Regression Discontinuity

History

The RD design was introduced in 1960 by Donald Thistlewaite and Donald Campbell to evaluate the effects of compensatory educational interventions. Since that time, it has been used in areas such as criminal justice, social welfare, economics, and pharmaceutical health services research, although it is still a relatively rarely used design. Indeed, a recent search of the psychological and medical literature since 1960 turned up only 59 references, of which 34 were about the technique and only 25 actually used it. Of the latter, about one third were dissertations, and nearly three quarters were in the field of education.

Method

The unit of assignment can be individuals or groups, such as classes within a school, entire schools, or even school districts. The units are assigned to the treatment or control conditions on the basis of their score on some measure. The cutoff score can be determined in one of two ways: either on the basis of knowledge of the properties of the score (e.g., a failing grade, the proportion of high school students not graduating, or blood pressure above some threshold for hypertension) or on the basis of the resources that are available (e.g., the program is able to handle only 25% of the people). In the simplest case, the grouping variable is the outcome variable itself, which is measured before and after the intervention for the experimental group and at equivalent times for the comparison group. Hence, the basic design looks like

and the results are plotted with the pretest on the x-axis and the posttest on the y-axis.

For example, people can be chosen for a weight-loss program if their body mass index (BMI) is more than 30, which is regarded as the cut point for obesity. At the end of a 6-month intervention, the BMIs for all people are measured again, and the hypothetical results are shown in Figure 1.

If the intervention has no effect, then the regression lines will be continuous at the cut point. The discontinuity at the cutoff, seen in Figure 1, indicates that the program was effective in that the postintervention scores for the experimental group tend to be lower than the prescores, whereas the before and after BMIs for those in the comparison group are about the same.

It is not necessary that the pre- and posttreatment variables be the same or that the cut point be [Page 1235]based on the pretest. For example, the pretreatment and cutoff scores could be the BMI, and the posttest score could be caloric intake. Similarly, caloric consumption could be measured before and after the intervention, but the cutoff could be based on the BMI; or caloric intake could be measured at Time 1, weight measured at Time 2, and BMI used to determine the cut point.

...