Copyright by Xue Gao 2019
The Dissertation Committee for Xue Gao Certifies that this is the approved version of the following Dissertation:
Essays on Innovation Policy, Knowledge Networks, and Cost Reductions in Deployment-Related Technologies in the Solar PV Industry
Committee:
Varun Rai, Supervisor
Sheila Olmstead
Kenneth Flamm
Gregory Nemet
Essays on Innovation Policy, Knowledge Networks, and Cost Reductions in Deployment-Related Technologies in the Solar PV Industry
by
Xue Gao
Dissertation
Presented to the Faculty of the Graduate School of The University of Texas at Austin in Partial Fulfillment of the Requirements for the Degree of
Doctor of Philosophy
The University of Texas at Austin August 2019
Dedicated to my parents, my husband Fei, and our daughter Adeline. Acknowledgements
I would like to express the deepest appreciation to my committee chair Dr. Varun Rai, who gave me continuous guidance and unlimited support to help me go through all the difficulties I came across throughout my Ph.D. career. He encouraged me to explore big questions and learn cutting-edge methodologies. He guided me on how to write and publish academic papers. He supported me to present our research in front of experts at academic conferences. More importantly, he always had confidence in me and never doubted my abilities. Under his supervision and support, I gradually built my confidence in being a good researcher and teacher. Without his support, I could not have done what I was able to do.
I would also like to extend my sincere gratitude to my committee members, Dr. Sheila Olmstead, Dr. Kenneth Flamm, and Dr. Gregory Nemet, for their valuable advice and unwavering support. It is very fortunate of me to have them as my committee members, who have always been generous with their time, ideas, support, and recommendation letters. I am also grateful to Dr. Joshua Busby, who was always eager to help. He was the most influential person in developing my qualitative skills. I have greatly appreciated his time, support, guidance, and goodwill throughout my job search. I also want to thank Dr. David Eaton for providing me with encouragement and guidance throughout the duration of the title insurance project.
I’m extremely grateful to Dr. Victoria Rodriguez. She is a motherly figure for me and helped me go through many difficulties I had during my Ph.D. journey. I would also
v like to thank the professors at the LBJ school who have taught and helped me: Dr. Peter Ward, Dr. Bill Spelman, Dr. Chandler Stolp, Dr. Paul von Hippel, and Dr. Carolyn Heinrich. They were always there to provide helpful advice, unparalleled support, and academic resources.
Special thanks to outstanding members of the Rai group, Cale Reeves, Ariane Beck, Mark Hand, Erik Funkhouser, Noorah Alhasan, Ashok Sekar, Vivek Shastry, Tiffany Wu, and Sindhu Maiyya. They have been friends, colleagues, and mentors whom I have treasured so much. I have greatly appreciated their professionalism and support as we worked together. The pleasant memories of weekly group lunch and many discussions and chats with them will be hard to forget.
I must thank collogues at the LBJ school, Cale Reeves, Francisca Bogolasky, Jodi Rosenstein, Nisha Krishnan, Miha Vindis, Noah Dust, Ilse Oehler, Abby Lane, Santiago Tellez, Christina Caramanis, Jiameng Zheng, Mark Hand, Rachael Singer, Selena Caldera, Ashlyn Webb, Diana Bolsinger, Eun Young Kim, Alfonso Alvarez, Julian Plough, and Regina Buono. Being part of the LBJ school has given me many pleasant memories to cherish. I want to thank them for their kindnesses and unwavering support.
Finally, to thank the people who shaped me into the person I am. To my parents, I’m extremely grateful that you’ve always inspired me to pursue my dream, encouraged my ambitions, stood behind me, provided continued support, and guided me in the right direction. I salute you for the selfless love and sacrifice you did for me. You mean the world to me, and I will never be able to thank you enough. To my husband Fei, you were always there for me without a doubt. You provided me with endless love, support, and vi companionship throughout the whole process. I am blessed to have you by my side. To my beloved daughter Adeline, I am forever grateful to have you in my life. Thank you, my sweet girl, for joining us when I started writing the dissertation. Thank you for giving me immense happiness and for motivating me to keep reaching for excellence.
