Cronbach’s alpha
is one of the most widely reported statistics relating to reliability
in the social sciences. It can be interpreted as an estimate of all
possible split-half statistics. It can also be interpreted as the
percent of variance that is “true score” variance. Thus,
if you have a measure with α = 0.80 in a particular sample,
there is about 80% that is “true score” and about 20%
error in the measurement for that sample.
Unfortunately, alpha
is still not reported in a majority of articles in modern, high-quality
research journals. This might stem from the anachronistic assertion
that scales, particularly those that are frequently used, are more
or less reliable. However, as we demonstrated throughout this chapter,
alpha can be influenced by a number of different factors that arespecific
to the sample (e.g., sample size, sample bias, etc.). This variability
across samples is not something often discussed in the psychometrics
literature, but it is important.
It highlights the sample-dependent nature of reliability and emphasizes
the need to attend to reliability and report it in each study. Estimation
of 95% confidence intervals can further aid in the interpretation
reliability and offer guidance about the replicability and precision
of the estimate.
Although alpha has been
the main focus of this chapter, there are modern alternatives to alpha.
Rasch modeling, for example, produces interesting information about
reliability from that perspective, as does Item Response Theory analysis.
Structural equation modeling (SEM) allows us to explicitly model the
latent construct and directly analyze it, eliminating the need for
alpha altogether. Where possible, it seems desirable to use modern
measurement methodologies.