1 Factor Analysis and Factor Analysis and Inference for Structured Inference for Structured Covariance Matrices Covariance Matrices Shyh-Kang Jeng Shyh-Kang Jeng Department of Electrical Engineeri Department of Electrical Engineeri ng/ ng/ Graduate Institute of Communicatio Graduate Institute of Communicatio n/ n/ Graduate Institute of Networking a Graduate Institute of Networking a nd Multimedia nd Multimedia
83
Embed
1 Factor Analysis and Inference for Structured Covariance Matrices Shyh-Kang Jeng Department of Electrical Engineering/ Graduate Institute of Communication
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
11
Factor Analysis and Inference for Factor Analysis and Inference for Structured Covariance MatricesStructured Covariance Matrices
Shyh-Kang JengShyh-Kang JengDepartment of Electrical Engineering/Department of Electrical Engineering/Graduate Institute of Communication/Graduate Institute of Communication/
Graduate Institute of Networking and MultiGraduate Institute of Networking and Multimediamedia
22
HistoryHistory
Early 20Early 20thth-century attempt to define -century attempt to define and measure intelligenceand measure intelligence
Developed primarily by scientists Developed primarily by scientists interested in psychometricsinterested in psychometrics
Advent of computers generated a Advent of computers generated a renewed interestrenewed interest
Each application must be examined Each application must be examined on its own meritson its own merits
33
Essence of Factor AnalysisEssence of Factor AnalysisDescribe the covariance among Describe the covariance among many variables in terms of a few many variables in terms of a few underlying, but unobservable, underlying, but unobservable, random random factorsfactors..A group of variables highly correlated A group of variables highly correlated among themselves, but having among themselves, but having relatively small correlations with relatively small correlations with variables in different groups variables in different groups represent a single underlying represent a single underlying factorfactor
44
Example 9.8Example 9.8Examination ScoresExamination Scores
55
Orthogonal Factor ModelOrthogonal Factor Model
variablesrandom leunobservab
:,,,,,,,
factorth on the eth variabl theof loading:
2121
2211
2222212122
1121211111
pm
ij
pmpmpppp
mm
mm
FFF
ji
FFFX
FFFX
FFFX
εFLμX
66
Orthogonal Factor ModelOrthogonal Factor Model
loadingfactor ofmatrix :
components th vector wirandom
variationof source factors, specificor errors :
components th vector wirandom
oft independen factors,common :
ofmatrix covariance andmean :,
components th vector wirandom :
L
ε
XF
XΣμ
X
p
m
p
77
Orthogonal Factor ModelOrthogonal Factor Model
0εFFε
Ψ
Ψεεεε
IFFFF
)'(),Cov(
00
00
00
)'()Cov(,0)(
)'()Cov(,0)(
2
1
E
EE
EE
p
88
Orthogonal Factor ModelOrthogonal Factor Model
LεFFFLFμXFX
εFLFFFεLFFμX
ΨLL
εεFεLLεFLFFL
μXμXXΣ
εεLFεLεFLLFF
εLFεLFμXμX
)'()'('),Cov(
