Structural Equation Modeling

using AMOS

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

1
 Contents
T1: Introduction
T2: Linear Regression & Correlation in SEM
T=
T3: Measurement & Structural models To
pic
T4: AMOS graphic & practice designing models
T5: AMOS output and interpretation
T6: Hypothesis testing
T7: Comparison of models
T8: Managing groups
T9: If model doesn’t fit …
T10: Specification search (exploratory analysis)
T11: Normality and outliers
T12: Bootstrapping
T13: Bayesian’s Estimation
T14: Others (missing data, sample size issues)
2
 Structural Equation Modeling
using AMOS

T1: INTRODUCTION

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

3
 Contents

• What is Structural Equation Modeling?

• What is Amos?

4
 Structural Equation Modeling
• Linear Modeling (basic analyses involved are
correlation and regression)
• Model for numerical outcome
• Variables that we don’t measure (construct)
can be included.
• Confirmatory analyses can be done.
• Two types of models involved … “structural
model” and “measurement model”

5
 Structural Equation Modeling
• Various names:
– Structural Equation Modeling
– Analysis of Moment Structure (AMOS)
– Analysis of Covariance Structure
– Causal modeling!!!

6
 Structural Equation Modeling
• How we approach to learn …
– Linear regression using SEM
– Correlation using SEM
– Structural model in SEM
– Measurement model in SEM
– More complex models

7
 About AMOS
• AMOS software has graphic interface –
makes very user friendly
• Use to analyze the advanced level of SEM

8
 About AMOS
• Some basic rules in AMOS
– (A) Single-headed arrow (unidirectional) is “regression”.
– (B) Double-headed arrow (bidirectional) is “correlation”.
– (C) A variable in square box is “measured” variable.
– (D) A variable in oval or rounded box is “unmeasured”
variable.

Knowledge, value and
C satisfaction are exogenous
variable (predictor or
A independent variable).
B
Performance is endogenous
variable (response,
D outcome or dependent
variable)

9
 About AMOS
What is ‘moment structure’?
Central moment in Math

µn = E (( X - µ )n )
n = 1, the first moment, µ1= 0
n = 2, the second moment, µ2= variance
n = 3, the third moment, µ3
skewness
n = 4, the fourth moment, µ4
kurtosis
In SEM, variance & covariance are basic
analyses.
10
 Structural Equation Modeling
using AMOS

T2: Linear Regression &

Correlation in SEM

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

11
 Linear Regression
• Simple Linear Regression

Y = a + bX + e

Performance = a + b*Knowledge + e

b
Knowledge Performance e

12
 Linear Regression
• Multiple Linear Regression

Y = b0 + b1X1 + b2X2 + b3X3……. + e

Performance = b0 + b1*Knowledge + b2*Value +
b3*Satisfaction + error

Knowledge b1
b2
Value Performance e
b3
Satisfaction

13
Linear Regression
Using kvsperform1.sav
Run SPSS and see!

Unstandardized

Standardized
14
 Correlation
• Simple Correlation (Bivariate)

Correlation between X and Y

Correlation between Knowledge and
Performance
r
Knowledge Performance

15
 Correlation
• Multiple (Bivariate) Correlations

Using kvsperform1.sav

Run SPSS and see!

16
 Structural Equation Modeling
using AMOS

T3: Measurement & Structural

models

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

17
 Measurement model
Ex08.amw
• Factor analysis  Showing 2 measurement models
 Latent variable or unmeasured variable
 Arrow is from 'unmeasured' to
'measured' variable
 Notice the regression coeff. '1'

Psyco. Test Explanation

visperc Visual perception scores
cubes Test of spatial visualizations
lozenges Test of spatial orientation
paragrap Paragraph comprehension score
sentence Sentence completion score
wordmean Word meaning test score

18
 Measurement model
Ex08.amw
• Factor analysis  Notice 'arrow' from 'unmeasured' to
'measured' variable
Using 3 tests, we are measuring a construct,
which is called 'spatial'.
'Spatial' is a construct which we cannot
measure directly.

However, ‘spatial’ comes out to surface as 3

Test 3 Aspects
aspects that are measurable by 3 tests.
visperc Visual perception
cubes Spatial visualizations It is considered that 'spatial' determines the
lozenges Spatial orientation 3 aspects. Therefore, measuring the 3
aspects, we can indirectly measure ‘spatial.’

19
 Measurement model
Ex08.amw
• Factor analysis  Notice the regression coeff. '1'

Expenditure:
A Book RM100 RM 100
A Pen US$ 2 RM 7
A Ruler S$ 2 RM 5
What is the total cost? RM 112

Total cost: You may calculate in

RM or US$ or S$
But only one scale, we can use to
total-up.

