A peer-reviewed electronic journal.

Copyright is retained by the first or sole author, who grants right of first publication to the Practical Assessment, Research & Evaluation.
Permission is granted to distribute this article for nonprofit, educational purposes if it is copied in its entirety and the journal is credited.
PARE has the right to authorize third party reproduction of this article in print, electronic and database forms.
Volume 18, Number 6, March 2013 ISSN 1531-7714

Practical Considerations for Using Exploratory Factor Analysis in

Educational Research
Amy S. Beavers, John W. Lounsbury, Jennifer K. Richards,
Schuyler W. Huck, Gary J. Skolits, and Shelley L. Esquivel
The University of Tennessee

The uses and methodology of factor analysis are widely debated and discussed, especially the issues
of rotational use, methods of confirmatory factor analysis, and adequate sample size. The variety
of perspectives and often conflicting opinions can lead to confusion among researchers about best
practices for using factor analysis. The focus of the present review is to clarify terminology,
identify key issues, and clarify areas of debate regarding best practices and functions of factor
analytic procedures. The conclusions and implications drawn should be useful to researchers in
education, psychology, and cognate social fields who employ factor analytic procedures or evaluate
research using factor analytic methods.

Factor analytic procedures are statistical methods used applications (Furr & Bacharach, 2008). Regardless of
for examining the relationships within a group of the setting, within each analysis there are a range of
observed variables, as measured through questions or choices and decisions the researcher must make to
items. It is important to note that factor analysis is not improve the accuracy of the factor analysis they use and
a singular statistical method, but rather a group of to enhance the quality of the resulting solution
statistical analyses that share similar methodology and (Fabrigar, Wegener, MacCallum & Strahan, 1999). A
functionality. The theoretical and mathematical commonly cited limitation of exploratory factor
variations among the processes allow the analyses to analysis (EFA) is its level of subjectivity stemming from
accommodate breadth of purpose and theory in the many methodological decisions a researcher must
research and result in the widespread use of the tool make to complete a single analysis, with the accuracy of
across disciplines and applications; however, it is the the results largely dependent upon the quality of these
flexibility of the statistical methods that fuel ongoing decisions (Henson & Roberts, 2006; Tabachnick &
debate about the appropriate applications of these Fidell, 2001).
methods. For over sixty years, researchers from varied To further compound the complexity of the
social science disciplines have saturated the factor analysis, factor analysis is a cyclical process of
analysis literature with definitions, discussions, and continually refining and comparing solutions until the
debates concerning factor analysis approaches, most meaningful solution is reached (Tabachnick &
applications, and recommendations for most Fidell, 2001). This commonsense approach to
appropriate usage (cf.,Garson, 2010; Loo, 1979; Pett, interpreting the analysis in light of theory and
Lackey & Sullivan, 2003; Tabachnick & Fidell, 2001; conceptual foundation should be accompanied by a
Velicer & Fava, 1990). strong theoretical and mathematical justification for the
Examination of factors, or dimensions, is most methodological choices and decisions, yet because of
often applied in the development and validation of the abundance of sources and opinions within the
measures such as personality scales (Schonrock-Adema, factor analysis literature, it is often difficult for a
Heijne-Penninga, Van Hell & Cohen-Schotanus, 2009); researcher to determine the most accurate use of this
however, it can be used in a variety of measurement tool within a given research context. The purpose of
Practical Assessment, Research & Evaluation, Vol 18, No 6 Page 2
Beavers, Lounsbury, Richards, Huck, Skolits & Esquivel; EFA

the present overview is to provide a step-by-step enough” to produce a reliable factor analytic solution? The
approach to factor analytic procedures and to offer an literature contains copious amounts of information in
evaluation of the theoretical and practical merits response to this question; however, criteria provided
associated with four common areas of debate. for determining the sufficiency of a sample for factor
Recommendations are offered for best factor analytic analytic procedures vary greatly and include a plethora
practice for educational researchers as framed by the of differing criteria. Sample size requirements may
following three questions: generally be categorized in two ways; a minimum
number of cases or a subjects-to-variables ratio (STV)
1. How large should the sample be to know if the
required to achieve an adequate sample. For example,
it is “large enough” to produce a reliable factor
selected criterion suggests the sample size should have:
analytic solution?
2. What is the difference between Component  51 more cases than the number of variables
Analysis and Common Factor Analysis? (Lawley & Maxwell, 1971).
3. Is it necessary to rotate the initial factor pattern  At least 10 cases for each item, and the
matrix in order to achieve an interpretable and subjects-to-variables [STV] ratio should be no
meaningful solution? lower than 5 (Bryant & Yarnold, 1995).
In the journey to understand factor analysis so that  At least 100 cases and a STV ratio of no less
responsible, methodologically-sound decisions could be than 5 (Suhr, 2006).
made, approximately 45 manuscripts were reviewed.  At least 150 - 300 cases (Hutcheson &
Initially, books and book chapters were consulted to Sofroniou, 1999).
provide a basic overview of the technique and its  At least 200 cases, regardless of STV (Gorsuch,
various components. From these sources, a number of 1983).
questions and controversies remained. Academic  At least 300 cases (Norušis, 2005).
Search Premier was used to identify peer-reviewed
journal articles that included overviews, examinations, Similar guidelines are provided throughout the
and/or applications of factor analytic methods and literature without clear consensus (Tabachnick & Fidell,
procedures. This process allowed the authors to 2001; Zhao, 2009). There is, however, general
identify seminal articles as well as trace evidence-based agreement that an inadequate sample size can be
decision making in factor analysis over time. While this detrimental to the factor analytic process and produce
does not constitute a meta-analysis, it does provide a unreliable, and therefore, non-valid results (Osborne &
broad foundation to ground the recommendations Costello, 2004; Pett et al., 2003). How, then, can
made by the authors for best practices use in EFA. anyone use factor analysis methods with confidence,
Review of these sources provided additional depth of assuming that a sample of insufficient size will
understanding and allowed the authors to draw undermine any meaning produced, without a realistic
conclusions as to general rules-of-thumb for many guide to what sample criteria is “best” or the “right”
decision-making controversies often encountered in one? Initial review of the literature suggests that a base
data analysis of research problems. These guidelines number of cases is required and that a ratio of cases to
formed the basis for the present article. Where variables should be considered once the base number is
apparent contradictions still existed in the literature, the met.
authors based decisions regarding recommendations on Upon closer inspection of the literature, a general
the methodological soundness of the studies consulted opinion has emerged, suggesting that ratio criteria do
and the mathematical foundations of the techniques not provide an accurate guide (Guadagnoli & Velicer,
employed. 1988; Hogarty, Hines, Kromrey, Ferron & Mumford,
SAMPLE SIZE 2005; MacCallum, Widaman, Preacher & Hong, 2001;
Osborne & Costello, 2004; Zhao, 2009). Guadagnoli
The first question comes during the planning and Velicer (1988) suggest, what has been largely
stages: How large should the sample be to know if it is “large
Practical Assessment, Research & Evaluation, Vol 18, No 6 Page 3
Beavers, Lounsbury, Richards, Huck, Skolits & Esquivel; EFA

