Optimal Model Selection in RDD and Related Settings Using Placebo Zones
Nathan Kettlewell and Peter Siminski.
Policy rules frequently create discontinuous ‘jumps’ in exposure to policies and programs, such as unemployment insurance, legal drinking age, pensions and so on. The regression discontinuity design (RDD) and related methods have become key tools for empirical researchers in these settings, allowing them to causally evaluate treatment effects.
We propose a new model-selection algorithm for RDD and related estimators. Candidate models are assessed within a ‘placebo zone’ of the running variable, where the true effects are known to be zero. The approach yields an optimal combination of bandwidth, polynomial, and any other choice parameters. It can also inform choices between classes of models and any other choices, such as covariates, kernel, or other weights.
We use the approach to evaluate changes in Minimum Supervised Driving Hours in the Australian state of New South Wales. Our results indicate that going from 0-50 hours reduces the probability a driver is involved in a motor vehicle accident by 1.4 percentage points (21%) in the first year of independent driving. However, we find no effect going from 50-120 hours.
We also re-evaluate evidence on the effects of Head Start (a comprehensive education, parenting and health program targeted at low income families in the U.S.) and Minimum Legal Drinking Age. We conclude with practical advice for researchers who want to use our method, including implications of treatment effect heterogeneity.
September 11, 2020