DOI 10.

1515/ijdhd-2014-0305 Int J Disabil Hum Dev 2014; 13(2): 191–204

Review

Daniel T.L. Shek* and Lu Yu

Confirmatory factor analysis using AMOS:

a demonstration
Abstract: The purpose of this paper is to demonstrate the identify/confirm a relatively small number of factors/
process of using AMOS to test first- and higher-order con- latent variables that can explain approximately the same
firmatory factor analysis (CFA) models. We performed the amount of variances in the observed variables. In this
analyses with the AMOS 17.0 statistic package, a very user- way, the overall complexity of the large set of original data
friendly program for structural equation modeling. In this (observed variables) can be reduced and simplified [1]. The
paper, we describe the concepts, theories, and basic steps procedure of factor analysis was developed a long time
of conducting CFA as well as provide a general introduc- ago [2]. However, due to the complex steps involved in its
tion to the software AMOS. The process of conducting two process, researchers only began using it when several sta-
different types of CFA within the framework of AMOS pro- tistical software packages (e.g., SPSS, SAS, LISREL, Mplus,
gram (first-order CFA and higher-order or hierarchical CFA) etc.) became commercially available in recent decades [3].
are illustrated based on the data collected from 604 sec- The utility of factor analysis has become very popular in
ondary school teachers involved in the Project P.A.T.H.S. in different contemporary scientific research fields. It has
Hong Kong. The factor structure of a subjective outcome also been described [4] as “one of the most powerful tools
evaluation form developed to assess program implement- yet devised for the study of complex areas of behavioral
ers’ subjective evaluation about the project was examined. scientific concern” (p. 689).
Confirmatory factor analysis (CFA) is one of two basic
Keywords: AMOS; confirmatory factor analysis; first-order types of factor analysis, with another type being explora-
factor model; hierarchical model; research methods. tory factor analysis (EFA). Compared with EFA, CFA is
theory-driven and aims to determine the ability of a pre-
defined factor model (specified on the basis of theory) to
*Corresponding author: Professor Daniel T.L. Shek, PhD, FHKPS,
BBS, SBS, JP, Associate Vice President (Undergraduate Programme),
fit an observed data set. In other words, CFA examines
and Chair Professor of Applied Social Sciences, Faculty of Health whether a specified set of factors (i.e., determined by a
and Social Sciences, Department of Applied Social Sciences, The theory in an a priori manner) really determines the vari-
Hong Kong Polytechnic University, Room HJ407, Core H, Hunghom, ances in the observed variables in the way that previously
Hong Kong, P.R. China, E-mail: [email protected] hypothesized. Gorsuch [5] commented that “confirmatory
Daniel T.L. Shek: Department of Applied Social Sciences, The Hong
factor analysis is powerful because it provides explicit
Kong Polytechnic University, Hong Kong, P.R. China; Centre for
Innovative Programmes for Adolescents and Families, The Hong hypothesis testing for factor analytic problems”, and con-
Kong Polytechnic University, Hong Kong, P.R. China; Kiang Wu sidered it to be “more theoretically important and should
Nursing College of Macau, Macau, P.R. China; and Department of be the much more widely used of the two major factor ana-
Social Work, East China Normal University, P.R. China lytic approaches” (p. 134).
Lu Yu: Department of Applied Social Sciences, The Hong Kong
In another paper of this special issue, basic concepts
Polytechnic University, Hong Kong, P.R. China
and theories with regard to CFA and EFA as well as their
similarities and differences have been introduced [6],
which are not repeated in this paper. Instead, the current
Introduction paper attempts to provide a more practical guide in using
CFA modeling. In particular, two different applications
Factor analysis refers to a set of mathematical methods of CFA in human development research are described.
that have been developed to analyze the interrelationships The procedures of testing for the factorial validity of both
among a series of observed variables and their underly- first-order CFA model and higher-order CFA model (a sec-
ing constructs, called factors or latent variables. By ana- ond-order model in this paper) are demonstrated with a
lyzing the pattern of correlations/covariances among a popular structural equation modeling software (AMOS).
large number of observed variables, factor analysis helps The testing of factorial invariance using AMOS 17.0 (SPSS

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM
192 Shek and Yu: Confirmatory factor analysis

Inc., Chicago, IL, USA) [7] is demonstrated in another ultimately rests on the underlying theory and the mean-
paper in this special issue. ingfulness of the model.