confirmed in the literature, that the needed sample size methods of extraction in factor analysis (Tabachnick &
is conditional upon the strength of the factors and the Fidell, 2001).
items. They provide a new criterion operationalizing
EVALUATE FACTORABILITY OF MATRICES
these relationships. If the factors have four or more
items with loadings of .60 or higher, then the size of Correlational Values
the sample is not relevant. If the factors have 10 to 12 In addition to meeting assumptions before the
items that load moderately (.40 or higher), then a factorization of a set of variables, the strength of the
sample size of 150 or more is needed to be confident in relationships and linear relationships are evaluated by
the results. Finally, if factors are defined with few reviewing the correlation matrix produced from the
variables and have moderate to low loadings, a sample data. Generally, correlations exceeding .30 provide
size of at least 300 is needed. Fabrigar et al. (1999) and enough evidence to indicate that there is enough
MacCallum et al. (2001), further support that stable commonality to justify comprising factors (Tabachnick
solutions can be reached with samples as low as 100 & Fidell, 2001). If intercorrelations are unexpectedly
when three to four strong items (loadings of .70 or low, it may be a result of low variance. Samples that are
greater) comprise a factor, suggesting that weaker too homogenous can exhibit low variance;
relationships need a larger sample size. consequently, the correlation will be low potentially
A strong solution, made up of stable factors, failing to reveal a factor, or common relationship, that
reduces the influence of the sample size; however, a does exist (Fabrigar et al., 1999).
larger sample size decreases sampling error resulting in
more stable solutions (Hogarty et al., 2005).
Determination of sample size sufficiency is dependent Table 1 includes the correlation matrix for a
upon the stability of the solution; therefore, the sample of 5 items in a Teacher Satisfaction and Sense
adequacy of a sample cannot be fully determined until of Belonging Scale (TSSBS).
the analysis has been conducted. While the final factor Table 1. Example of a Correlation Matrix of Five
solution can provide enough evidence to suggest that a Items in the TSSBS
sample is sufficient, one is still left with the question of
how much is enough when collecting the original Item 1 2 3 4 5
sample? Because the family of factor analysis 1 1
procedures are multivariate tools, and multivariate 2 .642 1
methods require larger sample sizes than do univariate 3 .542 .801 1
methods, one should plan for a sample of at least 150 4 .355 .539 .682 1
cases for initial structure exploration. 5 .244 .355 .475 .510 1
CHECK ASSUMPTIONS OF DATA Note. Pearson’s r Correlational values are
reported.
After the sample data has been obtained, the data
used must satisfy the assumptions required of With one exception, the intercorrelations exceed
multivariate statistical techniques, including: large .30. The correlation between Item 1 and Item 5 is .244.
sample size, linearity, absence of outliers, continuous While this value is below .30, the relationships with
data, lack of extreme multicollinearity, and low other items exceed .30. Therefore, this one correlation
percentage of missing data (Comrey, 1985; Pett et al., is not enough evidence to suggest that factoring would
2003). Factor analysis differs from other multivariate not be beneficial.
procedures in that there is no separate identification of Determinant of the Matrix
dependent or independent variables. The relationships
between variables are examined without specification An additional assessment of factorability of the
of one variables’ influence upon another. As a result, data comes from the determinant of the correlation
multivariate normality is not required within all matrix. The determinant of a matrix is a single value
Practical Assessment, Research & Evaluation, Vol 18, No 6 Page 4
Beavers, Lounsbury, Richards, Huck, Skolits & Esquivel; EFA

