii) Theerrortermhasastandarddeviationof2.65(measuredinlogpoints).
(iii) Yes.Howevertheregressiondoesnotcontrolformanyfactors(sizeoffirm,industry,
profitability,experienceandsoforth).
(iv) No.Inisolation,theseresultsdonotimplygenderdiscrimination.Genderdiscrimination
meansthattwoworkers,identicalineverywaybutgender,arepaiddifferentwages.Thus,
itisalsoimportanttocontrolforcharacteristicsoftheworkersthatmayaffecttheir
productivity(education,yearsofexperience,etc.)Ifthesecharacteristicsaresystematically
differentbetweenmenandwomen,thentheymayberesponsibleforthedifferenceinmean
wages.(Ifthisweretrue,itwouldraiseaninterestingandimportantquestionofwhywomen
tendtohavelesseducationorlessexperiencethanmen,butthatisaquestionabout
somethingotherthangenderdiscriminationintopcorporatejobs.)Thesearepotentially
importantomittedvariablesintheregressionthatwillleadtobiasintheOLScoefficient
estimatorforFemale.Sincethesecharacteristicswerenotcontrolledforinthestatistical
analysis,itisprematuretoreachaconclusionaboutgenderdiscrimination.
(b) (i) IfMarketValueincreasesby1%,earningsincreaseby0.37%

(ii) Femaleiscorrelatedwiththetwonewincludedvariablesandatleastoneofthevariablesis
importantforexplainingln(Earnings).Thustheregressioninpart(a)sufferedfromomitted
variablebias.
(c) ForgettingabouttheeffectorReturn,whoseeffectsseemssmallandstatisticallyinsignificant,
theomittedvariablebiasformula(seeequation(6.1))suggeststhatFemaleisnegatively
correlatedwithln(MarketValue).

8.9. Notethat
 Y 0 1 X 2 X 2
0 ( 1 21 2 ) X 2 ( X 2 21X ).

Defineanewindependentvariable Z X 21X , andestimate

2

Y 0 X 2 Z ui

196 SE .
Theconfidenceintervalis

dE (Y | X )
1
8.11. Linearmodel:E(Y|X)01X,sothat dX andtheelasticityis
X 1 X
1
E (Y | X ) 0 1 X

LogLogModel:E(Y|X)

E e 0 1 ln( X ) u | X e 0 1 ln( X ) E (eu | X ) ce 0 1 ln( X )
,wherec
E(e |X),whichdoesnotdependonXbecauseuandXareassumedtobeindependent.
u

dE (Y | X ) 1 0 1 ln( X ) E (Y | X )
ce 1
Thus dX X X ,andtheelasticityis1.

Chapter9
AssessingStudiesBasedon
MultipleRegression

9.1. Asexplainedinthetext,potentialthreatstoexternalvalidityarisefromdifferencesbetweenthe
populationandsettingstudiedandthepopulationandsettingofinterest.Thestatisticalresults
basedonNewYorkinthe1970sarelikelytoapplytoBostoninthe1970sbutnottoLosAngeles
inthe1970s.In1970,NewYorkandBostonhadlargeandwidelyusedpublictransportation
systems.AttitudesaboutsmokingwereroughlythesameinNewYorkandBostoninthe1970s.In
contrast,LosAngeleshadaconsiderablysmallerpublictransportationsystemin1970.Most
residentsofLosAngelesreliedontheircarstocommutetowork,school,andsoforth.Theresults
fromNewYorkinthe1970sareunlikelytoapplytoNewYorkin2010.Attitudestowards
smokingchangedsignificantlyfrom1970to2010.

9.3. Thekeyisthattheselectedsamplecontainsonlyemployedwomen.Considertwowomen,Beth
andJulie.Bethhasnochildren;Juliehasonechild.BethandJulieareotherwiseidentical.Both
 canearn$25,000peryearinthelabormarket.Eachmustcomparethe$25,000benefittothecosts
ofworking.ForBeth,thecostofworkingisforgoneleisure.ForJulie,itisforgoneleisureandthe
costs(pecuniaryandother)ofchildcare.IfBethisjustonthemarginbetweenworkinginthe
labormarketornot,thenJulie,whohasahigheropportunitycost,willdecidenottoworkinthe
labormarket.Instead,Juliewillworkinhomeproduction,caringforchildren,andsoforth.
Thus,onaverage,womenwithchildrenwhodecidetoworkarewomenwhoearnhigherwagesin
thelabormarket.

1 0 0 1 1u 1v
Q .
9.5. (a) 1 1 1 1
0 0 uv
P .
and 1 1 1 1
1 0 0 1 0
E (Q) , E ( P) 0
(b) 1 1 1 1
2
1
var(Q) ( 1 u 1 v ), var( P )
2 2 2 2

1 1
2
1
( u v ),
2 2

1 1
(c)
and
2
1
cov( P, Q) ( 1 u 1 V )
2 2

1 1
p cov(Q, P ) 2 1 v2 cov( P, Q )
1 1 u2 , 0
p
E (Q ) E ( P )
(d) (i) var( P ) u 2
v var( P)

( )
p 2
1 1 u 2 1 2 1 0,
(ii) u v usingthefactthat10(supplycurvesslopeup)and10
(demandcurvesslopedown).

9.7. (a) True.CorrelationbetweenregressorsanderrortermsmeansthattheOLSestimatoris

inconsistent.
(b) True.

9.9. Bothregressionssufferfromomittedvariablebiassothattheywillnotprovidereliableestimates
ofthecausaleffectofincomeontestscores.However,thenonlinearregressionin(8.18)fitsthe
datawell,sothatitcouldbeusedforforecasting.

9.11. Again,therearereasonsforconcern.Hereareafew.
Internalconsistency:Totheextentthatpriceisaffectedbydemand,theremaybesimultaneous
equationbias.

Externalconsistency:TheinternetandintroductionofEjournalsmayinduceimportantchangesin
themarketforacademicjournalssothattheresultsfor2000maynotberelevantfortodaysmarket.
 ( X% %
300
i X )(Yi Y )
1 i 1
( X% %2
300
i X)
9.13. (a) i 1 .BecausealloftheXisareused(althoughsomeareusedfor

thewrongvaluesofYj), X% X ,and i 1
n
( X i X )2
.Also, Yi Y 1 ( X i X ) ui u .
Usingtheseexpressions:

( X% ( X%i X )(ui u )
0.8 n n n
( X i X )2 i X )( X i X )
1 1 in1 1 i 0.8 n 1
i 1

n n
i 1
( X i X )2 i 1
( X i X )2 i 1
( X i X )2
1 0.8 n 1 n 1 n %
n
i 1
( X i X )2
n
i 0.8 n 1
( X% i X )( X i X )
n
i 1 ( X i X )(ui u )
1 1
1 n 1 n 1 n
n
i 1
( X i X ) 2

n
i 1
( X i X ) 2
( X i X )2
n i 1

wheren300,andthelastequalityusesanorderingoftheobservationssothatthefirst
%
240observations(0.8n)correspondtothecorrectlymeasuredobservations( X i Xi).

Asisdoneelsewhereinthebook,weinterpretn300asalargesample,soweusethe
approximationofntendingtoinfinity.Thesolutionprovidedherethusshowsthatthese
expressionsareapproximatelytruefornlargeandholdinthelimitthatntendstoinfinity.

Eachoftheaveragesintheexpressionfor 1 havethefollowingprobabilitylimits:

1 n p

n
i 1
( X i X ) 2
X2
,

1 0.8 n p

n i 1
( X i X ) 2
0.8 X2
,

1 n % p

n i 1
( X i X )( u i u ) 0
,and

1 n p

n i 0.8n 1
( % X )( X X ) 0
X i i
,
 %
wherethelastresultfollowsbecause X i XiforthescrambledobservationsandXjis
p

independentofX forij.Takentogether,theseresultsimply 1 0.81 .
i
p p

(b) Because 1 0.81 , 1 / 0.8 1 ,soaconsistentestimatorof1istheOLSestimator
dividedby0.8.

(c) Yes,theestimatorbasedonthefirst240observationsisbetterthantheadjustedestimator
frompart(b).Equation(4.21)inKeyConcept4.4(page129)impliesthattheestimatorbased
onthefirst240observationshasavariancethatis

1 var ( X i X )ui
var( 1 (240obs ))
var( X i )
2
240
.
Frompart(a),theOLSestimatorbasedonalloftheobservationshastwosourcesofsampling
( X% X )(u u )
300
i 1 i i

300
( X i X )2
error.Thefirstis i 1 whichistheusualsourcethatcomesfromthe
( X% X )( X X )
300
i i
1 i 241

i 1 ( X i X )2 ,whichisthesourcethatcomes
300

omittedfactors(u).Thesecondis
fromscramblingthedata.Thesetwotermsareuncorrelatedinlargesamples,andtheir
respectivelargesamplevariancesare:

( X% 1 var ( X i X )ui
300
i X )(ui u )
var i 1
i 1 ( X i X ) 300 var( X i )
300 2 2

and

( X%i X )( X i X )
300

var 1 i 241 12 0.2

i 1 ( X i X )
300 2 300
.
Thus

1 (300obs) 1 1 var ( X i X )ui 2 0.2

var

0.8 var( X i )
2 1
0.64 300 300

whichislargerthanthevarianceoftheestimatorthatonlyusesthefirst240observations.

