Sifting Statistical Significance From the Artifact of Regression– Discontinuity Design

When the covariate of an evaluation study using regression analysis is falhbly measured, the statistical test of program effectiveness is biased. In programs where the target population is “below average,” the bias tends to suppress the beneficial effects of the program, while raising them in programs designed for those “above average.” The magnitude of this bias is calculated, and a correction method is derived and illustrated. In many cases, these biases will make practical differences in program assessment, lending further support to the general distrust of nonexperimental methods. Nonetheless, our approach improves the reliability of nonexperimen tal methods by correcting one potential source of their bias.