vii Abstract
Essays on Innovation Policy, Knowledge Networks, and Cost Reductions in Deployment-Related Technologies in the Solar PV Industry
Xue Gao, Ph.D. The University of Texas at Austin, 2019
Supervisor: Varun Rai
Abstract: Deployment policies have stimulated a large number of renewable energy investment and deployment on a global scale, but they still need to be scaled up at least six times faster to keep global warming below hazardous levels. Therefore, particular emphasis should be placed on understanding the impact of deployment policy on technological innovations and cost reductions, because being more cost-effective and efficient can further facilitate scaling-up. The overarching aim of the dissertation is to understand whether and how deployment policy can lead to technological changes and cost reductions. To be specific, the first essay focuses on whether and how demand-pull policy can impact technological innovation. The second essay explores why deployment policy is essential to promote technological innovations through studying firms’ knowledge acquisition behaviors when conducting innovations. The third essay investigates three primary learning mechanisms through which deployment policy can facilitate cost reductions and contributes to separating these three mechanisms, namely, learning by doing, learning by searching, and learning by interacting. The particular focus of this dissertation is the role of geography. The importance of geography as an element of technological learning and viii innovation continues to be debated by researchers. The salience of this topic is ever greater today, when governments across the world are struggling to balance the opposing pulls of inward-looking, protective approaches to economic growth (e.g., “local jobs” and “preferential procurement”) and outward-looking, collaborative, and open approaches to economic growth (e.g., “open innovation” and “imported talent”). Additionally, this dissertation is related to a broader question, which also sums up my main research topic: what combination of innovation ecosystem features – including, local deployment policies, technology characteristics, social and human capital, demand factors, and supply conditions -- translate into persistent localized innovation and economic benefits.
ix Table of Contents
List of Tables ...... xiv
List of Figures ...... xvii
Chapter 1: Introduction ...... 1
Background ...... 1
Research objectives ...... 4
Empirical focus and research questions ...... 6
Brief summary of each essay ...... 8
Chapter 2: Relevant literature ...... 13
Technological innovation and innovation policy ...... 13
Technological learning ...... 20
Why and how geography matters in innovation ...... 22
Soft costs reductions and technological innovations in solar energy ...... 26
Patent data and citation analysis ...... 29
Chapter 3: Local demand-pull policy and energy innovation: Evidence from the solar photovoltaic market in China ...... 33
Introduction ...... 33
Background and related literature ...... 36
Related literature ...... 36
Background on distributed PV policies in China ...... 39
Data and methodology ...... 42
Data ...... 42
Measurement ...... 45
Empirical models ...... 47 x Omitted variables and potential endogeneity ...... 51
Results ...... 51
Descriptive analysis of PV BOS patents in China ...... 51
Regression model results ...... 55
Discussion and limitations ...... 61
Conclusion and policy implications ...... 63
Chapter 4: Does Geography Matter? A Study of Technological Learning and Innovation in the Solar Photovoltaic Balance-of-Systems (PV BOS) Industry ...... 68
Introduction ...... 68
Theoretical background ...... 70
Data and methodology ...... 74
Data ...... 74
Methodology ...... 76
Gravity model ...... 76
T-tests ...... 78
Regression model ...... 79
Results and discussion ...... 82
Descriptive analysis ...... 82
Results and discussions ...... 87
Distance and border effect on citation networks ...... 87
T-test results ...... 90
The effect of local/nonlocal knowledge on patent quality ...... 92
Conclusion and policy implications ...... 97
xi Chapter 5: Separating multiple learning mechanisms that facilitate cost reductions: Evidence from the U.S. solar PV industry ...... 101
Introduction ...... 101
Theoretical background ...... 104
Learning curve and learning mechanism ...... 104
Supply chain collaboration ...... 106
Data and methodology ...... 108
Data ...... 108
Econometric models ...... 110
Results and discussion ...... 117
Network characteristics ...... 117
Regression results ...... 124
Conclusion and policy implications ...... 132
Chapter 6: Conclusion ...... 137
Contributions to the literature ...... 137
Policy implications ...... 139
Future directions ...... 141
Appendice ...... 143
Appendix A: Supplemental information for Chapter 3 ...... 143
Robustness check ...... 143