Thus
 1 (300obs) 1 1 var ( X i X )ui 2 0.2

var
0.8 var( X i )
2 1
0.64 300 300

whichislargerthanthevarianceoftheestimatorthatonlyusesthefirst240observations.

Chapter10
RegressionwithPanelData

10.1. (a) Witha$1increaseinthebeertax,theexpectednumberoflivesthatwouldbesavedis

0.45per10,000people.SinceNewJerseyhasapopulationof8.1million,theexpected
numberoflivessavedis0.45810364.5.The95%confidenceintervalis(0.451.96
0.22)810[15.228,713.77].
(b) WhenNewJerseylowersitsdrinkingagefrom21to18,theexpectedfatalityrateincreasesby
0.028deathsper10,000.The95%confidenceintervalforthechangeindeathrateis0.028
1.960.066[0.1014,0.1574].Withapopulationof8.1million,thenumberoffatalities
willincreaseby0.02881022.68witha95%confidenceinterval[0.1014,0.1574]810
[82.134,127.49].
(c) WhenrealincomepercapitainNewJerseyincreasesby1%,theexpectedfatalityrate
increasesby1.81deathsper10,000.The90%confidenceintervalforthechangeindeathrate
is1.811.640.47[1.04,2.58].Withapopulationof8.1million,thenumberoffatalities
willincreaseby1.818101466.1witha90%confidenceinterval[1.04,2.58]810
[840,2092].
(d) Thelowpvalue(orhighFstatistic)associatedwiththeFtestontheassumptionthattime
effectsarezerosuggeststhatthetimeeffectsshouldbeincludedintheregression.
(e) Defineabinaryvariablewestwhichequals1forthewesternstatesand0fortheotherstates.
Includetheinteractiontermbetweenthebinaryvariablewestandtheunemploymentrate,
west(unemploymentrate),intheregressionequationcorrespondingtocolumn(4).Suppose
thecoefficientassociatedwithunemploymentrateisandthecoefficientassociatedwith
west(unemploymentrate)is.Thencapturestheeffectoftheunemploymentrateinthe
easternstates,andcapturestheeffectoftheunemploymentrateinthewesternstates.
Thedifferenceintheeffectoftheunemploymentrateinthewesternandeasternstatesis .
Usingthecoefficientestimate ( ) andthestandarderror SE( ), youcancalculatethet

statistictotestwhetherisstatisticallysignificantatagivensignificancelevel.

10.3. Thefivepotentialthreatstotheinternalvalidityofaregressionstudyare:omittedvariables,
misspecificationofthefunctionalform,imprecisemeasurementoftheindependentvariables,
sampleselection,andsimultaneouscausality.Youshouldthinkaboutthesethreatsonebyone.
Arethereimportantomittedvariablesthataffecttrafficfatalitiesandthatmaybecorrelatedwith
 theothervariablesincludedintheregression?Themostobviouscandidatesarethesafetyofroads,
weather,andsoforth.Thesevariablesareessentiallyconstantoverthesampleperiod,sotheir
effectiscapturedbythestatefixedeffects.Youmaythinkofsomethingthatwemissed.Since
mostofthevariablesarebinaryvariables,thelargestfunctionalformchoiceinvolvestheBeerTax
variable.Alinearspecificationisusedinthetext,whichseemsgenerallyconsistentwiththedata
inFigure8.2.Tocheckthereliabilityofthelinearspecification,itwouldbeusefultoconsidera
logspecificationoraquadratic.Measurementerrordoesnotappeartoaproblem,asvariableslike
trafficfatalitiesandtaxesareaccuratelymeasured.Similarly,sampleselectionisanotaproblem
becausedatawereusedfromallofthestates.Simultaneouscausalitycouldbeapotentialproblem.
Thatis,stateswithhighfatalityratesmightdecidetoincreasetaxestoreduceconsumption.Expert
knowledgeisrequiredtodetermineifthisisaproblem.

10.5. LetD2i1ifi2and0otherwise;D3i1ifi3and0otherwiseDni1ifinand0
otherwise.LetB2t1ift2and0otherwise;B3t1ift3and0otherwiseBTt1iftT
and0otherwise.Let011;i i 1andt t 1.

10.7. (a) Averagesnowfalldoesnotvaryovertime,andthuswillbeperfectlycollinearwiththe

statefixedeffect.
(b) Snowitdoesvarywithtime,andsothismethodcanbeusedalongwithstatefixedeffects.

u2
1 T Y .
10.9. (a) i T t 1 it whichhasvariance T BecauseTisnotgrowing,thevarianceisnot
gettingsmall. i isnotconsistent.

(b) Theaveragein(a)iscomputedoverTobservations.InthiscaseTissmall(T4),sothe
normalapproximationfromtheCLTisnotlikelytobeverygood.

10.11 Usingthehint,equation(10.22)canbewrittenas

1 1
X i 2 X i1 Yi 2 Yi1 X i 2 X i1 Yi 2 Yi1
n
i 1
1DM 4 4
n 1 1 2
i 1 4 X i 2 X i1 4 X i 2 X i1
2

X X Y Y
n
i2 i1 i2 i1
i 1 BA

X X Chapter11
n 2 1
i 1 i2 i1

RegressionwithaBinary
DependentVariable
11.1. (a) ThetstatisticforthecoefficientonExperienceis0.031/0.0093.44,whichissignificant
atthe1%level.
(b) zMatthew 0.712 0.031 10 1.022; (1.022) 0.847
z 0.712 0.031 0 0.712; (0.712) 0.762
(c) Christopher
(d) z Jed 0.712 0.031 80 3.192; (3.192) 0.999, thisisunlikelytobeaccuratebecausethe
sampledidnotincludeanyonewithmorethat40yearsofdrivingexperience.

11.3. (a) ThetstatisticforthecoefficientonExperienceist0.006/0.0023,whichissignificant

athe1%level.
ProbMatther0.7740.006100.836
ProbChristopher0.7740.00600.774
(b)

Theprobabilitiesaresimilarexceptwhenexperienceinlarge(40years).Inthiscasethe
LPMmodelproducesnonsensicalresults(probabilitiesgreaterthan1.0).

11.5. (a) (0.8060.041100.17410.015110)0.814

(b) (0.8060.04120.17400.01502)0.813
(c) Thetstatontheinteractiontermis0.015/0.0190.79,whichisnotsignificantatthe10%
level.

11.7. (a) ForablackapplicanthavingaP/Iratioof0.35,theprobabilitythattheapplicationwillbe

1
F ( 4.13 5.37 0.35 1.27) 27.28%.
deniedis 1 e0.9805
 (b) WiththeP/Iratioreducedto0.30,theprobabilityofbeingdeniedis
1
F (4.13 5.37 0.30 1.27) 22.29%
1 e1.249 .Thedifferenceindenialprobabilities
comparedto(a)is4.99percentagepointslower.
(c) ForawhiteapplicanthavingaP/Iratioof0.35,theprobabilitythattheapplicationwillbe
1
F (4.13 5.37 0.35) 9.53%.
deniedis 1 e2.2505 IftheP/Iratioisreducedto0.30,the
1
F (4.13 5.37 0.30) 7.45%.
probabilityofbeingdeniedis 1 e2.519 Thedifferencein
denialprobabilitiesis2.08percentagepointslower.
(d) Fromtheresultsinparts(a)(c),wecanseethatthemarginaleffectoftheP/Iratioonthe
probabilityofmortgagedenialdependsonrace.Inthelogitregressionfunctionalform,
themarginaleffectdependsonthelevelofprobabilitywhichinturndependsontherace
oftheapplicant.Thecoefficientonblackisstatisticallysignificantatthe1%level.Thelogit
andprobitresultsaresimilar.