First-order model and higher-order Steps in CFA modeling

model Similar to the procedures for conducting structural equa-
tion modeling (SEM), CFA typically proceeds through the
In a CFA model, if the variances of the observed vari- following steps: literature review, model specification,
ables are believed to be influenced by only one level of model identification, data collection and primary analy-
latent variables (first-order factors), this model is called ses, parameter estimation, model fit assessment, model
primary factor model or first-order model. In reality, as comparison and modification, and presentation and
models often allow the first-order factors to be corre- interpretation of results. In the introductory paper on SEM
lated, it sometimes makes sense to consider a second- of this special issue, basic steps from model specifica-
order or even higher-order model, which describes the tion to goodness-of-fit assessment have been thoroughly
correlations among the first-order factors. These CFA described. Here, we will not discuss these procedures
models that comprise second or higher-order factors are again. Instead, we discuss in detail the process of “model
called higher-order factor model or hierarchical model. modification and comparison”.
In a higher-order model, the first-order factors are further
explained by higher-order factors. Figure 1 shows an
example of a second-order model. While the three first- Model modification
order factors, F1, F2 and F3, explain the observed vari-
ables M1 to M3, M4 to M6 and M7 to M10, respectively; F1 If the initially hypothesized model does not fit the data
and F2 are explained by a second-order factor, F4. well, researchers often want to re-specify the model in a
In fact, the higher-order model is a special case of post-hoc fashion. Based on fit diagnostic information and
the first-order model, with the additional restriction that theoretical justification, the model is revised and fit to the
structure be imposed on the correlational pattern among data again in order to improve its goodness of fit or the par-
the first-order factors [8]. Given the same number of esti- simony and interpretability of the model. One approach to
mable parameters, fit statistics related to a model speci- model modification is to delete existing parameters – a
fied either as a first-order model or as a higher-order model process known as model trimming. Paths in the model that
will basically be equivalent. It should be highlighted that are not significant (p > 0.05) often indicate a wrong factor
the decision as to whether or not a measuring instrument loading. With theoretical interpretation, these paths can
shall be modeled as a first-order or second-order structure be deleted. The advantage of model trimming is that we

Figure 1 A second-order CFA model.

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM
 Shek and Yu: Confirmatory factor analysis 193

can obtain a simple model as well as a larger degree of Nested model comparison
freedom (df), which result in better model fit.
Another more common method of model modifica- Two models are nested when a subset of free param-
tion is to add more parameters into the model, namely, eters in one model is constrained in the other model.
model building. Two sets of statistics help researchers For example, the two models in Figure 2 are nested. The
decide which parameters shall be added. The first one is unrestricted model (Mu) becomes the restricted model
called “Modification Indices” (MIs). All freely estimated (Mr) by fixing the factor correlation between F1 and F2
parameters are set as having MI values equal to zero by as zero. The unrestricted model always fits better than
default. For each fixed parameter specified, there is an MI the restricted model (X2_Mu < X2_Mr), because the unre-
and its value represents the expected drop in chi-square stricted model always requires estimating more parame-
if the parameter is to be freely estimated. According to ters. The difference on X2 of nested models (X2_Mr-X2_Mu)
Jöreskog and Sörbom [9], a fixed parameter with the is also distributed as a X2 distribution with degrees of
largest modification index should be freed provided that freedom = df_Mr-df_Mu, called X2 difference test. If the
this parameter can be theoretically justified. For each MI, X2 difference test is non-significant, the parsimonious
there is an expected parameter change (EPC) value [10]. model (Mu) is preferred. The X2 difference test is valid for
This parameter represents the predicted change for each nested models only.
fixed parameter in the model. The larger the values of
MI and EPC, the more substantive sense the associated
relationships would make when they are added into the
Non-nested model comparison
model for a better fit. An important issue in model modi-
fication is that the modification is theoretically justified.
Non-nested model comparisons are used for models that
Furthermore, the modified model usually needs to be
are not nested, such as Models 1 and 2 in Figure 3. Two
replicated in another sample, i.e., cross-validation. There
popular methods are the Akaike Information Criterion
are different criteria [11, 12] as to what MI value indicates
(AIC) [13] and the Expected Cross-Validation Index (ECVI)
significant specification error (e.g., 3.84, 50, 100, etc.). In
[14]. These indices take into account both model fit (X2)
this study, covariances with MI values > 80 and regression
and model complexity (df). The ECVI also considers
weights with MI values > 50 are considered as indicating
sample size by incorporating a greater penalty function
misspecification, thus requiring further attention.
for fitting a non-parsimonious model in a smaller sample.
AMOS provides both AIC and ECVI values.

Model comparison
When there are several theoretically competing models, AMOS
researchers must identify which model fits the data better.
This requires model comparison. There are basically two AMOS, which stands for “analysis of moment structures”,
types of model comparison: nested model comparison is one of the most popular software packages devel-
and non-nested model comparison. oped for performing SEM [15, 16]. Compared with other

Unrestricted model (Mu) Restricted model (Mr)

Figure 2 Two nested models.

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM
194 Shek and Yu: Confirmatory factor analysis

Model 1: a three-factor model Model 2: a two-factor model

Figure 3 Two non-nested models.

programs (e.g., LISREL and Mplus), AMOS is commonly aspects of implementers’ perceptions about the program:
regarded as more user-friendly. AMOS has two different content, instructors’ performance, and the overall effec-
modes of model specification: AMOS Graphics and AMOS tiveness of the program. Thus, there are theoretically three
Basics. AMOS Graphics adopts a graphical interface, by factors inherent in the scale, including views towards
which people work directly from a path diagram using program (VC: 10 items), views towards implementers (VI:
the drawing tools designed with SEM conventions in 10 items), and views about program effectiveness (EF: 16
mind. Meanwhile, AMOS Basics is based on syntax and items).
all models can be specified using an equation format. The data were collected during the 2008–2009 school
For both AMOS Graphics and Basics, options related to year based on the responses of 2007 program implement-
the analyses are available from drop-down menus, and ers who taught the Project P.A.T.H.S. in the year of evalu-
estimates can be presented in text or table format. AMOS ation. There were 699, 704, and 604 participants teaching
Graphics also produces readily publishable output figures the program at Secondary 1 (S1), Secondary 2 (S2) and
and allows for the estimates to be displayed graphically. Secondary 3 (S3) levels, respectively. In this study, we only
AMOS is the most commonly used software by researchers use a subsample of program implementers taught at the
in published social work research using SEM in top-tier S3 level. CFA was first conducted to test a primary-order
journals [17]. In the next part of this paper, we will focus factor model of Form B. Second, a higher-order factor
on the general format of the AMOS Graphics program and model was further examined. Data collected from S1 and
demonstrate the process of conducting CFA using this S2 teachers shall be used in another paper in this issue to
approach. demonstrate how to test invariance of the factorial struc-
ture of Form B.

