Estimation of Cronbach’s alpha is extremely easy
in SAS. It is done using the CORR
procedure.
The ALPHA
option must be included on the CORR
statement
to request SAS to output estimates of alpha, and all the variables
to be included in the alpha estimate should be listed on the VAR
statement
(or they can be referenced by a prefix, e.g., EngProb: ). An example
of the syntax to produce estimates for the engineering problem-solving
items in the engineering data is presented below. It is followed by
a screenshot of the output produced by the ALPHA
option.
proc corr data=engdata alpha;
var EngProb:;
run;
As you can see in
Figure 11.1 PROC CORR output requested by the ALPHA option, the
ALPHA
option
produces a lot more than a single alpha estimate! The first table
provides a raw and standardized estimate of Cronbach’s alpha.
You should use the raw alpha if you are inputting your original data
into the
CORR
procedure, and you should use
the standardized alpha if you are inputting standardized versions
of your variables (e.g., z-scores) into the
CORR
procedure.
You would use standardized versions of your variables if some of the
items in the scale contained different numbers of response options
(e.g., three dichotomous items and five Likert-type 5-point items).
In the current example, we input the unstandardized engineering items
into the scale. Thus, we will interpret the raw alpha.
The second table provides
item-total correlations and revised estimates of alpha based on removal
of items. The item-total correlations are the correlations between
each individual item and the total composite score computed as the
average or sum of the items. We expect items to be highly correlated
with the total score as it represents the construct the item is intended
to measure. Items with low item-total correlations can be scrutinized
for potential removal. The revised alpha estimates represent the Cronbach’s
alpha if a particular item was removed from the scale. A revised alpha
estimate that is larger than our alpha indicates that an item contributes
unique variance that reduces our overall reliability. This unique
variance can in fact be unwanted error associated with an item., In
that case you might consider removing the item, or it could be construct-relevant
variance that is not captured by the other items, in which case you
would want to keep the item. Knowledge of the construct must be used
to guide decisions about item removal. Finally, similar to our interpretation
of the first table, we would interpret the raw or standardized versions
of these estimates based on the format of our data.
The engineering problem-solving
scale demonstrates good internal consistency. It has a strong alpha
of .95. The item-total correlations range from .80 to .88, and none
of the revised alphas exceeds the original estimate. Unless we had
a content-based reason for removing an item (e.g., an item is related
but does not directly reflect problem solving), we would recommend
making no changes to this scale.