20
 Structural model
• Regression & correlation

21
 Mixed model
Ex06a

• Mixed model There were 3 measurement models,

and 1 structural model.

22
 Mixed model
Ex06a

• Mixed model There were 3 measurement models,

and 1 structural model.

23
 Can we use categorical variable?
Experimental study

1. Treatment is coded ‘1’ for

experiment and ‘0’ for
control
2. For independent
(exogenous) variable, it
can be categorical
variable (as in multiple
linear regression).

Exa9a.amw 3. However, dependent

(endogenous / outcome)
But it must be 0/1 dichotomous variable. variables must be
numerical variable.

24
 Structural Equation Modeling
using AMOS

T4: AMOS GRAPHICS & PRACTICE

DESIGNING MODELS

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

25
 Start Menu bars
Working
Programs place

Manage Groups
Amos 7/16 Manage Models
Display parameter format
Amos
Graphics
Computation summary

Files in the current directory

Tour around AMOS

interface.
Tool bars
26
Practice designing a model

Using kvsperform2.sav Five Tasks

1. Design the model
2. Connect to the data file
3. Select some options for
analysis
4. Run the analysis
5. Check AMOS output

27
Practice designing a model
Using kvsperform1.sav

Five Tasks
1. Design the model
2. Connect to the data file
3. Select some options for
analysis
4. Run the analysis
5. Check AMOS output

28
Practice designing a model
Using c:\......\Grnt_fem.sav

Five Tasks
1. Design the model
2. Connect to the data file
3. Select some options for
analysis
4. Run the analysis
5. Check AMOS output

29
 Structural Equation Modeling
using AMOS

T5: AMOS OUTPUT AND

INTERPRETATION

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

30
 Important Results

• Unstandardized / Standardized statistics

• What is “model-fit measure”?
• Chi-square and other model fit measures
• Total/direct/indirect effects

31
 Unstandardized / Standardized
Unstandardized Standardized
S.H. arrow b (reg. weight) Standardized b
D.H. arrow Covariance Correlation
Variables Variance r2 (Sq Multiple Corr.)a
a on endogenous variables (response, dependent variable)

Using kvsperform1.sav

Unstandardized Standardized
32
 Model-fit measures
 Saturated model and the tested (default) model were compared.
 If it is not significant, the two models are not different.
 Then, the tested (default) model is considered 'fit' in SEM language.
Using kvsperform2.sav

Saturated Default (tested)

It means that the two bidirectional

lines are not necessary. Without them,
the model performs as similar as the
saturated model. (Adequate)
33
 Model-fit measures
 As the chi-square test for model-fit is influenced by sample
sizes, several other model fitness statistics have been
developed.
AMOS Parameters Criteria for fit
\cmin Chi Square test P ≥ 0.05
\cmindf Relative Chi Square 2 or less
\gfi GFI (goodness-of-fit index) 0.95 and above
\agfi AGFI (adjusted goodness-of-fit index) 0.90 and above
\cfi CFI (comparative fit index) 0.90 and above
\pratio PRATIO (parsimony ratio) higher is better fit
\rmsea RMSEA (RMSE of approximation) 0.08 or less
* Please refer to AMOS 6 Users' Guide (Appendix C)

34
 Total / Direct / Indirect effects
 Unstandardized total / direct / indirect effects
Using kvsperform1.sav
Knowledge
Performance

Direct = 1.026
Indirect = (0.968*0.978)+(0.861*1.117)
= 1.908
Total = 2.934

It means that ...

If 1 unit increase in Knowledge
score, Performance score will
increase for 2.934.
The direct effect is 35.0%, whereas
32.3% is through value, and 32.8%
Unstandardized
is through satisfaction.

35
 Total / Direct / Indirect effects
Visit e
x ample
 Standardized total/direct/indirect effects p aper
Using kvsperform1.sav
Knowledge
Performance

Direct = 0.246
Indirect = (0.708*0.321)+(0.569*0.406)
= 0.458
Total = 0.704

It means that ...

If 1 SD increase in Knowledge
score, Performance score will
increase for 0.704 SD.
The direct effect is 34.9%, whereas
32.3% is through value, and 32.8%
Standardized
is through satisfaction.

36
 Try Factor Analysis
 Confirmatory common factor analysis
mw )
8.a .sav Two common factors:
x 0 m Spatial and verbal abilities
E t_fe
rn
(G
Two common factors are correlated
(if uncorrelated, it is orthogonal).

Error terms (unique factors) are

uncorrelated.

Regression coeff. are 'factor

loadings'.

