Geographic Boundaries as Regression Discontinuities.

We explore the use of geographic boundaries as regression discontinuities, studying designs where the assignment variable is distance to a boundary and subjects on either side of this boundary are compared. We develop the identification assumptions behind RD designs of this type and suggest that the key assumption is more likely to be violated, since agents are better able to sort around the discontinuity. Moreover, we show that geographic RD designs that employ a naive notion of distance as the assignment variable fail to recover the treatment effects of interest. We develop an new estimator that is faithful to the inherently spatial qualities of the design and a method to test between identification strategies. We illustrate our argument and method with an application on whether ballot initiatives increase turnout. Specifically, we focus on a 2008 initiative to provide paid sick leave that was on the ballot in the city of Milwaukee but not the rest of the state. We find no evidence that turnout increased due to the ballot initiative. [ABSTRACT FROM AUTHOR]