128 Teachers’ Learning, Curriculum Innovations and Knowledge Applications

10
Exploratory Factor Analysis versus
Confirmatory Factor Analysis
Siti Aishah Hassan and Kaseh Abu Bakar

INTRODUCTION
Nowadays, most empirical researchers read, hear, and use the various available
multivariate data analysis techniques, such as Multiple Regression Analysis (MRA),
Multiple Discriminant Analysis (MDA), Multiple Analysis of Variance
(MANOVA), Exploratory Factor Analysis/ Principle Component Analysis (PCA),
Confirmatory Factor Analysis (CFA), and Structural Equation Modelling (SEM).
Some of these data analyses are frequently used by education researchers. However,
other data analyses still remain novel to the researchers. Particularly, exploratory
factor analysis (EFA) is widely taught in statistics at the level of postgraduate
studies but confirmatory factor analysis (CFA) still remains exclusive to advanced
researchers.
Nevertheless, with the advancement in computer technology, postgraduate
students and researchers are expected not to rely on EFA only. More and more
user friendly software are available in the market to handle the needs of the more
advanced data analysis techniques. CFA, is predominantly beyond the capacity
provided by Statistical Package for Social Science (SPSS). Therefore, a number of
software programs such as AMOS, SAS, EQS, LISREL, SEPATH, CALIS along
with others are available to serve the needs of CFA. Yet, before researchers embark
on the more advanced technique of data analysis, they need to conceptually
understand the relationships between the two techniques. They also need to compare
the similarities and the differences of these two data analyses. Accordingly, they
will understand the link between the two techniques and able to use the techniques
appropriately.
Therefore, this chapter seeks to answer four germane questions: What is factor
analysis? Why conduct factor analysis? What are the similarities and differences
between exploratory and confirmatory factor analysis? What kind of research
questions may employ exploratory or confirmatory factor analysis? Finally, this
chapter provides examples of research report writing that employs both exploratory
and confirmatory factor analyses.
 Exploratory Factor Analysis versus Confirmatory Factor Analysis 129

DEFINITION OF TERMS
In order to better comprehend this chapter, definitions of relevant terms are provided
below.
‘Factor Loadings’ are multiple correlations between factors and variables (items).
High loading indicates the factor significantly explained the variables. A
cutting point of .40 for loading is recommended.
‘Factor matrix’ is the table that displays the factor loadings of all variables on each
factor of orthogonal rotation.
‘Factor pattern matrix’ is the table that displays the factor loadings of all variables
on each factor of oblique rotation.
‘Orthogonal rotation’ is an extraction method of rotation for uncorrelated or very
low correlation among the factors.
‘Oblique rotations’ is an extraction method of rotation for highly correlated factors.
‘Squared multiple correlations’ (SMC) is the proportion variance explained of the
item by the factor. High SMC indicates a high reliability of the item.
‘Eigenvalue’ is the value that indicates the amount of variance accounted for by
each factor. For eigenvalues greater than 1, the total variance explained by
the emerging factors is at the optimum point. It is a value that normally
represents a sharp “elbow” or break point in the scree plot.
‘Scree plot’ is a graphical representation of the relationships between the number
of extracted factors and eigenvalues.
‘Bartlett test of sphericity’ is a test for overall significant correlation matrix. A
significant result indicates that the variables are highly correlated and hence
suitable for factor analysis study. In other words, some degree of
multicollinearity (intercorrelation) is desired.
‘Manifest variable’ is a directly measured variable. It is the item of the instruments.
‘Latent construct’ is the unobserved variable. It is the extraction of the hypothesized
factor. It is also known as dimension, factor, component, or underlying
structure.
‘Exogenous variable’ is a variable that may roughly be equated with an independent
or predictor variable, an arrow pointing outward.
‘Endogenous variable’ is a variable which may roughly be equated with a dependent
or criterion variable, an arrow pointing towards it.

WHAT IS FACTOR ANALYSIS?

