Walking the green line: Government sponsored R&D and clean technologies

Rentocchini, Francesco; Vezzani, Antonio; Montresor, Sandro Working Paper Walking the green line: Government sponsored R&D and clean technologies JRC Working Papers on Corporate R&D and Innovation (CoRDI), No. 01/2023 Provided in Cooperation with: Joint Research Centre (JRC), European Commission Suggested Citation: Rentocchini, Francesco; Vezzani, Antonio; Montresor, Sandro (2024) : Walking the green line: Government sponsored R&D and clean technologies, JRC Working Papers on Corporate R&D and Innovation (CoRDI), No. 01/2023, European Commission, Joint Research Centre (JRC), Seville This Version is available at: https://hdl.handle.net/10419/311180 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/ EUR XXXXX XX Walking the Green Line: Government Sponsored R&D and Clean Technologies JRC Working Papers on Corporate R&D and Innovation No 01/2023 2024 Rentocchini, F., Vezzani, A., Montresor, S. This document has been produced within the context of the Global Industrial Research & Innovation Analyses (GLORIA) activities that are jointly carried out by the European Commission's Joint Research Centre –Directorate Innovation and Growth and the Directorate General for Research and Innovation-Directorate F, Prosperity. GLORIA has received funding from the European Union's Horizon 2020 research and innovation programme. Any comments can be sent by email to: JRC-B6-[email protected], More information, including activities and publications, is available at: https://iri.jrc.ec.europa.eu/home/. Contact information E-mail: JRC-B6-[email protected] EU Science Hub https://joint-research-centre.ec.europa.eu JRC133670 Seville: European Commission, 2024 © European Union, 2024 The reuse policy of the European Commission documents is implemented by the Commission Decision 2011/833/EU of 12 December 2011 on the reuse of Commission documents (OJ L 330, 14.12.2011, p. 39). Unless otherwise noted, the reuse of this document is authorised under the Creative Commons Attribution 4.0 International (CC BY 4.0) licence (https://creativecommons.org/licenses/by/4.0/). This means that reuse is allowed provided appropriate credit is given and any changes are indicated. How to cite this report: European Commission, Joint Research Centre, Rentocchini, F. Vezzani, A., Montresor, S., Walking the Green Line: Government Sponsored R&D and Clean Technologies, European Commission, Seville, 2024, JRC133670. This document is a publication by the Joint Research Centre (JRC), the European Commission’s science and knowledge service. It aims to provide evidencebased scientific support to the European policymaking process. The contents of this publication do not necessarily reflect the position or opinion of the European Commission. Neither the European Commission nor any person acting on behalf of the C ommission is responsible for the use that might be made of this publication. For information on the methodology and quality underlying the data used in this publication for which the source is neither Eurostat nor other Commission services, users should co ntact the referenced source. The designations employed and the presentation of material on the maps do not imply the expression of any opinion whatsoever on the part of the European Union concerning the legal status of any country, territory, city or area or of its authorities, or concerning the delimitation of its frontiers or boundaries. 1 Contents Abstract ....................................................................................................................................................................................................................................................................... 2 Acknowledgements .......................................................................................................................................................................................................................................... 3 Executive summary .......................................................................................................................................................................................................................................... 4 1 Introduction..................................................................................................................................................................................................................................................... 5 2 Data and descriptive evidence ..................................................................................................................................................................................................... 8 3 Estimation strategy and variables ......................................................................................................................................................................................... 11 4 Results .............................................................................................................................................................................................................................................................. 13 4.1 Core estimations ....................................................................................................................................................................................................................... 13 4.2 Disentangling the spillovers of government sponsored R&D ........................................................................................................ 15 4.3 Robustness tests ....................................................................................................................................................................................................................... 17 4.4 The distributional effects of government sponsored R&D ............................................................................................................... 21 5 Concluding remarks ............................................................................................................................................................................................................................. 24 References ............................................................................................................................................................................................................................................................. 25 List of figures ..................................................................................................................................................................................................................................................... 28 List of tables ........................................................................................................................................................................................................................................................ 