Conducting CFA with the AMOS 17.0

program Examining the primary-order factor
structure of Form B
In this section, we focus on illustrating the process of per-
forming CFA with the AMOS 17.0 program. The purpose The hypothesized first-order factorial model consists of
is to provide the readers with a step-by-step instruction three factors, VC, VI and EF, which are graphically shown
on conducting CFA without difficulties. The example we in Figure 4. It is assumed that the three factors are inter-
used aims to testing the factor structure of a scale (Form correlated because they measure different but related
B) developed by Shek et al. [18] to measure program aspects of one construct (program implementers’ subjec-
implementers’ subjective evaluation of a positive youth tive evaluation of the program). Each item (observed vari-
development program (the Project P.A.T.H.S.) in Hong able) is regressed onto its respective factor. Reliabilities of
Kong. Specifically, Form B was designed to measure three the three subscales are affected by random measurement

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM
 Shek and Yu: Confirmatory factor analysis 195

Figure 4 Hypothesized primary-order three-factor structure of Form B.

errors, as indicated by the error terms associated with displayed on the left side. First, in the drop-down menu,
each item. In the following sections, we demonstrate the we clicked “Data Files”, chose “File Name”, and selected
process of performing the analyses with the use of AMOS the dataset we planned to analyze. If we click “View Data”,
17.0 program. As the output is very lengthy, all appendices the data we selected will be opened separately in its own
are placed in the website of the P.A.T.H.S. Project (www. format (e.g., SPSS, EXCEL, etc.). In other words, we could
paths.hk). directly import different data files in AMOS.
The next step is to draw the path diagram, so we
clicked the “Draw a Latent Variable or Add an indicator
Step 1: Data screening to a latent variable” icon once. We placed the cursor
where we wanted to draw the ellipse for the first latent
The first step in CFA is to conduct data screening. We scru- variable and clicked again. After drawing the ellipse rep-
tinized the sample data with SPSS before conducting CFA resenting the first latent factor, we added the observed
to check whether the basic assumptions of conducting variables by clicking the icon again with the cursor on
SEM can be met. We found that all variables were normally the center of the first latent variable. This action produced
distributed using the criterion that the univariate skew- a rectangle (representing a single observed variable), an
ness and kurtosis must be lower than 2 and 7, respectively arrow pointing from the latent factor to the observed vari-
[19]. No substantial outliers were identified. The sample able (representing a regression path), and a small circle
size of 604 was sufficient to perform CFA. with an arrow pointing toward the observed variable (rep-
resenting a measurement error term). Given that the factor
VC has 10 items, we added these additional observed vari-
Step 2: Model specification with AMOS ables with several more clicks. The AMOS then produced
Graphics additional indicator variables, each with its associated
error term. With the same procedure, the other two factors
After data screening, we can begin drawing the path and their related items can be drawn.
diagram and specifying the model within the framework To add the relationships among factors to the path
of AMOS Graphics. Once the AMOS program is opened diagram, we used the “Covariance” icon to draw
(Supplemental Data, Appendix 1, which accompanies double-headed arrows among the latent variables.
the article online at http://www.degruyter.com/view/j/ We drew three double-headed arrows because it was
ijdhd.2014.13.issue-2/issue-file = s/ijdhd.2014.13.issue-2. hypothesized that the three factors are correlated with
xml), one can see a blank rectangle, which is the space one another. The last step was to label each variable.
where the path diagram should be drawn. The tools are This was achieved by right clicking the mouse when

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM
196 Shek and Yu: Confirmatory factor analysis