11.9. (a) Thecoefficientonblackis0.084,indicatinganestimateddenialprobabilitythatis

8.4percentagepointshigherfortheblackapplicant.
(b) The95%confidenceintervalis0.0841.960.023[3.89%,12.91%].
(c) Theanswerin(a)willbebiasedifthereareomittedvariableswhichareracerelatedandhave
impactsonmortgagedenial.Suchvariableswouldhavetoberelatedwithraceandalsobe
relatedwiththeprobabilityofdefaultonthemortgage(whichinturnwouldleadtodenialof
themortgageapplication).Standardmeasuresofdefaultprobability(pastcredithistoryand
employmentvariables)areincludedintheregressionsshowninTable9.2,sotheseomitted
variablesareunlikelytobiastheanswerin(a).Othervariablessuchaseducation,marital
status,andoccupationmayalsoberelatedtheprobabilityofdefault,andthesevariablesare
omittedfromtheregressionincolumn.Addingthesevariables(seecolumns(4)(6))have
littleeffectontheestimatedeffectofblackontheprobabilityofmortgagedenial.

11.11. (a) Thisisacensoredortruncatedregressionmodel(notethedependentvariablemightbe

zero).
(b) Thisisanorderedresponsemodel.
(c) Thisisthediscretechoice(ormultiplechoice)model.
(d) Thisisamodelwithcountdata.
Chapter12
InstrumentalVariablesRegression

ln( Pi ,cigarettes ) ln( Pi ,1985

cigarettes
),
12.1. (a) Thechangeintheregressor, 1995
froma$0.50perpackincreasein
theretailpriceisln(8.00)ln(7.50)0.0645.Theexpectedpercentagechangeincigarette
demandis0.940.0645100%.07%.The95%confidenceintervalis(0.941.96
0.21)0.0645100%[8.72%,3.41%].
(b) Witha2%reductioninincome,theexpectedpercentagechangeincigarettedemandis
0.53(0.02)100%1.06%.
(c) Theregressionincolumn(1)willnotprovideareliableanswertothequestionin(b)when
recessionslastlessthan1year.Theregressionincolumn(1)studiesthelongrunpriceand
incomeelasticity.Cigarettesareaddictive.Theresponseofdemandtoanincomedecrease
willbesmallerintheshortrunthaninthelongrun.
(d) Theinstrumentalvariablewouldbetooweak(irrelevant)iftheFstatisticincolumn(1)was
3.6insteadof33.6,andwecannotrelyonthestandardmethodsforstatisticalinference.Thus
theregressionwouldnotprovideareliableanswertothequestionposedin(a).

2 1 n (Y 0TSLS 1TSLS X i )2
isnotconsistent.Writethisas a
12.3. (a) Theestimator a n 2 i 1 i
2
in1 (ui 1TSLS ( X i X i )) 2 , TSLS TSLS TSLS
where ui Yi 0 1 X i . Replacing 1 with1,
1
n2

2 1 n (u 1 ( X i X i )) 2 1n in1 ui2
assuggestedinthequestion,writethisas a n i 1 i
1
in1[ 12 ( X i X i )2 2ui 1 ( X i X i )].
n Thefirsttermontherighthandsideoftheequation
convergesto u , butthesecondtermconvergestosomethingthatisnonzero.Thus a is
2 2
notconsistent.

b2 1
in1 (Yi 0TSLS 1TSLS X i ) 2
(b) Theestimator n 2 isconsistent.Usingthesamenotationasin
u ,
andthisestimatorconvergesinprobabilityto u .
2 1 n 2 2
(a),wecanwrite b n i 1 i

12.5. (a) Instrumentrelevance. Z i doesnotenterthepopulationregressionfor X i

 (b) Zisnotavalidinstrument. X willbeperfectlycollinearwithW.(Alternatively,thefirststage
regressionsuffersfromperfectmulticollinearity.)
(c) Wisperfectlycollinearwiththeconstantterm.
(d) Zisnotavalidinstrumentbecauseitiscorrelatedwiththeerrorterm.

(a) Underthenullhypothesisofinstrumentexogeneity,theJstatisticisdistributedasa 1
2
12.7.
randomvariable,witha1%criticalvalueof6.63.Thusthestatisticissignificant,andinstrument
exogeneityE(ui|Z1i,Z2i)0isrejected.
(b) TheJtestsuggeststhatE(ui |Z1i,Z2i)0,butdoesntprovideevidenceaboutwhetherthe
problemiswithZ1orZ2orboth.

12.9. (a) Thereareotherfactorsthatcouldaffectboththechoicetoserveinthemilitaryandannual

earnings.Oneexamplecouldbeeducation,althoughthiscouldbeincludedintheregression
asacontrolvariable.Anothervariableisabilitywhichisdifficulttomeasure,andthus
difficulttocontrolforintheregression.

(b) Thedraftwasdeterminedbyanationallotterysothechoiceofservinginthemilitarywas
random.Becauseitwasrandomlyselected,thelotterynumberisuncorrelatedwithindividual
characteristicsthatmayaffectearningandhencetheinstrumentisexogenous.Becauseit
affectedtheprobabilityofservinginthemilitary,thelotterynumberisrelevant.

Chapter13
ExperimentsandQuasiExperiments

13.1. Forstudentsinkindergarten,theestimatedsmallclasstreatmenteffectrelativetobeingina
regularclassisanincreaseof13.90pointsonthetestwithastandarderror2.45.The95%
confidenceintervalis13.901.962.45[9.098,18.702].
Forstudentsingrade1,theestimatedsmallclasstreatmenteffectrelativetobeinginaregular
classisanincreaseof29.78pointsonthetestwithastandarderror2.83.The95%confidence
intervalis29.781.962.83[24.233,35.327].

Forstudentsingrade2,theestimatedsmallclasstreatmenteffectrelativetobeinginaregular
classisanincreaseof19.39pointsonthetestwithastandarderror2.71.The95%confidence
intervalis19.391.962.71[14.078,24.702].
 Forstudentsingrade3,theestimatedsmallclasstreatmenteffectrelativetobeinginaregular
classisanincreaseof15.59pointsonthetestwithastandarderror2.40.The95%confidence
intervalis15.591.962.40[10.886,20.294].

X X Control
13.3. (a) Theestimatedaveragetreatmenteffectis TreatmentGroup 1241120140
points.
(b) Therewouldbenonrandomassignmentifmen(orwomen)haddifferentprobabilitiesofbeing
assignedtothetreatmentandcontrolgroups.LetpMendenotetheprobabilitythatamaleis
assignedtothetreatmentgroup.RandomassignmentmeanspMen0.5.Testingthisnull
p Men 0.5 0.55 0.50
tMen 1.00,
1 1
p Men (1 p Men ) 0.55(1 0.55)
n Men 100
hypothesisresultsinatstatisticof
sothatthenullofrandomassignmentcannotberejectedatthe10%level.Asimilarresultis
foundforwomen.

13.5. (a) Thisisanexampleofattrition,whichposesathreattointernalvalidity.Afterthemale

athletesleavetheexperiment,theremainingsubjectsarerepresentativeofapopulationthat
excludesmaleathletes.Iftheaveragecausaleffectforthispopulationisthesameasthe
averagecausaleffectforthepopulationthatincludesthemaleathletes,thentheattritiondoes
notaffecttheinternalvalidityoftheexperiment.Ontheotherhand,iftheaveragecausal
effectformaleathletesdiffersfromtherestofpopulation,internalvalidityhasbeen
compromised.
(b) Thisisanexampleofpartialcompliancewhichisathreattointernalvalidity.Thelocalarea
networkisafailuretofollowtreatmentprotocol,andthisleadstobiasintheOLSestimatorof
theaveragecausaleffect.
(c) Thisposesnothreattointernalvalidity.Asstated,thestudyisfocusedontheeffectofdorm
roomInternetconnections.Thetreatmentismakingtheconnectionsavailableintheroom;the
treatmentisnottheuseoftheInternet.Thus,theartmajorsreceivedthetreatment(although
theychosenottousetheInternet).
(d) Asinpart(b)thisisanexampleofpartialcompliance.Failuretofollowtreatmentprotocol
leadstobiasintheOLSestimator.

13.7. Fromthepopulationregression
 Yit i 1 X it 2 ( Dt Wi ) 0 Dt vit ,
wehave

Yi 2 Yi1 1 ( X i 2 X i1 ) 2 [( D2 D1 ) Wi ] 0 ( D2 D1 ) (vi 2 vi1 ).

BydefiningYiYi2Yi1,XiXi2Xi1(abinarytreatmentvariable)anduivi2vi1,andusing
D10andD21,wecanrewritethisequationas
Yi 0 1 X i 2Wi ui ,

whichisEquation(13.5)inthecaseofasingleWregressor.

13.9. Thecovariancebetween 1i X i andXiis

cov( 1i X i , X i ) E{[ 1i X i E ( 1i X i )][ X i E ( X i )]}
E{1i X i2 E ( 1i X i ) X i 1i X i E ( X i ) E ( 1i X i ) E ( X i )}
E ( 1i X i2 ) E (1i X i ) E ( X i )

BecauseXiisrandomlyassigned,Xiisdistributedindependentlyof1i.Theindependencemeans
E ( 1i X i ) E ( 1i ) E ( X i ) and E ( 1i X i2 ) E ( 1i ) E ( X i2 ).