Factor analysis is a generic name given to a group of multivariate statistical
techniques whose primary purpose is to define the underlying dimension in a data
matrix. It is a technique used to analyze a large number of highly correlated
variables. This large number of variables is hypothesized to be explained by or
belong to certain common factors. With factor analysis, the researcher can first
130 Teachers’ Learning, Curriculum Innovations and Knowledge Applications

identify separate dimensions of the structure, which later can be used for two
purposes - summarization and data reduction. Summarization is to describe the
data using a much smaller number of concepts than the original individual data.
Data reduction is to have a composite measure for each dimension and to substitute
them in place of individual data. This is particularly important in order to further
analyze data by using other statistical techniques such as multiple regressions or
path analysis (Abdul Hair et al., 1998).

Why Conduct Factor Analysis?

The main reason to conduct factor analysis is to condense a large number of
observable variables into a smaller set of variables without forfeiting the important
information each variable carries. In other words, it is to extract the common factors,
structures, components, constructs, underlying dimensions, and latent variables.
The factors summarize a large number of variables into a few logical concepts.
We label the factors according to logical concepts. We can reduce the data by
using the composite measure of the variables under the same dimensions. This
type of factor analysis is known as exploratory factor analysis. It is conducted
without theoretical constraints imposed upon the solution. Specifically, exploratory
factor analysis is conducted to explore how many latent factors underlie a set of
variable scores when there are no previous studies conducted on the particular
newly adapted or developed instrument.
Principle component analysis that is embedded in SPSS is commonly used
for exploratory factor analysis. Although the mathematical formula for principle
component analysis and exploratory factor analysis are not the same, both methods
often yield similar results (Stevens, 1996, p. 362). Hence, principle component
analysis is interchangeably used for exploratory factor analysis.
However, when there is an established theory or a previous empirical study
conducted to identify the common factor for the large number of variables under
investigation, then the main reason to conduct factor analysis is to test the
measurement model. The theory explains how many factors exist and what the
factors are. There is a theoretical constraint that imposes items to load only on a
specific factor. Exclusively, confirmatory factor analysis is conducted to test the
construct validity of the instrument. Construct validity means how well the construct
explains the variables under the construct. An instrument is said to have construct
validity when the items are highly correlated with one another within the same
construct and the loading or squared multiple correlation of the item is significantly
correlated to the assigned construct.
In short, as the name implies, exploratory factor analysis is conducted to
explore the underlying dimensions of the measured variables. Confirmatory is used
to “confirm” the hypothesized measurement model, i.e. to establish the construct
validity of the instrument.
 Exploratory Factor Analysis versus Confirmatory Factor Analysis 131

What are the Similarities and Differences between Exploratory and

Confirmatory Factor Analysis?
Since both data analysis techniques are in the same family of data analysis, the
relationships between the two are summarized in Table 1.

Table 1 Similarities and Differences Between EFA and CFA

EFA CFA
Similarity To condense a large number of variables into a smaller
number of factors
PURPOSE
Difference To explore the underlying To test the construct
construct validity (to “confirm”
the construct)
Similarity Application on new population
APPLICABILITY Newly developed Has conducted EFA
instrument Replication of the
Difference Newly adapted instrument construct validity test
onto a different sample
THEORY ON Similarity Factors conceptually define variables that are highly
FACTOR correlated to each other
“Grounded” theory - Theory grounded- test
generate theory theory
Difference No predetermined number The number of factors
of factors is fixed
Items free to load on any Items are forced to load
factor only on specific factors.
INTER FACTOR Similarity A concept or theory explains whether the factors are
CORRELATION correlated or uncorrelated
Difference Orthogonal rotation for Single order construct
uncorrelated factors for uncorrelated /low
inter factor correlation
Parallel Oblique rotation for Second order construct
correlated factors for highly inter-
correlated factor
SOFTWARE Similarity Both use SPSS
Difference SPSS only AMOS embedded in SPSS
132 Teachers’ Learning, Curriculum Innovations and Knowledge Applications

What kind of research questions that may employ exploratory or confirmatory

factor analysis?
Since there are differences between the two techniques of factor analyses,
the research questions that they employ can be differentiated accordingly.

Research Questions for EFA/PCA

1. What are the underlying factors of ...?
2. How many factors explained the concept of...?
3. What do the total proportions variance explain...?