the cursor was placed on the variable of interest, choos- Step 4: Read AMOS text output
ing the option “Object Properties” from the right-click
menu, and then entering the related names of each vari- The AMOS output mainly includes two parts, namely,
able in the object properties dialog box. Alternatively, (A) model, parameters and estimation summary; and (B)
we were able to display the list of variables in the data model assessment.
by clicking the icon , and then dragging each vari-
able to its respective rectangle until each variable was
Summary of model, parameters, and estimation
labeled, as shown in Figure 4.
At this point, the drawing of path diagram related to
First, a summary of analysis and notes for group are pro-
the hypothesized primary-order CFA model (model speci-
vided in “Analysis Summary” and “Notes for Group”,
fication) was basically complete. The next task is to test
including information about the time of data analysis,
how well the hypothesized model fits with the real data,
title of the file, group number, model nature (recursive
i.e., parameter estimation. Prior to discussing how the
or non-recursive), and sample size. Second, all variables
model should be tested, we first review the different parts
in the model are listed in “Variable Summary” accom-
of the model and their relationships with one another.
panied by their categorization as observed, unobserved,
First, there are three latent variables, as indicated by
endogenous, and exogenous variables. This is followed by
the three ellipses labeled VC, VI, and EF. Second, the three
a summary of the total number of variables in the model
factors are inter-correlated, as indicated by the three two-
and the number in each of the four categories.
headed arrows. Third, there are 36 observed variables,
The Program next provides a summary of the param-
reflected by the 36 rectangles (VC_1 to VC10, VI_1 to VI_10,
eters in the model (“Parameter Summary”). As shown in
and EF_1 to EF_16); they represent item pairs from the
Figure 5B, there are 72 regression weights in total, 39 of
three subscales of VC, VI and EF, respectively. Fourth, the
which are fixed and 33 of which are estimated (“Unla-
observed variables load on the factors in the pattern: VC_1
beled”). The 39 fixed regression weights include the first
to VC10 load on Factor 1, VC; VI_1 to VI_10 load on Factor
of each set of three factor loadings and the 36 error terms.
2, VI; and EF_1 to EF_16 load on Factor 3, EF. Fifth, each
There are three covariances and 39 variances (3 factors
observed variable loads on only one factor. Sixth, errors of
and 36 error terms), which are all estimated.
measurement associated with each observed variable are
The next section, “Notes for Model” focuses on an
uncorrelated with each other.
overall summary of the model to be tested and the data to
be used. In Figure 5C, there are 702 pieces of known infor-
mation (“Number of distinct sample moments”), in which
Step 3: Parameter estimation
111 parameters are to be estimated, leaving 591 degrees of
freedom. This suggests that this model is over-identified.
To execute the data analyses process, we simply clicked
Finally, the chi-square value (2832.206), together with its
the “Calculation Estimates” icon . Before this, however,
degrees of freedom (591) and probability value (0.000) are
one must check the names of observed variables care-
provided as an overview of the model fit.
fully and ensure their consistency with the names in the
dataset. Any typographical error in the process of entering
variable labels would lead to the failure of the calculation
Model assessment
process.
To evaluate data with missing values, we checked
The primary interest in CFA is to determine the extent to
the option “Estimate means and interpret” in the “Anal-
which the hypothesized model could adequately describe
ysis Properties” window (Supplemental Data, Appendix
the sample data. AMOS provides evaluation of model fit
2a), which can be activated by clicking the icon on
from both the adequacy of the parameter estimation and
the palette. To obtain more results in the output, we
the model as a whole.
checked the corresponding blocks in the same window
under the sub-window of “Output” (Supplemental Data,
Appendix 2b). In this case, we selected “standardized Individual parameter estimation
estimation” to obtain the standardized regression coef-
ficients for each path. Once the analyses are completed, Individual parameter estimates for the current CFA model
the results can be reviewed by clicking the “View Text” in AMOS output are fully listed in Supplemental Data,
icon . The output is shown as in Figure 5A. Appendix 3. Results are presented separately for the factor

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM
 Shek and Yu: Confirmatory factor analysis 197

AMOS text output windows

A

AMOS text output: parameter summary

B

AMOS text output: notes for model

C

Figure 5 AMOS text output: summary.

loadings (“Regression Weights”), intercepts, covariances, considered unimportant to the model or indicating an
correlations, and variances. A selective part is shown insufficient sample size [21]. An examination of the output
in Table 1 for illustration. The fit of individual param- (Supplemental Data, Appendix 3) reveals all estimates to
eters in the model can be evaluated from three aspects: be reasonable and statistically significant; furthermore,
(a) whether the parameter estimates exhibit the correct all standard errors appear to be in a correct manner.
sign and size, which usually indicates either the model is
wrong or the input matrix lacks sufficient information; (b)
whether there is any excessively large or small standard Parameter estimation for the whole model
error, an indicator of poor model fit [9, 20]; and (3) whether
the parameter estimates are statistically significant. Non- The overall X2 value and its df and probability values are
significant parameters, except for error variances, are provided in the previous session. In addition to this, AMOS

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM
198 Shek and Yu: Confirmatory factor analysis

Table 1 AMOS output: parameter estimates (a selective part).

Estimate S.E. C.R. p Label

Regression weights
EF_16 < --- EF 1.000
EF_1 < --- EF 0.856 0.041 21.037 *** par_15
VC_1 < --- VC 1.000
VC_2 < --- VC 1.512 0.102 14.894 *** par_16
VI_1 < --- VI 1.000
VI_2 < --- VI 0.955 0.051 18.751 *** par_25
Standardized regression weights
EF_16 < --- EF 0.842
VC_10 < --- VC 0.873
VI_10 < --- VI 0.813
Intercepts
EF_15 3.050 0.033 93.623 *** par_38
VC_1 4.688 0.031 152.681 *** par_53
VI_1 4.304 0.035 124.415 *** par_63
Covariances
VI < -- > EF 0.235 0.235 10.677 *** par_34
EF < -- > VC 0.211 0.211 10.834 *** par_35
VI < -- > VC 0.204 0.020 10.123 *** par_36
Correlation
VI < -- > EF 0.598
EF < -- > VC 0.755
VI < -- > VC 0.706
Variances
EF 0.380 0.030 12.745 *** par_73
VC 0.205 0.026 7.964 *** par_74
VI 0.407 0.038 10.589 *** par_75
e12 0.215 0.013 16.109 *** par_77