Thus cov( 1i X i , X i ) canbefurthersimplified:

cov( 1i X i , X i ) E ( 1i )[ E ( X i2 ) E 2 ( X i )]
E ( 1i ) X2 .

So
cov( 1i X i , X i ) E ( 1i ) X2
E ( 1i ).
X2 X2

13.11. FollowingthenotationusedinChapter13,let1idenotethecoefficientonstatesalestaxinthe
firststageIVregression,andlet1idenotecigarettedemandelasticity.(Inbothcases,suppose
thatincomehasbeencontrolledforintheanalysis.)From(13.11)
 p
E ( 1i 1i ) cov( 1i , 1i ) cov( 1i , 1i )
TSLS E ( 1i ) Average Treatment Effect ,
E ( 1i ) E ( 1i ) E ( 1i )
wherethefirstequalityusestheusespropertiesofcovariances(equation(2.34)),andthesecond
equalityusesthedefinitionoftheaveragetreatmenteffect.Evidently,thelocalaveragetreatment
effectwilldeviatefromtheaveragetreatmenteffectwhen cov( 1i , 1i ) 0.Asdiscussedin
Section13.6,thiscovarianceiszerowhen1ior1iareconstant.Thisseemslikely.But,forthe
sakeofargument,supposethattheyarenotconstant;thatis,supposethedemandelasticitydiffers
fromstatetostate(1iisnotconstant)asdoestheeffectofsalestaxesoncigaretteprices( 1iisnot
constant).Are1iand1irelated?Microeconomicssuggeststhattheymightbe.Recallfromyour
microeconomicsclassthattheloweristhedemandelasticity,thelargerfractionofasalestaxis
passedalongtoconsumersintermsofhigherprices.Thissuggeststhat 1iand1iarepositively
related,sothat cov( 1i , 1i ) 0. BecauseE( )0,thissuggeststhatthelocalaveragetreatment
1i

effectisgreaterthantheaveragetreatmenteffectwhen 1ivariesfromstatetostate.

Chapter14
IntroductiontoTimeSeriesRegression
andForecasting

14.1. (a) SincetheprobabilitydistributionofYtisthesameastheprobabilitydistributionofYt1

(thisisthedefinitionofstationarity),themeans(andallothermoments)arethesame.
(b) E(Yt)01E(Yt1)E(ut),butE(ut)0andE(Yt)E(Yt1).ThusE(Yt)01E(Yt),and
solvingforE(Yt)yieldstheresult.

14.3. (a) Totestforastochastictrend(unitroot)inln(IP),theADFstatisticisthetstatistictesting

thehypothesisthatthecoefficientonln(IPt1)iszeroversusthealternativehypothesisthatthe
0.018
t 2.5714.
coefficientonln(IPt1)islessthanzero.Thecalculatedtstatisticis 0.007
FromTable14.4,the10%criticalvaluewithatimetrendis3.12.Because2.57143.12,
thetestdoesnotrejectthenullhypothesisthatln(IP)hasaunitautoregressiverootatthe10%
significancelevel.Thatis,thetestdoesnotrejectthenullhypothesisthatln(IP)containsa
stochastictrend,againstthealternativethatitisstationary.
(b) TheADFtestsupportsthespecificationusedinExercise14.2.Theuseoffirstdifferencesin
Exercise14.2eliminatesrandomwalktrendinln(IP).

E[(W c )2 ] E{[W W ) ( W c )]2 }

E[(W W ) 2 ] 2 E (W W )( W c) ( W c) 2
W ( W c) .
2 2
14.5. (a)
(b) Usingtheresultinpart(a),theconditionalmeansquarederror
 E[(Yt f t 1 ) 2 | Yt 1 , Yt 2 ,...] t2|t 1 (Yt |t 1 f t 1 ) 2

t2|t 1 E[(Yt Yt |t 1 )2 ].
withtheconditionalvariance Thisequationisminimizedwhenthe
f Y .
secondtermequalszero,orwhen t 1 t |t 1 (Analternativeistousethehint,andnoticethat
theresultfollowsimmediatelyfromexercise2.27.)
(c) ApplyingEquation(2.27),weknowtheerrorutisuncorrelatedwithut1ifE(ut |ut1)0.
FromEquation(14.14)fortheAR(p)process,wehave
ut 1 Yt 1 0 1Yt 2 2Yt 3 L pYt p 1 f (Yt 1 , Yt 2 ,..., Yt p 1 ),

afunctionofYt1anditslaggedvalues.Theassumption E (ut |Yt 1 , Yt 2 ,...) 0 meansthat

conditionalonYt1anditslaggedvalues,oranyfunctionsofYt1anditslaggedvalues,uthas
meanzero.Thatis,
E (ut | ut 1 ) E[ut | f (Yt 1 , Yt 2 ,..., Yt p 2 )] 0.

Thusutandut1areuncorrelated.Asimilarargumentshowsthatutandutjareuncorrelated
forallj1.Thusutisseriallyuncorrelated.

14.7. (a) FromExercise(14.1)E(Yt)2.50.7E(Yt1)E(ut),butE(Yt)E(Yt1)(stationarity)and

E(ut)0,sothatE(Yt)2.5/(10.7).Also,becauseYt2.50.7Yt1ut,var(Yt)
0.72var(Yt1)var(ut)20.7cov(Yt1,ut).Butcov(Yt1,ut)0andvar(Yt)var(Yt1)
(stationarity),sothatvar(Yt)9/(10.72)17.647.
 (b) The1stautocovarianceis
cov(Yt , Yt 1 ) cov(2.5 0.7Yt 1 ut , Yt 1 )
0.7 var(Yt 1 ) cov(ut , Yt 1 )
0.7 Y2
0.7 17.647 12.353.

The2ndautocovarianceis
cov(Yt , Yt 2 ) cov[(1 0.7)2.5 0.72 Yt 2 ut 0.7ut 1 , Yt 2 ]
0.72 var(Yt 2 ) cov(ut 0.7ut 1 , Yt 2 )
0.72 Y2
0.72 17.647 8.6471.

(c) The1stautocorrelationis
cov(Yt , Yt 1 ) 0.7 Y2
corr (Yt , Yt 1 ) 0.7.
var(Yt ) var(Yt 1 ) Y2

The2ndautocorrelationis
cov(Yt , Yt 2 ) 0.7 2 Y2
corr (Yt , Yt 2 ) 0.49.
var(Yt ) var(Yt 2 ) Y2

(d) Theconditionalexpectationof YT 1 givenYTis

2.50.7YT2.50.7102.374.11.
YT 1/T

14.9. (a) E(Yt)0E(et)b1E(et1)bqE(etq)0[becauseE(et)0forallvaluesoft].

var(Yt ) var(et ) b12 var(et 1 ) L bq2 var(et q )

2b1 cov(et , et 1 ) L 2bq 1bq cov(et q 1 , et q )
e2 (1 b12 L bq2 )
(b)

wherethefinalequalityfollowsfromvar(et) e foralltandcov(et,ei)0forit.
2
 (c) Yt0etb1et1b2et2bqetqandYtj0etjb1et1jb2et2j
q q b b cov(et k , et j m ),
bqetqjandcov(Yt,Ytj) k 0 m 0 k m whereb01.Noticethat
cov(etk,etjm)0foralltermsinthesum.
var(Yt ) e2 1 b12 , cov(Yt , Yt j ) e2b1 , cov(Yt , Yt j ) 0
(d) and forj1.

14.11. WritethemodelasYtYt101(Yt1Yt2)ut.Rearrangingyields
 Yt0(11)Yt11Yt2ut.

Chapter15
EstimationofDynamicCausalEffects

15.1. (a) Seethetablebelow.iisthedynamicmultiplier.Withthe25%oilpricejump,the

predictedeffectonoutputgrowthfortheithquarteris25ipercentagepoints.

Dynamic 95%confidence
Periodahead multiplier Predictedeffecton interval25[ i1.96SE
(i) ( i) outputgrowth(25 i) ( i)]
0 0.055 1.375 [4.021,1.271]

1 0.026 0.65 [3.443,2.143]

2 0.031 0.775 [3.127,1.577]
3 0.109 2.725 [4.783,0.667]
4 0.128 3.2 [5.797,0.603]
5 0.008 0.2 [1.025,1.425]
6 0.025 0.625 [1.727,2.977]
7 0.019 0.475 [2.386,1.436]
8 0.067 1.675 [0.015,0.149]
 (b) The95%confidenceintervalforthepredictedeffectonoutputgrowthfortheithquarterfrom
the25%oilpricejumpis25[i1.96SE(i)]percentagepoints.Theconfidenceintervalis
reportedinthetablein(a).