Research Questions for CFA

1. Is the hypothesized measurement model of...supported by the observed
data?
2. Does the two-factor or three-factor model of...better explain the measured
variables?
3. Is there any significant higher order construct that significantly explains the
factors of...?

CONCLUSION
This chapter provides examples from our own research that have been published
in journals and the unpublished dissertation. It is worthy to note that due to the
limited number of pages that are allowed in journals, the details have been reduced.
For dissertation writing however, lengthy report writing is preferred. Hence, we
provide both styles of writing. The first example is a Principle Component Analysis
(PCA) with orthogonal rotation. The second example is a PCA with an oblique
rotation. The third example is a Confirmatory Factor Analysis (CFA) with a single
order construct. The fourth example is a CFA with a second order construct.
Extensive report writing is provided for each type of analysis.
In summary, exploratory factor analysis is conducted based on three focal
principles: to determine how many factors are present to label the factors, and to
allow the items the freedom to load on any factors. On the other hand, confirmatory
factor analysis is conducted to test the construct validity; the researcher’s a priori
sets the number of factors and forces the items to load only on a specified factor.
Both analyses are conducted to condense data and increase reliability and validity
of the instruments. Hence, prudent use of both techniques of factor analyses is
ensured when their relationships are understood.
 Exploratory Factor Analysis versus Confirmatory Factor Analysis 133

Example 1: Principle Component Analysis with Orthogonal Rotation An Extensive

Report Writing for Exploratory Factor Analysis

Preliminary Analysis
In order to justify the use of principle component analysis, one of the important
assumptions to be assessed is the intercorrelations among the items. Barlett
Sphericity Test was statistically significant, χ2 (36) = 219.28, p = .001 the
variables were higly correlated to one another. Besides that, Kaiser-Meyer-
Olkin Measure of Sampling Adequacy was .650. Thus, there was evidence
for the overall measurement of sampling adequacy fulfilling the requirement
of at least .50. The individual measure of sampling adequacy is listed in
Table 1. The individual measure of sampling adequacy revealed that item 10
was low, MSA = .492. Thus, this item was considered as a candidate to be
deleted. Next, communality was assessed. It revealed that there were three
items needed to be deleted – items 12, 13, and 14. As a result, four items
were deleted from the original list of maternal involvement. Hence, parental
involvement inventory left with ten items.

Table
Table 1 Preliminary
1: Prelim analysis
inary analysis and and descriptive
descriptive statistics
statistics ofitem
of all all items
s
Item 1 2 3 4 5 6 7 8 9 10 11 12 13 14
1 .597
2 .360 .667
3 .464 .373 .507
4 .191 .116 .131 .641
5 .096 .131 .064 .553 .550
6 -.054 .198 .035 .208 .293 .682
7 -.109 .102 -.060 .172 .207 .271 .637
8 -.030 .226 .026 .158 .115 .163 .629 .544
9 .092 .290 .186 .132 .146 .059 .264 .327 .738
10 -.050 .158 .214 -.064 -.176 .147 .082 .025 .231 .492
11 -.081 .211 .008 .023 .084 .196 .303 .228 .314 .531 .630
12 .010 .017 -.029 .321 .237 .107 .281 .218 .141 -.061 .197 .684
13 .005 .129 .045 -.029 -.105 .050 .208 .089 .206 .197 .224 .250 .493
14 .060 .198 -.126 .134 .065 .256 .111 -.082 .219 .184 .196 .030 .014 .387
Commu .643 .557 .655 .650 .644 .416 .736 .812 .469 .689 .628 .363 .393 .232
M ean 3.03 3.65 3.33 .639 1.05 .98 .62 .69 1.97 2.99 2.63 2.20 2.53 2.38
SD .81 .83 .85 .92 .62 .74 .80 .93 1.12 1.2 1.13 1.3 0 1.01 1.06
Note: The diagonal ent ry is M SA