''''
'
)'()'(')'(')'(
')Cov(
''''''
''
EEE
EEEE
E
99
Orthogonal Factor ModelOrthogonal Factor Model
ijji
kmimkiki
i
imiii
iimiiiii
FX
XX
X
h
X
),Cov(
),Cov(
),Cov(
variancespecific y communalit)Var(
ycommunalit
)Var(
')Cov(
11
222
21
2
222
21
LFX
ΨLLX
1010
Example 9.1: VerificationExample 9.1: Verification
21
21
211
2221 21719,1714
3000
0100
0040
0002
8621
1174
81
61
27
14
68472312
473852
2355730
1223019
hh
Σ
1111
Example 9.2: No SolutionExample 9.2: No Solution
toryunsatisfac),Var(575.01
ioncontradict ,1|),(Co||),Cov(|||
1)Cov(,1)Cov(,575.1,7.0/4.0
1,4.0,1
7.0,9.0,1
00
00
00
14.07.0
4.019.0
7.09.01
,,
12111
111111
112111121
323131212
221
311121111211
3
2
1
312111
31
21
11
313133212122111111
FXrrFX
XF
FXFXFX
Σ
1212
Ambiguities of Ambiguities of L When L When mm>1>1
X
T
T
ΨLLΨL'LTT'ΨLL'Σ
ITT'TFTF
FTF
εFLεFLTT'εLFμX
ITTT
**
*
*
**
for system coordinate
in the factors ofrotation :matrix orthogonal
of choice by the unaffected are iescommunalit
'
)Cov(')Cov(
0)(')(
' matrix, orthogonal :
EE
mm
1313
Principal Component SolutionPrincipal Component Solution
m
jijiii
pmm
mm
ppp
1
2
2
1
'
'22
'11
2211
''222
'111
00
00
00
ΨLL'
e
e
e
eee
eeeeeeΣ
1414
Principal Component SolutionPrincipal Component Solution
22
221
2
1
221
2211
21
~~~~:iescommunalit estimated
~,,,,
~
: variancespecific estimated
ˆˆˆˆˆˆ~~
:loadingsfactor estimated ofmatrix
,ˆˆˆ
or of pairsr eigenvecto-eigenvalue:ˆ,ˆ
imiii
m
jijiiip
mmij
p
ii
h
sdiag
pm
Ψ
eeeL
RSe
1515
Residual MatrixResidual Matrix
1111121
221
211
2211
21
221
ˆˆˆ'ˆˆ~~~
:factorcommon first thefrom
)tr(
variancesample total theon tocontributi
~:factorcommon first thefrom on tocontributi
ˆˆ
matrix residual of entities squared of summed
matrix residual:~~~
ee
S
Ψ'LLS
p
pp
iii
pm
sss
s
λλ
1616
Determination of Number of Determination of Number of Common FactorsCommon Factors
factored is if 1an greater th of seigenvalue
ofnumber the toequal set :criterionAnother
explainedbeen has variance
sample total theof "proportion suitable" a until
increased is retained factorscommon ofnumber
for ˆ
for ˆ
factorth to
due variancesample
totalof Proportion2211
RR
R
S
m
p
sss
j j
pp
j
1717
Example 9.3Example 9.3Consumer Preference DataConsumer Preference Data
1818
Example 9.3Example 9.3Determination of Determination of mm
2 take
variancesample (standard) total theof
93.0ˆˆ
proportion cumulative afor account They
1an greater th
seigenvalueonly theare 81.1ˆ,85.2ˆ
21
21
m
p
1919
Example 9.3Example 9.3Principal Component SolutionPrincipal Component Solution
2020
Example 9.3Example 9.3FactorizationFactorization
1
81.01
11.053.01
91.079.011.01
00.044.097.001.01
07.00000
011.0000
0002.000
00012.00
000002.0
54.010.075.053.082.0
80.094.065.078.056.0
54.080.0
10.094.0
75.065.0
53.078.0
82.056.0
~~~Ψ'LL
2121
Example 9.4Example 9.4Stock Price DataStock Price Data
Weekly rates of return for five stocksWeekly rates of return for five stocks– XX11: Allied Chemical: Allied Chemical– XX22: du Pont: du Pont– XX33: Union Carbide: Union Carbide– XX44: Exxon: Exxon– XX55: Texaco: Texaco