29 Annexes .................................................................................................................................................................................................................................................................... 30 A Data Construction ...................................................................................................................................................................................................................................... 31 B. Robustness Checks and Specifications ................................................................................................................................................................................ 33 Misspecification of outcome and treatment models: doubly robust estimators ...................................................................... 33 Results from the IPW treatment model and alternative weights ............................................................................................................ 35 Alternative distributional effects of government sponsored clean technologies..................................................................... 41 2 Abstract We examine whether government sponsored R&D induces the development of clean technologies with a high impact on subsequent technological development. The analysis uses information on USPTO patents granted between 2005 and 2015 and combines different methods to control for possible sorting of projects into public funding and for non-random (public) treatment. We also assess the distributional effect of government sponsored R&D. Results show that patents from public funded projects have a significantly higher impact and that this is particularly true for highly cited patents, thus supporting a role for technology-push policies in determining a clean technological transition. 3 Acknowledgements Authors are listed in random order following the American Economic Association guidelines on random order of co-authors. A consent form has been signed by all co-authors, digital confirmation code: _5DcvWHF8lZK. A previous version of this paper has been presented at the following events: GDS17 study group seminar, Roma Tre University, 9th May 2022; EC-JRC IID seminar series, 16th May 2022; 1st annual workshop of the eco-innovation society, University of Ferrara, 10-11 November 2022; INNOVA MEASURE V Final Workshop, JRC-ISPRA, 18th April 2023; IV workshop on eco-innovation, circular economy, and firm performance, University of Barcelona, 20-21 April 2023. We also thank two anonymous reviewers and the editor of the working paper series for their helpful suggestions. Francesco Rentocchini is currently an employee of the European Commission. The views expressed are purely his personal views and may not in any circumstances be regarded as stating an official position of the European Commission. The usual caveats apply. Authors Francesco Rentocchini: European Commission, JRC Seville and DEMM, University of Milan. Corresponding author at: European Commission, Joint Research Centre (JRC). Office 00/017, Edificio Expo, calle Inca Garcilaso 3, 41092 Sevilla, Spain. E-mail: [email protected] Antonio Vezzani: Rennes School of Business, Department of Strategy and Innovation Sandro Montresor: Gran Sasso Science Institute. The JRC Working Papers on Corporate R&D and Innovation are published under the editorial supervision of Alexander Tübke and James Gavigan in collaboration with Sofia Amaral-Garcia, Fernando Hervás, Koen Jonkers, Francesco Rentocchini at the European Commission – Joint Research Centre, and in cooperation with Sara Amoroso (German Institute for Economic Research, DEU), Michele Cincera (Solvay Brussels School of Economics and Management, Université Libre de Bruxelles, BEL), Alex Coad (Waseda University, Tokyo, JAP), Enrico Santarelli (University of Bologna, ITA), Daniel Vertesy (International Telecommunication Union, CHE – and UNU-MERIT, NLD), Antonio Vezzani (Roma Tre University, ITA); Marco Vivarelli (Università Cattolica del Sacro Cuore, Milan, ITA) and Zoltan Csefalvay (Mathias Corvinus Collegium, HUN). 4 Executive summary This work presents a critical analysis of the role of public policy in the development and adoption of clean technologies, particularly the influence of government-sponsored Research and Development (R&D) programs. The study aims to provide empirical evidence of the effect of public R&D policies on advancing clean technologies with substantial knowledge spillovers. Despite the global commitment to reduce greenhouse gas emissions, the level of public support for clean technologies has been inconsistent across OECD countries since 2011. The development of clean technologies has stagnated, and the private sector's incentive to innovate in this area seems to have decreased. This calls for a reassessment of the effectiveness of R&D policies in promoting the development of clean technologies. The work underscores the need for public action in the economics of climate change. Market-based policies, such as carbon-tax, are not enough as they fail to internalise the long-term benefits of superior, clean technologies. Therefore, R&D subsidies are necessary to redirect innovation from dirty to clean technologies. The study posits that science-push policies, such as public R&D, are expected to have a profound impact on new clean technologies due to their novelty and role as foundational elements for subsequent technological advancements. Public R&D is crucial in promoting the development of high-impact clean technologies, which form the basis of a new technological paradigm. In our empirical investigation, we use patents granted by the USPTO between 2005 and 2015 linked to procurement contracts or research grants with a US funding agency to examine the effect of technology-push policy. The results reveal a significant impact of government-supported clean technologies on subsequent innovations, with supported technologies receiving about 26% more citations than non-supported ones within a 5-year period. This effect was noted among clean technologies with the highest impact on subsequent technological development. This analysis provides two significant implications:  Climate change policy modelling should acknowledge the potential influence of policies on the knowledge spillovers of technologies rather than treating them as exogenous.  In the implementation of climate change policies, R&D support should accompany standard market pull interventions to expedite technical change towards sustainable growth. This argument provides a rationale for reversing the declining trend in technology support policies observed in OECD countries since 2011. In conclusion, this work highlights the critical role of government-sponsored R&D in fostering impactful clean technologies. The findings provide a compelling argument for policy makers to reinforce their R&D programs to achieve sustainable growth and mitigate climate change. 