Abstrakt: We examine the impact of R&D tax credits and direct R\&D
subsidies on Norwegian firms' innovation activities, as measured by
patent and trademark applications. To address the problem of endogenous
selection, we apply machine learning methods to estimate average
treatment effects, which are applicable to situations where there is a
huge number of potential confounding factors (p) relative to number of
observations (n) (possibly p > n), or equivalently, the "true" control
function is unknown and cannot be estimated consistently. We extend the
literature originally developed by Belloni et al. (2014, 2016) and
Chernozhukov et al. (2018), by using statistical learning methods in the
context of an event study design, where treatments are sequential and
possibly repeated. Our results show that both direct subsidies
and tax credits have significant positive effects on innovation
activities. However, the magnitude of the effects depend critically on
firms' pre-treatment characteristics. In particular, the statistically
significant estimates are all related to firms without prior
innovations. Moreover, our results suggest that R\&D support should be
directed to promote innovations at the extensive margin, i.e. to firms
with a high potential of becoming innovative rather than to firms with a
record of being innovative.