Peter Siminski - Choosing the Best Specification in Regression Discontinuity Designs, Regression Kink Designs and Regression Probability Jump and Kink Designs

Melbourne Institute Seminar Series

Melbourne Institute Seminar Room
Room 6.05, FBE Building
111 Barry St, Carlton

Map

More Information

mi-seminars@unimelb.edu.au

Title: Choosing the Best Specification in Regression Discontinuity Designs, Regression Kink Designs and Regression Probability Jump and Kink Designs

Abstract: When using Regression Discontinuity Design (RDD), Regression Kink Design (RKD), and Regression Probability Jump and Kink Design (RPJKD), applied researchers must choose between many specifications, which vary according to bandwidth, functional form and possibly other factors. Currently, there is no accepted method for choosing between these specifications, with much emphasis on robustness testing. All of these modelling choices involve a trade-off between efficiency and bias. We propose that the set of candidate models can be greatly reduced through Monte Carlo simulation, using a data generating process which closely resembles the true data. This process can be used instead of optimal bandwidth selection algorithms. We illustrate this with an application in which we evaluate changes to motor vehicle licensing policy in the Australian state of New South Wales (NSW). Graduated driver licensing (GDL) systems have been widely introduced internationally, and have successfully reduced motor vehicle accidents. But little is known about which components of GDLs are most effective and whether they could be refined. Minimum supervised driving hours (MSDH) are a key component of GDLs. We study two increases to MSDHs in NSW – from zero to 50 hours, and 50 to 120 hours. The discontinuities create curious ‘first stage’ relationships between date of birth and ‘treatment’ which could be seen as discontinuities, or kinks, or both. We estimate the effects using many regression discontinuity, regression kink, and regression probability jump and kink design estimators. We employ a Monte Carlo procedure using simulated data to choose between these estimators. We conclude that the first policy change (from zero to 50 hours) had a substantial effect on reducing crashes amongst young drivers. The second policy change does not seem to have had an effect.

Presenter: Peter Siminski, UTS

The program coordinator of these seminars is Jiao Wang. If you would like to subscribe to the Melbourne Institute Seminar Series email list, please contact Jiao.