The Factor Solutions

Beginning with 14 items, a varimax rotation was conducted, resulting in four-
factor solutions with eigenvalues greater than 1. The total variance explained
was 56.32%. However, due to the constraint that we adopted for the standard
requirement of MSA and communality greater than 0.5, thus this four factor-
solutions were not acceptable.
Hence, we decided to conduct a second varimax rotation with the 10 items
left. This time, the result showed three factors were extracted with 58.17% of
134 Teachers’ Learning, Curriculum Innovations and Knowledge Applications

total variance explained. Factor 1 consisted of items 8, 9, 10, and 12. Whereas
factor 2 consisted of items 1, 2, and 3 and finally factor 3 consisted of items
4, 5, and 6. Analyzing the items of each factor, we noticed that factor 1 could
not be logically labeled, which consisted of item 11. This item is related to
Islamic dimension and does not logically related to the existing factor.
Thus, we finally conducted the third varimax rotation for the 9 items left
after discarding item 11. The results of the analysis showed there were three
latent variables emerged. This indicated that the three underlying dimensions
accounted for 61.82% of the total variance explained of nine items and found
to be higher as compared to the first (56.32%) and the second (58.17%) times
varimax rotations were conducted with fourteen and ten items respectively.
Thus, the solution of three latent structures of maternal involvement seemed
to be the best. The dimensions can be logically labelled as reading
encouragement for factor 1, homework monitoring for factor 2, and school
relationship for factor 3. The variance of the first dimension, the largest
eigenvalue was 2.47, while the other subsequence eigenvalues were 1.76 and
1.34, respectively. Inspection on the scree plot also pointed out the 9-items
measured three factors, with the sharp break point after the three factors. All
estimated factor loadings were large enough to be of practical significance at
p = .05; even the weakest loading extracted was .549. In addition, all of the
directions of loadings were positive and free from factorial complexity. The
results of the final factor solution are illustrated in Table 2.

Table 2 Rotated Component Matrix

In summary, the study suggests that there are three underlying dimensions of
maternal involvement construct among mothers of a selected Integrated
Primary Islamic Schools. The dimensions are maternal reading
encouragement, maternal homework monitoring, and maternal school
relationship.
 Exploratory Factor Analysis versus Confirmatory Factor Analysis 135

Example 2: Principle Component Analysis with Oblique Rotation

A brief report writing style for exploratory factor analysis

Example 3: Confirmatory Factor Analysis - Single Order Measurement Model

A brief report writing style for confirmatory factor analysis

Figure 3 shows the results of the three-factor measurement model of Maternal

Accountability. All fit indices exceeded the recommended threshold values
GFI, AGFI, IFI, TLI, CFI >.90, RMSEA<.08, indicating that the model fit
the data. The inter-factor correlations were r = -.03, .19 and .35, substantiated
the hypothesis that the three factors were distinct. The negative correlations
were due to the negatively worded items for factor named Reliable. The
loadings range was .49 to .82. Succinctly, construct validity for maternal
accountability is supported.

Figure 3 Measurement Model for Maternal Accountability

136 Teachers’ Learning, Curriculum Innovations and Knowledge Applications

Confirmatory Factor Analysis -Second Order Measurement Model

An extensive report writing style for confirmatory factor analysis

Model Testing Results for Maternal Piety

Maternal Piety (MP) was hypothesized to be a second order latent construct

for three underlying dimensions, namely Faith in Allah and the Hereafter
(Faith), Call for Virtue (Virtue), and Forbidding F-Vices (F-Vices). The results
as illustrated in Figure 4, γ 2 (31) = 38.87, p = .156, suggested that there was
no significant difference between the hypothesized model and the observed
model. GFI, AGFI, IFI, TLI, CFI >.90 and RMSEA <.08 indicated that the
model fits the data. Moreover, there were no offending estimates, suggesting
that the hypothesized model of Maternal Piety was admissible. The result
above was achieved after taking into consideration the Modification Index
(MI). We allowed the residuals for items MSC9 and MSC7 to correlate as
suggested by MI.