calculated using the values within a square matrix, assessing the strength of the relationships and
revealing the presence or absence of possible linear suggesting factorability of the variables.
combinations within the matrix. The determinant of a 2
Bartlett’s Test of Sphericity
x 2 matrix is ad – cb. The determinant of the 2
The determinant value in the TSSB example is
4 1 very close to zero. To evaluate if this determinant value
x 2 matrix is equal to 4(5) - 3(1), which equals
3 5 is statistically different from zero, Bartlett’s Test of
17. When the determinant does not equal zero, the Sphericity is used. The null hypothesis of Bartlett’s test
matrix can be explained by linear combinations; states that the observed correlation matrix is equal to
however, if the determinant of a matrix equals zero it is the identity matrix, suggesting that the observed matrix
described as a singular matrix. A singular matrix has is not factorable (Pett et al., 2003). In the example used,
either an infinite number of linear combinations or Bartlett’s test produced a significant test result,
there are no possible linear combinations within the rejecting the null hypothesis. Bartlett’s Test provides
2 5 evidence that the observed correlation matrix is
matrix. For example, in the matrix , the
6 15 statistically different from a singular matrix, confirming
determinant equals 2(15) – 6(5) = 0. Row one, [2, 5],
that linear combinations exist.
does not contribute any unique value to the solution
and can be expressed entirely as a linear combination of Kaiser-Meyer-Olkin Test of Sampling
other rows: 3(2x + 5y) = 6x + 15y. Adequacy
With the exception of cases where the determinant The Kaiser-Meyer-Olkin Test of Sampling
is zero, the values can be arranged into linear Adequacy (KMO) is a measure of the shared variance
combinations. In factor analysis, these linear in the items. Kaiser, Meyer, and Olkin suggest the
combinations are considered factors. A non-zero following guideline for assessing the measure (Friel,
determinant indicates that a factor or component is n.d.):
mathematically possible; however, it does not offer any
Table 3. Interpretation Guidelines for the Kaiser-Meyer-Olin
indication of the practical meaning or significance of
Test
the factors. Because the values of a correlation matrix
are restricted to values between – 1 and 1, the values KMO Value Degree of Common Variance
for the determinant of a correlation matrix range from 0.90 to 1.00 Marvelous
0 to 1. The values seen most often are very small, 0.80 to 0.89 Meritorious
suggesting that a few linear combinations exist (Pett et
0.70 to 0.79 Middling
al., 2003). For example, Table 2 reports the measures of
factorability for the TSSBS correlation matrix used 0.60 to 0.69 Mediocre
previously (see Table 1). 0.50 to 0.59 Miserable
Table 2. Example of Measures for Assessing the 0.00 to 0.49 Don’t Factor
Correlation Matrix
Measure Value
Determinant 1.14 E -11 INITIAL EXTRACTION
Bartlett’s Test of Sphericity p < .0001 Factoring the matrix begins with the initial
Kaiser-Meyer-Olkin Test of extraction of linear combinations. Matrix algebra is
.889
Sampling Adequacy used to create linear combinations of items that explain
the greatest amount of variance1 among the items. The
In addition to using the determinant of a matrix,
initial extraction assumes that each combination is
Bartlett’s Test of Sphericity and the Kaiser-Meyer-
orthogonal (independent or uncorrelated) to the others.
Olkin Test of Sampling Adequacy (KMO) are
commonly used to provide more complex measures for
1 The type of variance included varies by extraction method.
Practical Assessment, Research & Evaluation, Vol 18, No 6 Page 5
Beavers, Lounsbury, Richards, Huck, Skolits & Esquivel; EFA

These linear combinations are called factors or scores are a result of an underlying factor or construct
components. The first factor extracted accounts for the (DeCoster, 1998). A measure to assess the student’s
greatest percentage of variance in the items. The next attitudes about math would be an illustration of this
linear combination attempts to account for the relationship. The student’s responses to the items are
maximum amount of remaining variance that is not thought to reflect their underlying attitudes about math.
included in the first factor. This process continues until Component Analysis Common Factor Analysis
all of the variance in the sample is explained (Suhr,
2006). The two most common terms used to discuss
the initial extraction are component analysis and common
factor analysis.
What is the difference between Component
Analysis and Common Factor Analysis? It is
Figure 1. Directional Relationships in Component
commonly accepted practice to use factor analysis as a
Analysis and Common Factor Analysis. This figure
broad heading for two distinct statistical techniques:
illustrates the differences of the causal relationships in
component analysis and common factor analysis
component analysis and common factor analysis
(Tabachnick & Fidell, 2001). These terms, as well as
other related terms such as component and factor, are Mathematically, component analysis and common
often used interchangeably within the literature causing factor analysis differ in the amount of variance included
confusion for the reader (Garson, 2010; Furr & in the solution. There exist multiple types of variance:
Bacharach, 2008; Tabachnick & Fidell, 2001). Both common (shared) variance, specific (unique) variance, and error
theoretical and mathematical differences exist between variance (measurement error). Common variance is the
component analysis and common factor analysis. The variability present in an item that is shared with other
failure to make these distinctions clear leads to items and factors. Specific variance is the variance
difficulty interpreting the context and diminishes the resulting from the unique attributes of an item that
researcher’s ability to make theoretically sound cannot be explained by other variables or factors. Error
decisions. variance is associated with the measurement process
and is an indication of unreliability.
Component analysis serves as a means to
accurately report and evaluate a large number of Component analysis includes all three types of
variables using fewer components, while still preserving variance and does not partial out any variance from the
the dimensions of the data. It is widely described as a items when examining the relationships. Because the
data reduction method used to summarize a large set of total variance is included in components analysis, some
variables (Costello & Osborne, 2005; Velicer & argue that the estimates provided reflect inflated values
Jackson, 1990b). Theoretically, component analysis (Costello & Osborne, 2005). By contrast, common
assumes that the component is a composite of the factor analysis removes specific variance and error
observed variables, or that the individual item scores variance from the calculations, including only common
cause or define the component (DeCoster, 1998). See variance to extract the factor solution.
Figure 1. A student’s score on a math test is an The literature contains a multitude of studies
illustration of this causal relationship in component whose purpose is to determine if the mathematical
analysis. The student’s performance on each item difference between component analysis and common
comprises the overall test score. factor analysis result in a practically different solution
By contrast, common factor analysis allows the (Osborne & Costello, 2004; Costello & Osborne, 2005;
exploration of underlying constructs, which cannot be Fabrigar et al., 1999). In fact, a special issue of
measured directly, through items thought to be Multivariate Behavioral Research (1990) focuses entirely on
reflective measures of the construct (Byrne, 2001). this topic and suggests that the alignment of the
Common factor analysis assumes that individual item solutions are dependent upon the strength and stability
Practical Assessment, Research & Evaluation, Vol 18, No 6 Page 6
Beavers, Lounsbury, Richards, Huck, Skolits & Esquivel; EFA

