Subjects with extreme scores on a first measure of the dependent variable tend to have scores closer to the mean on a second measure. According to Campbell (1969, p. 414): "Take any dependent measure that is repeatedly sampled, move along it as in a time dimension, and pick a point that is the highest (lowest) so far. On the average, the next point will be lower (higher), nearer the general trend." Suppose that the dependent variable is measured twice for a group of subjects, once at Time A and later at Time B, and that the independent variable is introduced in the interim. Suppose also that value observed for subjects at Time A is considerably higher (lower) than would typically be the case. If scores on the dependent measure differ at these two times, it may be due to the independent variable or to a regression artifact.
One day at school, the children viewed the 20-minute cartoon (Control condition). Two days later, the Generalization Probe was conducted. Then, in a class the following week, the children viewed the 20-minute interactive video (Experimental condition). The plan was to administer a second Generalization Probe two days after that. However, at this point, the experimenter realized that she had insufficient funding to complete the study and would only be able to retest ten children. She selected the ten poorest performing children on the first Generalization Probe, the mean score of which was 0.1. Their mean score on the second Generalization Probe was 2.5. We conclude that the 20-minute interactive video improved the children's self-protection skills in a potential abduction situation.
One day at school, the children viewed the 20-minute cartoon (Control condition). Two days later, the Generalization Probe was conducted. Then, in a class the following week, the children viewed the 20-minute interactive video (Experimental condition). The plan was to administer a second Generalization Probe two days later. However, at this point, the experimenter realized that she had insufficient funding to complete the study and would only be able to retest ten children. She wrote the name of each child on a separate slip of paper, put all the slips in a bowl, and the first ten names she pulled out were selected for the second Generalization Probe. Their mean score on the first Generalization Probe was 1.1 and their mean score on the second Generalization Probe was 2.5. We conclude that the 20-minute interactive video improved the children's self-protection skills in a potential abduction situation.
The first item is an example in which regression to the mean is a threat to internal validity. The children were selected for retesting on the basis of their extremely low scores on the first Generalization Probe. Based on the principle of statistical regression, these children will tend to score higher on the second Generalization Probe. Their improvement across the two Generalization Probes may be due to intervening exposure to the interactive video or to a regression artifact.
In the second item, because the children were selected for re-testing on the basis of chance, their mean score on the first Generalization Probe more likely represents a value closer to the true mean of the population of children in the Control condition. We can be more confident that the improved score across the two Generalization Probes was not the result of regression to the mean.