2222
Example 9.4Example 9.4Stock Price DataStock Price Data
385.0176.0400.0676.0451.0ˆ,343.0ˆ
382.0472.0662.0206.0387.0ˆ,452.0ˆ
435.0541.0335.0178.0612.0ˆ,540.0ˆ
582.0526.0260.0509.0240.0ˆ,809.0ˆ
421.0421.0470.0457.0464.0ˆ,587.2ˆ
1523.0426.0322.0462.0
1436.0389.0387.0
1599.0509.0
1577.0
1
0037.00063.00057.00048.00054.0'
'55
'44
'33
'22
'11
e
e
e
e
e
R
x
2323
Example 9.4Example 9.4Principal Component SolutionPrincipal Component Solution
2424
Example 9.4Example 9.4Residual Matrix for Residual Matrix for mm=2=2
0
232.00
017.0019.00
012.0055.0122.00
017.0069.0164.0127.00
~~~Ψ'LLR
2525
Maximum Likelihood MethodMaximum Likelihood Method
matrix diagonal a
condition Uniqueness
where
2
2),(
is likelihood thenormal, then are
normaljointly are and that Assume
1
1
1
'2/12/
2/'tr2/)1(2/)1(
ΔLΨL'
ΨLL'Σ
Σ
ΣΣμ
εLFμX
εF
1
μxΣμx
xxxxΣ
np
npn
jjj
jj
e
eL
n
jjj
2626
Result 9.1Result 9.1
pp
pjjj
imiii
pn
sssj
h
N
2211
222
21
222
21
2
1
21
ˆˆˆ
factorth todue variance
sample totalof Proportion
so
ˆˆˆˆ
are iescommunalit theof estimates likelihood maximum The
diagonal being
ˆˆ'ˆ subject tofunction likelihood themaximize
ˆ and,ˆ,ˆ estimators likelihood maximum The
),,( from sample random :,,,
LΨL
xμΨL
ΨLL'ΣΣμXXX
2727
Factorization of Factorization of RR
pj
h
njjj
imiii
222
21
222
21
2
2/12/12/12/12/12/1
2/12/12/12/12/12/1
2/1
ˆˆˆ
factorth the todue variancesample
zed)(standardi totalof Proportion
ˆˆˆˆ
ˆˆˆ
ˆˆˆ'ˆˆˆˆˆˆˆˆ
estimator likelihood maximum
'
: variableedstandardiz
z'zz
z'zz
ΨLL
VΨVLVLVVΣVρ
ΨLL
ΨVVLVLVΣVVρ
μXVZ
2828
Example 9.5: Factorization ofExample 9.5: Factorization ofStock Price DataStock Price Data
2929
Example 9.5Example 9.5ML Residual MatrixML Residual Matrix
0
000.00
004.0031.00
000.0004.0003.00
004.0024.0004.0005.00
ˆˆˆ Ψ'LLR
3030
Example 9.6Example 9.6Olympic Decathlon DataOlympic Decathlon Data
3131
Example 9.6Example 9.6FactorizationFactorization
3232
Example 9.6Example 9.6PC Residual MatrixPC Residual Matrix
3333
Example 9.6Example 9.6ML Residual MatrixML Residual Matrix
3434
A Large Sample Test for A Large Sample Test for Number of Common FactorsNumber of Common Factors
pn
H
H
n
H
nne
H
H
n
n
n
n
n
nnpn
n
pppmmppp
SΣS
Σ
SΨ'LLΨ'LL
Ψ'LLΣ
SSS
Σ
ΨL'LΣ
1
2/
0
0
12/
0
2/2/
1
)()()()(0
ˆtrˆ
ln2
likelihood maximized
under likelihood maximizedln2ln2
gfor testim statistic ratio likelihood
ˆˆˆtr2
expˆˆˆ toalproportion
ˆˆˆˆ:under function likelihood Maximum
/)1(, toalproportion
:case generalfor function likelihood Maximum
matrix definite positiveother any :
:
3535
A Large Sample Test for A Large Sample Test for Number of Common FactorsNumber of Common Factors
)1812(2
1
and large are and that provided
)(ˆˆˆ
ln)6/)542(1(
if cesignifican of level at the reject
ionapproximat square-chi sBarlett'
ˆlnln20ˆtr
2
2/])[(
0
1
2
ppm
pnn
mpn
H
n
mpmpn
nn
S
Ψ'LL
S
ΣSΣ
3636
Example 9.7Example 9.7Stock Price Model TestingStock Price Model Testing
