Osborne 2015

Practical Assessment, Research, and Evaluation
Volume 20 Volume 20, 2015 Article 2
2015
What is Rotating in Exploratory Factor Analysis?

Jason W. Osborne
Follow this and additional works at: https://scholarworks.umass.edu/pare
Recommended Citation
Osborne, Jason W. (2015) "What is Rotating in Exploratory Factor Analysis?," Practical Assessment,
Research, and Evaluation: Vol. 20 , Article 2.
DOI: https://doi.org/10.7275/hb2g-m060
Available at: https://scholarworks.umass.edu/pare/vol20/iss1/2
This Article is brought to you for free and open access by ScholarWorks@UMass Amherst. It has been accepted for
inclusion in Practical Assessment, Research, and Evaluation by an authorized editor of ScholarWorks@UMass
Amherst. For more information, please contact scholarworks@library.umass.edu.
Osborne: What is Rotating in Exploratory Factor Analysis?
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 20, Number 2, January 2015 ISSN 1531-7714
What is Rotating in Exploratory Factor Analysis?

Jason W. Osborne, University of Louisville
Exploratory factor analysis (EFA) is one of the most commonly-reported quantitative methodology
in the social sciences, yet much of the detail regarding what happens during an EFA remains
unclear. The goal of this brief technical note is to explore what “rotation” is, what exactly is
rotating, and why we use rotation when performing EFAs. Some commentary about the relative
utility and desirability of different rotation methods concludes the narrative.
Those of us who regularly use exploratory factor Spearman’s seminal work in this area, few statistical
analysis (EFA), one of the most commonly-used techniques have been so widely used and, often, subject
statistical techniques reported in the social sciences to misperceptions (see, for example, Costello &
literature (e.g., Fabrigar, Wegener, MacCallum, & Osborne, 2005; Osborne, Costello, & Kellow, 2008).
Strahan, 1999; Osborne, Costello, & Kellow, 2008), Principal components analysis (PCA) is a
know that rotation happens, that there are different computationally simplified version of the general class
types of rotation, and hopefully that the goal of all of dimension reduction analyses. PCA computes the
rotation methods is to clarify results. But what exactly analysis without regard to the underlying latent
is rotating during an EFA? The goal of this article is to structure of the variables, using all the variance in the
answer that simple question. manifest variables. PCA was developed decades ago
Let us look “under the hood” at the mysterious when analyses were mostly computed by hand and thus
inner workings of EFA.1 shortcuts that did not substantially diminish the
outcome were valuable. Now with popular statistical
Basics of exploratory factor analysis
software packages and the readily available processing
Exploratory factor analysis is a statistical tool used power in even the cheapest laptop computers, PCA is
for many purposes. It was originally developed in the probably not necessary. It is also not considered a true
early 1900s to attempt to establish intelligence as a method of factor analysis and there is disagreement
unitary or multidimensional construct (Spearman, among statisticians about when it should be used, if at
1904), and is a general-purpose dimension reduction all. There are some situations where PCA might be an
tool with many applications. In the modern social appropriate option, but more often than not,
sciences, it is perhaps most frequently used to explore researchers use PCA when EFA would be appropriate
the psychometric properties of an instrument or scale. and preferable (for example, see Ford, MacCallum, &
Exploratory factor analysis examines all the pairwise Tait, 1986; Gorusch, 1983; Widaman, 1993). It is still
relationships between individual variables (e.g., items the default dimension reduction procedure in many
on a scale) and seeks to extract latent factors from the statistical analysis software packages despite it usually
measured variables. During the 110 years since not being the desirable choice. We will not discuss
PCA further, as the goal of the paper is to focus on
1 EFA and rotation.
In this article the discussion is limited to exploratory factor
analysis as there is no rotation analogue in confirmatory factor
analysis.
Published by ScholarWorks@UMass Amherst, 2015 1
Practical Assessment, Research, and Evaluation, Vol. 20 [2015], Art. 2
Practical Assessment, Research & Evaluation, Vol 20, No 2 Page 2
Osborne, Rotation in EFA
Exploratory factor analysis. The second method of 5. Assessing and passing judgment on a possible
dimension reduction is common factor analysis or or planned solution to the problem
exploratory factor analysis. Common factor analysis 6. Comparing and contrasting two solutions to
recognizes that model variance contains both shared the problem on a particular dimension such
and unique variance across variables. EFA examines as cost
only the shared variance from the model each time a 7. Selecting one idea or solution to the problem
factor is created, while allowing the unique variance from among those considered
and error variance to remain in the model. 8. Communicating elements of the solution in
sketches, diagrams, lists, and written or oral
The general process for conducting exploratory
reports
factor analysis is briefly outlined in Table 1. Many of
these steps are discussed in depth elsewhere (Osborne
Interest in engineering:3
et al., 2008; Osborne & Fitzpatrick, 2012; Thompson,
1. I find many topics in engineering to be
2004). Because this is intended to be a brief technical
interesting
note, let us focus on step #4, rotation.
2. Solving engineering problems is interesting to
me
Table 1. Steps To Follow When Conducting 3. Engineering fascinates me
EFA 4. I am interested in solving engineering
1. Data cleaning problems
2. Deciding on extraction method to use 5. Learning new topics in engineering is
3. Deciding how many factors to retain interesting to me
4. Deciding on a method of rotation (if 6. I find engineering intellectually stimulating
desired)
5. Interpretation of results (return to #3 if In this example I will use principal axis factoring
solution is not ideal) extraction, specifying two factors to be extracted, and
6. Replication will request a factor loading plot.4 In this example, the
analysis produces two strong factors (extracted
A pedagogical example eigenvalues of 7.42 and 3.28, which accounts for
76.42% of the variance). The extracted (un-rotated)
To illustrate the points in this paper, I will use data factor matrix is presented in Table 2, below. I also
from a study on engineering majors conducted several present the scree plot from this analysis showing that
years ago. For this example, we will examine two scales the two-factor solution is strongly supported.
that should be minimally correlated: Engineering
problem solving and interest in engineering. The items As you can see in Table 2, although we expect two
from the relevant subscales are listed below: very clear factors, the factor loadings are not
immediately identifiable as two separate factors prior to
Engineering problem solving items:2 rotation. To be sure, looking only at Factor 1 loadings,
all fourteen items seem to be similar. It is only in
How well did you feel prepared for: combination with the loadings on Factor 2 where the
1. Defining what the problem really is two factors separate. If one plots each item in two-
2. Searching for and collecting information dimensional space (Factor 1 on the X axis, and Factor 2
needed to solve the problem on the Y axis), we see clearly the separation, as
3. Thinking up potential solutions to the presented in Figure 1, below:
problem
4. Detailing how to implement the solution to
3
the problem Assessed on a seven point Likert type scale anchored by
“strongly disagree” and “strongly agree”
4
In many analyses of these data, no other eigenvalue rose
above 0.50 in value and thus for the purposes of this
2
Assessed on a seven point Likert type scale anchored by “did increasingly inaccurately labeled brief technical note is
not prepare me at all” to “prepared me a lot” discounted as a legitimate possibility.
https://scholarworks.umass.edu/pare/vol20/iss1/2
DOI: https://doi.org/10.7275/hb2g-m060 2
Table 2. Unrotated Factor Matrix and scree plot