provides many other fit statistics in the section “Model mean square error of approximation (RMSEA). According
Fit”. The goodness-of-fit statistics for the current hypoth- to Hu and Bentler [22], the CFI, TLI, and NFI values should
esized three-factor CFA model are fully presented in Sup- be equal to or above 0.90 to indicate a satisfactory model
plemental Data, Appendix 4. A selective part of the results fit, with values close to 1 suggesting a very good fit on any
is shown in Figure 6. of the indexes; for RMSEA, a value of < 0.11 indicates a
For each set of fit statistics, there are three rows. First, reasonable fit and a value of < 0.05 or less indicates good
“Default model” represents the statistics for the hypoth- model fit in relation to the degrees of freedom.
esized model; the second row, “Saturated model”, repre- As can be seen in Table 2, the indexes suggest an
sents for the statistics for a just-identified model in which insufficient fit of the model to the current data (CFI = 0.868,
the number of estimated parameters equals to the number TLI = 0.860, NFI = 0.839, and RMSEA = 0.079). In other
of data points; and the third row, “Independence model”, words, the hypothesized model was only fairly fitted and
focuses on a model that hypothesizes all variables are model modification might be needed. Therefore, the next
uncorrelated (i.e., Null Model). task is to identify the areas of misfit in the model and
In the paper on SEM of this special issue, the authors modify the current model.
have introduced common goodness-of-fit assessment and a
process to determine which indicators shall be included in
a report [6]. It has been recommended that researchers shall Step 5: Model modification
report the X2 test statistic, some incremental fit indices,
and some residual based indices. For the current model To detect model misspecification, AMOS yields important
testing, five goodness-of-fit indexes are reported in Table 2, diagnostic information, which can be obtained by clicking
including the X2, comparative fit index (CFI), Tucker and on the “Analysis Properties” icon and checking the item
Lewis’s index of fit (TLI), normed fit index (NFI), and root “Modification Indices” in the “Output” session. Then, the

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM
 Shek and Yu: Confirmatory factor analysis 199

Figure 6 AMOS output: goodness-of-fit statistics.

AMOS output file generates a set of MIs and their asso- items VC_5 and VC_6 (e41 and e42: MI = 98.132), and
ciated EPCVs, which provide important information for between items EF_8 and EF_7 (e19 and e20: MI = 88.466).
model modification. As mentioned, MIs of values > 80 and These MI values represent misspecified error covari-
50 deserve our attention for covariances and regression ances, which refer to systematic measurement errors in
weights, respectively. item responses. Such error may come from characteristics
The MIs and EPCV for possible covariances and specific to the items or to the respondents [23]. For the
regression weights for the present model are selectively regression weights, although seven items showed cross-
listed in Table 3 (complete results can be seen in Supple- loadings, the cross-loading of item VI_6 on VC factor
mental Data, Appendix 5). In reviewing the parameters in had the highest MI value (51.961), which accounted for
the covariance part, four parameters showed particularly substantial misspecification of the hypothesized factor
large values, representing the covariances between items loading. This may indicate that item VI_6, in addition to
VI_7 and VI_8 (e53 and e54: MI = 228.462), between items measuring views towards program implementers, also
EF_15 and EF_14 (e12 and e13: MI = 214.544), between measures views towards the program.

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM
200 Shek and Yu: Confirmatory factor analysis

Table 2 Goodness-of-fit statistics for the three-factor CFA model.

Model tested χ2 Df CFI TLI NFI RMSEA

Model performance 2832.21 591 0.868 0.860 0.839 0.079

Criterion for goodness of fit – – ≥ 0.90 ≥ 0.90 ≥ 0.90 ≤ 0.10

Note: CFI, comparative fit index; TLI, Tucker and Lewis’s index of fit; NFI, normed fit index; RMSEA, root mean square error of approximation.

A review of the items in the questionnaire revealed between errors of item VC_5 and item VC_6; Model M5
several observations. First, item VI_6, read as “I gained a included a correlated error between items EF_8 and
lot during the course of instruction”, appeared to measure EF_7. The specification of these models simply entails
not only how the implementer views his/her own perfor- the inclusion of one regression path and/or a few cor-
mance, but also the implementers’ perception about the related errors, which can be quickly done by using the
program, i.e., the factor VC. Therefore, it would be mean- “Draw path” and “Draw covariances” functions in AMOS
ingful to add a regression path from VC to VI_6. Second, graphics.
items VI_7 and VI_8 both talk about teacher’s caring The goodness-of-fit indexes of the five modified
and willingness to help students; items VC_5 and VC_6 models as well as the initial model are summarized in
are related to students’ participation and interaction in Table 4. As can be seen, almost all indexes supported a
class; items EF_15 and EF_14 are about promoting stu- good fit of M5 to the current data (CFI = 0.92, TLI = 0.91,
dents’ prosocial attributes; and items EF_8 and EF_7 both NFI = 0.89, IFI = 0.92, and RMSEA = 0.06). Given that the six
linked with students’ behavioral competence and self- models were nested, X2 difference tests were conducted.
determination. Obviously, there may be some overlaps in The results showed that the differences in X2 values
the meanings of these pairs of items. As such, it seemed to between M5 and the other five models (M0 to M4) are all
make sense that we include these error correlations in the statistically significant (p < 0.01). The AIC value for M5
re-specified model. is also much smaller than the original model and other
Thus, the next step was to re-specify the original modified models. Based on these findings, it may be con-
model (named as M0). Five modified models were pro- cluded that the revised three-factor model, M5, provides a
posed based on the above analyses. In Model M1, the significantly better fit to the data than other models. The
path from factor VC to item VI_6 was specified as a free final model, M5, is displayed in Figure 7.
parameter; Model M2 included a correlated error between Finally, the results of the CFA for the final model can
errors of item VI_7 and item VI_8 based on Model M1; be easily presented schematically with AMOS Graphics.
Model M3 further included a correlation between errors After completion of the analyses, we just needed to click
of item EF_15 and item EF_14; Model M4 added the path the “View the Output Path Diagram” icon (Supplemental