model)factor -(two reject tofail
84.3)05.0()05.0(
62.0)0065.1ln(6/)5810(1100
ˆˆˆln]6/)542(1[
0065.1193163.0
194414.0ˆˆˆ
ˆˆ
ˆˆˆˆˆˆˆ
ˆ
ˆˆˆˆ
ˆ
ˆˆ
0
21
2
2/])[(
2/12/1
2/12/12/12/1
2/1
2/1
2/1
2/1
2
H
mpn
mpmp
n
nnn
S
Ψ'LL
R
ΨLL
VSV
VΨVV'LLV
V
V
S
Ψ'LL
V
V
S
Σ
z'zz
3737
Example 9.8Example 9.8Examination ScoresExamination Scores
3838
Example 9.8Example 9.8Maximum Likelihood SolutionMaximum Likelihood Solution
3939
Example 9.8Example 9.8Factor Rotation Factor Rotation
4040
Example 9.8Example 9.8Rotated Factor LoadingRotated Factor Loading
4141
Varimax CriterionVarimax Criterion
possible asmuch as
factoreach on loadings theof squares theout" spreading" i.e.,
factorth for loadings
(scaled) of squares of variance
:tionInterpreta
possible as large as
/~~1
makes that select
ˆ/ˆ~,,ˆˆ
1
1 1
2
1
2*4*
**
m
j
m
j
p
i
p
iijij
iijij
jV
pp
V
h
T
ITT'TLL*
4242
Example 9.9: Consumer-Example 9.9: Consumer-Preference Factor AnalysisPreference Factor Analysis
4343
Example 9.9Example 9.9Factor RotationFactor Rotation
4444
Example 9.10 Example 9.10 Stock Price Factor AnalysisStock Price Factor Analysis
4545
Example 9.11Example 9.11Olympic Decathlon Factor AnalysisOlympic Decathlon Factor Analysis
4646
Example 9.11Example 9.11Rotated ML LoadingsRotated ML Loadings
4747
Factor ScoresFactor Scores
data original
theof nsnsformatiolinear tra Involve 2.
values true theas ˆ and ˆTreat 1.
sestimationour of Essences
analysisnext
toinputs and purposes diagnosticfor used
case)th (
by attained values theof estimateˆ
iij
j
jj
j
F
ff
4848
Weighted Least Squares MethodWeighted Least Squares Method
z'zz
z'zzz
'zzz
'z
ΨLLρxxDz
zΨLΔzΨLLΨLf
xxΨLΔxxΨLLΨLf
μxΨLLΨLf
LfμxΨLfμxεΨε
εLFμX
ˆˆˆˆ),(
ˆˆˆˆˆ)ˆˆˆ(ˆ
matrixn correlatiofor
)(ˆ'ˆˆ)(ˆ'ˆ)ˆˆ'ˆ(ˆ
estimates likelihood maximul use
)(')'(ˆ
)()'(' minimize
2/1
11111
11111
111
11
1
2
jj
jjj
jjj
p
j i
i
4949
Factor Scores of Principal Factor Scores of Principal Component MethodComponent Method
Ifff
xxe
xxe
xxe
f
eeeL
zLLLfxx'LL'Lf 'zz
'z
n
jjj
n
jj
jm
m
j
j
j
mm
jjjj
i
nn
ψ
1
'
1
'
'2
2
'1
1
2211
11
ˆˆ1
1,0ˆ1
,
ˆˆ
1
ˆˆ
1
ˆˆ
1
ˆ
ˆˆˆˆˆˆ~
~~~ˆor ~~~ˆ
equal)nearly are ( method squareleast unweighted Use
5050
Orthogonal Factor ModelOrthogonal Factor Model
0εFFε
Ψ
Ψεεεε
IFFFF
)'(),Cov(
00
00
00
)'()Cov(,0)(
)'()Cov(,0)(
2
1
E
EE
EE
p
5151
Regression ModelRegression Model
Rj
Rj
Rjj
LSj
jj
mp
p
E
N
N
fΔIfLΨ'LI
fLΨ'LILΨ'LxxΨ'LLΨ'Lf
Ψ'LLΨ'LIΨ'LL'LxxΨ'LL'Lf
LΨLL'L'ILΣL'IxF
μxΨLL'L'μxΣL'xF
IL'
LΨLL'ΣΣ
Σ0FμX
ΨLL'0εLFμX
*
*
ˆ)ˆ(ˆ))ˆˆˆ((
ˆ)ˆˆˆ()ˆˆˆ()(ˆˆ)ˆˆˆ(ˆ
ˆˆ)ˆˆˆ()ˆˆˆ(ˆ),()ˆˆˆ(ˆˆ
)()|Cov(
)()()()|(
), (:normaljoint and
),(:
111
111111
11111
11
11
5252
Factor Scores by RegressionFactor Scores by Regression
z'zz
'z
ΨLLρxxDz
zRLf
xxS'Lf
ˆˆˆˆ),(
ˆˆ
or
)(ˆˆ
2/1
1
1
jj
jj
jj
5353
Example 9.12Example 9.12Stock Price DataStock Price Data
'4.12.1ˆˆ
regression
'0.28.1ˆˆˆˆˆˆ
squaresleast weighted
40.170.020.040.150.0
18.00000
061.0000
0047.000
00025.00
000050.0
ˆ,
883.0208.0
507.0365.0
335.0643.0
164.0850.0