Figure 4.1 Measurement Model for Maternal Piety

 Exploratory Factor Analysis versus Confirmatory Factor Analysis 137

Precisely, Maternal Piety was statistically significant, as indicated by the

first latent construct named Virtue (γ12 = .69, R2 = .48, p = 0.001). Faith, the
second latent construct, was a priori set to be the reference indicator due to
the conceptually sound indicator of Piety. Hence, the p-value was not estimated
and presumed that Faith was significant indicator of Piety. This was
substantiated when the estimated parameter (γ12 = .90, R2 = .80) indicated
80% of total variance for Faith. Finally, the third latent construct F-Vices
was statistically significant as well (γ31 = .74, R2 = .55, p =. 001). In addition,
all of the items (except item MSC 11) were statistically significant at α =.01,
with loadings greater than 0.40. The details are outlined in Table 4.
Table 4:
Table 4 Results
Results of
of the
theCFA
CFAofofMaternal
MaternalPiety
Piety

In summary, the hypothesized model of Maternal Piety was a second order

construct for tri-dimensional measurement model; comprising of the Call for
Virtue (Virtue), Forbidding Vices (F-Vices), and Faith in hereafter (Faith).
It was supported by all the estimated fit statistics. Faith was found to be the
best indicator for Maternal Piety, followed by F-Vices and finally Virtue.
Illustratively, Faith best indicated by item MSC4, mom reminds me about
hereafter. F-Vices were best indicated by item MSC7; mom forbids me from
watching indecent TV programs. Virtue was equally best indicated by items
MSC2 and MSC8, congregational prayer and reading Al-Qur’an respectively.

REFERENCE
Arbuckle, J.L. (2007). AmosTM 16.0 User’s Guide. Bethelhem Pike, PA: Amos
Development Corporation.
Arbuckle, J.L. and Wothke, W. (1999). AMOS 4.0 User’s Guide. Chicago, IL:
Small Waters Corporation.
Armsden, G.C. and Greenberg, M.T. (1987). The inventory of parent and peer
attachment: Individual differences and their relationship to psychological well-
being in adolescence. Journal of Youth and Adolescence, 16, 427-425.
138 Teachers’ Learning, Curriculum Innovations and Knowledge Applications

DeVellis, R. (1991). Scale development: Theory and Applications. Newbury Park,

California: Sage Publication.
Hair, J.F., Black, W.C., Babin, B.R., Anderson, R.E. and Tantham, R.L. (2006).
Multivariate Data Analysis (6th Edn.). New Jersey: prentice Hall.
Kline, P. (1994). An Easy Guide to Factor Analysis. London: Routledge.
Lantos, E. and Rezmovic, V. (1981). A confirmatory factor analysis approach to
construct validation. Educational and Psychological Measurement, 41, 61-
72.
Sidek Mohd.Noah, Jamiah Manap, Azimi Hamzah, Hasnan Kasan, Turiman Suandi,
Kairul Anwar Mastor, Rumaya Juhari, Azma Mahmood, Zanariah Mohd Nor
and Steven Eric Krauss. (2006). Pembinaan inventori personaliti muslim
(InPm) untuk kegunaan belia Malaysia. Jurnal Pekama, 12, 53-69.
Siti Aishah Hassan, Abdullah Seif Abdullah, Noriah Ishak and Hassan Langgulung.
(2008). Measuring the unmeasurable: Maternal piety scales. Pertanika Journal
of Social Sciences & Humanities, 6(1), 95-105.
Siti Aishah Hassan. (2006). Maternal quality time, children’s emotional intelligence
and their academic performance: A structural equation modeling analysis.
Unpublished Doctoral Dissertation, International Islamic University of
Malaysia.
Stevens, J. (2002). Applied Multivariate Statistics for The Social Sciences. New
Jersey: Lawrence Erlbaum Associates, Publishers.
Stevens, J. (1996). Applied Multivariate Statistics for The Social Sciences. New
Jersey: Lawrence Erlbaum Associates, Publishers.
Thompson, B. (2004). Exploratory and Confirmatory Factor Analysis:
Understanding Concepts and Applications. Washington, DC: American
Psychological Association.
Thorndike, R. (1997). Measurement and Evaluation in Psychology and Education
(6th Edn.). New Jersey: Prentice Hall.