Page 468 - Encyclopedia of Nursing Research
P. 468
RElIABIlITy n 435
a bivariate correlation as it accounts for more mean differences in a similar manner to sta-
of the error variance inherent in any measure. bility. Consistency is assumed if the scores
A bivariate correlation tells the investigator are equivalent. Assessment with alternative/ R
whether individuals who scored high on the parallel forms is not comparison with two
first administration also scored high on the different measures of the concept. It is com-
second, but it does not provide information parison of two essentially identical tests that
on whether the scores are the same. were developed at the same time through
The problem with stability is that it is the same procedures. Therefore, a difficulty
not always reasonable to assume that the with this approach to equivalent reliability is
concept will remain unchanged over time. If obtaining a true parallel or alternative form
the person’s true score on a concept changes of an instrument.
within 2 weeks, instability and high random A more common way to look at equiva-
error will be assumed—when, in effect, it is lence is through internal consistency proce-
possible that the instrument is consistently dures. The assumption underlying internal
measuring change across time. Reliance on a consistency is that the response to a set of
2-week interval for measuring stability may scale items should be equivalent. All inter-
be faulty and must be directly related to the nal consistency approaches are based in
theoretical understanding of the concept correlational procedures. An earlier form of
being measured. internal consistency is split-half reliability, in
A special case of stability occurs with which responses to half the items’ on a scale
instruments that are completed by raters on are randomly selected and compared with
the basis of their observations. Intrarater reli- responses on the other half.
ability refers to the need for ratings to remain Currently Cronbach’s (1951) alpha reli-
stable across the course of data collection and ability coefficient is the most prevalent
not change due to increased familiarity and technique for assessing internal consis-
practice with the instrument. The ICC assess- tency. Developed in the 1950s, the formula
ment procedures can be used for intrarater basically computes the ratio of variability
reliability as for test–retest reliability but will between individual responses to the total
utilize slightly different formula looking at variability in responses, with total variabil-
absolute agreement versus consistency. A ity being a composite of the individual var-
kappa statistic also can be calculated when iability and the measurement error. As with
dealing with agreement among observers. the ICC, Cronbach’s alpha is a ratio ranging
However, the ICC is adequate to deal with from 0 to 1, with the values closer to 1 indi-
most of these situations, and the kappa sta- cating less measurement error. The ratio
tistic has no clear advantage over the ICC. reflects the proportion of the total variance
Equivalence is evaluated in two major in the response that is due to real differences
ways. The first of these predated the avail- between subjects. A general guideline for
ability of high-speed computers and eas- use of Cronbach’s alpha to assess an instru-
ily accessed statistical packages. This set of ment is that well-established instruments
techniques deals with the comparison of must demonstrate a coefficient value above
scores on alternate or parallel forms of the .80, whereas newly developed instruments
instrument to which the subject responds at should reach values of .70 or greater. This
the same point in time. Parallelism means should not be taken to indicate that the higher
that an item on one form has a comparable the coefficient, the better the instrument.
item on the second form, indexing the same Excessively high coefficients indicate redun-
aspect of the concept, and that the means dancy and unnecessary items. A special case
and variances of these items are equal. These of alpha is the Kuder–Richardson 20, which
scores are compared through correlation or is essentially alpha for dichotomous data.