377.0601.0
ˆ
fromsolution likelihood maximum
1'
1'11'
zRLf
zΨLLΨLf
z'
ΨL
R
*z
z*z
*zz
*z
z*z
5454
Example 9.12Example 9.12Factor Scores by RegressionFactor Scores by Regression
5555
Example 9.13: Simple Summary Example 9.13: Simple Summary Scores for Stock Price DataScores for Stock Price Data
)~
(ˆ,ˆ
)~
(ˆ,ˆ
scoressummary
854.0226.0
815.0258.0
316.0766.0
128.0889.0
323.0746.0
~~,
524.0712.0
473.0712.0
234.0795.0
458.0773.0
216.0784.0
~
loadingfactor component principal
5423211
2542543211
*
*
L
L
TLLL
xxfxxxf
xxxfxxxxxf
5656
A Strategy for Factor AnalysisA Strategy for Factor Analysis
1. Perform a principal component factor 1. Perform a principal component factor analysisanalysis– Look for suspicious observations by plotting Look for suspicious observations by plotting
the factor scoresthe factor scores– Try a varimax rotationTry a varimax rotation2. Perform a maximum likelihood factor 2. Perform a maximum likelihood factor analysis, including a varimax rotationanalysis, including a varimax rotation
5757
A Strategy for Factor AnalysisA Strategy for Factor Analysis3. Compare the solutions obtained from 3. Compare the solutions obtained from the two factor analysesthe two factor analyses– Do the loadings group in the same manner?Do the loadings group in the same manner?– Plot factor scores obtained for PC against Plot factor scores obtained for PC against
scores from ML analysisscores from ML analysis
4. Repeat the first 3 steps for other 4. Repeat the first 3 steps for other numbers of common factorsnumbers of common factors5. For large data sets, split them in half 5. For large data sets, split them in half and perform factor analysis on each part. and perform factor analysis on each part. Compare the two results with each other Compare the two results with each other and with that from the complete data setand with that from the complete data set
5858
Example 9.14Example 9.14Chicken-Bone DataChicken-Bone Data
1
937.01
894.0874.01
878.0877.0926.01
450.0482.0467.0422.01
603.0621.0602.0569.0505.01
length ulna length, humerus :Wing
length tibia length,femur :Leg
breadth skull length, skull :Head
dimensions boneon tsmeasuremen 276
65
43
21
R
XX
XX
XX
n
5959
Example 9.14:Principal Component Example 9.14:Principal Component Factor Analysis ResultsFactor Analysis Results
6060
Example 9.14: Maximum Likelihood Example 9.14: Maximum Likelihood Factor Analysis ResultsFactor Analysis Results
6161
Example 9.14Example 9.14Residual Matrix for ML EstimatesResidual Matrix for ML Estimates
000.0000.0000.0001.0001.0004.0
000.0000.0000.0000.0001.0
000.0000.0000.0000.0
000.0001.0003.0
000.0000.0
000.0
ˆˆˆz
'zz ΨLLR
6262
Example 9.14Example 9.14Factor Scores for Factors 1 & 2Factor Scores for Factors 1 & 2
6363
Example 9.14Example 9.14Pairs of Factor Scores: Factor 1Pairs of Factor Scores: Factor 1
6464
Example 9.14Example 9.14Pairs of Factor Scores: Factor 2Pairs of Factor Scores: Factor 2
6565