of the items. Removing sources of unreliability suggests however, the benefit of using ML is that in addition to
that common factor analysis would produce more the correlational estimates, it produces significance
accurate solutions; however, it is commonly reported tests for each item as well as fit statistics for the
that the results of both processes are similar (Fava & structure.
Velicer, 1992; Tabachnick & Fidell, 2001; Velicer & Example. To extract the initial factor solution of
Jackson, 1990a). This proves to be true when the the example TSSB scale, Principal Axis Factoring
measures used are reliable. The mathematical proofs (PAF) is used because the items are believed to reflect
and research on both sides of this well-established the underlying satisfaction and sense of belonging
debate are beyond the scope of this review. experienced by teachers. In Table 4, the PAF solution
In a practical research context, the researcher is compared to the Principal Component Analysis
should be aware that component analysis and common solution as a means to evaluate the difference between
factor analysis function similarly and can produce common factor analysis and component analysis in this
comparable results. The mathematical and theoretical data. The amount of variance in an item that can be
foundations of the two methods vary; however, the explained by the factor is displayed in a factor pattern
practical sequencing of steps and processes are the matrix. The columns of the factor loading matrix
same (Pett et al., 2003). Component analysis includes represent the factor (component) and the rows display
the total variance in the items and has no underlying each item or variable.
structural assumptions. Common factor analysis only Table 4. Example of Initial Extraction for Five Items in the TSSB
includes the common variance and hypothesizes that Principal Component
the item responses are a product of an underlying Principal Axis Factoring
Analysis
construct. Both component analysis and common Factors Components
factor analysis are mathematically able to reduce ITEM 1 2 3 4 5 1 2 3 4 5
variables to a smaller number of components or 1 .498 .402 .510 .338 -.351 .404
factors; however, the precise interpretability and
2 .650 .346 .659 .383 -.374
understanding of these values vary by the method used
to extract these linear combinations. 3 .726 -.400 .729 -.407
4 .758 -.360 .760 -.404
Methods of Initial Extraction
5 .585 .365 .595 -.357 .437
Component analysis includes the total variance in Note. Loadings of less than |.32| were suppressed.
the initial extraction. Principal Component Analysis (PCA)
is the most widely used extraction method of Even though the results produced are similar, it is
component analysis and is most appropriate when the important to consider the best method that most
purpose is to reduce the number of items to a smaller accurately depicts the purpose and needs of the
number of representative components (Costello & research hypotheses. Because more variance is
Osborne, 2005; DeCoster, 1998), whereas common included, the PCA solution on the right has more items
factor analysis only includes the common (shared) with cross loadings, meaning an item’s variance can be
variance in the extraction. The two most commonly explained by multiple factors. In the PAF solution,
used extraction methods of common factor analysis are Item 1 only loads on factors one and five; whereas, in
Principal Axis Factoring (PAF) and Maximum Likelihood the PCA solution, Item 1 loads on components one,
Estimation (ML)2. PAF is appealing because it requires three and four. Pett et al. (2003) suggest that it is best
no distributional assumptions and may be used if data to compare the PCA solution to the PAF solution, and
are not normally distributed (Fabrigar et al., 1999). ML then use the one that makes the most intuitive sense.
requires multivariate normality (Pett et al., 2003); This further reinforces the responsibility of the
researcher to be thoughtful when making processual
2Additional component and factor extraction methods include choices, having both theoretical and conceptual
image, alpha, unweighted least squares and generalized least justification for those decisions. Regardless of choice,
squares.
Practical Assessment, Research & Evaluation, Vol 18, No 6 Page 7
Beavers, Lounsbury, Richards, Huck, Skolits & Esquivel; EFA