Summary statistics ...... 153
Three ways of grouping the 31 provinces in China ...... 155
Specific items for PV BOS classification ...... 157
Appendix B: An analysis of US BOS innovations ...... 158
xii Appendix C: Supplemental information for Chapter 4 ...... 164
Robustness check ...... 164
Summary statistics ...... 169
Appendix D: Supplemental information for Chapter 5 ...... 170
Robustness check: excluding vertically integrated installers ...... 170
Robustness check: Instrumental variables ...... 175
Summary statistics ...... 177
References ...... 178
xiii List of Tables
Table 1: Summary of relevant empirical studies to test whether the location of
demand matters for inducing innovation...... 38 Table 2: Search terms for PV BOS patents in Chinese and English...... 44 Table 3: Regression results for Equations 1 and 2 using different estimation methods. ...56 Table 4: Regression results for different grouping methods (Equation 3), based on competitiveness index according to the city ranking in the “Blue book of
China’s provincial competitiveness,” geography, and GDP...... 57 Table 5: Regression results for differentiation between inverter and non-inverter
patents. Note that Model 7 is the same as Model 2...... 59 Table 6: Regression results for including only inventions in the dependent variable...... 60 Table 7: The specific items and keywords for PV BOS patent data collection ...... 75 Table 8: Comparison between the percentage of knowledge source and percentage of
knowledge diffusion of U.S. PV BOS patents across different countries ....85 Table 9: Comparison between the percentage of knowledge source and percentage of knowledge diffusion of U.S. PV BOS patents across different states in
the U.S...... 86 Table 10: Regression results of the gravity model ...... 89 Table 11: T-test results of the proportion of total backward citations (added by both applicants and examiners) that are in the same country as the focal
patents for different categories ...... 91 Table 12: T-test results of the proportion of total backward citations (added by both applicants and examiners) that are in the same state as the focal patents
for different categories ...... 91
xiv Table 13: T-test results of the proportion of backward citations (only added by applicants) that are in the same country as the focal patents for different
categories ...... 92 Table 14: T-tests results of the proportion of the backward citations (only added by applicants) that are in the same state as the focal patents for different
categories ...... 92 Table 15: Regression results of equation 3 and equation 4 using different windows for
receiving forward citations (exclude self-citations) ...... 95 Table 16: Regression results of equation 5 using different windows for receiving
forward citations (exclude self-citations) ...... 96 Table 17: The specific items and keywords for PV BOS patent data collection ...... 109 Table 18: The annual number of installers, panel manufacturers, and inverter
manufacturers ...... 119 Table 19: The regression results of the effect of learning by doing, learning by
searching, and learning by interacting ...... 128 Table 20: The regression results of including total installed capacity, market share
and competition ...... 131 Table A1: Regression results of including global demand ...... 144 Table A2: Regression results of excluding colinear independent variable ...... 146 Table A3: Regression results of including utility-scale installed capacity ...... 149 Table A4: Regression results of equation 1 and 2 using Poisson regression ...... 150 Table A5: Regression results of equation 3 using Poisson regression ...... 151 Table A6: Regression results for including only inventions in the dependent variable
and differentiating inverters and non-inverters using Poisson regressions .152 Table A7: Summary statistics for key province-level variables...... 153 xv Table A8: Correlation table for independent variables ...... 154 Table A9: Scope of the technologies included within each of the four main BOS
components analyzed in the main chapter...... 157 Table B1: Data sources and descriptions of dependent variables ...... 162 Table B2: Regression results for the equation1 using different year lags ...... 163 Table C1: Regression results of equation 3 and equation 4 using different windows
for receiving forward citations (include self-citations) ...... 165 Table C2: Regression results of equation 5 using different windows for receiving
forward citations (include self-citations) ...... 166 Table C3: Regression results of equation 3 and 4 using Poisson regression ...... 167 Table C4: Regression results of equation 5 using Poisson regression ...... 168 Table C5: Summary statistics for key variables...... 169 Table D1: Installed capacity and the percentage of vertically integrated installers that