Example 9.14Example 9.14Pairs of Factor Scores: Factor 3Pairs of Factor Scores: Factor 3
6666
Example 9.14Example 9.14Divided Data SetDivided Data Set
1940.0927.0894.0386.0598.0
1911.0909.0420.0587.0
1950.0406.0587.0
1352.0572.0
1366.0
1
,139
1931.0863.0866.0584.0660.0
1835.0844.0606.0694.0
1901.0575.0639.0
1540.0588.0
1696.0
1
,137
22
11
R
R
n
n
6767
Example 9.14: PC Factor Analysis Example 9.14: PC Factor Analysis for Divided Data Setfor Divided Data Set
6868
WOW CriterionWOW Criterion
In practice the vast majority of In practice the vast majority of attempted factor analyses do not attempted factor analyses do not yield clear-cut resultsyield clear-cut results
If, while scrutinizing the factor If, while scrutinizing the factor analysis, the investigator can shout analysis, the investigator can shout “Wow, I understand these factors,” “Wow, I understand these factors,” the application is deemed successfulthe application is deemed successful
Sets of linear equations to specify Sets of linear equations to specify phenomena in terms of their phenomena in terms of their presumed cause-and-effect variablespresumed cause-and-effect variables
In its most general form, the models In its most general form, the models allow for variables that can not be allow for variables that can not be measured directlymeasured directly
Particularly helpful in the social and Particularly helpful in the social and behavioral sciencebehavioral science
7070
LISREL (LISREL (LiLinear near SStructural tructural RelRelationships) Modelationships) Model
rnonsingula ,0,
measured becan :,
),Cov(ed,uncorrelatmutually ,,
),Cov(,)(
),Cov(,)Cov(,)(
observeddirectly not , variableseffect)-and-(causelatent :,
)Cov(,)(,
)Cov(,)(,
)Cov(,)(,
equationst measuremen
BIB
Y X
0ζξδεζ
0εη0η
0δξΦξ0ξ
ηξ
Θδ0δδξΛX
Θε0εεηΛY
Ψζ0ζζΓξBηη
δ
ε
iiik
x
y
BB
E
E
E
E
E
7171
ExampleExample: performance of the firm: performance of the firm
: managerial talent: managerial talent
YY11: profit: profit
YY22: common stock price: common stock price
XX11: years of chief executive experience: years of chief executive experience
XX22: memberships on board of directors: memberships on board of directors
7272
Linear System in Control TheoryLinear System in Control Theory
DuCxy
BuAxx
7373
Kinds of VariablesKinds of Variables
Exogenous variables: not influenced Exogenous variables: not influenced by other variables in the systemby other variables in the system
Endogenous variables: affected by Endogenous variables: affected by other variablesother variables
Residual: associated with each of the Residual: associated with each of the dependent variablesdependent variables
7474
Construction of a Path DiagramConstruction of a Path DiagramStraight arrowStraight arrow– to each dependent (endogenous) to each dependent (endogenous)
variable from each of its sourcevariable from each of its source
Straight arrowStraight arrow– also to each dependent variables from also to each dependent variables from
