Although alpha is generally the preferred method of estimating
reliability (particularly in the social sciences), it is a fairly
basic method that can be easily influenced. It is not robust with
respect to many potential characteristics of a sample or instrument.
It is important to be aware of these characteristics so that appropriate
steps can be taken to reduce their effect and improve our estimates
of reliability. Each characteristic is described below.
High average
inter-item correlation. All other things being equal,
alpha is higher when the average correlation between items is higher.
But even Cronbach specifically pointed out that when inter-item correlations
are low, alpha can be high with enough items with low intercorrelations.
This is one of the chief drawbacks to interpretability of alpha—that
with enough mostly unrelated items, alpha will move into the “reasonable”
range that most researchers use as a rule of thumb.
Length
of the scale. As mentioned above, all other things being
equal, longer scales will have higher alphas.
Reverse coded
items (negative item-total correlations). Many scales
are constructed with reverse-coded items. However, alpha cannot provide
accurate estimates when the analysis includes items with negative
item-total correlations. Thus, any item that is expected to have a
negative item-total correlation (e.g., if the factor loading is negative
when most others are positive) should be reversed prior to analysis.
Random responding or response
sets. Random responding (discussed in Chapter 8) tends
to attenuate all of these estimates because it primarily adds random
error. Thus, failure to identify this issue in your data will lead
to under-estimation of the internal consistency of the data. Response
sets can have a variety of effects, depending on the response set.
Some types of response sets will inflate alpha estimates and some
can attenuate alpha (for an overview of response sets, and how one
can identify them, you might see Osborne & Blanchard, 2011).
Multidimensionality. The
assumption of alpha is that all items within a particular analysis
represent a single dimension, or factor. To the extent that assumption
is violated, the estimate of alpha will be misestimated. Thus, the
factor structure of the scale should be considered before submitting
items to this type of analysis.
Outliers. Outliers (inappropriate
values, also discussed in Chapter 8 usually have the effect of increasing
error variance. This generally has the effect of attenuating the estimate
of alpha. Thus, data should be carefully screened prior to computing
alpha.
In addition to the above
factors that can influence the estimate of alpha, a theoretical issue
can hinder its interpretation. Alpha was built for a time when researchers
often summed or averaged items on scales. Many researchers do this
today. Of course, when summing or averaging items in a scale, you
are making an assumption that all items contribute equally to the
scale—that the weighting of each item is identical. Alpha also
assumes that all items contribute equally. Yet from what we have seen
in earlier chapters, that might not be a valid assumption, and it
is not the assumption made by EFA (or confirmatory factor analysis,
latent variable modeling, IRT, Rasch measurement, etc.). For example,
in Chapter 4 we saw that the pattern loadings for the engineering
problem-solving items ranged from 0.79 to 0.91, and for the GDS the
loadings ranged from 0.23 to 0.68 when only one factor was extracted.
If you square the loadings to estimate the unique proportion of the
shared variance that is attributable to each item, this amounts to
a range of 0.62 to 0.82 (for engineering problem-solving) and from
0.05 to 0.46 for GDS. Thus, each item contributes a different proportion
of variance to each factor.
Historically, this led
researchers to create “factor scores,” which weighted
each item by the factor loading to more closely approximate what a
latent variable score might be. As we discussed in Chapter 9, we do
not believe this is a good idea. However, there is not currently a
good way to deal with this theoretical issue when using alpha (in
our opinion). Those interested in estimating internal consistency
via alpha must simply be aware of its violation of the assumption
of unique item contribution and live with it.