Factor
1 2
EngProbSolv1 .759 -.389
INTERESTeng1 .630 .521
Extraction Method: Principal Axis Factoring.
less subjective or exploratory (e.g., Horst, 1941),

leading to initial algorithms such as Quartimax (Carroll,
1953) and Varimax (Kaiser, 1958).5
Quite simply, we use the term “rotation” because,
historically and conceptually, the axes are being rotated
so that the clusters of items fall as closely as possible to
them.6 Looking at Figure 1, for example, if you
imagine rotating the axes so that they intersect the
centroid of each cluster variables, you get the essence
of rotation.
As you can see in Figure 2 (and Table 3), following
rotation of the axes (or items), the items now fall
closely about each axis line. This has the effect of
making the factor loading pattern much clearer as one
Figure 1: Unrotated solution from initial extraction of the two pairs of coordinates for each item tends to
be close to 0.00, as you can see in Table 3. In this
What exactly is rotation, and what is being example analysis, the factors were correlated r = 0.37.
rotated?
Unrotated results from a factor analysis – as
presented above- is not easy to interpret, although the
plot helps. Simply put, rotation was developed not 5
Note that rotation does not alter the basic aspects of the
long after factor analysis to help researchers clarify and analysis, such as the amount of variance extracted from the
simplify the results of a factor analysis. Indeed, early items. Indeed, although eigenvalues might change as factor
methods were subjective and graphical in nature loadings are adjusted by rotation, the overall percentage of
(Thurstone, 1938) because the calculations were labor variance accounted for will remain constant.
6
intensive. Later scholars attempted to make rotation Alternatively, you could imagine rotating each cluster of items
toward the axis. It really works out to be functionally the same.