attention must be given to the method in which the or equal to one (Costello & Osborne, 2005). In a
variance is accounted for and considered at each step in component analysis extraction, such as PCA, where the
decision making and interpretation. total variance is accounted for, every item has one unit
of variance. If a single component could explain 100%
Determine the Number of Factors to Retain
of the variance for all of the items, the eigenvalue for
Using the results from the initial extraction, the that component would be equal to the total number of
researcher must then determine how many factors items. The reasoning behind the Kaiser Criterion is that
should be retained in order to best represent the data a component having an eigenvalue greater than one
and the existing relationships. The first factor accounts accounts for more variance than would a single item,
for the most variance. The amount of variance thus suggesting merit for combining those items into a
explained by each subsequent factor continually factor or component; however, this is only true if each
decreases (Tabachnick & Fidell, 2001). The objective is item contributes one unit of variance. Pett et al. (2003)
to choose enough factors to adequately represent the indicate that the Kaiser Criterion should only be used
data, while eliminating factors that are not statistically in PCA when the total variance is accounted for in the
or theoretically relevant (Fabrigar et al., 1999). The extraction.
body of literature suggests that choosing to retain more
Eigenvalues can be useful if interpreted with an
factors than are needed is less detrimental to the
understanding of their conceptual meaning regardless
analysis than eliminating factors that are needed;
of how much variance was extracted; however, the “cut
however, retaining too many factors can deplete the
value” should also reflect this consideration. In
solution erroneously resulting in weak factor loadings
common factor analysis extractions, where only
(Pett et al., 2003). Additionally, the literature cautions
common (shared) variance is used, the variance
against using a solution with only one or two factors, as
included for each item is less than one. In this case, if a
it may not provide an accurate representation of the
single factor could explain all of the variance in the
structure (Fava & Velicer, 1992; Pett et al., 2003).
items, the eigenvalue would still not equal the total
In addition to general recommendations, there are number of items. If the Kaiser criterion was used, a
multiple criteria methods to further inform the factor factor could account for significant variance but not be
selection using eigenvalues and extracted variance. The retained because the eigenvalue was less than one,
eigenvalue is a value associated with each factor resulting in the underextraction of factors.
describing the amount of variance in the items that can
Example. Using the all items of the TSSBS as an
be explained by that factor (Pett et al., 2003). Every
example, the factors were extracted using Principal
factor or component has an eigenvalue. This principle
Axis Factoring, a method that does not include all of
can be observed in mathematics when simplifying
the variance in the extraction. Table 5 represents a
multinomial expressions. For example, the binomial 6x
sample of the variance explained.
+ 15y can be factored or simplified to 3(2x + 5y). A
value of “3” is the maximum common amount that can Notice that the total initial eigenvalue estimates are
be extracted, or explained. Although the process for different than the total extraction sums. The initial
determining the eigenvalue based on common variance eigenvalue estimates include all of the variance. The
is significantly more complex, the conceptual principle PAF solution removes the shared and error variances
is the same. in extraction, reducing the eigenvalues and percent of
variance explained. In a PCA extraction where the total
Kaiser Criterion
variance is included, no change occurs between the
How much variance does a factor have to explain estimated values and the extracted values. For this
in order to warrant the retention of a factor example, using the Kaiser Criterion, five factors should
(component)? The most commonly used eigenvalue be retained in order to sufficiently represent the TSSB
criteria is the Kaiser Criterion, which states that factors scale; however, considering that not all the variance is
should be retained if their eigenvalues are greater than
Practical Assessment, Research & Evaluation, Vol 18, No 6 Page 8
Beavers, Lounsbury, Richards, Huck, Skolits & Esquivel; EFA

included, factors six and seven could also be viable limited to those occurring before the bend in the elbow
linear combinations of the items. (Fabrigar et al., 1999).
Table 5. Example of Total Variance Explained for a Principal
Axis Factoring of the TSSBS
Extraction Sums of
Initial Eigenvalues Squared Loadings
% of Cumulative % of Cumulative
Factor Total Variance % Total Variance %
1 13.107 46.810 46.810 12.769 45.603 45.603
2 2.057 7.347 54.157 1.737 6.202 51.805
3 1.808 6.459 60.616 1.491 5.326 57.131
4 1.335 4.767 65.383 1.020 3.644 60.774
5 1.230 4.394 69.777 .903 3.226 64.000
6 .963 3.440 73.217
7 .859 3.069 76.286
Note. The factors and initial eigenvalues continue until 100% of
the variance is accounted for. The full results were not needed for
the purposes of this discussion.
Figure 2. SPSS Scree Plot. This figure demonstrates the
The Kaiser Criterion method is often criticized scree plot of the eigenvalues and factors from the TSSB
and stated to be used beyond its capabilities resulting in extraction seen in Table 5.
inaccurate determination of factor retention (Costello
& Osborne, 2005; Velicer & Jackson, 1990a). Many This subjectivity is apparent when examining the
believe that the Kaiser Criterion tends to overextract scree plot of the eigenvalues of the TSSB extraction
factors and cite that it is also capable of underextracting (See Figure 2). Where does the “bend” occur? Should
factors as well (Fabrigar et al., 1999; Henson & the cutoff be at the third factor, the fourth, or even the
Roberts, 2006; Schonrock-Adema et al., 2009). Fabrigar sixth? The difficulty identifying the precise cut point
et al. (1999) criticizes that the criterion of identifying most often leads to overextraction of factors (Henson
one as the cut is quite arbitrary. Based on an & Roberts, 2006).
understanding that the eigenvalue represents the Variance Extracted
maximum variance that a single linear combination
(factor or component) can statistically explain, using A third selection method based on similar
the eigenvalues as an indication of value for retaining conceptual structure is to retain the number of factors
the factor is conceptually sound; however, the Kaiser that account for a certain percent of variance extracted.
Criterion should only be used in component analysis. The literature varies on how much variance should be
explained before the number of factors is sufficient.
Scree Plot The majority suggest that 75 – 90% of the variance
Cattell’s Scree Plot is a graphical representation of should be accounted for (Garson, 2010; Pett et al.,
the factors and their corresponding eigenvalues. The x- 2003); however, some indicate as little as 50% of the
axis represents the factors (components) and the variance explained is acceptable. As with any criteria
eigenvalues are along the y-axis. Because the first method solely depending on variance, this seemingly
component accounts for the greatest amount of broad standard must be viewed in relation to the
variance, it has the highest eigenvalue. The eigenvalues foundational differences between extraction methods.
continually decrease resulting in a picture that is often The amount of variance that was included for
called the “elbow” shape. The scree plot cutoff is quite extraction must be considered when interpreting the
subjective, requiring that the number of factors be value of percent of variance extracted. Component
analysis includes more variance to be explained,
Practical Assessment, Research & Evaluation, Vol 18, No 6 Page 9
Beavers, Lounsbury, Richards, Huck, Skolits & Esquivel; EFA

