A data-driven procedure to determine the bunching window: an application to the Netherlands

We extend the bunching approach introduced by Saez (Am Econ J Econ Policy 2:180–212, 2010) by proposing an intuitive, data-driven procedure to determine the bunching window. By choosing the bunching window ad hoc, researchers throw away informative data points for estimating the counterfactual income distribution in the absence of the kink. Assuming a descending bunching mass to both sides of the threshold, the proposed algorithm produces a distribution of lower and upper bounds for the bunching window. In each iteration, the bunching window is defined as all contiguous bin midpoints around the threshold that lie outside of the confidence band resulting from running a local regression through all data points outside of the excluded region. Monte Carlo simulations provide evidence that our data-driven procedure outperforms larger bunching windows in terms of bias and efficiency. In our application for the Netherlands, we find clear evidence of bunching behaviour at all three thresholds of the Dutch tax schedule with a precisely estimated elasticity of 0.023 at the upper threshold, which is driven by self-employed, women and joint tax filers.