In Table 4, below, loadings less than 0.10 are

suppressed, allowing easy visual confirmation that the
EFA with oblique rotation produced our expected
result.
Table 4. Pattern Matrix with loadings < 0.10
suppressed
Factor
1 2
EngProbSolv1 .859
EngProbSolv2 .841
EngProbSolv3 .879
EngProbSolv4 .909
EngProbSolv5 .886
EngProbSolv6 .869
EngProbSolv7 .868
Figure 2: Rotated factor solution following Principal
EngProbSolv8 .790
Axis Factoring extraction and oblique (oblimin)
rotation INTERESTeng1 .801
INTERESTeng2 .921
INTERESTeng3 .922
Table 3. Pattern Matrix7 INTERESTeng4 .904
INTERESTeng5 .897
Factor INTERESTeng6 .864
1 2 Extraction Method: Principal Axis Factoring.
Rotation Method: Oblimin with Kaiser Normalization.
Different types of rotations
EngProbSolv4 .909 -.025 There are many choices of rotation method,
EngProbSolv5 .886 .021 depending on what software you are using. Each uses
EngProbSolv6 .869 .020 slightly different algorithms or methods to achieve the
EngProbSolv7 .868 .033 same broad goal- simplification of the factor structure.
EngProbSolv8 .790 .072 Rotation methods fall into two broad categories:
INTERESTeng1 .042 .801 orthogonal and oblique (referring to the angle
INTERESTeng2 -.023 .921 maintained between the X and Y axes). Orthogonal
INTERESTeng3 -.014 .922
rotations produce factors that are uncorrelated (i.e.,
maintain a 90o angle between axes); oblique methods
allow the factors to correlate (i.e., allow the X and Y
axes to assume a different angle than 90o).
INTERESTeng6 .009 .864 Traditionally, researchers have been guided to
Extraction Method: Principal Axis Factoring.
orthogonal rotation because (the argument went)
Rotation Method: Oblimin with Kaiser Normalization.
uncorrelated factors are more easily interpretable.
There is also an argument in favor of orthogonal
Most statistical packages will allow small loadings rotation as the mathematics are simpler, and that made
to be suppressed following rotation, so that the results a significant difference during much of the 20th century
become even more obvious and immediately apparent. when EFA was performed by hand calculations or
7
much more limited computing power. Orthogonal
When an oblique rotation is performed, where factors are rotations are generally the default setting in most
allowed to correlate, the factor loadings are contained in a
table called the “pattern matrix.” More on this below.
statistical computing packages.
There does not seem to be a compelling reason for The mathematical algorithms for each rotation are
modern researchers to default to orthogonal rotations. different, and beyond the scope of this brief technical
In the social sciences (and many other sciences, such as note. Note that for all rotations, the goal is the same:
biomedical sciences) we generally expect some simplicity and clarity of factor loadings. For details on
correlation among factors, since behavior is rarely how they achieve these goals, you should refer to the
partitioned into neatly packaged units that function manual for your statistical software (e.g., IBM SPSS
independently of one another. Therefore using base statistics manual p. 97,8 or Gorusch, 1983; for a
orthogonal rotation potentially results in a less useful good overview of the technical details of different
solution where factors are correlated. Remembering versions of varimax rotation, see Forina, Armanino,
that EFA is an exploratory technique (not a Lanteri, & Leardi, 1989).
confirmatory technique), we should be looking for the Do orthogonal and oblique rotations produce
clearest solution possible. Further, there does not noticeable differences?
appear to be a drawback to using oblique rotation even
if the factors are truly uncorrelated. Oblique rotations Orthogonal and oblique rotations will produce
do not force factors to be correlated, and so in that virtually identical solutions in the unlikely case where
instance, the factors would be allowed to assume a factors are perfectly uncorrelated. As the correlation
correlation of zero, and the solution would be the same between latent variables diverges from r = 0.00, then
as that of an orthogonal rotation. the oblique solution will produce increasingly clearer
results. Looking at the same data after orthogonal
Oblique rotation output is only slightly more (Varimax) rotation (Figure 3 and Table 5, below), one
complex than orthogonal rotation output, but should
yield either identical or superior results to that of
orthogonal rotations. In SPSS output the rotated factor
matrix is interpreted after orthogonal rotation; the
rotated factor matrix represents both the loadings and
the correlations between the variables and factors. In
contrast, when using oblique rotation the pattern matrix
is examined for factor/item loadings and the factor
correlation matrix reveals any correlation between the
factors. The pattern matrix holds the loadings (which
are of most interest), and each row of the pattern
matrix can be thought of as a regression equation
where the standardized observed variable is expressed
as a function of the factors, with loadings as the
regression coefficients. The structure matrix holds the
correlations between the variables and the factors,
which are generally of less interest in exploratory
applications (e.g., Gorusch, 1983). Figure 3: Same data using orthogonal rotation
(Varimax)
There are a variety of choices in each category.
Varimax rotation is by far the most orthogonal
rotation, likely because it is the default in many can see that this outcome still provides a similar
software packages, but also because it was developed as conclusion, but with a slightly less clear solution. This
an incremental improvement upon prior algorithms is because these factors are modestly correlated, but the
quartimax, and equamax. There is no widely preferred mandate to maintain a 90o angle between axes means
method of oblique rotation; all tend to produce similar
results (Fabrigar et al., 1999), and it seems generally fine
8
to use the default settings in software packages. Retrieved from
Common oblique rotations you will see include: direct ftp://public.dhe.ibm.com/software/analytics/spss/documentati
oblimin, quartimin, and promax. on/statistics/22.0/en/client/Manuals/IBM_SPSS_Statistics_Bas
e.pdf