5 1 Introduction The reduction of greenhouse gas emissions is the most important mean to mitigate climate change (IPCC, 2022). Innovation policy packages have enabled cost reductions and supported global adoption of low-emission technologies. However, an important part of emission reduction will depend on new clean technologies that are still embryonic and marked by high technological and market uncertainty (IEA, 2021). This represents an obstacle to private initiatives and makes the support of public policy crucial for their development. To this end, some governments have reinforced their Research and Development (R&D) programs to foster environmental innovation, like in the US with the ARPA-E scheme and in the EU with the Innovation Fund. Nevertheless, the public support to clean technologies is less systematic than it may appear. Across the OECD, from 2011 the (stringency) level of technology support policies has declined until 2016 and has then experienced a scattered increase, but without reaching the 2011-peak (Kruse et al., 2022). This has occurred while the development of clean technologies, as revealed by environmental patents, has stopped growing and embarked along a continuous slow down until the most recent years (Dechezleprêtre and Kruse, 2022; IEA, 2020). Private incentives to develop new clean technologies might have decreased and evidence about the effectiveness of R&D policies in restoring them is thus needed to justify their budgeting. The relevance of public action is a well-recognised intrinsic feature of the economics of climate change (Stern, 2008; Nordhaus, 2019). Among the different leverages, the role of public support to R&D has been less scrutinized compared to market-based and regulatory approaches. Despite its ascertained role in directing technical change towards sustainable growth, a gap remains about the strength of public R&D in playing this role: does it facilitate environmental innovations that act as steppingstones to subsequent technological developments? From a theoretical point of view, a recent stream of endogenous growth models applied to the environment have shown that policy is crucial for the development of clean technologies, given the path-dependent nature of technical change (Acemoglu et al., 2012; 2016; Hémous and Olsen, 2021). A sole market-based policy, such as carbon-tax, is not enough (Acemoglu et al., 2012). As the market keeps on allocating resources to innovation by looking at immediate profits, without retaining the discounted benefits that superior technologies will bring over the long run, the dirty technology sector may remain the first best allocation for incumbents even with a carbon tax. In order to redirect innovation from dirty to clean technologies, R&D subsidies are needed to make the market internalise the higher returns, private and social, of clean technologies in the long run. Following the same background, science-push policies like public R&D can be expected to have a deeper impact on new clean technologies, related to their novelty and their role as basic components for subsequent technological development (Trajtenberg et al., 1997). Unlike marketpull policies, which act on private incentives in the short run, science-push policies increase the expected social value of clean technologies, whose time horizon is longer and provide inventors with incentives to work on more radical innovations. Similarly to what has been found for other technologies (Acemoglu and Linn, 2004; Dranove et al., 2020; Dubois et al., 2015; Finkelstein, 2004), public R&D is arguably needed to spur the development of high impact clean technologies, which represent basic steppingstones in the unfolding of a new technological paradigm. Despite the intuition behind this argument, its supporting theoretical mechanisms have not been fully addressed yet. Furthermore, the extent to which this specific impact of R&D policy actually happens still lacks systematic empirical evidence. To fill this gap, we examine whether government sponsored R&D facilitates the development of clean technologies with large knowledge spillovers on subsequent innovations. Given the pathdependency that characterises technical change (Acemoglu et al., 2012), an important part of the policy impact in fact passes through the (knowledge) value that newly developed green technologies have for the development of subsequent ones. Consistently with the path-dependency hypothesis, 6 should policy induced clean technologies be marked by larger spillovers, they could foster the diffusion of clean knowledge through their influence on subsequent technological developments. We do expect this to happen by referring to firms’ decisions to invest in radically new research projects (Azoulay et al., 2019), which yield innovative outcomes of high impact in terms of knowledge spillovers. These projects are typically risky and early-stage, and are thus marked by marginal costs that overcome their marginal benefits to a larger extent than lower impact projects. This is due to different mechanisms. To start with, high impact innovative outcomes increase the risk of being imitated and this stimulates the innovator to delay their realisation over time (Mukherjee and Pennings, 2004). Furthermore, having a larger impact naturally increases the externalities that early innovators can dynamically have on later innovators, decreasing the returns that the former can appropriate and thus the incentives to undertake the relative investment (Scotchmer, 1991). For these reasons, firms generally find unprofitable to invest in high impact projects and their realisation is thus crucially linked to the public support. By the same token, the marginal technology that government sponsored R&D supports, can be expected to have higher knowledge spillovers than non-supported ones. This differential effect has been found by Azoulay et al. (2019) looking at the impact of scientific grants on firms’ patenting in the pharmaceutical and biotechnology industries. The underlying mechanisms leading to their results are expected to hold also with respect to clean technologies. These are technologies whose development relies on the combination of more diverse and novel technological components than non-clean ones, and which thus require a larger and more uncertain cognitive effort (Barbieri et al., 2020). Furthermore, investments in clean technologies have been proved to yield positive returns to firms only in the presence of high energy costs, thus increasing their market uncertainty (Popp, 2002). This further constrains the incentives for firms to invest in new, high-impact clean technologies. Such technologies, arguably, are more likely to receive support from government-sponsored R&D. We investigate the extent to which this is actually the case, by filling a gap in the empirical research about the development of clean technologies, mainly focused on environmental policies that act on the (private) market side, like: shocks inducing changes in energy prices (Noailly and Smeets, 2015; Hassler et al., 2021), emission trading systems (Calel and Dechezleprêtre 2016), changes in emission standards (Rozendaal and Vollebergh, 2021), international environmental agreements (Dugoua, 2021), and carbon and environmental taxes (Aghion et al., 2016). Empirical analyses of R&D policies are instead more scattered. With respect to the automobile industry, Aghion et al. (2016) showed that temporary R&D subsidies designed to increase energy efficiency can favour the development of incremental clean (grey) innovations, while radically clean innovations remain unaffected. Working on the R&D grants issued by the US Department of Energy, Howell (2017) shows that recipient small businesses in clean energy sectors increase their patenting, VC financing, and survival rate, while these effects are non-significant in conventional (dirty) energy technologies like natural gas and coal. Additional evidence regarding the impact of other types of technology support policies, particularly those focusing on demand-side strategies such as public green procurement, is limited and primarily found in a few select works (Ghisetti 2017; Krieger and Zipperer, 2022). We add to this stream of empirical research focusing on government sponsored R&D in the US and provide evidence of its role in fostering the development of clean technologies with a high impact on subsequent innovations. We rely on patents granted at the USPTO between 2005 and 2015 to applicants linked to at least one procurement contract or research grant with a US funding agency, and investigate the impact of technology-push policy through a quasi-experimental estimation framework. Using citations from other patents to proxy the impact on subsequent technological development, we show that the effect is remarkable in size: in a 5-years window government supported clean technologies have about 26% more citations than non-supported ones. The size of the effect remains sizeable also when we further disentangle citations to consider different types of 13 4 Results 4.1 Core estimations The primary objective of this paper is to identify parameter 𝛽𝛽, which represents the effect of government R&D support on the development of impactful clean technologies. However, before discussing the main results we present estimations for the whole sample, including clean and nonclean technologies. The purpose of this analysis is to investigate whether government R&D support has a differentiated effect between the two in terms of subsequent technological development. Table 2 reports the results of equation (1) using the full sample of patents, including a dummy for clean technologies and its interaction with GvtRD; citations are considered for a time span of 5 years. Results in column 1 do not include applicant fixed effects, which are instead included in column 2. With the inclusion of applicant fixed effects the coefficient attached to government sponsored R&D change sign, from positive to negative. Due to the heterogeneity of the estimation sample in terms of technologies and applicants, we do not think the results should be taken as overall evidence of the effect of government support. Instead, the results suggest that the government might tend to factor in the applicants’ potential when evaluating projects: this confirms that intra-applicant comparisons are a key element to properly assess the effect of government R&D support on subsequent inventions. Table 2 - Effect of government sponsored R&D on subsequent innovation – all technologies. # five-year forward citations (1) (2) Gvt R&D 0.971** -1.134** [0.262] [0.336] Clean tech 4.375** 4.287** [0.261] [0.261] Gvt R&D x clean tech 3.184* 2.278* [1.267] [0.935] originality 6.775** 5.189** [0.266] [0.260] team size 1.896** 1.804** [0.071] [0.071] # of applicants 0.010 -0.074 [0.070] [0.069] Filing year FE Yes Yes 14 Technology FE Yes Yes Applicant FE No Yes F-test 71.030 40.983 R sq 0.047 0.117 N (Patents) 464,123 462,745 The unit of observation is patent application for the full sample. Dependent variable in all columns is the number of five-year forward citations. All estimates are OLS and include technology fixed effect at the 3-digit CPC level. Robust standard errors in parenthesis. + p<0.1, * p<0.05, ** p<0.01. Table 2 confirms previous evidence showing that clean technologies have, on average, a higher impact on subsequent inventions compared to other technologies (Barbieri et al., 2020; Dechezlepretre et al., 2017). The table also points to a higher citation premium for government sponsored clean technologies compared to non-government sponsored ones. Overall, these preliminary estimates reveal that clean technologies behave differently than non-clean ones, also when considering the role of public support. We now focus on clean technologies and present the main results of the paper. Table 3 reports the estimates of equation (1) with respect to clean technologies, showing the absolute and relative effect of government R&D support on subsequent inventions. In column 1 we report the results of a linear specification without using inverse probability weighting scheme and applicant fixed effects. We then report results obtained by applying the inverse-probability weighting (column 2) and by adding applicant fixed effects (column 3). The comparison of the relative Average Treatment Effect (ATE) across columns, reported at the bottom of the table, allows us to evaluate the bias reduction deriving from the enrichment of the estimation approach. Table 3 - Effect of government sponsored R&D on subsequent innovation – clean technologies. # five-year forward citations (1) (2) (3) Gvt R&D 6.472** 6.107** 4.427** [1.346] [1.341] [1.109] originality 26.221** 32.572** 16.635** [1.409] [2.322] [1.819] Team size 4.272** 5.103** 3.520** [0.384] [0.639] [0.420] # of applicants -0.951** -1.420* -0.864* [0.362] [0.589] [0.436] 15 Filing year FE Yes Yes Yes Technology FE Yes Yes Yes Applicant FE No No Yes Weighting scheme No IPW IPW F test 29.197 20.829 16.011 Relative ATE 0.394** 0.360** 0.258** [0.084] [0.082] [0.066] R sq 0.053 0.052 0.421 N (Patents) 38,729 36,726 36,245 The unit of observation is patent application for the sample of cleantech patents. Dependent variable in all columns is the number of five-year forward citations. Relative ATEs are computed as the relative difference between potential outcome means for treated and untreated groups (∑𝑦𝑦𝚤𝚤 � 𝑁𝑁𝑡𝑡 𝑖𝑖𝑁𝑁𝑡𝑡 �-∑𝑦𝑦𝚤𝚤 � 𝑁𝑁𝑢𝑢 𝑖𝑖𝑁𝑁𝑢𝑢 �)/ ∑𝑦𝑦𝚤𝚤 � 𝑁𝑁𝑢𝑢 𝑖𝑖𝑁𝑁𝑢𝑢 � where 𝑦𝑦𝚤𝚤 � is the predicted value from the relevant regression model. Columns 2 and 3 are estimated using 𝑝𝑝(𝑥𝑥𝑖𝑖)/(1 − 𝑝𝑝(𝑥𝑥𝑖𝑖)) to weight untreated observations and 1 otherwise. 