Despite the growing literature on tax incentives for research and development (R&D), little is known about how the design aspects of the fiscal policies shape the effectiveness. This paper applies meta-regression techniques to assess the effect of various design attributes on innovation inputs. Using 496 estimates from 37 studies exploiting direct approaches, the results indicate that the base definition of tax incentives substantially drive the heterogeneity found in the literature. MetaForest, a novel machine learning algorithm, confirms these results. Furthermore, I find significant publication bias in favor of reporting positive effects of fiscal incentives, which is more prevalent among studies published in peer-reviewed journals.