Table 3 Modification indices of covariances for the hypothesized model.

Covariances MI EPCV Covariances MI EPCV Covariances MI EPCV

e53 < -- > e55 20.84 0.06 e41 < -- > e55 27.04 0.09 e35 < -- > e40 4.15 –0.02
e53 < -- > e54 228.46 0.15 e41 < -- > e48 639 –0.04 e35 < -- > e39 6.66 –0.03
e52 < -- > e54 4.37 –0.03 e41 < -- > e42 98.13 0.17 e33 < -- > e50 4.48 –0.02
e46 < -- > e53 21.99 –0.05 e40 < -- > e42 54.58 0.12 e31 < -- > e54 5.23 –0.02
e44 < -- > e52 42.38 0.12 e38 < -- > e53 4.44 0.03 e19 < -- > e20 88.47 0.08
e44 < -- > e45 25.71 0.07 e38 < -- > e49 6.47 –0.04 e18 < -- > e52 5.12 –0.03
e15 < -- > e32 5.65 –0.03 e13 < -- > e17 5.06 0.03 e12 < -- > e13 214.54 –0.15
e14 < -- > e52 5.48 –0.03 e13 < -- > e14 47.72 0.07 e11 < -- > EF 4.24 –0.02

Regression Weight MI EPCV Regression Weight MI EPCV Regression Weight MI EPCV

VI_8 < --- VC 5.53 –0.11 VI_4 < --- VC 11.60 –0.14 VC_7 < --- VI 5.19 –0.11
VI_8 < --- EF 5.35 –0.08 VI_4 < --- EF 11.29 –0.10 VC_2 < --- EF 4.26 –0.09
VI_6 < --- VC 51.96 0.43 VI_1 < --- VC 7.35 0.15 VC_1 < --- VI 8.84 –0.12
VI_6 < --- EF 39.65 0.27 VI_1 < --- EF 5.03 0.09 EF_14 < --- EF_15 4.34 –0.01
VI_6 < --- VC_8 4.81 0.01 VC_10 < --- VI 8.57 –0.10 EF_15 < --- EF_14 5.22 –0.01

Note: the bold entries are with large values and referred to in the text.

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM
 Shek and Yu: Confirmatory factor analysis 201

Table 4 Goodness-of-fit statistics for the modified primary-order CFA models.

Model Modification X2 Df CFI TLI NFI IFI RMSEA AIC

M0 Original model 2832.21 591 0.87 0.86 0.84 0.87 0.08 3054.21
M1 M0 + path VC to VI_6 2706.07 590 0.88 0.87 0.85 0.88 0.08 2930.07
M2 M1+ correlated errors of VI_7 and VI_8 2435.94 589 0.89 0.88 0.86 0.89 0.07 2661.94
M3 M2+ correlated errors of VC_5 and VC_6 2327.61 588 0.90 0.89 0.87 0.90 0.07 2555.61
M4 M3+ correlated errors of EF_13 and EF_14 2080.25 587 0.91 0.91 0.88 0.91 0.07 2310.25
M5 M4+ correlated errors of EF_7 and EF_8 1993.33 586 0.92 0.91 0.89 0.92 0.06 2225.33
Criterion for goodness of fit – – ≥ 0.90 ≥ 0.90 ≥ 0.90 ≥ 0.90 ≤ 0.10 –

Note: CFI, comparative fit index; TLI, Tucker and Lewis’s index of fit; NFI, normed fit index; IFI, incremental fit index; RMSEA, root mean
square error of approximation; AIC, Akaike information criterion.

Data, Appendix 6), and then a readily publishable figure and perceived effectiveness of the program (EF), are
describing the CFA model can be produced. We can either explained by one higher-order factor of program imple-
choose to present the unstandardized estimates (Supple- menter’s general evaluation about the project. Therefore,
mental Data, Appendix 7a) or the standardized estimates the Form B data may be presented by a hierarchical facto-
(Supplemental Data, Appendix 7b). rial structure.
Specifically, the proposed hierarchical model of
Form B has the following four hypotheses based on
the results of the primary-order factor model. First,
Examining the higher-order factor responses to items of Form B can be explained by three
structure of form B first-order factors (VC, VI and EF), and one second-
order factor (Implementer’s subjective evaluation, ISE).
As discussed earlier, a model is called high-order factor Second, except for item VI_6, each item of Form B has a
mode if the first-order factors are allowed to be correlated non-zero loading on the first-order factor as it is designed
and can be further explained by higher-order factors. to measure zero loadings on the other two first-order
In our example, it is reasonable to hypothesize that the factors; VI_6 has a cross-loading on VC. Third, error
three first-order factors, namely, views towards program terms associated with four pairs of items are correlated
content (VC), views towards program implementer (VI) (VI_7 and VI_8; EF_15 and EF_14; VC_5 and VC_6, EF_8

Figure 7 The final three factor model of Form B.

