RDP 2018-07: The GFC Investment Tax Break Appendix B: Extra RD Results

As noted in Section 5, for the RD specification we make a number of choices regarding the bandwidth, kernel, polynomial order and revenue ceiling. Below we show that the results are reasonably robust to these choices.

Figures B1 and B2 show the estimated treatment effects and 95 per cent confidence intervals using polynomials of order 1–4, with differing bandwidths. The results using a linear model are unsurprisingly somewhat sensitive to the choice of bandwidth, though they are not knife edge, as there is a moderate region around $0.2 million where the results are significant. Using higher-order polynomials we also find evidence of a significant treatment effect. These tend to be at higher bandwidths, reflecting the fact that we need a more flexible model once we look at a larger region. Importantly, the point estimates tend to be around 0.5–0.7, as in the baseline model. This gives us confidence that the results are reasonably robust.

Figure B3 shows the estimated treatment effects and confidence intervals using the same bandwidth as in the baseline model, but with differing ceilings for revenue in 2008/09. Overall, the results are reasonably robust to this choice, as long as the ceiling is not set very low.