that the centroids of the clusters cannot move closer to was to reduce groups of variables to conceptually
the axis lines. In this case, the difference is not great, important latent variables. Given this, the overriding
but noticeable in Table 5, where again loadings less goal of EFA must be to make sense of data. A clear,
than 0.10 were suppressed. You can see that only one easily- interpreted structure to the EFA is the goal,
loading in Table 5 is less than the suppression cutoff of therefore.
0.10, leaving a less clear result than the oblique rotation Many authors have written on guidelines for
from Table 4. This is a small but clear example of the extraction and rotation of factors, focusing on
higher efficacy of oblique rotations to create clear eigenvalues, scree plots, parallel analysis, replication,
patterns of results in EFA where the factors are indeed and so on. It is my belief that the over-arching value
correlated. has to be theoretical framework and an easily-
Table 5. Rotated Factor Matrix with loadings < interpretable factor structure. Absent this, which we
0.10 suppressed use to make sense of data, none of the technical details
seem important.
Factor
1 2 In this brief technical note I used two-dimensional
plots as an example. If an instrument is uni-
EngProbSolv1 .830 .153
dimensional no rotation is possible. If an instrument is
EngProbSolv2 .809
three- (or more) dimensional, then items are plotted in
EngProbSolv3 .861 .157 multidimensional space, and three (or more) axes are
EngProbSolv4 .888 .146 rotated within this multidimensional space with the
EngProbSolv5 .879 .185 same goals.
EngProbSolv7 .862 .189 I also suggest that in the modern era of high-
power computing, orthogonal rotations are probably
not a best practice, as oblique rotations can accurately
model uncorrelated and correlated factors, whereas
orthogonal rotations cannot handle correlated factors
as effectively. Thus, there is little cost to using oblique
INTERESTeng4 .164 .897 rotations regardless of the underlying relatedness of the
INTERESTeng5 .165 .868 factors.
Extraction Method: Maximum Likelihood. REFERENCES
Rotation Method: Varimax with Kaiser Normalization. Carroll, J. B. (1953). An analytical solution for
approximating simple structure in factor analysis.
Summary Psychometrika, 18(1), 23-38.
The enigmatic and widely-used technique of Costello, A. B., & Osborne, J. W. (2005). Exploratory Factor
Analysis: Four recommendations for getting the most
exploratory factor analysis is a complex class of from your analysis. Practical Assessment, Research, and
procedures with many options. Many of us, even those Evaluation, 10(7), 1-9.
with decades of experience using EFA, remain unclear
Fabrigar, L. R., Wegener, D. T., MacCallum, R. C., &
on some of the nuances and details of what exactly is Strahan, E. J. (1999). Evaluating the use of exploratory
happening “under the hood” when we perform this factor analysis in psychological research. Psychological
analysis. Rotation is literally a rotation of X- and Y- Methods, 4(3), 272.
axes in order to align clusters of variables plotted in Ford, J. K., MacCallum, R. C., & Tait, M. (1986). The
two-dimensional space with the axis lines, which has application of exploratory factor analysis in applied
the effect of clarifying the loading patterns in tables psychology: A critical review and analysis. Personnel
(making larger numbers larger, and smaller numbers Psychology, 39, 291-314.
closer to zero).9 Recall that the original goal of EFA Forina, M., Armanino, C., Lanteri, S., & Leardi, R. (1989).
Methods of varimax rotation in factor analysis with
9
Or, if you prefer, rotation of the clusters of items to more applications in clinical and food chemistry. Journal of
closely align with the axis lines.
Chemometrics, 3(S1), 115-125. doi: Osborne (Ed.), Best Practices in Quantitative Methods (pp.
10.1002/cem.1180030504 205-213). Thousand Oaks, CA: Sage Publishing.
Gorusch, R. L. (1983). Factor Analysis (2nd ed.). Hillsdale, Osborne, J. W., & Fitzpatrick, D. (2012). Replication
NJ: Lawrence Erlbaum Associates. Analysis in Exploratory Factor Analysis: what it is and
Horst, P. (1941). A non-graphical method for transforming why it makes your analysis better. Practical Assessment,
an arbitrary factor matrix into a simple structure factor Research, and Evaluation, 17(15), 1-8.
matrix. Psychometrika, 6(2), 79-99. doi: http://pareonline.net/getvn.asp?v=17&n=15
10.1007/BF02292176 Spearman, C. (1904). "General Intelligence," Objectively
Jones, B. D., Osborne, J. W., Matusovich, H. M., & Paretti, Determined and Measured. The American Journal of
M. C. (2014). Relationships among students’ perceptions of a Psychology, 15(2), 201-292. doi: 10.2307/1412107
first-year engineering design course and their engineering Thompson, B. (2004). Exploratory and confirmatory factor
identification, motivational beliefs, course effort, and academic analysis: Understanding concepts and applications: American
outcomes. Unpublished manuscript, Virginia Tech, Psychological Association.
Blacksburg, VA. Thurstone, L. (1938). A new rotational method in factor
Kaiser, H. (1958). The varimax criterion for analytic rotation analysis. Psychometrika, 3(4), 199-218.
in factor analysis. Psychometrika, 23(3), 187-200. doi: Widaman, K. F. (1993). Common factor analysis versus
10.1007/BF02289233 principal component analysis: Differential bias in
Osborne, J. W., Costello, A. B., & Kellow, J. T. (2008). Best representing model parameters? Multivariate Behavioral
Practices in Exploratory Factor Analysis. In J. W. Research, 28(3), 263-311.
Acknowledgements and Author Notes

