This journey started over a decade ago when one of the first author’s students (Blandy
Costello) walked into his office and shared her frustration over conflicting advice
and directives from different doctoral committee members. That initial discussion
highlighted the lack of consensus on best practices related to exploratory factor
analysis, and ended with us deciding that there were empirical ways to explore best
practices. After many simulations and further discussions, two articles were published
related to these issues: Osborne & Costello (2004) and Costello & Osborne (2005).
Both have received significant attention in the literature, and the latter has been
cited about 3600 times as we write this. Obviously, we recognized that there is some
utility in attempting to explicate best practices in quantitative methods, and that
has led to other articles, books, and finally, to this project.
EFA is such a confounding, controversial, and misused (yet valuable and interesting)
technique that it has provided lots of fun and fodder for this type of endeavor. We
hope you agree it has been worthwhile. Our goal is to collect and elaborate on evidence-based
best practices that were published previously, to put them in a single place that
is easily accessible, and to model how to implement them within SAS. For those of
you who have persevered and have reached this part of the book, we hope that you have
drawn the following conclusions:
-
Keep the “E” in EFA! Many researchers have attempted to perform “confirmatory” analyses through exploratory
analyses. Many researchers use confirmatory language and draw confirmatory conclusions
after performing exploratory analyses. This is not appropriate. EFA is a fun and important
technique, but we need to use confirmatory techniques (e.g., CFA) when we desire to
draw those types of conclusions.
-
EFA is a large-sample technique. We hope that through the course of the book you have become convinced that the best
results from EFA come when the sample is appropriately large. There are examples in
this book and elsewhere in the literature of the volatile and nonrepresentative results
that can happen in small-sample EFA. A reasonable rule of thumb, if one is intent
on robust analyses, would be a minimum of 20 cases for each variable in the analysis. We have had students and colleagues
show us analyses that had fewer cases than variables. That is rarely a good state
of affairs, in our opinion.
-
Clean your data, and deal with missing data appropriately. Garbage in, garbage out. We won’t belabor this point — but we hope you take it seriously.
If we don’t see you address whether you checked your data, tested assumptions, and
dealt appropriately with missing data, we might wonder whether anything else you report
matters.
-
Useful results are those that are precise and generalizable. In our mind, the most useful results are those that we can generalize to other samples,
or use to draw good inferences about the population as a whole. Likewise, the worst
use of anyone’s time is to publish or present results that are not replicable, or
that are so imprecise that we cannot draw any conclusions. Large samples and clean
data (in addition to strong factor loadings and larger numbers of strongly loading
variables per factor) contribute to this mission. Small samples and weak loadings
(and few variables per factor) make for messy, conflicting, and useless results.
-
Principal components analysis is not exploratory factor analysis. We have seen endless debate amongst a small number of partisans regarding PCA vs
EFA. Almost nobody else seems to care about this debate, caring rather for whether
they can trust their results and interpret them sensibly. If you feel some compelling
reason to use PCA (and we do not see one at present), then we hope this book can guide
you as well. Most of the best practices that we have covered in this book also apply
to PCA. If you insist on using PCA, at least do it with large samples, clean data,
and with the limitations of the procedure clearly and overtly admitted.
-
If you use EFA, don’t use the defaults! If you want to consider yourself to be modeling and exploring latent variables in
the best way possible, you want to use ML, iterated PAF, or ULS extraction (depending
on whether your data meets the assumptions of ML), and we think you want to use oblique
rotation methodologies (either oblimin or Promax seems to work fine in most cases—if
one doesn’t work, try the other). Scholars in this area spend so much energy arguing
about which extraction or rotation technique is best. But keep our mantra in mind—this is just an exploration. Thus, it should be considered a low-stakes endeavor. Whatever you find from EFA
should subsequently be confirmed in a large sample confirmatory analysis.
-
Use multiple decision rules when deciding how many factors to extract. Another point of constant argument in this field seems to be what decision rule is
best in guiding someone on how many factors to extract. We reviewed several, and none
are perfect. Just in our three examples, one had a clearly uninterpretable scree plot,
one parallel analysis produced what we consider to be questionable guidance, and one
MAP analysis was clearly (to our eye, anyway) ambiguous and unhelpful. The other criteria
were also at times confusing and problematic. The best guide is theory, and beyond
that, choose whatever provides the results that make the most sense. If you cannot make sense of the results—in other words, if you cannot easily explain
to someone what each factor means—then you need to go back to exploring. Because any
model that you produce has to be confirmed with CFA in the context of a new sample,
this seems to us the most sensible approach. Thanks to Brian O’Connor, we have easily
accessible ways of trying out modern decision criteria (MAP, parallel analysis). Use
them, but realize that no one decision rule will be perfect in all situations.
-
Replicate your results. If you have two good samples, you can present replication statistics like those we
reviewed in Chapter 5, or you can put a single sample to work in bootstrap analysis,
like those we explored in Chapter 6. It’s not easy, nor is it automatic, but with
the syntax and macros we share, it is not too difficult. And we think that it provides
invaluable perspective on your results. We wish this mandate to replicate results
would permeate every research lab and statistics class, regardless of what statistical
techniques they use. The lessons that are contained in these chapters are equally
valid if you are performing ANOVA or regression analyses, hierarchical linear modeling,
or nonparametric techniques. Replicate your results, bootstrap your analyses, and
report (and interpret) confidence intervals for important effects so we, as readers,
can get more out of the hard work that you put into your research.
-
Have fun! The ability and training to perform research like this is a wonderful gift. The first
author has been lucky enough to spend the last 25 years doing quantitative research,
and the second author is at the beginning of her journey, but we have enjoyed every
minute of it. Those of us who perform data analysis are the ones who are present at the moment each tiny bit of knowledge is created.
We create knowledge—we ask questions and find answers. Sometimes those answers are
not what we expect, which is an opportunity to ask better questions or learn something
unexpected. We cannot think of a more rewarding way to spend our career, and we hope
each one of you experiences the same joy and thrill from your research.