(c) ThepredictedcumulativechangeinGDPgrowthovereightquartersis
25(0.0550.0260.0310.1090.1280.0080.0250.019)8.375%.

(d) The1%criticalvaluefortheFtestis2.407.SincetheHACFstatistic3.49islargerthanthe
criticalvalue,werejectthenullhypothesisthatallthecoefficientsarezeroatthe1%level.

15.3. ThedynamiccausaleffectsareforexperimentA.Theregressioninexercise15.1doesnotcontrol
forinterestrates,sothatinterestratesareassumedtoevolveintheirnormalpatterngivenchanges
inoilprices.

15.5. Substituting

X t X t X t 1
X t X t 1 X t 2
L L
X t X t 1 L X t p 1 X t p

intoEquation(15.4),wehave

Yt 0 1 X t 2 X t 1 3 X t 2 L r 1 X t r ut
0 1 ( X t X t 1 L X t r 1 X t r )
2 ( X t 1 L X t r 1 X t r )
L r ( X t r 1 X t r ) r 1 X t r ut
0 1 X t ( 1 2 ) X t 1 ( 1 2 3 ) X t 2
L ( 1 2 L r ) X t r 1
( 1 2 L r r 1 ) X t r ut .

ComparingtheaboveequationtoEquation(15.7),wesee 00,11,212,
3123,,andr112rr1.

Write ut i 0 1 ut i

15.7.
i
%
 (a) Because E (u% i | X t ) 0 foralliandt,E(u |X )0foralliandt,sothatX isstrictlyexogenous.
i t t

t 1 ) 0
E (ut j | u%
forj0,Xtisexogenous.HoweverE(ut+1 | u% %
t 1 ) ut 1 sothatX is
(b) Because t
notstrictlyexogenous.

15.9. (a) Thisfollowsfromthematerialaroundequation(3.2).

(b) QuasidifferencingtheequationyieldsYt1Yt1(11)0 u% t , andtheGLSestimatorof

1
T (Y Y )
(11)0isthemeanofYt1Yt1 T 1 t 2 t 1 t 1 .Dividingby(11)yieldstheGLS
estimatorof0.
(c) Thisisarearrangementoftheresultin(b).
1 T Y 1 (Y Y1 ) TT1 T11 Tt 21 Yt , 0GLS T1 (YT Y1 ) T1 T11 Tt 21 Yt
(d) Write 0 T t 1 t T T sothat 0
1
.
1 T 1 (YT Y1 )
1 1
2
andthevarianceisseentobeproportionalto T

Chapter16
AdditionalTopicsinTime
SeriesRegression

0
Y E (Yt ) ,
16.1. Y followsastationaryAR(1)model, Yt 0 1Yt 1 ut . ThemeanofY is 1 1
t t

and E (ut |Yt ) 0.

 Yt , Yt h|t E (Yt h |Yt , Yt 1 ,K ),
(a) Thehperiodaheadforecastof is
Yt h|t E (Yt h |Yt ,Yt 1 ,K )
E ( 0 1Yt h 1 ut |Yt , Yt 1 ,K )
0 1Yt h 1|t
0 1 ( 0 1Yt h 2|t )
(1 1 ) 0 12Yt h 2|t
(1 1 ) 0 12 ( 0 1Yt h 3|t )
(1 1 12 ) 0 13Yt h 3|t
L L
1 1 L 1h 1 0 1hYt
1 1h
0 1hYt
1 1
Y 1h (Yt Y ).
(b) Substitutingtheresultfrompart(a)intoXtgives

X t iYt i |t i [ Y 1i (Yt Y )]
i 0 i 0

Y i (Yt Y ) ( 1 )i
i 0 i 0

Y Y Y
t .
1 1 1

utfollowstheARCHprocesswithmeanE(ut)0andvariance t 1.0 0.5ut 1.

2 2
16.3.

(a) ForthespecifiedARCHprocess,uthastheconditionalmean E (ut |ut 1 ) 0 andthe

conditionalvariance.
var (ut | ut 1 ) t2 1.0 0.5ut21 .
TheunconditionalmeanofutisE(ut)0,andtheunconditionalvarianceofutis

var (ut ) var[ E (ut |ut 1 )] E[var (ut | ut 1 )]

0 1.0 0.5E (ut21 )
1.0 0.5var (ut 1 ).

Thelastequationhasusedthefactthat E (ut ) var(ut ) E (ut )] var(ut ), whichfollows

2 2

becauseE(ut)0.Becauseofthestationarity,var(ut1)var(ut).Thus,var(ut)1.0
0.5var(ut)whichimplies var(ut ) 1.0 / 0.5 2.
 (b) When ut 1 0.2, t 1.0 0.5 0.2 1.02. Thestandarddeviationofutist1.01.Thus
2 2

3 ut 3
Pr ( 3 ut 3) Pr
1.01 t 1.01
(2.9703) (2.9703) 0.9985 0.0015 0.9970.

Whenut12.0, t 1.0 0.5 2.0 3.0. Thestandarddeviationofutist1.732.Thus

2 2

3 u 3
Pr ( 3 ut 3) Pr t
1.732 t 1.732
(1.732) (1.732) 0.9584 0.0416 0.9168.

16.5. Because Yt Yt Yt 1 Yt 1 Yt 1 Yt ,
T T T T T

Yt 2 (Yt 1 Yt )2 Yt 21 (Yt )2 2Yt 1Yt .

t 1 t 1 t 1 t 1 t 1

So
1 T 1 1 T T T

Yt 1Yt T 2
T t 1
Y
t 1
t
2
Yt 21 (Yt ) 2 .
t 1 t 1

Tt 1 Yt 2 Tt 1 Yt 21 Tt 11 Yt 2 YT2 Y02 Tt 11 Yt 2 Y 2 Y 2 Y 2
Notethat T 0 T becauseY 0.Thus:
0

1 T 1 1 2 T

T t 1
Yt 1 Yt YT (Yt )2
T 2 t 1
1 YT
2
1 T

2 T
T t 1
(Yt ) 2 .

1 T

T
Y Xt Y Y
T
t 1Yt Yt 1
T
t 1 t t 1 t t 1
.
Y 1 T
T T 2
X t2
t 1
2
t 1 t 1 t 1 Y
16.7. T t 1 Followingthehint,thenumeratoristhesame
u2
1
Tt 1 Yt Yt 1
d
( 12 1).
expressionas(16.21)(shiftedforwardintime1period),sothat T 2 The
Tt 1 (Yt 1 ) 2 T1 Tt 1 ut21 u2
1 p
denominatoris T bythelawoflargenumbers.Theresultfollows
directly.

16.9. (a) Fromthelawofiteratedexpectations

 E (ut2 ) E t2
E 0 1ut21
0 1 E ut21
0 1 E ut2
2
wherethelastlineusesstationarityofu.Solvingfor E (ut ) givestherequiredresult.

(b) Asin(a)

E (ut2 ) E t2
E 0 1ut21 2ut2 2 L put2 p
0 1 E ut21 2 E ut2 2 L p E ut2 p
0 1 E ut2 2 E ut2 L p E ut2

0
E (ut2 )
sothat 1 tp1 i .

2
(c) Thisfollowsfrom(b)andtherestrictionthat E (ut ) >0.
(d) Asin(a)

E (ut2 ) E t2
0 1 E ut21 1 E t21
0 (1 1 ) E ut21
0 (1 1 ) E ut2
0

1 1 1

E ut2 0.
(e) Thisfollowsfrom(d)andtherestrictionthat
Chapter17
TheTheoryofLinearRegression
withOneRegressor

17.1. (a) Supposetherearenobservations.Letb1beanarbitraryestimatorof1.Giventhe

estimatorb1,thesumofsquarederrorsforthegivenregressionmodelis
n

(Y b X ) .
i 1
i 1 i
2

1RLS , therestrictedleastsquaresestimatorof ,minimizesthesumofsquarederrors.That

1
RLS
is, 1 satisfiesthefirstorderconditionfortheminimizationwhichrequiresthedifferential
ofthesumofsquarederrorswithrespecttob1equalszero:

2(Y b X )( X ) 0.
i 1
i 1 i i

Solvingforb1fromthefirstorderconditionleadstotherestrictedleastsquaresestimator

XYn
1RLS i n1 i 2 i .
i 1 X i

(b) Weshowfirstthat 1RLS isunbiased.Wecanrepresenttherestrictedleastsquaresestimator

1RLS intermsoftheregressorsanderrors:

n X Y n X ( X ui ) n X u
1RLS i n1 i 2 i i 1 i n 1 2i 1 i n1 i 2 i .
i 1 X i i 1 X i i 1 X i

Thus

in1 X i ui in1 X i E (ui | X 1,K , X n )

E ( 1RLS ) 1 E 2
1 E 1 ,
i 1 X i in1 X i2
n

wherethesecondequalityfollowsbyusingthelawofiteratedexpectations,andthethird
equalityfollowsfrom

in1 X i E (ui | X 1 ,K , X n )
0
in1 X i2
becausetheobservationsarei.i.d.andE(ui |Xi)0.(Note,E(ui |X1,,Xn)E(ui |Xi)because
theobservationsarei.i.d.
 RLS
Underassumptions13ofKeyConcept17.1, 1 isasymptoticallynormallydistributed.
RLS
Thelargesamplenormalapproximationtothelimitingdistributionof 1 followsfrom
considering
n X u 1
n X u
1RLS 1 i n1 i 2 i n1 i n1 i 2 i .
i 1 X i n i 1 X i

ConsiderfirstthenumeratorwhichisthesampleaverageofviXiui.Byassumption1ofKey
Concept17.1,v hasmeanzero: E ( X i ui ) E[ X i E (ui | X i )] 0. Byassumption2,v isi.i.d.By
i i

v 1 n
i 1 X i ui , then /n.
2 2
assumption3,var(vi)isfinite.Let n v v
Usingthecentrallimit
theorem,thesampleaverage

n
1
v / v
v n
v i 1
i
d
N (0, 1)

or

1 n

n i 1
X i ui
d
N (0, v2 ).

