Page 18 - IDEA Studie 07 2023 TACR

ARE SUBSIDIES TO BUSINESS R&D EFFECTIVE? REGRESSION DISCONTINUITY EVIDENCE FROM THE TA CR ALFA PROGRAMME
! = #$ + & '1 − $ *+ + & $ + + , - .' + / + / + 0 . !"#$ "% ""&"" '!"#!$$#!"#
')*
IDEA 2023
! is the outcome in year t for firm i participating in project p submitted to call c. Our !"#$
primary outcome of interest is the firm’s total R&D expenditure, but we also consider additional outcomes: privately-funded R&D expenditure, publicly-funded R&D expendi- ture (in total and individual components: domestic direct funding, EU direct funding, R&D tax relief), current R&D expenditure, capital R&D expenditure, a dummy for having filed a patent, and employment, sales, and labour productivity. With the exception of the patenting dummy, all outcome variables are included as natural logarithms.9
$ is a dummy variable marking whether project p received an ALFA grant, and + is ""
the running variable, given by each project’s average score (number of points) across 3 or 4 evaluators. We normalise the score so that it equals zero at the threshold, i.e., projects with a zero or a positive score were funded, and projects with a negative score were not.10 Use of higher degree polynomials in the running variable has been shown to lead to noisy estimates, to results that are highly sensitive to the degree of the polynomial, and to poor coverage of confidence intervals, frequently offering empirical support for evidently nonsensical results (Gelman & Imbens, 2019). For this reason, we use a linear polynomial in our running variable and test the robustness of the results to using a quadratic polynomial instead. As is standard in RD analysis, we use local polynomials that are independently estimated on each side of the threshold (Lee & Lemieux, 2010).
Consistent identification of causal effects in RD designs generally does not require, and is not helped by, inclusion of additional controls in regressions. Controlling for additional predetermined covariates can, however, increase the precision of estimates (Calonico
9 The individual components of the total R&D expenditure by the source of funding and the type of costs are equal to zero for many firms. For this reason, we calculate the logarithm for R&D variables other than the total R&D expenditure as !"# (&(') + '), where x is a given component of R&D expenditure and &(') is constant specific to variable x. Chen & Roth (2023) show that estimation results with this widely-used transformation are not scale-invariant (i.e., they depend on the value of &(')) and the transformation affects the relative weight of the extensive margin (e.g., firms with strictly positive expenditures on R&D-related buildings and machinery) and the intensive margin (the size of the expenditure) in the regressions. We take one of the approaches suggested by Chen & Roth to tackle this issue, which is to establish an explicit trade-off between the extensive and intensive margins. Specifically, we set &(') to the 5th percentile among all non-zero values of x as observed in 2010 (the year before the start of projects in the 1st call of the ALFA programme). This implies that going from zero expenditure to expenditure on the 5th percentile (among strictly positive values) increases the logarithmised value by 1, and is, thus, considered in the estimation as equivalent to an intensive-margin change of !"#(2) ≅ 70%.
10 4 projects in calls 1 and 2 received exactly the same threshold scores, but were not funded. To match the stated interpretation of the normalized score, we subtract 0.01 point from the normalized score from these 4 projects. Dropping these 4 projects, or dropping all projects, both funded and unfunded with scores exactly at the threshold in calls 1 and 2, has no material effects on the results (results available from the authors upon request).
(
16