Page 19 - IDEA Studie 07 2023 TACR

ARE SUBSIDIES TO BUSINESS R&D EFFECTIVE? REGRESSION
DISCONTINUITY EVIDENCE FROM THE TA CR ALFA PROGRAMME IDEA 2023
et al., 2019).11 For this reason, we include a set of controls .' in our baseline !"#!$
specification. All controls are measured in the pre-treatment year 2+, defined as the year in which the project was submitted and evaluated, or, equivalently, as the year before the year in which the project was to start, provided it won a grant. Firstly, they include pre-treatment values of all the outcome variables we examine, as listed above. This allows us to keep the list of controls fixed across various outcomes while ensuring that we always control for the pre-treatment value of a given outcome. In addition, the controls include demographic characteristics describing firms’ age, legal form, sector, and location, and the project-level characteristics. Finally, we control for year dummies /# and call dummies /$ .
The assumption that projects above and below the threshold are similar, conditional on their score, is unlikely to hold for projects further away from the threshold. Therefore, we restrict the analysis to projects with scores that lie within bandwidth h around the threshold. For the total R&D expenditures, our main outcome of interest, the mean square error (MSE) optimal bandwidth selection procedure with covariates by Calonico et al. (2019) suggests a bandwidth of 5.8 points for SMEs and 3.7 for large firms. To make the bandwidth consistent across firm-size samples and across outcomes, we use a fixed bandwidth of 5 points in our baseline specification. We estimate the equation above using weighted least squares, with weights given by a kernel function 3'+"⁄h*.12 As a baseline, we use a triangular kernel function, which assigns a linearly smaller weight to observations further away from the threshold. We test the robustness of the results to using alternative bandwidth values and kernel functions. We report bias-corrected RD estimates and robust standard errors clustered at the firm level (Calonico et al., 2019).13
Validity tests
The identification in our RD design rests on the assumption that scores were not manipulated around the cutoff. Such manipulation by the evaluators was made unlikely by the fact that the score received by each project was an average of points awarded independently by three or four evaluators, and that the exact location of the cutoff was not known at the time the points were assigned. That said, the Board of the Programme and the Board of TA CR, in principle, had the right to adjust the number of points allocated
11 For similar reasons, researchers often include pre-treatment covariates when analysing randomised experiments.
12 The estimation is performed in Stata using command rdrobust (Calonico et al., 2014, 2017).
13 To estimate the bias of the regression function estimator, we use a second order polynomial and bandwidth / = 10, which is close to the MSE-optimal bias bandwidth of 10.6 for SMEs and 7.1 for large firms. We test the robustness of the results to this choice.
17