Whenever
factors are correlated, there is, naturally, a question as to whether
there truly are several independent factors or whether there is a
single “higher-order” factor. This
has been a point of discussion for many decades, and is often conceptually
and theoretically important. For example, is self-concept a single
thing, or several separate things? Is depression a single construct
composed of several subconstructs, or is it really not a coherent
construct?
Scholars writing in
this area since the early 20th century have argued that when initial
factor analyses (we can refer to these as “first-order”
factors as they come from the first level of analysis) produce correlated
factors, researchers should explore whether there are second- or higher-order
factors in order to more fully explicate the model (e.g., Gorsuch,
1983; Thompson, 2004).
There are at least two
issues with higher-order factors that we need to address. The first
is how to perform the analysis and interpret the results. The second
issue is more conceptual: if initial EFA produces imperfect abstractions
(unobservable variables called factors) that might or might not be
precise representations of population dynamics, then higher-order
factor analysis would entail analyzing these imperfect abstractions
to create possibly more imperfect higher-order abstractions. This
makes us a bit uneasy, particularly in an exploratory framework. Given
how volatile and unpredictable the results of EFA can be, it seems
that taking those results and analyzing them again doubles (or raises
to a power) the risk of going awry and far afield of the true character
of the population dynamics.
Another issue is whether
there is a meaningful conceptual distinction between a single unitary
factor or a multifactor construct with a unitary higher-order factor.
We should seek to operate under principles of parsimony—in
other words, if you as a scholar cannot meaningfully communicate why
this distinction is important, you might want to ask whether it is
worthwhile to pursue.
Since almost all factors
are correlated in the population, the assertion that this analysis
needs to take place under these conditions should be regarded carefully.
Researchers must decide (a) whether higher-order exploratory analyses
are desirable and (b) how strong a correlation warrants this extra
exploration. If factors are correlated around r = 0.25 (about 6.25%
overlap), is that enough of a correlation to justify higher-order
analysis? What about r = 0.40, which equates to only 16% overlap?
In our opinion, it’s difficult to define what a second-order
factor is.