2. Purpose
To analyze a set of relationships between one or
more IVs, either continuous or discrete, and one
or more DVs, either continuous or discrete.
Multiple IVs
Multiple DVs
Confirmatory purpose
Advantage
Estimation of errors
Simultaneous test
3. Kinds of research questions
The goal of SEM is to examine whether the model
produce an estimated population covariance matrix
is fitted to the sample covariance matrix or not.
Adequacy of model
Testing theory
Amount of variance in the variables accounted for by the
factors
Reliability of the indicators
Parameter estimates
Intervening variables (indirect effect)
Group difference
Longitudinal difference
Multilevel modeling
4. Limitations to factor analysis
Theoretical issues
Theory
causality
Practical issues
Sample size and missing data
15* umber of variable (Stevens, 1996)
5* number of estimated parameter (Bentler & Chou, 1987)
100 (Loehiln, 1992)
Multivariate normality and outliers
Linearity
Absence of multicollinearity and singularity
Residual
5. Fundamental equation for SEM
Regression coefficients
Variance-covariance matrices
Data point
𝑌𝑗 = 𝛾𝑗𝑖 𝑋𝑖 + 𝜖1
𝐶𝑂𝑉 𝑋𝑖, 𝑌𝑗 = 𝐶𝑂𝑉(𝑋𝑖, 𝛾𝑗𝑖 𝑋𝑖 + 𝜖1)
= 𝛾𝑗𝑖 𝐶𝑂𝑉(𝑋𝑖, 𝑋𝑖)
= 𝛾𝑗𝑖 𝛿 𝑋 𝑖 𝑋 𝑖
η = 𝐵η + 𝛾ξ + ζ
𝑝 𝑝+1
2
variance +covariance
6. Graphical representation for SEM
x1
x2
x3
ξ belief η behavior
y1
y2
y3
δ1 ε1
λx1
γ
ζ
ε2
ε3
δ2
δ3
Error
Exogenous
observed
variable
Factor loading
Exogenous
latent variable
Structural parameter
Endogenous
latent variable
Factor loading
Endogenous
observed
variable
Structural
model
measurement
model
Error
λx2
λx3
λy1
λy2
λy3
measurement
model
8. Some important issue
Model identification
A unique numerical solution for each of the
parameters in the model
Data point
Overidentified
Data point >parameters (df>0)
Just identified
Data point =parameters (df=0)
Underidentified (X)
Data point <parameters (df<0)
𝑝 𝑝 + 1
2
Model
identification
Measurement
model
Structural
model
9. Assessing the fit of the model
Category Index
Absolute fit index
χ2, χ2/df
GFI
AGFI
Comparative fit index
RMSEA
CFI
NFI
NNFI
IFI
Parsimonious fit
index
PGFI
PNFI
CN
Residual
RMR
SRMR
10. Absolute fit index
Chi-square test
N.S.
Chi-square / df < 2 (3)
Indices of proportion of variance accounted
𝑔𝑜𝑜𝑑𝑛𝑒𝑠𝑠 − 𝑜𝑓 − 𝑓𝑖𝑡 𝑖𝑛𝑑𝑒𝑥 𝑮𝑭𝑰 =
𝑡𝑟( 𝜎′ 𝑊 𝜎)
𝑡𝑟(𝑠′ 𝑊𝑠)
𝑎𝑑𝑗𝑢𝑠𝑡𝑒𝑑 𝑔𝑜𝑜𝑑𝑛𝑒𝑠𝑠 − 𝑜𝑓 − 𝑓𝑖𝑡 𝑖𝑛𝑑𝑒𝑥 𝑨𝑮𝑭𝑰 =
1 − 𝐺𝐹𝐼
1 −
𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑒𝑠𝑡. 𝑝𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟𝑠
𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑑𝑎𝑡𝑎 𝑝𝑜𝑖𝑛𝑡
11. Comparative fit index
Nested model
Independent
model
Saturated
model
Unrelate
d
variable
s
df=0
13. Nested model
Saturated model
Model 1
Model 2
Model 3
Parent model or
Full model
all parameters are freely
estimated
restriction nested in
restriction nested in
restriction nested in
χ2 、df ↑
14. Comparative fit index
Assessing the fit of the model
𝑐𝑜𝑚𝑝𝑎𝑟𝑎𝑡𝑖𝑣𝑒 𝑓𝑖𝑡 𝑖𝑛𝑑𝑒𝑥 𝑪𝑭𝑰 = 1 −
𝜏 𝑒𝑠𝑡.𝑚𝑜𝑑𝑒𝑙
𝜏𝑖𝑛𝑑𝑒𝑝.𝑚𝑜𝑑𝑒𝑙
𝜏 𝑒𝑠𝑡.𝑚𝑜𝑑𝑒𝑙 = χ𝑖𝑛𝑑𝑒𝑝.𝑚𝑜𝑑𝑒𝑙
2
− 𝑑𝑓𝑖𝑛𝑑𝑒𝑝.𝑚𝑜𝑑𝑒𝑙
𝜏 𝑒𝑠𝑡.𝑚𝑜𝑑𝑒𝑙 = χ 𝑒𝑠𝑡.𝑚𝑜𝑑𝑒𝑙
2
− 𝑑𝑓𝑒𝑠𝑡.𝑚𝑜𝑑𝑒𝑙
𝜏𝑖 = 𝑣𝑎𝑙𝑢𝑒 𝑜𝑓 𝑚𝑖𝑠𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛
𝑟𝑜𝑜𝑡 𝑚𝑒𝑎𝑛 𝑠𝑞𝑢𝑎𝑟𝑒 𝑒𝑟𝑟𝑜𝑟 𝑜𝑓 𝑎𝑝𝑝𝑟𝑜𝑥𝑖𝑚𝑎𝑡𝑖𝑜𝑛 𝑹𝑴𝑺𝑬𝑨 =
𝐹0
𝑑𝑓 𝑚𝑜𝑑𝑒𝑙
𝐹0 =
χ 𝑚𝑜𝑑𝑒𝑙
2
− 𝑑𝑓 𝑚𝑜𝑑𝑒𝑙
𝑁
Compare to
independent
model
Compare to
saturated
model
15. Comparative fit index
𝑛𝑜𝑟𝑚𝑒𝑑 𝑓𝑖𝑡 𝑖𝑛𝑑𝑒𝑥 𝑵𝑭𝑰 =
χ𝑖𝑛𝑑𝑒𝑝
2
− χ 𝑚𝑜𝑑𝑒𝑙
2
χ𝑖𝑛𝑑𝑒𝑝
2
𝑛𝑜𝑛 − 𝑛𝑜𝑟𝑚𝑒𝑑 𝑓𝑖𝑡 𝑖𝑛𝑑𝑒𝑥 𝑵𝑵𝑭𝑰/𝑻𝑳𝑰 =
χ𝑖𝑛𝑑𝑒𝑝
2
−
𝑑𝑓𝑖𝑛𝑑𝑒𝑝
𝑑𝑓 𝑚𝑜𝑑𝑒𝑙
χ 𝑚𝑜𝑑𝑒𝑙
2
χ𝑖𝑛𝑑𝑒𝑝
2
− 𝑑𝑓𝑖𝑛𝑑𝑒𝑝
Underestimate with small N
Too sensitive to stable
𝑖𝑛𝑐𝑟𝑒𝑚𝑒𝑛𝑡𝑎𝑙 𝑓𝑖𝑡 𝑖𝑛𝑑𝑒𝑥 𝑰𝑭𝑰 =
χ𝑖𝑛𝑑𝑒𝑝
2
− χ 𝑚𝑜𝑑𝑒𝑙
2
χ𝑖𝑛𝑑𝑒𝑝
2
− 𝑑𝑓 𝑚𝑜𝑑𝑒𝑙
29. Nested model
Saturated model
Model 1
Model 2
Model 3
Parent model or
Full model
all parameters are freely
estimated
restriction nested in
restriction nested in
restriction nested in
χ2 、df ↑
30. Nested model
fixed path in structure model
x1
x2
x3
perceived
control
behavior
y1
y2
y3
x4 x5 x6
intention
C, 0
31. Nested model
fixed loading in measurement model
x1
x2
x3
ξ perceived
control
x4
x5
x6
ξ intention
x1
x2
x3
ξ perceived
control
x4
x5
x6
ξ intention
a
a
b
b
32. Nested model
fixed correlation between latent variables
in measurement model
x1
x2
x3
ξ perceived
control
x4
x5
x6
ξ intention
x1
x2
x3
ξ perceived
control
x4
x5
x6
ξ intention
C, 0
33. Which one fits better?
Nested model
△χ2 test
Target coefficient
Initial model χ𝑖
2
Restricted model χ 𝑟
2
T= χ𝑖
2
/ χ 𝑟
2
Non-nested model
AIC
CAIC
34. Parsimonious purpose
General concept and multiple dimensions
1 order
2 order
Nested model
Required number of lower order dimension
>3
3 (equivalent model)
38. Generalization
model
data 1
Population 1
fit
fit
Can we generalize the
psychological
phenomenon of P1 to P2 ?
Population 2
data 2
are the constructs comparable
between the two populations?
(mean, variance, covariance or
correlation)