Let us now consider some of the consequences of less than perfect reliability. Suppose we use our scale of prejudice against foreign-made cars (see Basic ideas) to predict some other criterion, such as subsequent actual purchase of a car. If our scale correlates with such a criterion, it would raise our confidence in the validity of the scale, that is, that it really measures prejudices against foreign-made cars, and not something completely different. In actual test design, the validation of a scale is a lengthy process that requires the researcher to correlate the scale with various external criteria that, in theory, should be related to the concept that is supposedly being measured by the scale.

How will validity be affected by less than perfect scale reliability? The random error portion of the scale is unlikely to correlate with some external criterion. Therefore, if the proportion of true score in a scale is only 60% (that is, the reliability is only .60), then the correlation between the scale and the criterion variable will be attenuated, that is, it will be smaller than the actual correlation of true scores. In fact, the validity of a scale is always limited by its reliability.

Given the reliability of the two measures in a correlation (i.e., the scale and the criterion variable), we can estimate the actual correlation of true scores in both measures. Put another way, we can correct the correlation for attenuation:

rxy,corrected = rxy /(rxx*ryy)½

In this formula, rxy,corrected stands for the corrected correlation coefficient, that is, it is the estimate of the correlation between the true scores in the two measures x and y. The term rxy denotes the uncorrected correlation, and rxx and ryy denote the reliability of measures (scales) x and y. The Reliability module provides an option to compute the attenuation correction based on user-specified values, or based on actual raw data (in which case the reliabilities of the two measures are estimated from the data).