Split-Half Reliability

Basic Principles and Estimation Procedures

According to classical test theory, variations in examinees’ test scores are due to (a) variations in the test takers’ true ability or trait and (b) error. The proportion of variation in the total score, [Page 1573]resulting from variations in the examinees’ true ability, is defined as test reliability. Traditionally, researchers have estimated reliability by administering a test twice to the same examinees and correlating their scores obtained at the 2 times (test–retest reliability) or administering two parallel forms of a test to test takers and correlating their scores on the two forms (parallel forms reliability). Both of these methods have limitations because it is not always feasible to administer a test multiple times and not all tests have multiple forms. One convenient alternative is to split a test in half and use each half as a parallel form of the other. Comparing scores on the two halves is then another way to measure test reliability. This method, referred to as split-half reliability, is considered a measure of the internal reliability of a test or how consistently the items perform within a test. The underlying assumption is if a test measures a single construct, then individuals should perform equally well on both halves of the test.

To estimate split-half reliability, the first step is to split the test in half and administer the two halves to the examinees. If there are multiple subscales or content areas assessed within a single test, split-half reliability should be calculated for each subscale or content area separately. When a test has more than 2 items, there are apparently multiple ways to split it. The principle is to obtain two halves as equivalent as possible. One could consider using the middle item (e.g., the fifth item on a 10-item scale) as the dividing point or randomly divide the test items into two groups which would represent the two halves of the test. However, because many tests organize items by difficulty level, these methods could easily lead to nonequivalent halves, which would result in an underestimation of the test reliability. Furthermore, attention and fatigue may affect individuals’ performance differently at the beginning and at the end of the test. That is, individuals may be more alert when they take the first half of a test, but more tired when they take the second half of the test. For these considerations, a commonly used approach is to split the test by even- and odd-number items. Another approach is to manually balance the difficulty level in the two halves of the test. Once the test is split in half and administered to the examinees, a Pearson correlation coefficient is calculated between the scores of the two halves of the test.

...