𝑝𝑝(𝑥𝑥𝑖𝑖) is the propensity score calculated as per Table B.2 in appendix B. Figures B.1 and B.2 in Appendix B report statistics relative to the propensity score procedure used to compute weights, showing a good performance in terms of bias reduction. All estimates are OLS and include technology fixed effect at the 4-digit CPC level. Robust standard errors in parenthesis. + p<0.1, * p<0.05, ** p<0.01. The effect of government R&D support is high and significant in all the specifications. Interestingly, the bias reduction is particularly strong when including applicant fixed effects: the use of propensity score lowers the relative ATE (the ATE in terms of potential outcome) of public support by 5.7 percentage points, while when adding the fixed effects this reduces by an additional 10.7 percentage points. According to our preferred specification (column 3), government supported clean technologies have about 26% more citations than non-supported ones within a 5-year window. In line with expectations, patents with a higher originality and a larger team size receive on average more citations, while somehow unexpectedly patents collaborated between more applicants show a lower number of citations. This last result may suggest that applicants tend to develop their more promising R&D projects (in terms of potential future impact) alone or in small collaborative settings. 4.2 Disentangling the spillovers of government sponsored R&D In this section, we further disentangle the impact of government support in the development of impactful clean technologies by considering different types of knowledge spillovers. First, we consider the knowledge spillovers generated by public R&D support outside the sphere of the focal applicant, by excluding those generated by its own citations. To clean our dependent variable from self-citations we use two different approaches: i) we eliminate a citation only if it is completely determined by the same applicant of the cited patent (noself_nostric); ii) we eliminate a citation if at least one applicant in the citing patent is also in the cited patent (noself_strict). In the former case, we rule out only sharp “intra-applicant” spillovers, but still allow for knowledge spillovers deriving from collaborations of the same applicant in subsequent projects (if a patent with applicants A and B cite a patent of applicant A, we still count the citation). In the second case, we capture pure knowledge spillovers, i.e. inter-applicant spillovers, because none of the citing 16 applicants should be involved in the development of the cited patent. The results are reported in columns (1) and (2) of Table 4. Second, we focus on the knowledge spillovers that flow from the applicants that have received government R&D support to the rest of the world. To do this, we exclude from our dependent variable all the citations from patents registered by applicants that have received at least one US public contract as recorded by the 3PFL. Like in the case of self-citations, we compute these spillovers using a non-strict and a strict definition; results are reported in columns (3) and (4). Finally, we assess whether government R&D support helps generate clean technologies with higher international knowledge spillovers, thus having a higher impact on the subsequent technological development in other economies. To do so, we build two variables: i) cit_US, a binary variable taking value 1 is all citations of a given patent are generated only by US applicants (column 5); ii) geo_breadth, counting the number of applicant’s countries of the citing patents (column 6). The results reported in Table 4 suggest that spillovers effects are rather strong and not localized among the applicants receiving public support. In fact, the coefficient attached to government sponsored R&D is statistically significant at the usual level in columns 1 to 4. The magnitude of the coefficient decreases when using the stricter definition of spillovers, which is consistent with the reduced number of citations considered. However, the relative average treatment effects do not vary substantially from the main estimations (see Table 3, column 3), confirming the sizeable effect of government support on follow up inventions. Table 4 - Effect of government sponsored R&D on different kinds of knowledge spillovers, clean technologies (1) (2) (3) (4) (5) (6) noself nostrict noself strict out nostrict out strict cit US geo breadth Gvt R&D 4.376** 3.151** 4.350** 2.176** 0.006 0.137+ [1.017] [0.866] [1.051] [0.672] [0.007] [0.073] originality 15.795** 12.856** 16.104** 9.132** 0.171** 1.453** [1.764] [1.655] [1.800] [1.241] [0.018] [0.119] Team size 3.049** 2.773** 3.259** 2.110** 0.010** 0.238** [0.370] [0.360] [0.389] [0.262] [0.002] [0.019] # of applicants -0.692+ -0.824* -0.783+ -0.562* - 0.003* -0.051** [0.380] [0.372] [0.401] [0.268] [0.002] [0.019] Filing year FE Yes Yes Yes Yes Yes Yes Technology FE Yes Yes Yes Yes Yes Yes Applicant FE Yes Yes Yes Yes Yes Yes 17 Weighting scheme IPW IPW IPW IPW IPW IPW Relative ATE 0.286** 0.256** 0.274** 0.230** 0.007** 0.041+ [0.067] [0.071] [0.067] [0.072] [0.009] [0.022] N (Patents) 36,245 36,245 36,245 36,245 36,245 36,245 The unit of observation is patent application for the sample of cleantech patents. Relative ATEs are computed as relative difference between potential outcome means for treated and untreated groups (∑𝑦𝑦𝚤𝚤 � 𝑁𝑁𝑡𝑡 𝑖𝑖𝑁𝑁𝑡𝑡 �-∑𝑦𝑦𝚤𝚤 � 𝑁𝑁𝑢𝑢 𝑖𝑖𝑁𝑁𝑢𝑢 �)/ ∑𝑦𝑦𝚤𝚤 � 𝑁𝑁𝑢𝑢 𝑖𝑖𝑁𝑁𝑢𝑢 � where 𝑦𝑦𝚤𝚤 � is the predicted value from the relevant regression model. All columns are estimated using 𝑝𝑝(𝑥𝑥𝑖𝑖)/(1 − 𝑝𝑝(𝑥𝑥𝑖𝑖))to weight untreated observations and 1 otherwise. 𝑝𝑝(𝑥𝑥𝑖𝑖) is the propensity score calculated as per Table B2 in Appendix B. For the regression adjustment via propensity score, we enforce a common support by removing the 5% of the treatment observations at which the propensity score density of the control observations is at a minimum. All estimates are OLS and include technology fixed effect at the 4-digit CPC level. Robust standard errors in parenthesis. + p<0.1, * p<0.05, ** p<0.01. Results reported in columns 5 and 6 show that the spillovers generated by government supported clean patents are not more localized within the US than privately funded ones but have a higher geographical breadth: the technology is further developed by applicants from a larger basket of countries. This suggests that spillovers from government supported clean technologies may differ from those in other areas of intervention. Moreover, in line with models of growth emphasizing the strategic complementarities in clean research across countries (Aghion et al., 2015; Dechezlepretre et al., 2017), the larger geographical scope of citations is consistent with the idea that policies supporting the development of clean technologies can have an effect in fostering the global generation of such technologies. 