An early version of this paper was presented as a talk at the 2013 American Statistical Association
Conference on Statistical Practice in New Orleans, LA. Many thanks to Elizabeth Hoag-Carhart (Old
Dominion University) who helped with that presentation.
Readers interested in receiving a free PDF of the author’s new book (Best Practices in Exploratory Factor
Analysis) can email the author with PAREonline in the subject to request a copy.
Citation:
Osborne, Jason W. (2015). What is Rotating in Exploratory Factor Analysis? Practical Assessment, Research &
Evaluation, 20(2). Available online: http://pareonline.net/getvn.asp?v=20&n=2
Author:
Jason W. Osborne
Educational and Counseling Psychology, Counseling, and College Student Personnel
College of Education and Human Development
University of Louisville
Louisville, KY, USA
jason.osborne [at] louisville.edu

Osborne 2015

Uploaded by

Copyright:

Available Formats

Osborne 2015

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Osborne 2015

Uploaded by

Copyright:

Available Formats

Practical Assessment, Research, and Evaluation

Volume 20 Volume 20, 2015 Article 2

What is Rotating in Exploratory Factor Analysis?

Follow this and additional works at: https://scholarworks.umass.edu/pare

A peer-reviewed electronic journal.

What is Rotating in Exploratory Factor Analysis?

Table 2. Unrotated Factor Matrix and scree plot

less subjective or exploratory (e.g., Horst, 1941),

Published by ScholarWorks@UMass Amherst, 2015 3

In Table 4, below, loadings less than 0.10 are

Published by ScholarWorks@UMass Amherst, 2015 5

Acknowledgements and Author Notes

Published by ScholarWorks@UMass Amherst, 2015 7

You might also like