suggesting that higher percentages of explained  Tabachnick and Fidell (2001) state that “none
variance are expected than would be required when of the extraction techniques routinely provide
only common variance is included. an interpretable solution without rotation” (p.
Practical Implications 601).
 Fabrigar, Wegener, MacCallum, & Strahan
It is best to evaluate the initial extraction with (1999) provide that it is “usually necessary for a
multiple criterion methods and by comparing the researcher to select a method for rotating the
factors suggested to retain (Costello & Osborne, 2005; initial factor analytic solution to a final solution
Schonrock-Adema et al., 2009). Using the eigenvalues, that can be more readily interpreted” (p. 273).
the scree plot, and the percent of variance extracted,
 Child (1990) explains, “Most factor analysts
the TSSB example could require the retention of
agree that direct solutions are not sufficient.
between two and eight factors (see Table 5). It is
Adjustment to the frames of reference by
acceptable practice to vary the number of factors
rotation methods improves the interpretation
retained and compare the solutions. Ultimately, the
of factor loadings by reducing some of the
decision of how many factors to retain should be made
ambiguities which accompany the preliminary
based on comprehensibility and interpretability in the
analysis” (as cited in Suhr, 2006, p. 3)
context of the research (Suhr, 2006). Are the factors
represented by multiple variables that share a These comments are representative examples of
conceptual meaning? This evaluation is more relevant the information found within the literature. The use of
after the matrix has been rotated; however, it should be words that allow for exceptions, such as “routinely” in
considered when retaining factors as well. Tabachnick and Fidell (2001) and “usually” in Fabrigar
et al. (1999) leave question as to whether is it ever
FACTOR ROTATION appropriate to use only the initial factor pattern matrix.
Factor rotation is readily accepted as a sequential In his 1947 work, Thurstone (as cited in Fabrigar et al.,
step when conducting a factor analysis. As discussed 1999) suggested that the most easily interpretable
previously, the factors, or components, and the factor solution is the “simple structure” solution. He also
loadings of each variable are linear combinations of indicated that, because there are an infinite number of
these relationships. The mathematical purpose of factor solutions, the component matrix should be rotated in
analysis is to summarize the relationships among order to produce a solution with this simple structure.
variables and the factors. These linear combinations do
By understanding how the variance is removed,
not have a single, unique solution (Fabrigar et al.,
the initial solution can be interpreted and examined as a
1999). There exist an infinite number of rotations
series of linear combinations. However, the initial
(alternative solutions) that all explain the same amount
solution may not be the most useful. Rotations create a
of variance (DeCoster, 1998; Tabachnick & Fidell,
statistically comparable solution that is usually more
2001). After the number of factors to include has been
meaningful and easy to interpret.
determined, all other factors are discarded. The items
are factored again, forced into a specified number of Simple Structure
factors. That solution is then rotated. This is called Simple structure is achieved when each factor is
factor rotation. represented by several items that each load strongly on
Is it necessary to rotate the initial factor that factor only (Pett et al., 2003; Tabachnick & Fidell,
pattern matrix in order to achieve an interpretable 2001). Practically, “several items” is generally
and meaningful solution? The literature frequently considered to be at least three to five items with strong
suggests that rotating the initial factor solution is loadings (Guadagnoli & Velicer, 1988). An item is
critical for interpretation of the factors and indicator considered to be a good identifier of the factor if the
variables. This stance is presented quite consistently, loading is .70 or higher and does not significantly cross
and is widely followed without much explanation: load on another factor greater than .40 (Garson, 2010).
Practical Assessment, Research & Evaluation, Vol 18, No 6 Page 10
Beavers, Lounsbury, Richards, Huck, Skolits & Esquivel; EFA

These guidelines vary slightly within the literature. might be better described and interpreted in terms of
Tabachnick & Fidell (2001) suggest that the secondary oblique solutions” (p. 763).
loading (or cross-loading) should be no greater than Oblique rotations account for the relationships
.32. Costello and Osborne (2005) suggest that a loading between the factors, which often is more appropriate
of .50 is enough to be considered “strong,” while within social science research. Fabrigar et al. (1999)
Guadagnoli and Velicer (1988) state that the loading emphasize that oblique rotations can be used even
should be .60 or greater. Generally, a communality when the factors are not significantly correlated. If the
(loading) of .70 or greater is ideal because that suggests factors are not correlated, then the rotation will provide
that approximately 50% of the variance of that item is estimates of the factor correlations that are close to
accounted for by the factor. zero. Oblique rotational methods include direct
Rotations are very similar to many other more oblimin, promax, orthoblique and procrustes. There is
familiar mathematical concepts used more frequently in not a single best method recommended for oblique
basic math and algebra. An infinite number of rotations and the method choice often depends on the
numerical combinations could represent the ratio 2:3, options available through the software used (DeCoster,
such as 4:6, 10:15, or even 160:240, without altering the 1998; Fabrigar et al., 1999).
meaning of the relationship. In algebra and Example. To demonstrate the differences in an
trigonometry, this multiplicative property is often used orthogonal and oblique rotation, the TSSB solution was
to make the operation easier to solve. For example, the rotated using an orthogonal rotation and an oblique
equation -1/3x – 2/5y = -7/15 may be solvable, but rotation (See Table 6). The values in the orthogonal
multiplying the equation by -15 will simplify the factor matrix on the left represent maximized
equation without altering the values: -15(-1/3x – 2/5y relationships of each item with the factor. Each
= -7/15) = 5x + 6y = 7. Although it is much more relationship is assessed independently. Conceptually
intricate, the rotation of the initial factor solution is interpreted, in the orthogonal solution, factor one can
grounded in the same mathematical properties. Herein, explain .391 of the variance associated with the
the certain rotational techniques can serve to create a responses in item 2, and factor two is able to explain
more interpretable solution without altering the .497 of the variance, suggesting that factor two is more
structural relationships. Although the initial solution is representative of the item than factor one.
capable of being interpreted, the factors and
communalities are more easily identified through the The oblique rotation accounts for relationship
use of rotational methods to reach a simple structure between factors before determining an item’s
solution. relationship to the factor. This solution is presented
using two matrices: the pattern matrix and the structure
Orthogonal and Oblique Rotations matrix. The pattern matrix values reflect the
There are two main types of rotational methods: relationships between the item and the factor when the
orthogonal and oblique. Orthogonal rotations (varimax, variance of the other factors are removed. In the
quartimax, and equimax) are appropriate when the oblique pattern matrix in Table 6, once the relationship
purpose for the factor analysis is to generate factor between the factors is removed, factor two can
scores (PCA) or when the theoretical hypotheses additionally account for .429 of the variance associated
concern uncorrelated dimensions (Loo, 1979). Of the with the responses for item 2. The factor structure does
orthogonal types of rotations, varimax is generally not factor out influences of other factors and provides
regarded as best and is most widely used (Fabrigar et the correlations between item and factor without
al., 1999; Loo, 1979). Loo (1979) cautions that controlling for shared variance. The structure matrix
orthogonal rotations are not always theoretically demonstrates that all of the items are related to both
appropriate and do “not reflect interrelationships that factors.
probably exist in much clinical data, such relationships
Practical Assessment, Research & Evaluation, Vol 18, No 6 Page 11
Beavers, Lounsbury, Richards, Huck, Skolits & Esquivel; EFA