It is confirmatory because we will

say at the end the model fits or not?
(P value)
37
 Structural Equation Modeling
using AMOS

T6: Hypothesis testing

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

38
 Hypothesis testing
Using kvsperform2.sav

Let's test whether correlation between knowledge and value is 'zero' or not?

1. Click "view the input path diagram." 1

2. Right click on double-headed arrow object and select
2 "Object Properties."
3. Enter '0' to "Covariance" box in 'Parameters tab'.
4. Run the analysis
5. Chi-square = 0.16, df = 1, P = 0.693
6. It means that this model is not different from saturated
model which has correlation of -0.03.
It means that the correlation is not different from zero.

You may also test whether it is different from any other value say (0.5) or not?

39
 Hypothesis testing
Using kvsperform2.sav

Let's test whether the variances of knowledge and value are equal or not?

1
1. Click "view the input path diagram."
2. Right click on variable "knowledge" and select
"Object Properties."
3. Enter 'var1' to "Variance" box in 'Parameters tab.'
4. Do the same thing for "value."
5. Run the analysis
5. Chi-square = 0.18, df = 1, P = 0.675
6. It means that the variances are not significantly
different.
3
40
 Hypothesis testing
Using kvsperform2.sav

Let's test whether the means of knowledge and value equal or not?

1. View >>> Analysis Properties

2. Under Estimation, click "Estimate means and
intercept."
3. Right click on "knowledge" variable and select
"Object properties".
4. Enter "mean1" in "Mean" box in 'Parameters tab.'
5. Do the same thing for "value."
6. Run the analysis and interpret.
3
It is marginally significantly different (P=0.044).

41
 ch

r a ppr
o a
Hypothesis testing
t he
o
An

Using kvsperform2.sav

Let's test whether the means of knowledge and value equal or not?

1. View >>> Analysis Properties

2
2. Under “Output”, select “Critical ratios for
differences”
3. Now, assign “mean1” and “mean2” for 2
variable means 5
4. Run the analysis
5. AMOS output (Pairwise parameter
comparisons): Critical ratio = 2.028 (It
means ‘significant’, >1.96)
6. How to get P value? (P=0.043)
In Excel, [=(1-NORMSDIST(2.028))*2]
42
 Structural Equation Modeling
using AMOS

T7: Comparison of models

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

43
 av Comparison of models
2. s
rm
fo

A B
r
pe
vsk
ing
Us

Let's test whether the two models are different or not? What does it mean?
If we remove 2 double-headed arrows (between knowledge and value and between value and
satisfaction) from model A, does it change significantly?
1. We assign "cor1" and "cor2" for 2 correlation.
2. Analyse >>> Manage Models…
3. Change “Default model” to “Model A”
4. Click "New" 5
5. Type "Model B" in Model Name
6. Type "cor1=cor2=0" in "Parameter Constraints"
7. Run the analysis
8. AMOS output --> Model Comparison 6
9. Find the P value (0.439)
44
 av Comparison of models
2. s
rm
fo

A B
r
pe
vsk
ing
Us

Let's test whether the two models are different or not? What does it mean?
If we remove 1 double-headed arrow (between knowledge and value) from the model A, does
it change significantly?
1. We assign "cor1" between knowledge and value.
2. Analyse >>> Manage Models…
3. Change “Default model” to “Model A” 4
4. Click "New"
5. Type "Model B" in Model Name
6. Type "cor1=0" in "Parameter Constraints"
7. Run the analysis
8. AMOS output --> Model Comparison 5
9. Find the P value (0.558)
45
 av Comparison of models
2. s
rm
fo

A B
r
pe
vsk
ing
Us

This P value: Compairing with saturated model This P value: Compairing with saturated model

Model Comparison
between A & B

What is Saturated model?

All variables are connected.
We get chi-square=0, df=0.

46
 Structural Equation Modeling
using AMOS

T8: Managing groups

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

47
 av Managing Groups
2. s
rm
fo

To analyse subgroups: male and female

r
pe
vsk
ing

2
Us

1. Analyse >>> Manage Groups…

2. Type "Male" in "Group Name" 3
3. Click "New"
4. Type "Female" in "Group Name" & Click
"Close"
5. Click "Select data file" icon
6. Click "Grouping Variable" and select
"sex"
7. Click "Group Value" and select "1"
8. For female, select file name, grouping
variable and then group value as "2"
9. Then, run analysis. 6 7
48
 av Managing Groups
2. s
rm
fo

To analyse subgroups: male and female

r
pe
vsk

Now, we are interested to test the

ing

regression coeff., are the same between

Us

males and females or not?