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM
202 Shek and Yu: Confirmatory factor analysis

and EF_7), while other error terms linked with items are on the values of CFI, TLI, IFI and RMSEA, we can deter-
uncorrelated. Fourth, covariation among the three first- mine that the second-order factor model fits the current
order factors would be explained by their regression on data at a satisfying level. It should be noted that the esti-
the second-order factor, ISE. mates of this hierarchical model are the same as those
Several aspects must be considered while testing of the primary factor model (M5) we tested earlier (see
a hierarchical model. First, the single-headed arrows Table 5). Given the same number of estimable parameters,
from second-order factor to first-order factor represent goodness-of-fit statistics of a model specified as a three-
second-order loadings that are all freely estimated. For primary factor model would be equivalent to those of the
model identification, a constraint must be placed either model specified as a second-factor model. As Rindskoph
on the variance of an independent factor or on one of and Rose [8] explained, the second-factor model can be
the regression paths. The two parameters cannot be regarded a special case of the primary factor model; with
estimated simultaneously. Due to the fact that the rela- the added restriction that structure be imposed on the cor-
tionships between higher-order factor and each lower- relational pattern among the first-order factors. Neverthe-
order factors are more important than the variance of less, the decision of whether or not an instrument should
higher-order factor in the hierarchical model, usually the be structured as a first-order or second-order factor model
variance of the high-order factor is constrained to be 1.0. shall still depend on the underlying theory and meaning-
Second, the first-order factors serve as both independ- fulness of the model.
ent and dependent variables; thus, it follows that their In our example, the three first-order factors (views
respective variances and covariances are no longer esti- towards program, views towards implementer, and per-
mable parameters but presumed to be explained by the ceived program effectiveness) all reflect one underly-
higher-order factor. Therefore, there are no two-headed ing construct – the program implementer’s general
arrows connecting the first order factors. Third, errors subjective evaluation of the whole project. Therefore,
are assumed to exist in the prediction of each first-order the higher-order factor model is considered preferable to
factor from the second-order factor. Thus, each first-order the primary-order factor model. Given the above findings
factor is associated with a residual error. Figure 8 displays and theoretical consideration, it may be concluded that
the hierarchical model. the implementer’s subjective outcome evaluation form
Analyses based on the re-specified second-order (Form B) can be best represented as a second-order three-
model were conducted using the same procedure in testing factor structure. The goodness-of-fit of this model to the
the primary factor model. Goodness-of-fit statistics of the current data collected from a large sample of secondary
hierarchical model to the data are shown in Table 5. Based school teachers provides strong evidence for the construct

Figure 8 The hypothesized second-order factorial structure of Form B.

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM
 Shek and Yu: Confirmatory factor analysis 203

Table 5 Goodness-of-fit statistics for the second-order CFA model.

Model χ2 Df CFI TLI NFI IFI RMSEA AIC

Second-order factor model 1993.33 586 0.92 0.91 0.89 0.92 0.06 2225.33
Criterion for goodness of fit – – ≥ 0.90 ≥ 0.90 ≥ 0.90 ≥ 0.90 ≤ 0.10 –

Note: CFI, comparative fit index; TLI, Tucker and Lewis’s index of fit; NFI, normed fit index; IFI, incremental fit index; RMSEA, root mean
square error of approximation; AIC, Akaike information criterion.

validity of Form B. This finding suggests that this instru- The present findings strongly suggest that Form B, a
ment can effectively evaluate the subjective outcome of subjective outcome evaluation scale that has been used in
the Project P.A.T.H.S. from program implementers’ per- previous studies [24], is an instrument possessing factorial
spectives. However, it should be acknowledged that the validity. Establishing sound construct validity of a meas-
sample in this demonstration only involved S3 teachers. urement tool has always been an important issue in ques-
Therefore, it would be meaningful to investigate the gen- tionnaire development and application. However, such
eralizability of this instrument by examining its factorial analyses are rarely done in positive youth development
invariance across different teacher groups. The process of research, particularly in different Chinese contexts. In
testing for factorial invariance will be discussed in another view of the lack of validated positive youth development
paper of this special issue. programs in various Chinese contexts [25], the authors of
the present work argue that it is important to step up eval-
uation tasks that definitely require validated instruments.
The use of validated instruments in the Project P.A.T.H.S.
Summary is a good example in this field [24, 26, 27]. It is our modest
wish that this paper can facilitate future applications of
The present paper demonstrate the basic concepts of and
CFA. With the hands-on guide provided in this paper,
procedures in conducting CFA via AMOS, a popular struc-
future researchers may now find it easier to conduct statis-
tural equation modeling statistical package. Specifically,
tical analyses as well as examine and establish the meas-
the steps of examining both a first-order factorial model
urement properties of their proposed instruments.
and a second-order factorial model of a questionnaire
(Form B) were introduced and illustrated using AMOS
Acknowledgments: The preparation of this paper and the
Graphics 17.0 [7]. With these analyses, the factorial valid-
Project P.A.T.H.S. were financially supported by The Hong
ity of the second-order three-factor model of Form B (the
Kong Jockey Club Charities Trust. The authorship is car-
program implementer’s subjective evaluation form) was
ried equally between the first and second authors.
established in a large sample of secondary school teach-
ers in Hong Kong. Details of the output can be seen in the
website of the P.A.T.H.S. Project (www.paths.hk). Received January 3, 2013; accepted February 8, 2013

