Avoid: Regression to the mean. Because of the imperfect reliability of tests as discussed above, a phenomenon that has plagued adult education programs for decades is regression to the mean. This usually happens when a group of adults is administered as a pre-test, a standardized test that has been normed using traditional test development methods, and a part of the group is identified as low in ability and sent to a program. Then, later on, when just the low group is post-tested, it is found that the average post-test score is higher than the pre-test score. Under these circumstances, the program offers the gain between pre and post-test scores as evidence of the effectiveness of the program in bringing about achievement.

However, regression to the mean is a statistical process that generally operates under the foregoing conditions. Whenever a low-scoring group is separated off from the total group and then retested, the average score of the post-test will generally be larger than the average score of the pre-test. This is due to the fact that many people are in the low group on the pre-test because they guessed poorly or did not perform well due to anxiety, lack of recent practice in test-taking and so forth, as mentioned earlier. So, when they are retested, their average score moves up toward (that is, regresses toward) the mean (or average) score of the total group on which the test was normed. 3

Such warm-up and regression effects can be quite large. In one study, military recruits new to the service were tested with a standardized, grade-school normed reading test. Those scoring below the sixth grade level were retested two weeks later, with no intervening reading instruction, and those who scored above the sixth grade were excluded from the study. Two weeks later, the remaining recruits who scored below the sixth grade level were retested with a third form of the reading test, and those who scored above the sixth grade level were excluded. This process reduced the number of people reading below the sixth grade level by 40 percent (Sticht, 1975)!

Regression effects can be reduced in several ways. One is to use the retesting procedure discussed above. Obviously, this requires quite a commitment to testing. It also requires the use of standardized tests with at least three comparable forms, one for the first testing, a second for the next testing of the group identified as low on the first testing, and a third for the post-testing of the group identified in the second testing who were placed in the program of interest.

Regression effects can also be reduced by not testing learners until they have adjusted to the program and obtained some practice in test-taking as noted earlier.

In another approach to managing regression effects, scores on post-tests may be adjusted for regression by using the correlation between pre and post-test scores. This permits the prediction of post-test scores from pre-test scores. Then, actual post-test scores can be compared to the predicted scores. Only the gain that exceeds the predicted post-test scores is then used to indicate program effectiveness. This procedure requires technical assistance from a knowledgeable statistician or psychometrician.

Regression effects may also be estimated and adjusted for by comparing the program group to a group with similar pre-test scores which does not receive the educational program being evaluated (though note that the control group should receive some practice in test-taking, to offset the "warm-up," "surge" or "practice" effects discussed above). This "treatment" and "no treatment" groups comparison permits programs to adjust their gains for regression.