1. We assign b1m, b2m, b3m for male path diagram for 3

regression terms
2. We assign b1f, b2f, b3f for female path diagram for 3
regression terms
3. Analyse -> Manage Models…
4. Change the name of "Default model" to "Model A"
5. Click "New" and type "Model B" in "Model Name"
6. Type the following line by line in "Parameters constraints"
7. Then, run the analysis and see output (Model comparison) (A
vs B: P=0.333)
(Reg.coeff. are not different between males and females.)

49
 Structural Equation Modeling
using AMOS

T9: If model doesn’t fit …

(for Exploratory studies)

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

50
 If model doesn't fit?

For confirmatory analyses, if model doesn't fit,
we just report.

For exploratory analyses, we continue to
explore which model could fit better.

Approaches to improve the models:

1. Using Modification Index (to add terms / parameters)
2. Remove non-significant terms (AMOS output: Estimates)

51
 av
fo
rm
2. s
If model doesn't fit?

Using Modification index
r
pe
vsk
ing
Us

1. View -> Analysis Properties…

2. Go to "Output" tab, and select "Modification Indices"
(Leave the threshold as 4)
3. Run the analysis; Go to AMOS output;
4. Click on “Modification Indices”
Find that AMOS indicates you: If you add correlation between Knowledge
and Satisfaction, the model chi-square will be smaller (IMPROVED) for at
least 69.566 (Modification index, MI).
And the parameter will change to +3.674 (Par Change).
52
 If model doesn't fit?
v
. sa
ce
a cti
pr
g
c hin
eat
ing
Us

Say, this is an exploratory study. Logic, Formal, Intuitive, and DL_C are
attitude variables. These attitudes of teachers could determine their teaching
practice (Repeat, Exercise, DL_P, and Creative). Researchers consider that Repeat
and Exercise are “Old” teaching styles and DL_P and Creative are “New” teaching
styles.
Let’s explore this study ….. try using modification indices to improve the model.

53
 Structural Equation Modeling
using AMOS

T10: Specification search

(exploratory analysis)

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

54
 Specification Search

For exploratory study
Using teaching practice.sav

Let’s explore this study ….. try using "Specification Search"

55
Add all the terms that we would like to test …
Correlation between attitude variables, correlation between unique variables, and
more regression lines from attitudes to two latent variables.
1. Go to View -> Analysis Properties -> under Numerical tab, change iteration limit
to 50.
2. Go to Analyse -> Specification Search;
3. Click this icon ; Then click on the lines that we want to test;
4. Click "option" ; select "next search" and click "stepwise"
5. Run the analysis
56
The output can be sorted by, e.g. click on P value. The best model is underlined.
Double clicking on the row will show you the model on the path diagram. Decide
(Choose) the model carefully. It should be biologically plausible.
57
 Specification Search
mw v)
8.a .sa
0 m
Ex t_fe
G rn
(

Remove reg. coeff. 1 on 2 measurement model; set variance 1 for spatial and
verbal; make all regression lines to optional;
Go to option, next search, and use "All subsets" (more sure to get good model)

58
 Structural Equation Modeling
using AMOS

T11: Normality & Outliers

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

59
 Normality (univariate and multivariate)

Using kvsperform2.sav In Analysis Properties,

under “Output” tab,
check the “Tests for
normality and outliers”

c.r. = critical ratio

>1.96 : not normal
<=1.96 : normal

60
Mahalanobis distance (MD) to detect (multivariate) outliers

df CP on MD2 CP on MD
1 5.02 2.24
2 7.38 2.72
3 9.35 3.06
4 11.14 3.34
5 12.83 3.58
6 14.45 3.80
7 16.01 4.00
Bigger value of MD reflects more distant. 8 17.53 4.19
Generally, cut point for MD2 is X2df(0.975) . 9 19.02 4.36
(If df=1, all MD2 above 5.02 are outliers) 10 20.48 4.53
11 21.92 4.68
* p1 and p2 are the position of data point on 12 23.34 4.83
the two axes plot (like scatter diagram). 13 24.74 4.97
14 26.12 5.11
15 27.49 5.24
Ref: Wing-Kam Fung (1999). Outlier Diagnostics in Several Multivariate CP = cut point
Samples. The Statistician, Vol. 48, No. 1 (1999), pp. 73-84.

61
 Structural Equation Modeling
using AMOS

T12: Bootstrapping

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

62
 Bootstrapping
► SEM needs normal distribution assumption.
► All results obtained earlier based on formula (based on normality assumption)
► Bootstrapping is … not based on any distribution, not using any formula.
► In bootstrapping, repeated sampling is done (e.g. 500 times) from the sample
and calculate the Standard Errors which do not need any distribution
assumption.