use their own inverters ...... 171 Table D2: Installed capacity and the percentage of vertically integrated installers that
use their own panels ...... 172 Table D3: Regression results including the instrumental variable ...... 175 Table D4: Summary statistics for key variables...... 177
xvi List of Figures
Figure 1: Global greenhouse gas emissions by sector from 2000 to 2014 (MtCO2e)...... 2 Figure 2: The solar PV and wind energy capacity (GW) and renewable energy
investment (billion USD) from 2007 to 2017...... 4 Figure 3: The dissertation structure ...... 5 Figure 4: Residential Solar PV System Pricing. Source: SEIA/Wood Mackenzie
Power & Renewables U.S. Solar Market Insight ...... 27
Figure 5: Installed capacity (MW) of distributed PV in China, 2011-2014...... 40 Figure 6: Annual distribution of PV BOS applications, inventions, utility and design
(counts)...... 52 Figure 7: Annual number of PV BOS patents including invention application, invention, utility and design (count), by category and installed capacity
of installed distributed PV (MW)...... 53 Figure 8: Geographic distribution of the provinces with less than or equal to 10 PV BOS patents from 2012 to 2014 (blue) and the provinces with more than
or equal to 60 PV BOS patents from 2012 to 2014 (red)...... 54 Figure 9: Geographic distribution of the provinces with nearly zero MW installed capacity of distributed PV from 2012 to 2014 (blue) and the provinces with more than or equal to 500 MW installed capacity of distributed PV
(red)...... 54 Figure 10: Annual distribution of PV BOS patents by technological category (counts). ..84 Figure 11 Annual distribution of assignee numbers by technological category
(counts)...... 84 Figure 12: The annual number of total installers and the annual number of installers
with PV BOS patents from 2000 to 2014 ...... 118 xvii Figure 13 (a) : An example network structure between two installers and their inverter
manufacturers ...... 120 Figure 13 (b) : An example network structure between two installers and their panel
manufacturers ...... 120 Figure 14 (a) The distribution of installers’ all projects to panel manufacturers and
inverter manufacturers in the market ...... 123 Figure 14 (b) The distribution of installers’ principle projects to panel manufacturers
and inverter manufacturers in the market ...... 124 Figure A1: The geographic distribution of province groups based on geography...... 155 Figure A2: The geographic distribution of province groups based on competitiveness index according to the city ranking in the “Blue book of provincial
competitiveness”...... 155 Figure A3: The geographic distribution of province groups based on GDP...... 156 Figure B1: Annual U.S. PV BOS patents by technology category (count) ...... 159 Figure B2: Annual U.S. PV BOS assignees by technology category (count) ...... 160 Figure B3: The share of U.S. PV BOS patents by state ...... 161 Figure D1: The distribution of installers’ projects to panel manufacturers and inverter manufacturers in the market. Figure on the left includes all the installers
and figure on the right excludes the vertically integrated installers...... 173 Figure D2: The distribution of installers’ projects to panel manufacturers and inverter manufacturers in the market (excluding the vertically integrated
installers) ...... 174 Figure D3: The distribution of installers’ projects to panel manufacturers and inverter manufacturers in the market (including all the installers) (the same graph
in the main text) ...... 174 xviii Chapter 1: Introduction
BACKGROUND
The Special Report on Global Warming of 1.5 °C issued by Intergovernmental Panel on Climate Change (IPCC) points out that “human activities are estimated to have caused approximately 1.0°C of global warming above pre-industrial levels in 2017 and global warming is likely to reach 1.5°C between 2030 and 2052 if it continues to increase at the current rate” (IPCC, 2018). Therefore, to limit the global temperature below 1.5 °C by 2030, urgent and ambitious actions on global climate change mitigation are required. The energy sector accounts for more than 70% of the total global greenhouse gas emissions (Figure 1), and thus decarbonizing the energy sector is recognized as a critical way to achieve the long-term climate goals. Phasing out the use of fossil fuel and accelerating renewable energy deployment play a key role in decarbonizing the energy sector. However, achieving a rapid and global-scale deployment of renewable energy requires overcoming various economic, technological, institutional, political, social, and financial barriers. Facing these barriers, governments around the world have adopted various policies to support renewable energy deployment, including but not limited to carbon taxes, cap and trade schemes, feed-in tariff, renewable energy portfolios, and tax credits.