2
Forthedenominator, X i isi.i.d.withfinitesecondvariance(becauseXhasafinitefourth
moment),sothatbythelawoflargenumbers

1 n 2 p

n i 1
X i E ( X 2 ).

CombiningtheresultsonthenumeratorandthedenominatorandapplyingSlutskystheorem
leadto

1
in1 X i ui var( X i ui )
n ( 1RLS u ) n

d
N 0, .
1
n n
i 1 X i
2
E( X 2 )
RLS
(c) 1 isalinearestimator:

n X Y Xi
1RLS i n1 i 2 i i 1 aiYi ,
n
where ai .
i 1 X i in1 X i2

Theweightai(i1,,n)dependsonX1,,XnbutnotonY1,,Yn.

Thus
 Xun
1RLS 1 i n1 i 2 i .
i 1 X i
 1RLS isconditionallyunbiasedbecause

n X u
E ( 1RLS | X 1 ,K , X n E 1 i n1 i 2 i | X 1 ,K , X n
i 1 X i
Xu
n

1 E i n1 i 2 i |X 1 ,K , X n 1 .
i 1 X i
Thefinalequalityusedthefactthat

in1 X i ui in1 X i E (ui | X 1 ,K , X n )

E | X 1 , K , X n 0
i 1 X i in1 X i2
n 2

becausetheobservationsarei.i.d.andE(ui |Xi)0.

RLS
(d) Theconditionalvarianceof 1 , givenX1,,Xn,is
n X u
var( 1RLS | X1,K , X n ) var 1 i n1 i 2 i | X1 ,K , X n
i 1 Xi
X var(u | X ,K , X n )
n 2
i 1 i n i 2 12
( i 1 Xi )
in1 X i2 u2

( in1 Xi2 )2
2
u
.
n
i 1X 2
i

(e) TheconditionalvarianceoftheOLSestimator 1 is

u2
var( 1 |X1 ,K , X n ) .
in1 ( X i X )2
Since

n n n n n

( X i X )2 X i2 2 X X i nX 2 X i2 nX 2 X i2 ,
i 1 i 1 i 1 i 1 i 1

RLS
theOLSestimatorhasalargerconditionalvariance: var( 1 |X 1 ,K , X n ) var( 1 |X 1 ,K , X n ).
RLS
Therestrictedleastsquaresestimator 1 ismoreefficient.
 RLS
(f) Underassumption5ofKeyConcept17.1,conditionalonX1,,Xn, 1 isnormally
distributedsinceitisaweightedaverageofnormallydistributedvariablesui:
Xun
1RLS 1 i n1 i 2 i .
i 1 X i
 RLS
Usingtheconditionalmeanandconditionalvarianceof 1 derivedinparts(c)and(d)
RLS
respectively,thesamplingdistributionof 1 ,conditionalonX ,,X ,is 1 n

2
1RLS ~ N 1 , n u 2 .
i 1 X i
(g) Theestimator

in1 Yi in1 ( 1 X i ui ) in1 ui

%
1 1
in1 X i in1 X i in1 X i
Theconditionalvarianceis

in1 ui
var( %
1 | X 1 ,K , X n ) var 1 | X 1 ,K , X n
i 1 X i
n

var(ui | X 1 ,K , X n )
n
i 1
( in1 X i ) 2
n u2
.
( in1 X i ) 2
% RLS
Thedifferenceintheconditionalvarianceof 1 and 1 is

n u2 2
var( % RLS | X ,K , X )
1 | X 1 ,K , X n ) var( 1 1 n n u 2.
( X i )
n
i 1
2
i 1 X i
% RLS
Inordertoprove var( 1| X 1 ,K , X n ) var( 1 | X 1 ,K , X n ), weneedtoshow

n 1
n
( X i )
n 2
i 1 i 1 X i2
orequivalently

2
n
n

n X i2 X i .
i 1 i 1
ThisinequalitycomesdirectlybyapplyingtheCauchySchwartzinequality
 2
n
n n

( ai bi ) ai2 bi2
i 1 i 1 i 1

whichimplies

2 2
n
n n n n

X i 1 Xi 12 X i2 n X i2 .
i 1 i 1 i 1 i 1 i 1

% RLS
Thatis ni 1 X i ( x 1 X i ) , or var( 1| X 1 ,K , X n ) var( 1 | X 1 ,K , X n ).
n 2 n 2

%
Note:because 1 islinearandconditionallyunbiased,theresult
var( % RLS | X ,K , X )
1 | X 1 ,K , X n ) var( 1 1 n followsdirectlyfromtheGaussMarkovtheorem.
17.3. (a) UsingEquation(17.19),wehave

1
n ( X X )ui
n ( 1 1 ) n n1 in1 i
n
i 1 ( X i X )2
1
in1[( Xi X ) ( X X )]ui
n n

n i 1 ( X i X )
1 n 2

1
in1 ( X i X )ui (X X ) 1
in1 ui
n
n
1
n
in1 ( X i X )2 1
n
in1 ( X i X )2
1
in1 vi ( X X ) 1
in1 ui
n
n
1
n
in1 ( X i X )2 1
n
in1 ( X i X )2

bydefiningvi(XiX)ui.

(b) Therandomvariablesu1,,unarei.i.d.withmeanu0andvariance 0 u . Bythe

2

centrallimittheorem,

n (u u ) 1
in1 ui
n

d
N (0, 1).
u u

X X 2 , or X X 0.
p p

Thelawoflargenumbersimplies Bytheconsistencyofsample

1 n
( X X ) 2
variance, n i 1 i convergesinprobabilitytopopulationvariance,var(Xi),whichis
finiteandnonzero.TheresultthenfollowsfromSlutskystheorem.

(c) Therandomvariablevi(XiX)uihasfinitevariance:
var(vi ) var[( X i X ) i ]
E[( X i X ) 2 ui2 ]
E[( X i X ) 4 ]E[(ui ) 4 ] .

TheinequalityfollowsbyapplyingtheCauchySchwartzinequality,andthesecondinequality
followsbecauseofthefinitefourthmomentsfor(Xi,ui).Thefinitevariancealongwiththefact
thatvihasmeanzero(byassumption1ofKeyConcept15.1)andviisi.i.d.(byassumption2)
impliesthatthesampleaverage v satisfiestherequirementsofthecentrallimittheorem.Thus,

v 1
in1 vi
n

v v
satisfiesthecentrallimittheorem.

(d) Applyingthecentrallimittheorem,wehave
1
in1 vi
n

d
N (0, 1).
v
 Becausethesamplevarianceisaconsistentestimatorofthepopulationvariance,wehave
1
in1 ( X i X ) 2 p
n
1.
var( X i )
UsingSlutskystheorem,

in1 vt
1
n
v

d
N (0,1),
1
n n
i 1 ( X t X ) 2

X2
orequivalently

1
in1 vi var(vi )
n

d
N 0, .
1
n (Xi X )
n
i 1
2
[var( X i )]2

Thus

1
in1 vi ( X X ) 1
in1 ui
n ( 1 1 ) n
n
1
n in1 ( X i X )2 1
n in1 ( X i X )2
var(vi )

d
N 0,
[var( X i )]2

sincethesecondtermfor n ( 1 1 ) convergesinprobabilitytozeroasshowninpart(b).