4.3 Robustness tests In this section we present a battery of robustness checks to further corroborate our main results. First, while in our analysis we control for different types of spillovers and for observable differences in potential outcomes, with the data at stake we are not able to allocate budget at the patent level. The budget of government supported projects is not homogeneous and can be allocated to different activities other than R&D for technological development; moreover, there are cases of multiple patents linked to the same government supported project. This mean that the treatment might be not homogeneous, and we cannot directly model it. To assess whether the heterogeneity in the government R&D support can be an issue when trying to identify its effect on follow up citations, we run two ancillary regressions at the applicant level. In the first, we assess whether higher amounts of government support induce more clean tech patents. In the second, we assess whether higher amounts of government support lead to more citations to the overall portfolio of supported patents once controlling for their number. If the coefficient attached to government support is statistically significant in the former but not in the latter, this would be indirect evidence that the heterogeneity in the treatment is not a major issue when assessing its effect on the impact of supported clean tech on follow up innovations. The results reported in Table 5 suggest that this is the case. Higher amounts of government R&D support lead to a higher number of clean tech patents, but once controlling for clean tech patents it has not effect on the number of citations received. 18 Table 5: The effect of government sponsored R&D on the quantity and quality of inventions - cleantech sample (1) # green Gvt patents (2) # citations to green patents Amount of Gvt R&D 0.082** -1.497 [0.028] [0.996] # green Gvt patents 31.500** [11.378] Avg. team size 0.128 37.426** [0.123] [11.398] Avg. # of applicants 0.098 -0.014 [0.089] [10.099] Avg. originality -0.165 170.284 [1.185] [88.706] Constant 2.483* -260.113* [1.197] [103.976] R sq 0.320 0.328 N (applicants) 868 868 Regressions are based on the 868 applicants that received government sponsored R&D and developed green technologies. The dependent variable in column 1 is the number of green patents sponsored by public R&D and in column 2 the number of citations received by green government R&D sponsored patents. Gvt R&D amount is the total amount received by the applicant. Estimates in both columns include controls for the average number of inventors, average number of applicants, the average originality index and the share of patents in different cleantech technological classes. Column 2 includes also the number of number of green patents sponsored by public R&D. + p<0.1, * p<0.05, ** p<0.01 Second, we run a randomised falsification test by randomly assigning treatment across our sample of clean tech patents while keeping the share of treated patents constant. We build 100 different replications of our favourite specification (Column 3 of Table 3) to assess whether a placebo government support would still exert a positive effect on forward citations. Figure 2 displays the coefficients attached to the placebo government R&D support for the 100 replications: the figure shows that we cannot reject the hypothesis that the coefficient is equal to zero, providing further evidence supporting our results. 19 Figure 2: Randomised falsification test In the figure are reported the coefficients attached to GVT R&D (equation 1) when treatment is randomly distributed across clean patents. As in our preferred specification (Table 3, column 3), the estimations include applicant fixed effects and probability weights. Third, we re-run our favourite specification by weighting observations by the inverse of the size of applicants’ clean tech patent portfolios. In our sample, the number of applicants’ observations is proportional to the number of their clean tech patents; giving the same weight to each applicant, this reweighting scheme ensures that results are not driven by applicants with larger portfolios. The results, reported in Table 6 column 1, confirm the positive effect of government support. Table 2: Effect of green government sponsored R&D on subsequent innovation – robustness checks # five-year forward citations # seven-year forward citations # five-year forward citations (1) (2) (3) (4) (5) (6) Gvt sponsored R&D 6.593* 6.257** 6.411** 0.233** 3.733** 5.467* [2.872] [1.232] [1.441] [0.048] [1.091] [2.436] originality index 32.189** 9.774** 20.327** 1.435** 16.433** 18.574** [3.647] [1.287] [2.112] [0.134] [1.805] [4.651] 20 Team size 6.222** 1.961** 4.491** 0.127** 3.446** 1.671** [1.071] [0.254] [0.471] [0.009] [0.411] [0.349] # of applicants 0.056 -0.296 -1.078* -0.014 -0.900* 0.236 [1.193] [0.271] [0.478] [0.010] [0.424] [0.316] Filing year FE Yes Yes Yes Yes Yes Yes Technology FE Yes Yes Yes Yes Yes Yes Applicant FE No Yes Yes Yes Yes Yes Filing year X Tech FE No No No No Yes No Model OLS OLS OLS Poisson OLS OLS Weighting scheme GPAT IPW IPW IPW IPW IPW SEs Robust Robust Robust Robust Robust CPC-Year F test 5.826 15.888 17.994 . 52.906 5.103 Rel ATE 0.358* 0.445** 0.284** 0.262** 0.216** 0.345* Rel ATE SE [0.160] [0.090] [0.065] [0.060] [0.065] [0.155] R sq 0.085 0.287 0.423 . 0.425 0.424 N (Patents) 36,245 25,025 36,245 36,238 36,244 31,058 The unit of observation is patent application for the sample of cleantech patents. Dependent variables are the number of five-year forward citations in columns 1, 2, 4, 5 and 6 and the number of seven-year forward citations in column 3. Relative ATEs are computed as relative difference between potential outcome means for treated and untreated groups (∑𝑦𝑦𝚤𝚤 � 𝑁𝑁𝑡𝑡 𝑖𝑖𝑁𝑁𝑡𝑡 �- ∑𝑦𝑦𝚤𝚤 � 𝑁𝑁𝑢𝑢 𝑖𝑖𝑁𝑁𝑢𝑢 �)/ ∑𝑦𝑦𝚤𝚤 � 𝑁𝑁𝑢𝑢 𝑖𝑖𝑁𝑁𝑢𝑢 � where 𝑦𝑦𝚤𝚤 � is the predicted value from the relevant regression model. Column 1 regression is weighted by the inverse of the number of cleatech patents in the applicant portfolio. Columns 2 and 3 are estimated using 𝑝𝑝(𝑥𝑥𝑖𝑖)/(1 − 𝑝𝑝(𝑥𝑥𝑖𝑖))to weight untreated observations and 1 otherwise. 𝑝𝑝(𝑥𝑥𝑖𝑖) is the propensity score calculated as per Table B2 in Appendix B. For the regression adjustment via propensity score, we enforce a common support by removing the 5% of the treatment observations at which the propensity score density of the control observations is at a minimum. All estimates are OLS and include technology fixed effect at the 4-digit CPC level. Robust standard errors in parenthesis. + p<0.1, * p<0.05, ** p<0.01. Fourth, we test the robustness of our results against a more conservative definition of clean technologies. We build upon Dechezleprêtre et al. (2021) and define clean patents using 4-digit CPC technological classes Y02B, Y02C, Y02E and Y02T.2 Results from the estimation on this sub-set of clean technologies, reported in Table 6 column 2, still confirm our main results and possibly magnifies the effect of government support in terms of relative ATE. Column 3 in Table 6 further shows the results when using a 7-years window in the computation of the number of forward citations. While the coefficient attached to government support is higher than that for the 5-years window, the effect of government support in terms of relative ATE is in line with our main results. 