Why is public policy important in helping overcome these barriers and accelerate deployment? While the increasingly urgent climate change challenge may serve as the practical justification for policy intervention in renewable energy deployment, there are several theoretical rationales to justify deployment policy. In particular, policy interventions are needed to correct multiple market failures. The first market failure is environmental externalities. Greenhouse gas emissions have negative externalities, meaning that the social costs of emissions are higher than private costs. Therefore, private investments have insufficient incentives to adopt renewable energy technologies. The second market failure arises because knowledge has positive externalities and thus private benefits of developing new knowledge/technology are lower than
1 social benefits. This also leads to insufficient investment in developing new knowledge and technologies associated with renewable energy. The third market failure is related to adoption externalities, which means that the private costs of adopting new technologies are positively associated with the scale of technology adoptions through mechanisms such as learning by doing and learning by using (Jaffe et al., 2004).
60000
50000
40000
30000
20000
10000
0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014
Energy Agriculture Industrial Processes Land-Use Change and Forestry Waste Bunker Fuels
Figure 1: Global greenhouse gas emissions by sector from 2000 to 2014 (MtCO2e). Source: the CAIT Climate Data Explorer. 2017. Country Greenhouse Gas Emissions. Washington, DC: World Resources Institute.
Globally enforced and leakage-free emission taxes and trading system, while widely considered as the most efficient policy instruments to address environmental externalities, face great political barriers and design challenges. Consequently, in a second-best world, deployment policy plays a key role in correcting these market failures and directly increasing technology adoption and diffusion. Deployment policies or demand-pull policies are policy instruments that aim to stimulate investment and market demand. Two common deployment policy instruments for renewable energy technologies are feed-in tariffs (FITs) and renewable portfolio standard (RPS). The FIT is the most common policy instrument, which compensates power generators a fixed-rate
2 based on electricity generation. RPS requires power producers to generate a certain minimum share of their electricity from renewable energy. These deployment policies have largely stimulated renewable energy investment and deployment around the world (IEA, 2018).
In the last decade, the cumulative investment in renewable power and fuels has reached $2.7 trillion (Figure 2), which approximately equals three times the investment in fossil fuel generation over the same time (REN21, 2018). The investment mainly focuses on solar photovoltaics (PV) and wind energy, which accounts for more than 60% of the total investment. Between 2007 and 2017, the global capacity of solar PV technology and wind energy rapidly increased from 8 GW to 402 GW and from 94 GW to 539 GW, respectively. China and the U.S. are the two largest wind and solar PV markets in the world. China accounts for more than half of global demand and manufacturing of solar PV technologies and about 40% of global demand of wind energy. The US placed the second in solar and wind installed capacity. Solar PV became both China’s and the U.S.’s leading source of new generating capacity in 2016 and 2017 (REN21, 2018).
3 600 350
500 300
250 400 200 300 150 200 100 Installed capacity (GW) capacity Installed
100 50 New investment (billion USD)
0 0 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017
Previous year' wind capacity (GW) wind annual additions (GW) Previous year' solar PV capacity (GW) solar PV annual additions (GW) Annual new investment (billion USD)
Figure 2: The solar PV and wind energy capacity (GW) and renewable energy investment (billion USD) from 2007 to 2017. Source: renewables 2018 global status report from REN21
RESEARCH OBJECTIVES
Although deployment policies have stimulated substantial global renewable energy investment and deployment, the renewable energy investment and deployment are still far behind the levels needed to achieve the greenhouse gas emissions mitigation targets over the next 10-15 years (IRENA, 2014, IEA 2018). Therefore, particular emphasis should be placed on understanding how deployment policies can best promote technological innovations and reduce installation costs, making renewable energy technologies more cost-effective, thereby spurring further investments and deployment. It is clear that deployment policies enable scale-up of renewable energy installations, but their impact on technology innovations and cost reductions are much less clear.