How to do?
In Analysis Properties, under
“Bootstrap” tab, do the following:

Using kvsperform2.sav Then, run the analysis as usual.

Check the AMOS output.
63
 Check the AMOS output.
► Regression coeff. (mean) and its SE based on
500 bootstrapping samples.
► This approach does not need any distribution
assumption.
► Therefore, if the normality assumption is not
met, we can safely use bootstrapping.

Average of
regression coeff.

64
► Not only that, traditional methods do not have formula to calculate 95% CI for
some statistics (e.g. R2). Bootstrap can do this.

► In Analysis Properties, under “Bootstrap” tab, do the following:

► Then, in AMOS output:

65
 Structural Equation Modeling
using AMOS

T13: Bayesian Estimation

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

66
 Bayesian Estimation
► Inferential statistics using Maximum Likelihood Estimation or Least-square
estimation: Population parameter is totally unknown & fixed.
► But in Bayesian approach, it is considered unknown but random, and allows to
use whatever known information to add to the analysis.
► E.g. A test score is 0 to 100. We are sure that the mean score of a group must
be between 0 and 100, and the variance is between 0 & 502. If we put a little
more effort, we may come up with a narrower range than this. This is called
"Prior distribution" (known information before analysis).
► The parameter is not totally unknown but proposed with distribution.
► This "prior distribution", together with the information from the existing data,
the estimation is made (i.e. Posterior distribution).
► What is the benefit? Adding prior information into analysis, the estimation
becomes more efficient (smaller CI width for the same sample size) and more
power to reject the null with the same sample size.
► This facility is available in AMOS.

67
 Bayesian Estimation
kvsperform2.sav Traditional result:

► How to do?
► View -> Analysis properties -> Estimation -> Estimate
means and intercepts
► Analyze -> Bayesian estimation : then Bayesian estimation
starts.
68
 Number of samples taken
►Right click on one
row. Select "Show
prior"
►Enter the range of
prior distribution.

Note: SD is the real standard error in this table.

►E.g. select normal
distribution, mean=1,
SD=0.3, then "Apply"
and "Close"
►Then you may get
this output
►SE = 0.218 (traditional
SE 0.292)

Resampling method used is Monte Carlo Markov chain (MCMC) algorithm.

69
 Structural Equation Modeling
using AMOS

T14: Other issues

by
Lin Naing @ Mohd. Ayub Sadiq
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam

70
Sample size issue:
AMOS 6.0 guideline (p 70):
The table shows that we need
sample size of 200 and above in
order to get correct P value (in
testing correlation using t test.)

* How to read this table?

With sample size 10, AMOS (SEM) will
give P value 0.073 when the true P value
is 0.05. It will give 0.018 when the true
value is 0.01. AMOS (SEM) will over
estimate the P value if the sample size is
smaller.

Sample size should be also considered according to ….

• Factor analysis rule of thumb (5-10 per item) for measurement model
• Sample size calculation for regression analysis (e.g. using PS software).
71
Missing data handling (Ch 17, 18, 30 & 31):

► Listwise deletion
► Pairwise deletion
► Data imputation (missing replacement)
All methods assume that data are missing at random.

In AMOS, just by selecting “Estimate means and

intercepts”, the analysis will run it but no data
imputation is done (Just by deletion).
Ex
17
.am
However, this approach still has disadvantage. E.g. w
AMOS will not give ‘modification index’.

72
Missing data handling (Ch 17, 18, 30 & 31):
Ex
1 7. a
Available Data Imputation in AMOS: m
w

Regression imputation (no Multiple Imputation) (easy to use)

Stochastic regression imputation (with MI) (Difficult to use it)
Bayesian imputation (with MI) (Difficult to use it)

At least with single data imputation, we can

replace the missing and continue analysis as
usual.

Of course, multiple imputation is preferable but

still we have to do some manual calculation.

Watch out. AMOS may improve.

Rubin, D. B. 1987. Multiple imputation for nonresponse data in surveys. New

York: John Wiley & Sons, Inc..
73
Missing data handling (Ch 17, 18, 30 & 31):
Alternative way is that:

We can analyze to obtain correlation matrix, n, mean and SD of observed

variables (without replacing missing data).
Arrange the data in Excel as below. Use this Excel sheet as data for
AMOS. Then no more missing data problem …

74
 Thank you
very much.

Dr Lin Naing @ Mohd. Ayub Sadiq

Senior Lecturer
PAPRSB Institute of Health Sciences
Universiti Brunei Darussalam
Jalan Tungku Link, Gadong BE 1410
Negara Brunei Darussalam
Email: naing61@gmail.com
75