its residualits residual
Curved, double-headed arrowCurved, double-headed arrow– between each pair of independent between each pair of independent
(exogenous) variables thought to have (exogenous) variables thought to have nonzero correlationnonzero correlation
7575
Example 9.15Example 9.15Path DiagramPath Diagram
7676
Example 9.15Example 9.15Structural EquationStructural Equation
0),Cov(
),Cov(,),Cov(,),Cov(
0
0
00
0
21
331232121
2
1
3
2
1
43
21
2
1
2
1
7777
Covariance StructureCovariance Structure
'YXΓΦΛΛ
'δξΛεζΓξΛYX'XY
ΘΦΛΛΘΛξΛXX'X
ΘΛΨΓΦΓ'Λ
ΘΛηΛYY'Y
0B
XYX
XYY
ΣΣ
ΣΣΣ
X
Y
'xy
xy
δ'xxδ
'xx
ε'yy
ε'yy
)],[Cov(
))()(()(),Cov(
)Cov()()Cov(
)Cov()()Cov(
discussion hesimplify t to Assume
)Cov(),Cov(
),Cov()Cov(Cov
2221
1211
EE
E
E
7878
EstimationEstimation
parameters model theestimate
tocriterion "likelihood maximum" a andcriterion
squares"least " a usescurrently program LISREL The
esapproximat
closely that ˆmatrix a produce toused bemust estimates
parameter initial with begins that routinesearch iterativeAn
equations resulting thesolve and ,ˆ settingby
parameters model theestimate toused is
matrix covariance sample the
,,2,1 ,' nsobservatio Given
2221
1211
S
Σ
SΣ
SS
SSS
xy 'j
'j
njn
7979
Example 9.16Example 9.16Artificial DataArtificial Data
42
2322
1
2
222
12
1
211
2
4
3
2
1
2
11
2
1
2
1
12
1
)Cov(,),Cov(
)Cov(
0
0)Cov(,
0
0)Cov(
)(Var,)(Var
1,
1,
XXY
Y
δε
X
X
Y
Y
8080
Example 9.16Example 9.16Artificial DataArtificial Data
1.1ˆˆ,6.1ˆˆ
7.3ˆˆˆ,4.6ˆˆˆ,8.12ˆˆˆˆ
2.3ˆˆ,4.6ˆˆˆ,4.55ˆˆˆˆˆ
6.27ˆˆˆˆ,3.14ˆˆˆˆ
1.16.1|4.62.3
6.17.3|8.124.6
4.68.12|4.556.27
2.34.6|6.273.14
42
322121
2222
1
211
2
2221
1211
SS
SS
8181
Example 9.16Example 9.16Artificial DataArtificial Data
3.00
05.0ˆ,2.00
05.0ˆ
1
2ˆ,2
1ˆ
1ˆ,4ˆ,8.0ˆ,0.2ˆ2
δε
xy
ΘΘ
ΛΛ
8282
Assessing the Fit of the ModelAssessing the Fit of the Model
The number of observations The number of observations pp for for YY and and qq for for XX must be larger than the must be larger than the total number of unknown parameters total number of unknown parameters tt < < ((pp++qq)()(pp++qq+1)/2+1)/2
Parameter estimates should have Parameter estimates should have appropriate signs and magnitudesappropriate signs and magnitudes
Entries in the residual matrix Entries in the residual matrix SS – – should be uniformly smallshould be uniformly small
8383
Model-Fitting StrategyModel-Fitting StrategyGenerate parameter estimates using Generate parameter estimates using several criteria and compare the several criteria and compare the estimatesestimates– Are signs and magnitudes consistent?Are signs and magnitudes consistent?– Are all variance estimates positive?Are all variance estimates positive?– Are the residual matrices similar?Are the residual matrices similar?
Do the analysis with both Do the analysis with both SS and and RRSplit large data sets in half and Split large data sets in half and perform the analysis on each halfperform the analysis on each half