17.5. BecauseE(W 4)[E(W2)]2var(W2),[E(W2)]2E(W 4).ThusE(W2)<.

17.7. (a) Thejointprobabilitydistributionfunctionofui,uj,Xi,Xjisf(ui,uj,Xi,Xj).Theconditional

probabilitydistributionfunctionofuiandXigivenujandXjisf(ui,Xi |uj,Xj).Sinceui,Xi,i
1,,narei.i.d.,f(ui,Xi |uj,Xj)f(ui,Xi).Bydefinitionoftheconditionalprobability
distributionfunction,wehave
 f (ui , u j , X i , X j ) f (ui , X i | u j , X j ) f (u j , X j )
f (ui , X i ) f (u j , X j ).

(b) TheconditionalprobabilitydistributionfunctionofuiandujgivenXiandXjequals

f (ui , u j , X i , X j ) f (ui , X i ) f (u j , X j )
f (ui , u j | X i , X j ) f (ui | X i ) f (u j | X j ).
f (Xi , X j ) f (Xi ) f (X j )

Thefirstandthirdequalitiesusedthedefinitionoftheconditionalprobabilitydistribution
function.Thesecondequalityusedtheconclusionthefrompart(a)andtheindependence
betweenXiandXj.Substituting

f (ui , u j | Xi , X j ) f (ui | X i ) f (u j | X j )
 intothedefinitionoftheconditionalexpectation,wehave

E (ui u j | X i , X j ) ui u j f (ui , u j |X i , X j ) dui du j

ui u j f (ui | X i) f (u j | X j )dui du j
ui f (ui | X i) dui u j f (u j | X j ) du j
E (ui | X i ) E (u j | X j ).

(c) LetQ(X1,X2,,Xi1,Xi+1,,Xn),sothatf(ui|X1,,Xn)f(ui |Xi,Q).Write

f (ui , X i , Q)
f (ui | X i , Q)
f ( X i , Q)
f (ui , X i ) f (Q )

f ( X i ) f (Q)
f (ui , X i )

f (Xi )
f (ui | X i )
wherethefirstequalityusesthedefinitionoftheconditionaldensity,thesecondusesthefact
that(ui,Xi)andQareindependent,andthefinalequalityusesthedefinitionoftheconditional
density.Theresultthenfollowsdirectly.

(d) Anargumentlikethatusedin(c)implies
f (ui u j | X i , K X n ) f (ui u j | X i , X j )

andtheresultthenfollowsfrompart(b).

17.9. Weneedtoprove
1 n
[( X i X )2 ui2 ( X i X )2 ui2 ] 0.
p

n i 1

Usingtheidentity X X ( X X ),
 1 n 1 n 1 n

n i 1
[( X i X ) 2 ui2 ( X i X ) 2 ui2 ] ( X X ) 2 ui2 2( X X ) ( X i X )ui2
n i 1 n i 1
1 n
( X i X ) 2 (ui2 ui2 ).
n i 1

Thedefinitionof ui implies

ui2 ui2 ( 0 0 ) 2 ( 1 1 )2 X i2 2ui ( 0 0 )

2ui ( 1 1 ) X i 2( 0 0 )(1 1 ) X i .

1
in1[( X i X ) 2 ui2 ( X i X ) 2 ui2 ]
Substitutingthisintotheexpressionfor n yieldsaseriesofterms
p
b X u
eachofwhichcanbewrittenasanbnwhere an 0 and n
1 n r s
whererandsare
n i 1 i i

integers.Forexample, an ( X X ), an ( 1 1 ) andsoforth.Theresultthenfollowsfrom
p
in1 X ir uis d
wheredisafiniteconstant.Let wi X i ui andnotethatwi
1 r s
Slutksystheoremif n
isi.i.d.Thelawoflargenumberscanthenbeusedforthedesiredresultif E ( wi ) . Thereare
2

twocasesthatneedtobeaddressed.Inthefirst,bothrandsarenonzero.Inthiscasewrite

E ( wi2 ) E ( X i2 r ui2 s ) [ E ( X i4 r )][ E (ui4 s )]

andthistermisfiniteifrandsarelessthan2.Inspectionofthetermsshowsthatthisistrue.In
thesecondcase,eitherr0ors0.Inthiscasetheresultfollowsdirectlyifthenonzero
exponent(rors)islessthan4.Inspectionofthetermsshowsthatthisistrue.

17.11. Note:inearlyprintingofthethirdeditiontherewasatypographicalerrorintheexpressionfor Y|X.

Y ( XY / X2 )( x X )
Thecorrectexpressionis Y | X .
(a) Usingthehintandequation(17.38)
1
fY | X x ( y )
(1 XY
2
Y
2
)
1 x 2
x X y Y y Y
2
1 x X
2

exp X
2 XY .
2(1 XY ) X
2
X Y Y 2 X

Simplifyingyieldsthedesiredexpression.
(b) TheresultfollowsbynotingthatfY|X=x(y)isanormaldensity(seeequation(17.36))withT|
Y2|X
Xand 2
.

(c) LetbXY/ X andaYbX.

2

17.13 (a) Theanswerisprovidedbyequation(13.10)andthediscussionfollowingtheequation.

TheresultwasalsoshowninExercise13.10,andtheapproachusedintheexerciseis
discussedinpart(b).
(b) WritetheregressionmodelasYi01Xivi,where0E(0i),1E(1i),andviui
(0i0)(1i1)Xi.Noticethat
E(vi|Xi)E(ui|Xi)E(0i0|Xi)XiE(1i1|Xi)0
because0iand1iareindependentofXi.BecauseE(vi|Xi)=0,theOLSregressionofYionXi
willprovideconsistentestimatesof0E(0i)and1E(1i).Recallthattheweightedleast
i 0 1 X i2
squaresestimatoristheOLSestimatorofYi/ionto1/iandXi/i,where .
Writethisregressionas
Yi / i 0 (1 / i ) 1 ( X i / i ) vi / i .

Thisregressionhastworegressors,1/iandXi/i.Becausetheseregressorsdependonlyon
Xi,E(vi|Xi)0impliesthatE(vi/i|(1/i),Xi/i)0.Thus,weightedleastsquaresprovidesa
consistentestimatorof0E(0i)and1E(1i).
Chapter18
TheTheoryofMultipleRegression

18.1. (a) Theregressioninthematrixformis

YX U
with
TestScore1 1 Income1 Income12

TestScore 2 1 Income 2 Income 22
Y , X
M M M M

TestScore n 1 Income n Income 2n

U1
0
U
U 2 , 1 .
M
2
U n
(b) ThenullhypothesisisH0:R rversusH1:Rr,with
R = (0 0 1) and r = 0.
TheheteroskedasticityrobustFstatistictestingthenullhypothesisis
1
F ( R r) R R ( R r)/q

Withq1.Underthenullhypothesis,
F
d
Fq ,
.
WerejectthenullhypothesisifthecalculatedFstatisticislargerthanthecriticalvalueofthe
Fq ,
distributionatagivensignificancelevel.

Var (Q) E[(Q Q ) 2 ]

E[(Q Q )(Q Q )]
E[(cW c W )(cW c W )]
cE[(W W ) ( W W )]c
18.3. (a) c var( W ) c c w c
wherethesecondequalityusesthefactthatQisascalarandthethirdequalityusesthefact
thatQc w.

(b) Becausethecovariancematrix W ispositivedefinite,wehave c wc 0 foreverynon
zerovectorfromthedefinition.Thus,var(Q)>0.Boththevectorcandthematrix W are

finite,sovar(Q) c wc isalsofinite.Thus,0<var(Q)<.

18.5. PXX(XX)1X,MXInPX.

(a)PXisidempotentbecause
 PXPXX(XX)1XX(XX)1X X(XX)1XPX.
MXisidempotentbecause
M X M X (I n PX ) (I n PX ) I n PX PX PX PX
I n 2PX PX I n PX M X

PXMX0nxnbecause

PX M X PX (I n PX ) PX PX PX PX PX 0n n

1
(b) Because ( X X) X Y, wehave
X X (XX )1 X Y P Y
Y X

whichisEquation(18.27).Theresidualvectoris
YY
U Y P Y (I P ) Y M Y.
X n X X

WeknowthatMXXisorthogonaltothecolumnsofX:
MXX(InPX)XXPXXXX(XX)1XXXX0
sotheresidualvectorcanbefurtherwrittenas
M Y M ( X U ) M X M U M U
U X X X X X

whichisEquation(18.28).