References
1. Reymont R, Jöreskog KG. Applied factor analysis in the natural 6. Shek DTL, Yu L. Use of structural equation modeling in human
sciences. Cambridge: Cambridge University Press, 1993. development research. Int J Disabil Human Dev 2014;13:157–67.
2. Spearman C. General intelligence, objectively determined and 7. Arbuckle JL. Amos (Version 17.0) [Computer Program]. Chicago,
measured. Am J Psychol 1904;15:201–93. IL: SPSS, 2008.
3. Thompson B, Dennings B. The unnumbered graphic scale 8. Rindskoph D, Rose T. Some theory and applications of confirmatory
as a data-collection method: An investigation comparing second-order factor analysis. Multivar Behav Res 1988;23:51–67.
three measurement strategies in the context of Q-technique 9. Jöreskog KG, Sörbom D. LISREL 8: Structural equation modeling
factor analysis. Paper presented at the annual meeting of the with the SIMPLIS command language. Hillsdale, NJ: Lawrence
Mid-South Educational Research Association, New Orleans, 1993. Erlbaum, 1993.
4. Kerlinger FN. Foundations of behavioral research, 3rd ed. New 10. Saris WE, Satorra A, Sörbom D. The detection and correction
York, NY: Holt Rinehart Winston, 1986. of specification errors in structural equation models. In: Clogg
5. Gorsuch RL. Factor analysis, 2nd ed. Hillsdale, NJ: Erlbaum CC, editor. Sociological methodology. San Francisco, CA:
Associates, 1983. Jossey-Bass 1997:105–29.

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM
204 Shek and Yu: Confirmatory factor analysis

11. Aluja A, Rossier J, Garcia LF, Angleitner A, Kuhlman M, 20. Bentler PM. Comparative fit indexes in structural models.
Zuckerman M. A cross-cultural shortened form of the ZKPQ Psychol Bull 1990;107:238–246.
(ZKPQ-50-cc) adapted to English, French, German, and Spanish 21. Byrne BM. Structural equation modeling with AMOS: Basic
languages. Pers Indiv Differ 2006;41:619–28. concepts, applications, and programming. Mahwah, NJ:
12. Lei P, Wu Q. Introduction to structural equation modeling: issues Lawrence Erlbaum, 2001.
and practical considerations. Educ Meas 2007;26:33–43. 22. Hu LT, Bentler PM. Evaluating model fit. In: Hoyle RH, editor.
13. Akaike H. Factor analysis and AIC. Psychometrika 1987;52:317–32. Structural equation modeling: Concepts, issues, and
14. Browne MW, Cudeck R. Alternative ways of assessing applications. Thousand Oaks, CA: Sage, 1995:76–99.
model fit. In: Bollen KA, Long JS, editors. Testing structural 23. Jöreskog KG, Aish AM. Structural modeling by example
equation models. Newbury Park, CA: Sage, 1993:445–55. – applications in educational, sociological, and behavioral
15. Arbuckle JL. Amos 17.0 user’s guide. Chicago: SPSS, 2009. research. Contemp Psychol 1990;35:687–8.
16. Jöreskog KG, Sörbom D. LISREL 8.7 for Windows. Lincolnwood, 24. Shek DTL, Ma CM. Program implementers’ evaluation of the
IL: Scientific Software International, 2004. Project P.A.T.H.S.: findings based on different datasets over
17. Guo B, Perron BE, Gillespie DF. A systematic review of structural time. ScientificWorldJournal 2012, ArticleID 918437, 10 pages.
equation modelling in social work research. Br J Soc Work Doi: 10.1100/2012/918437.
2009;39:1556–74. 25. Shek DTL, Yu L. A review of validated youth prevention and
18. Shek DTL, Siu AM, Lui JH, Lung, DW. P.A.T.H.S. to Adulthood: A positive youth development programmes in Asia. Int J Adolesc
Jockey Club Youth Enhancement Scheme (evaluation manual). Med Health 2011;23:317–24.
Social Welfare Practice and Research Centre, The Chinese 26. Sun RC, Shek DTL. Positive youth development, life satisfaction
University of Hong Kong, 2006. and problem behaviour among Chinese adolescents in Hong
19. Chou CP, Bentler PM. Estimates and tests in structural equation Kong: a replication. Soc Indic Res 2012;105:541–59.
modeling. In: Hoyle RH, editor. Structural equation modeling: 27. Shek DTL, Siu AM, Lee TY. Subjective outcome evaluation of
concepts, issues, and applications. Thousand Oaks, CA: Sage, the Project P.A.T.H.S.: Findings based on the perspective of the
1995:37–55. program implementers. ScientificWorldJournal 2007;7:195–203.

Authenticated | [email protected] author's copy

Download Date | 6/14/17 1:35 PM