Reducing Bias and Increasing Precision in Rdds By Adding a Pretest or Nonequivalent Comparison Group
Friday, November 13, 2015 : 1:50 PM
Brickell South (Hyatt Regency Miami)
*Names in bold indicate Presenter
The regression discontinuity design (RDD) is now standard in education for identifying causal estimates despite its dependence on functional form estimation, its identification of effects only at the cutoff determining treatment assignment, and its lower precision than an experiment. To mitigate these weaknesses, we add no-treatment comparison functions to the basic RDD to create comparative regression discontinuity (CRD) designs. One addition is of a pretest measure of the study outcome, called CRD-Pre; the other addition is of a nonequivalent comparison group, called CRD-CG. We compare the performance of each to a randomized experiment (RCT) and to a basic RDD using data from the National Head Start Impact Study. The results show that (a) both the RDD and CRD designs produce unbiased estimates at the cutoff where CRD designs have greater power; and (b) both CRD designs also produce unbiased causal estimates above the cutoff and are nearly as powerful as the RCT there. Since basic RDDs lack power and cannot be relied on for unbiased causal estimation away from the cutoff, whenever its assumptions are met CRD design deserves to replace RDD.