18.7. (a) Wewritetheregressionmodel,Yi1Xi2Wiui,inthematrixformas

YXWU
with
Y1 1
X W1 u1

Y X W u2
Y 2 , X 2 , W 2 , U ,
M M M M

Yn X n Wn un
1 , 2 .
TheOLSestimatoris
1 1
X X XW X Y

W X WW
WY

2
1
X X X W X U
1
2 W X W W W U
1
1 1
X X 1
X W 1
X U
n n

n

2
1
n
WX 1
n
W W 1
n
W U
1
1
in1 X i2 1
in1 X iWi 1
n
in1 X i ui
1 n n
2 1 n
2 W u
n i 1 Wi X i n i 1 Wi
1 n 1 n
n i 1 i i
 1
in1 X i2
p
E ( X 2 ); 1n ni 1 Wi 2
p
E (W 2 ); 1
in1 X iWi
p
Bythelawoflargenumbers n n
E ( XW ) 0 (becauseXandWareindependentwithmeansofzero); n i 1 X i ui E ( Xu ) 0
1 n p

1
in1 X i ui
p
E ( Xu ) 0
(becauseXanduareindependentwithmeansofzero); n Thus
1
1 p 1 E( X 2 ) 0 0

2 0 2
E (W ) E (Wu )
2
1

2 EE ((WWu2 ))
.


E (Wu )
2
p
2 2
(b) Fromtheanswerto(a) E (W 2 ) ifE(Wu)isnonzero.
(c) ConsiderthepopulationlinearregressionuiontoWi:
uiWiai
whereE(Wu)/E(W2).Inthispopulationregression,byconstruction,E(aW)0.Usingthisequationfor
uirewritetheequationtobeestimatedas

Yi X i 1 Wi 2 ui
X i 1 Wi ( 2 ) ai
X i 1 Wi ai

where 2 . Acalculationlikethatusedinpart(a)canbeusedtoshowthat

1
n ( 1 1 ) 1
in1 X i2 1
in1 X iWi 1
in1 X i ai
n n
n

n ( 2 )
1
in1 Wi X i 1
in1 Wi 2 1 in1 Wu ai
n n n
1
E( X 2 ) 0 S1

d
2
0 E (W ) S 2
 whereS1isdistributed N (0, a E ( X 2 )). ThusbySlutskystheorem
2

a2
n ( 1 1 )
d
N 0,
E ( X 2 )
NowconsidertheregressionthatomitsW,whichcanbewrittenas:
Yi X i 1 d i
wherediWiai.Calculationslikethoseusedaboveimplythat
d2

n 1r 1
d

N 0,
E ( X 2 )
.

r
Since d a E (W ), theasymptoticvarianceof 1 isneversmallerthantheasymptotic
2 2 2 2

varianceof 1 .

( XM W X)1 XM W Y
( XM W X) 1 XM W ( X W U )
( XM W X) 1 XM W U.
18.9. (a)
ThelastequalityhasusedtheorthogonalityMWW0.Thus

( XM W X)1 XM W U (n 1 XM W X)1 (n 1 XM W U ).
(b) UsingMWInPWandPWW(WW)1Wwecanget

n 1XM W X n1X(I n PW )X
n 1XX n 1XPW X
n 1XX ( n 1XW)( n 1W W ) 1 ( n 1WX).
n 1XX 1n in1 X i Xi . 1
n X X .
Firstconsider The(j,l)elementofthismatrixis n i 1 ji li By
Assumption(ii),Xiisi.i.d.,soXjiXliisi.i.d.ByAssumption(iii)eachelementofXihasfour
moments,sobytheCauchySchwarzinequalityXjiXlihastwomoments:
E ( X 2ji X li2 ) E ( X 4ji ) E ( X li4 ) .
1
n X X
BecauseXjiXliisi.i.d.withtwomoments, n i 1 ji li obeysthelawoflargenumbers,so
1 n

n i 1
p
X ji X li E ( X ji X li ) .

Thisistrueforalltheelementsofn1XX,so
1 n
n 1XX Xi Xi p E (Xi Xi ) XX. .
n i 1
ApplyingthesamereasoningandusingAssumption(ii)that(Xi,Wi,Yi)arei.i.d.and
Assumption(iii)that(Xi,Wi,ui)havefourmoments,wehave
 1 n
n 1WW
n i 1
Wi Wi E ( Wi Wi ) WW ,
p

1 n
n 1 XW Xi Wi E ( Xi Wi ) XW ,
p

n i 1
and
1 n
n 1 WX
n i 1
Wi Xi E ( Wi Xi ) WX .
p

FromAssumption(iii)weknow XX , WW , XW , and WX areallfinitenonzero,Slutskys

theoremimplies
n 1XM W X n1XX ( n 1XW ) (n1W W )1 ( n1W X)
p
XX XW -1WW WX
whichisfiniteandinvertible.

(c) Theconditionalexpectation
E (u1| X, W ) E (u1| X1 , W 1 )

E (u2 | X, W ) E (u2 | X 2 , W2)
E (U|X, W )
M M

E (un | X, W) E (un | X n , W n )
W1 W 1

W W2
2 W .
M M

Wn W n

ThesecondequalityusedAssumption(ii)that ( X i , Wi , Yi ) arei.i.d.,andthethirdequality
appliedtheconditionalmeanindependenceassumption(i).
(d) Inthelimit
n 1 XM W U
p
E ( XM W U|X, W) XM W E (U|X, W) XM W W 0 k1 1

because M W W 0.
1
1

(e) n X M W X convergesinprobabilitytoafiniteinvertiblematrix,and n X M W U converges
inprobabilitytoazerovector.ApplyingSlutskystheorem,

(n 1XM W X) -1 (n1XM W U )
p
0.
Thisimplies

p
.

Ir 0 Q1 '
0 0 Q '
18.11. (a) UsingthehintC[Q1Q2] 2 ,whereQQI.TheresultfollowswithAQ1.
 (b) WAVN(A0,AInA)andtheresultfollowsimmediately.
(c) VCVVAAV(AV)(AV)WWandtheresultfollowsfrom(b).

18.13. (a) ThisfollowsfromthedefinitionoftheLagrangian.

(b) Thefirstorderconditionsare
%
(*)X(YX )R 0
and
%
(**)R r0
Solving(*)yields
%
(***) (XX)1R .

MultiplyingbyRandusing(**)yieldsrR R(XX)1R ,sothat

[R(X X)1R ]1(R r).
Substitutingthisinto(***)yieldstheresult.
%
(c) Usingtheresultin(b),YX (YX )X(XX)1R[R(XX)1R]1(R r), sothat
% %
(YX ) (Y X )(Y X )(YX )(R r) [R(X X)1R ]1(R r)

2(YX )X(XX)1R[R(XX)1R]1(R r).

But(YX )X0,sothelasttermvanishes,andtheresultfollows.

(d) Theresultin(c)showsthat(R r)[R(XX)1R]1(R r)SSRRestrictedSSRUnrestricted.Also
su2 SSR /(nkUnrestricted1),andtheresultfollowsimmediately.
Unrestricted

18.15. (a) Thisfollowsfromexercise(18.6).

% % %
(b) Yi Xi u i ,sothat
1 n
n
%

i 1
%' X
X i i

X%' u%
i 1
i i

1 n
n
%

i 1
%' X
X i i

X ' M ' Mu
i 1
i i

1 n
n
%

i 1
%' X
X i i

X ' M 'u
i 1
i i

1 n
n
%

i 1
%' X
X i i

X%' u
i 1
i i
 X% 1n in1 (T 1 Tt 1 ( X it X i ) 2 ), 1
where (T t 1 ( X it X i ) ) arei.i.d.withmean X%and
Q T 2
Q
(c)
finitevariance(becauseXithasfinitefourthmoments).Theresultthenfollowsfromthelawof
largenumbers.
(d) ThisfollowstheCentrallimittheorem.
(e) ThisfollowsfromSlutskystheorem.

(f) i arei.i.d.,andtheresultfollowsfromthelawoflargenumbers.
2

1/ 2 % 1/ 2 % %
(g) Let i T X i ' u%i i T ( ) X i ' X i .Then

%' u% 2 T 1 ( )2 ( X
i2 T 1/2 X %' X%)2 2T 1/ 2 ( ) X
% %
i i i i i i i ' Xi

1 %' X%) 2 2T 1/ 2 ( ) 1 n X % %
in1 i2 in1 i2 T 1 ( )2 1n in1 ( X
1
n i i n i 1 i i ' X i
and n

n ( X%' X%)2 E[( X%i ' X%i ) 2 ]

p
p
Because ( ) 0 ,theresultfollowsfrom(a) n i 1 i i
1
and(b)
in1 i X% % % %
p
i ' X i E (i X i ' X i ).
1
n Both(a)and(b)followfromthelawoflargenumbers;both
(a)and(b)areaveragesofi.i.d.randomvariables.Completingtheproofrequiresverifying
% %2 % %
that ( Xi ' Xi ) hastwofinitemomentsand i Xi ' Xi hastwofinitemoments.Theseinturn
followfrom8momentassumptionsfor(Xit,uit)andtheCauchySchwartzinequality.
Alternatively,astronglawoflargenumberscanbeusedtoshowtheresultwithfinite
fourthmoments.

18.17 Theresultsfollowfromthehintsandmatrixmultiplicationandaddition.