The regression discontinuity (RD) design has recently become a standard method for identifying
causal effects for policy interventions. We use an unusual “tie breaking” experiment, the
Kentucky Working Profiling and Reemployment Services, to investigate the performance of
widely used RD estimators. Two features characterize this program. First, the treatment
(reemployment services) is assigned as a discontinuous function of a profiling variable (expected
benefit receipt duration), which allows the identification of both experimental and
nonexperimental samples. Second, we deal with a discontinuity frontier rather than a
discontinuity point, which allows the identification of local average treatment effects over a wide
range of the support of the discontinuous variable. Using a variety of multivariate parametric and
nonparametric kernel estimators, we estimate the bias with respect to the benchmark
experimental estimates. In general, we find that local linear kernel estimates show the least bias,
but parametric estimates perform reasonably well. We also examine two alternative
discontinuities – geography and time – and find that they provide credible estimates as well.

(Joint paper with Dan Black (University of Chicago) and Jose Galdo (McMaster University), April 2007)