2 These four CPC codes groups technologies related to buildings, to GHG capture and storage, to reduction of GHG in energy production and distribution, and to transportation. In other words, from our definition of clean tech we drop patents with codes: Y02A, Y02D, Y02P, Y02W and Y04S (see figure 1 for short labels). 21 Column 4 in Table 6 presents the results for a (pseudo-)Poisson regression model with multiple high-dimensional fixed effects (Correia, 2020), ensuring the robustness of our findings to the count data format of our dependent variable (number of citations). Column 5 incorporates fixed effects for each combination of patent filing year and technological class (CPC at the subclass level, e.g. Y02A). Finally, Column 6 reports results with standard errors clustered at the same level (patent filing year-technological class level).3 Reassuringly, all robustness checks produce outcomes consistent with our baseline results. 4.4 The distributional effects of government sponsored R&D In this section, we assess the distributional effects of R&D government support on follow up innovations. To do so, we implement the recentered influence function (RIF) of the unconditional quantile (Firpo et al., 2009), to evaluate the impact of changes in the non-treated to treated (government supported) status on quantiles of the marginal distribution of forward citations. Previous results could be eventually interpreted as follows: the government is able to systematically define/select technological needs/solutions with a potential higher impact on follow up innovations, compared to other actors in the economy. Despite lock-in effects and the possible local search performed by private actors, this is a rather strong result to be put forward. One can instead expect that in most cases the government support will not have a real effect. Conversely, it can be expected to incentivise research in research areas with lower expected private returns in the short term and thus increase the probability that some key technologies are developed. In other words, the (average) impact of government R&D support discussed above may be driven by the highly cited patents in the sample. Therefore, the relevance of assessing the distributional aspects of the effect found in the previous sections derives from the assumptions underlying average effects and the interpretation of the relative results. Figure 3 shows the distribution of citations to clean tech patents comparing the group of patents resulting from projects supported by the government with that not originated by government supported R&D projects. The figure shows that the distribution of patents citations is rather skewed and suggests that differences between government supported and not supported patents are not constant along the distribution. At the bottom of the citation distribution supported and not supported patents do not differ much, with the latter showing a slightly higher fraction of patents. Instead, government sponsored cleantech patents are more frequent in the upper tail of the citation distribution, suggesting that the effect of government support operates through the development of most impactful clean technologies. 3 Recent discussions in the econometrics literature suggest that clustered standard errors may be overly conservative, and the appropriate level of clustering should be carefully chosen (Abadie et al., 2023). In our study, it is not advisable to cluster standard errors at the applicant level, due to the high number of clusters (1,400), many of which have few observations. Consequently, we have decided to cluster standard errors at the technology-year level, aligning with the evidence of a non-random distribution of government support at the technology level (see Figure 1) and potential yearly differences, resulting in 144 clusters. However, we had to exclude 5,187 patents because they were assigned to more than one 4-digit CPC code. For this reason, we do not apply clustered standard errors in all specifications. 22 Figure 3: Fraction of cleantech patents for government sponsored and non-sponsored patents, by number of citations. Table 7 reports the results of the RIF estimations on the quantiles of the citations distributions. Consistently with the argument and the descriptive evidence above, the effect of the government R&D support is visible only among the third and fourth quintile of the citations’ distribution, where it is quantifiable in about 8.1% and 8.6% citations with respect to the respective potential outcome. Interestingly, also the coefficients attached to originality sizably increases along the quintile of the distributions, confirming the general result that more original (and more risky) clean patents may have a higher impact on follow up innovation. Table 3: Effect of government sponsored R&D on subsequent innovation, distributional effects via RIF regressions for clean tech (1) (2) (3) (4) 20th quantile 40th quantile 60th quantile 80th quantile Gvt R&D 0.160 0.264 0.722* 1.610* [0.101] [0.177] [0.343] [0.821] originality 1.827** 3.120** 5.982** 15.097** [0.220] [0.311] [0.538] [1.320] 29 List of tables Table 1: Summary statistics .......................................................................................... 8 Table 2 - Effect of government sponsored R&D on subsequent innovation – all technologies. ................13 Table 3 - Effect of government sponsored R&D on subsequent innovation – clean technologies. ............14 Table 4 - Effect of government sponsored R&D on different kinds of knowledge spillovers, clean technologies ..........................................................................................................................16 Table 5: The effect of government sponsored R&D on the quantity and quality of inventions - cleantech sample ..................................................................................................................18 Table 6: Effect of green government sponsored R&D on subsequent innovation – robustness checks .......19 Table 7: Effect of government sponsored R&D on subsequent innovation, distributional effects via RIF regressions for clean tech ............................................................................................22 Table A.1: Correspondence between number of contractors in 3PFL and number of applicants in Patstat for clean tech patents. ....................................................................................................31 Table B.1: Effect of green government sponsored R&D on subsequent clean innovation – doubly robust estimators (AIPW and IPWRA) ........................................................................................33 Table B.2: Selection into green government sponsored R&D – propensity score. ...............................35 Table B.3: Effect of green government sponsored R&D on subsequent innovation – alternative weights for regression adjustment ................................................................................................40 30 Annexes 31 A Data Construction From the 3PFL database (de