Table 6. Example of Rotational Comparison of 10 Items in the exists to suggest the factors are not correlated,
TSSBS allowances should be made enabling the true
Rota- Orthogonal: relationships among factors to be reflected in the
Oblique: Direct Oblimin
tion Varimax solution. Fabrigar et al. (1999) suggest that if the
Rotated Factor Structure
Matrix
Matrix
Pattern Matrix
Matrix
relationship among factors is unknown an oblique
Factor Factor Factor rotation should be used first. Then, if the correlations
ITEM 1 2 1 2 1 2 are low, opt to use an orthogonal rotation. Regardless
of methodological choice, theoretical and mathematical
1 .571 .657 .367 .587
limitations and meaning must be acknowledged. Just as
2 .391 .497 .429 .556 .607 decisions for extracting initial factors and determining
3 .761 .818 .583 .820 the number of factors to retain, rotational type should
4 .350 .688 .699 .593 .768 be made based on theoretical purpose of the research.
5 .731 .776 .578 .796 If the factors are conceptually independent, then
6 .395 .660 .637 .624 .759
orthogonal rotation is acceptable; however, oblique
rotations are generally more appropriate for social
7 .417 .404 .336 .543 .530
science research where the factors are usually related.
8 .443 .374 .386 .555 .512
9 .844 .960 .888 .580
INTERPRETING THE FACTOR SOLUTION
10 .740 .838 .782 .515 Because factor analytic processes are iterative,
Note. Maximum Likelihood Estimation used for Initial Extraction
much of the evaluation has occurred throughout each
subsequent step. The rotated factor solution is useful to
The orthogonal rotation in Table 6 has more examine and further refine the factors. Mathematical
cross-loadings than the oblique solution. The oblique and conceptual examination is required for accurate
rotation controls for the shared variance between the interpretation of both the items and the factors. The
factors. By first acknowledging that most all of the items should possess a significant loading, indicating a
items and factors are related, as displayed in the statistically valued contribution; however, an item’s
structure matrix, oblique rotations allows more conceptual significance should be examined before an
apparent differences to emerge in the pattern matrix. item is removed from the set. Theoretical knowledge is
Because oblique solutions can incorporate the more relevant than a statistical measure. If an item is
relationships between the factors, they usually fit the not significantly correlated to any of the factors
data better than the orthogonal solution (Henson & (generally considered to be less than .30) and does not
Roberts, 2006). In this case where the factors are provide a conceptually vital dimension to the measure,
strongly correlated (r = .709), the oblique solution is the item should be removed. Additionally, a complex
much more representative of the theoretical variable, or a variable that loads on more than one
relationships. factor, should be removed if the cross-loading is greater
than .40 (Schonrock-Adema et al., 2009). Once the
Determining whether an orthogonal or oblique weak items have been removed, the data should be
rotation is most appropriate seems to be controversial factored again without the presence of that item for a
simply for the fact that much of the literature supports more refined solution (Pett et al., 2003). Interpretation
use of oblique rotations, yet orthogonal rotations are of the factor also requires that each factor be
still the most commonly used and reported in studies sufficiently identified. This means that a factor contain
using factor analysis (Costello & Osborne, 2005). at least three to five items with significant loadings in
Darlington (n.d.) suggests that this may be a result of order to be considered a stable and solid factor
more complex interpretation of the oblique solutions. (Costello & Osborne, 2005). More importantly, the
Tabachnick and Fidell (2001) recommend using items and the factors should make sense conceptually.
orthogonal rotations if the correlations between factors
are low; however, unless strong theoretical foundation
Practical Assessment, Research & Evaluation, Vol 18, No 6 Page 12
Beavers, Lounsbury, Richards, Huck, Skolits & Esquivel; EFA

