Four types of reliability will be discussed in terms of classical test theory. The four types discussed are: (1) split-half, (2), parallel forms, (3) test-retest, and (4) internal consistency. These four methods can be reduced to basically two methods: (1) the reliability coefficient (split-half, parallel forms, and test-retest) and the standard error of measurement (internal consistency). The reliability coefficient assesses the degree that one test or part of test can predict another and uses some correlational method. The standard error of measurement assesses the degree that an individuals scores varies over parallel tests and uses ANOVA methods (although it has been shown in chapters 2 and 3 that these methods are different parts of the general linear model).

A notion of parallel tests is needed understand this section. A test is made up of items that are designed to measure psychological attributes. A test designed to measure a single attribute is called a univariate test --all items on the test are intended to measure the same thing. The Beck Depression Scale is such a test -- it is designed to measure the attribute of depression. A test that is designed to measure more than one attribute is called a multivariate test -- the MMPI is such a test with its many subscales. The idea of parallel tests is that two tests measure the same attribute. Since they measure the same thing they are identical or parallel. You should note that the split-half and parallel tests are similar. When a tests is split into two part (split-half) the two halves become parallel tests (in this chapter they can be thought of as the same concept).  It is this notion of parallel items (or interchangable items) that we are testing when we assess reliability.

Reliability is a problem because psychological characteristics can't be measured perfectly. When considering psychological attributes, there is considerable unreliability. And the error in measurement is a problem that you must deal with in some way.

The Psychosocial Assessment Scale is a multivariate test since there are six subscales. But within those subscale, those items are univariate within those factors or subtests. There are six subscales on the PAS, however, 2 of the subtests have only 1 item. It may be debatable whether a subscale with only 1 item is really a subscale. The split-half and coefficient alpha cannot be used when there is only 1 item on a subscale. Each subscale is considered as a scale itself when assessing reliability.