4 Therefore, the overarching aim of my dissertation – a collection of three closely-coupled essays – is to understand whether and how deployment policy can lead to technological changes and cost reductions. Specifically, the first essay focuses on whether and how demand-pull policy can impact technological innovation; the second essay, through studying firms’ knowledge acquisition behaviors in the innovation process, explores why deployment policy is essential to promote technological innovations; and, the third essay investigates three major mechanisms through which deployment policy can facilitate cost reductions and contributes to separating these three mechanisms including learning by doing, learning by searching, and learning by interacting (Figure 3). Throughout these three essays, one important focus of my dissertation is the role of geography, the importance of which as an element of technological learning and innovation continues to be debated by researchers. Thus, the overarching focus of my dissertation may be summed up as addressing the following broad question: what combination of innovation ecosystem features – including, local deployment policies, technology characteristics, social and human capital, demand factors, and supply conditions – translate into persistent localized innovation and economic benefits? Essay 2 Essay 1 Demand-pull Technological Policy Innovations Essay 3 Cost Reductions
Supply chain network
Figure 3: The dissertation structure
5 EMPIRICAL FOCUS AND RESEARCH QUESTIONS
The empirical focus of my dissertation is balance-of-system (BOS) components in the solar PV industry. PV has been the fastest growing electricity generation technology globally in the last decade. Large-scale deployment of distributed solar PV technologies is widely considered to be an important piece in addressing the environmental impacts of the electricity sector, making PV an important technology to study. Solar panel/module and non-module components are two major types of hardware in a solar PV system. Solar panels directly convert sunlight into electricity through the photovoltaic effect. BOS components are the major non-module components, which include four technology categories: inverter, mounting, monitoring, and site assessment. Inverters convert power generated by solar PV cells and modules from direct current (DC) to alternating current (AC). Mounting is used to physically support solar PV modules and/or arrays. Monitoring includes methods and devices that monitor health, connectivity, and output of solar PV systems. Site-assessment includes methods and mechanisms for appraising the photovoltaic output potential of a land parcel, determining the ideal placement of PV systems on potential sites, and the extent to which obstructions or shade may impede the output potential of PV sites.
Most academic studies so far have focused on PV panels and modules, and only a small body of literature focuses on BOS. More recently, scholars have begun to emphasize that a wider deployment of solar PV in the future would depend heavily on innovations and cost reductions in
BOS (Ardani et al., 2013, 2012; Barbose and Darghouth, 2018; Seel et al., 2014, 2013). The price of solar panel/module has decreased dramatically over the last 15 years, falling from $3.8 per watt in 2000 to $0.8 per watt in 2015. Currently, panels account for only about 15% of the total price of an installed (distributed) solar PV system. However, BOS component prices haven’t changed substantially. Just by themselves BOS components account for about 20% of the total price of an installed PV system. However, BOS components are also closely related to several non-hardware costs such as installation costs, racking costs, and other installer soft costs (customer acquisition, financing, overhead, etc.). These non-hardware costs account for more than 60% of the total system
6 price. As such, there is a general consensus that the biggest barrier for a wider deployment of PV in the future is the high price of non-module components.
But why is it necessary to study BOS components separately? Firstly, the learning/innovation process for PV panels is different from the learning process of BOS components – a difference mainly driven by different technological characteristics (Lewis and Wiser, 2007; Neij et al., 2017; Shum and Watanabe, 2008; Wene, 2000). PV panels are pretty standard and globally traded products, but BOS components often have much more customized innovations that depend upon the local context such as local climate conditions, local building code, local rooftop structures, local customer’s preference, and local conditions of power grid. Thus, BOS components represent a good case involving the positive, interactive feedback between geography, innovation, and localized economic effects. Technological and operational aspects of BOS components make local conditions salient, thus affording a rich empirical setting to explore the geography-innovation-economy interactions at the local level.
Furthermore, quite interestingly, differences in learning/technology processes are expected to exist even among subcategories of PV BOS components (inverter, monitoring, mounting, and site assessment). Inverter and monitor technologies are expected to have the least local nature, compared to the other three subcategories, because inverters and monitors are closer to globally- traded products that are relatively easily adaptable to the local context. Mounting equipment are expected to have the most local dependence, because mounting techniques, methodologies, and equipment heavily depend on local climate, local building code, and local rooftop features. Site assessment systems and tools are also expected to depend on local conditions, such as local grid features and local customers’ preferences. As a result, different technology characteristics within the four BOS categories allow me to explore how the impact of deployment policy may depend on different technological characteristics and associated learning/innovation processes.