FINAL THOUGHTS DeCoster, J. (1998). Overview of factor analysis. Retrieved

March 5, 2013, from http://www.stat-
Every step of the process in a factor analysis help.com/notes.html
requires the researcher to be firmly grounded in Fabrigar, L.R., Wegener, D.T., MacCallum, R.C. & Strahan,
contextual theory and fundamental understanding of E.J. (1999). Evaluating the use of exploratory factor
factor analysis methodology. The greatest difference in analysis in psychological research. Psychological Methods,
methods centers on how the technique accounts for 4, 272-299.
variance and relationships between the factors. Fava, J.L. & Velicer, W.F. (1992). The effects of
Regardless of choice, decisions should be supported by overextraction on factor and component analysis.
strong theoretical and mathematical justification, Multivariate Behavioral Research, 27, 387-415.
providing credibility to the final outcome. Garson, D. (2010, February 11). Factor analysis. Retrieved
While factor analysis contains many variations, the March 5, 2013, from http://faculty.chass.ncsu.
process may be summarized as a series of mathematical edu/garson/PA765/factor.htm
iterations designed to create linear combinations in Gorsuch, R. (1990). Common factor analysis versus
order to explain the data. Each iteration reveals new component analysis: some well and little known facts.
information, further expanding the reseacher’s Multivariate Behavioral Research, 25, 33-39.
understanding of the relationships. Based on the new Guadagnoli, E. & Velicer, W.F. (1988). Relation of sample
perspective, the structure is refined until the solution size to the stability of component patterns. Psychological
reached is parsimonious, mathematically sound, and Bulletin, 103, 265-275.
theoretically grounded. The factor structure should Henson, R.K. & Roberts, J.K. (2006). Use of exploratory
continue to be tested and refined to more fully factor analysis in published research: common errors
understand the relationships in different contexts; and some comment on improved practice. Educational
and Psychological Measurement, 66, 393-416.
however, just because more testing of the structure is
desired, does not mean that the current solution is not Hogarty, K.Y., Hines, C.V., Kromrey, J.D., Ferron, J.M &
Mumford, K.R. (2005). The quality factor solutions in
useful. When evaluated in light of the strength of
exploratory factor analysis: the influence of sample size,
methodological purpose and use, the solution can be communality, and overdetermination. Educational and
very meaningful and contribute significantly to relevant Psychological Measurement, 65, 202-226.
research. It is important to remember that factor
Loo, R. (1979). The orthogonal rotation of factors in clinical
analysis is a mathematical process. While the matrix research: a critical note. Journal of Clinical Psychology, 35,
relationships are not elementary, they are still simply an 762-765.
indication of how the responses provided for each item
MacCallum, R.C., Widaman, K.F., Preacher, K.J & Hong, S.
relate to others. Factor analysis should always be (2001). Sample size in factor analysis: the role of model
interpreted in light of theory and common sense. error. Multivariate Behavioral Research, 36, 611-637.
REFERENCES Osborne, J.W. & Costello, A.B. (2004). Sample size and
subject to item ratio in principal components analysis.
Byrne, B. M. (2001). Structural equation modeling with AMOS - Practical Assessment, Research & Evaluation, 9 , 1-15.
Basic concepts, applications, and programming. LEA, ISBN 0- Retrieved March 5, 2013 from
8058-4104-0 http://pareonline.net/getvn.asp?v=9&n=11
Costello, A.B. & Osborne, J.W. (2005). Best practices in Pett, M., Lackey, N. & Sullivan, J. (2003). Making sense of
exploratory factor analysis: four recommendations for factor analysis. Thousand Oaks: Sage Publications, Inc.
getting the most from your analysis. Practical Assessment.
Research & Evaluation, 10, 1-9. Retrieved March 5, 2013 Schonrock-Adema, J., Heijne-Penninga, M., Van Hell, E.A.
from http://pareonline.net/getvn.asp?v=10&n=7 & Cohen-Schotanus, J. (2009). Necessary steps in
factor analysis: enhancing validation studies of
Darlington, R. (n.d.). Factor Analysis. Retrieved March 5, educational instruments. Medical Teacher, 31, e226-e232.
2013, from
http://psych.cornell.edu/Darlington/factor.htm Suhr, D. (2006). Exploratory or Confirmatory Factor
Analysis. SAS Users Group International Conference (pp. 1 -
17). Cary: SAS Institute, Inc.
Practical Assessment, Research & Evaluation, Vol 18, No 6 Page 13
Beavers, Lounsbury, Richards, Huck, Skolits & Esquivel; EFA

Tabachnick, B. & Fidell, L. (2001). Using multivariate statistics. Velicer, W.F. & Jackson, D.N. (1990b). Component analysis
Needham Heights: Allyn & Bacon. versus common factor analysis: some issues in selecting
Velicer, W.F. & Fava, J.L. (1990). An evaluation of the an appropriate procedure. Multivariate Behavioral Research,
effects of variable sampling on component, image, and 25, 1-28.
factor analysis. Multivariate Behavioral Research, 25, 193- Zhao, N. (2009, March 23). The minimun sample size in factor
209. analysis. Retrieved March 5, 2013, from
Velicer, W.F. & Jackson, D.N. (1990a). Component analysis http://www.encorewiki.org/display/~nzhao/The+Mi
versus common factor analysis: some further nimum+Sample+Size+in+Factor+Analysis
observations. Multivariate Behavioral Research, 25, 97-114.

Citation:
Beavers, Amy S., Lounsbury, John W., Richards, Jennifer K., Huck, Schuyler W., Skolits, Gary J. & Esquivel,
Shelley L. (2013). Practical Considerations for Using Exploratory Factor Analysis in Educational Research.
Practical Assessment, Research & Evaluation, 18(6). Available online: http://pareonline.net/getvn.asp?v=18&n=6

Corresponding Author:

Amy Beavers
The University of Tennessee
2509 River Drive
Knoxville, TN 37996-4539
Email: abeavers [at] utk.edu