Rassenfosse et al., 2019) we retrieve information on procurement contracts and research grants signed by the US government, and on the patents filed to protect the resulting inventions. The 3PFL database comprises information for 37,925 patents granted by the USPTO between 2005 and 2015. In order to create the control group of non-treated patents we proceeded as follows: i) The 3PFL does not provide a code that can be directly used in Patstat to identify the patent applicants. Therefore, we have used the patent_nr field (the number that identifies the publication of the granted application at the USPTO) from the 3PFL database to retrieve all the person identifiers associated to 3PFL patents from Patstat. This reduces the sample of patents to 37,003 as we excluded patents where the inventor was also the applicant, and no other applicant was reported in PATSTAT. Table A.1 shows, for each 3PFL patent the number of applicants in Patstat and the number of contractors in 3PFL. In most cases the correspondence was 1:1 and the applicant-contractor pair directly identified. Table A.1: Correspondence between number of contractors in 3PFL and number of applicants in Patstat for clean tech patents. Number of contractors in 3PFL 1 2 3 4 5 6 7 8 9 Number of applicants in PATSTAT 1 30,213 3,002 332 40 13 2 1 1 2 2,309 483 93 24 1 1 3 303 70 28 2 1 4 55 13 2 1 5 10 3 6 7 1 Note: the selection excluded patents where the inventor was also the applicant and no other applicant was reported. ii) In all the cases where the matching did not result in a 1:1 or 1:many correspondence we have disambiguated the matching using the names of applicants and contractors in the two databases. 32 In all the cases where the number of applicants and contractors was the same (first row and first column of Table A.1) we associated the relative entries using the Levenshtein distance, which provided a measure of similarity between the names reported in the two databases. All other cases have been manually disambiguated. As a result, we created a correspondence table between applicant codes (person_id) in Patstat and the contractor identifier reported in 3PFL. iii) We used the disambiguated list of person_id to retrieve all the patents granted to the same applicant over the 2005-2015 period. These patents represent the control group. 33 B. Robustness Checks and Specifications Misspecification of outcome and treatment models: doubly robust estimators To assess the robustness of our findings to possible issues related to misspecification of the selection into treatment model we rely on the augmented inverse-probability-weighted (AIPW) and the inverse-probability-weighted regression-adjustment (IPWRA) estimators. While the IPW estimator used in the main analysis models only the treatment probability, the AIPW estimator model both the outcome and the treatment probability. The advantage of the AIPW is that it is enough that only one of the two models is correctly specified to consistently estimate the treatment effect; for this reason, this type of estimator is known as a property known as being. The AIPW estimator includes an augmentation term that corrects the estimator when the treatment model is incorrect. This augmentation term vanishes when the treatment is properly specified, and the sample size is large. Similarly, inverse-probability-weighted regression-adjustment (IPWRA) estimators integrate models for the outcome and treatment status and possess the double robustness property. IPWRA estimators utilize the inverse of the estimated treatment-probability weights to estimate regression coefficients that correct for missing data. These coefficients are then used to calculate potential outcome means (Wooldridge, 2010). To the best of our knowledge, there is no literature that compares the relative efficiency of AIPW and IPWRA estimators, so we report results from both approaches in the table below. Results show high and significant coefficients. As expected, coefficients are higher than in our favorite specification (Column 3 in Table 3 in the main text) as AIPW and IPWRA estimators do not allow to control for the applicant fixed effects. Table B.1: Effect of green government sponsored R&D on subsequent clean innovation – doubly robust estimators (AIPW and IPWRA) (1) (2) Gvt R&D 9.074** 9.250** [1.801] [1.761] Controls Yes Yes Filing year FE Yes Yes Technology FE Yes Yes Applicant FE No No Estimator AIPW IPWRA 34 Relative ATE 0.566** 0.577** [0.114] [0.111] N (Patents) 36,726 36,726 + p<0.1, * p<0.05, ** p<0.01 35 Results from the IPW treatment model and alternative weights Table B.2: Selection into green government sponsored R&D – propensity score. (1) originality 0.458** [0.065] Team size 0.015** [0.006] # of applicants 0.014* [0.006] CC adaptation (Y02A) 0.591** [0.033] CCMT buildings (Y02B) -0.158** [0.047] GHG capture (Y02C) 0.040 [0.057] CCMT ICT (Y02D) -0.763** [0.051] GHG energy (Y02E) 0.297** [0.027] CCMT production (Y02P) -0.007 [0.027] CCMT transport (Y02T) -0.281** [0.030] CCMT waste (Y02W) 0.135+ [0.072] ICT for energy (Y04S) -0.320** 36 [0.069] Filing year FE Yes Technology FE Yes Chi2 test 1643.711 McFadden's R sq 0.070 N (Patents) 38657 + p<0.1, * p<0.05, ** p<0.01 37 Figure B.1: Variance ratio of residuals vs bias before and after matching 38 Figure B.2: Covariate bias before and after matching We test the robustness of our results to alternative weights used for regression adjustment. First, we implement two differently defined weights coming from the propensity score calculation. The first weight rebalances the treated group only: it takes value (1 − 𝑝𝑝𝚤𝚤 �)/𝑝𝑝 for the patents having received government R&D support and 1 otherwise (with 𝑝𝑝𝚤𝚤 �being the fitted value from Table B2 above. The second weight is a standard inverse probability weight taking value 1/𝑝𝑝 for cleantech patents receiving government R&D support and 1/(1 − 𝑝𝑝) otherwise. Following existing work (Hirano el al., 2003; Brunell and Di Nardo, 2004), both weights are computed preserving proportions between the treated and untreated group. Finally, we compute weights from a coarsened matching procedure (Iacus et al., 2012). Figures B.3 reports comparison between treated and untreated patents in relation to global and local imbalance measures. The global imbalance statistic is calculated as the local imbalance measures difference between the multidimensional histogram of pretreatment covariates in the treated group and the same in the control group. In our specific case, the value of 0.672 is the reference point for the unmatched data, and a decrease in the value after matching (0.597) indicates a reduction in the level of imbalance. Similar reductions in the local imbalance measures are found for the individual variables. Figure B.4 provides a comparison of variable means before and after matching and shows an important reduction in bias following the matching procedure. Overall, the two figures reassure us about the ability of the chosen approach to reduce bias from observables.