After you are done with the exploratory
factor analysis, your journey is just beginning, rather than ending.
The exploration phase might be drawing to a close, but the psychometric
evaluation of an instrument is merely starting. The process of performing
exploratory factor analysis is usually seeking to answer the question
of whether a given set of items forms a coherent factor (or several
factors). After we decide whether this is likely, evaluating how well
those constructs are measured is important. Along the way, we can
also ask whether the factor that is being examined needs all the items
in order to be measured effectively.
To fully evaluate an instrument, we should evaluate whether
the factors or scales that we derive from the EFA are reliable, confirmed
in a new sample, and stable (invariant) across multiple groups. In
this chapter, we will briefly look at the most common method of assessing
scale reliability, Cronbach’s coefficient
alpha.
Let us start with a
discussion of the modern view of reliability and validity. When developing
a scale to be used in research, there is a delicate dance between
focusing on creating a scale that is a “good” scale,
and the acknowledgment in modern research methods that things like
factor structure, reliability, and validity are joint properties of
a scale and of the particular sample being used (Fan & Thompson,
2001; Wilkinson, 1999). It should be self-evident to modern researchers
that a scale needs to be well-developed in order to be useful, and
that we do that in the context of a particular sample (or series of
samples, as we recommended when discussing replication). Thus, those
of us interested in measurement must hold two somewhat bifurcated
ideas in mind simultaneously —that a scale can be stronger
or weaker, and that scales are strong or weak only in the context
of the particular sample being used. This can lead to a nihilistic
mindset if carried too far, and so we recommend that we take a moderate
position in this discussion: that scales can be more or less strong,
but that all scales need to be evaluated in the particular populations
or data that they reside in.