7 In sum, speaking to these theoretical, empirical, and policy opportunities, my dissertation entitled “Essays on Innovation Policy, Knowledge Networks, and Cost Reductions in Deployment- Related Technologies in the Solar PV Industry” explores three major research questions: (1) the efficacy of local demand-pull policy in driving local innovation of PV balance-of-system (PV BOS) components, (2) the knowledge acquisition process that fuels firm-level innovation process in PV BOS, with a focus on linking knowledge networks and innovation quality at different geographic scales, and 3) the role of technological changes and supply-chain networks in cost reductions, with a focus on separating three mechanisms (learning by doing, learning by searching, and learning by interacting) through which deployment policies potentially facilitate cost reductions.
BRIEF SUMMARY OF EACH ESSAY
Essay one ---- My first essay empirically examines whether demand-pull policy can induce technology innovation and whether the locus of demand-pull policy matters for the engendered innovation. I used keyword search strings to develop two original databases of PV BOS patents in the field of distributed PV from the State Intellectual Property Office of The People’s Republic of China (SIPO) between 2005 and 2014 and from the United States Patent and Trademark Office (USPTO) between 2000 and 2014, respectively. The empirical design exploits the subnational variation in (policy-created) demand to study the impact of geographically-differential (local vs. non-local) demand on local innovation. Results from a range of econometric models support the demand-induced innovation hypothesis. More importantly and novel, the results also show that only local demand significantly induces PV BOS innovations in the distributed PV market. These results are consistent in both the U.S. and Chinese markets. At the same time, these results are contrary to what is known for PV modules. My novel focus on PV non-module components, thus, helps complete the picture of how to understand the policy-geography-innovation nexus in PV.
8 Moreover, I explored whether local demand has an indispensable role in inducing local innovation by differentiating between inverter and non-inverter technologies. The results emphasize that within the same technology (an integrated PV system, here) the technological trajectories of different subcomponents (modules, inverters, etc.) potentially interact differently with local market conditions. Local demand has a stronger positive effect on technologies that have a significant local nature. But for relatively standard and globally traded products (such as panels and inverters), technological innovations are more likely to be driven by global markets. These findings suggest that whether local demand is an essential precondition to local industry development (i.e., innovations) depends on technological characteristics and the nature of associated learning processes.
Overall, results from my research emphasize the importance of local demand and show that depending upon the technology involved, local demand-pull policies could translate into local innovative activity. The technological learning fueling innovation in PV BOS results from an interactive process between technology and the local context. In this case, market size is akin to a locally bounded asset, and thus innovation induced by the market is a region-specific competitive advantage. In other words, local demand-side policy can help integrate multiple policy goals including environmental protection, technological innovation, and economic competitiveness. The findings also inform the discussion on ‘green industry development’. The results show that technology characteristics are critical factors in explaining the role of local demand in green industry development, and thus if learning and innovation processes depend on local context, then local demand is a prerequisite for developing a competitive local industry. Additionally, the success of subnational governments’ efforts in expanding market demand and translating it into local innovative activities also underscores the potential viability of a bottom-up approach to renewable energy governance, at least for those technologies that have strong local dependence.
9 Essay two ---- Inspired by my first essay, my second essay explores why innovation is localized and the extent to which localized learning is important for innovation. I answer these questions from the perspective of the underlying learning process with a focus on the characteristics of knowledge acquisition and knowledge networks of PV BOS innovations using patent citation data. For this work, I extended my original U.S. PV BOS database by including backward citations and forward citations for each U.S. PV BOS patent. Backward citations refer to patents that are cited by the focal patent and forward citations include all the patents that cite the focal patent in the future. Using three different quantitative approaches to triangulate the problem, one main contribution of this essay is to quantify the influence of geographic distance on knowledge acquisition, which provides a concrete understanding regarding the extent to which the localized learning is happening in the U.S. PV BOS industry and its importance in solar PV deployment. The results show that citation-related linkages have a strong localization effect at the state level, but not at the country level. Moreover, this study provides empirical evidence showing that knowledge acquisition is associated with technology characteristics. I find that technologies that are more related to local context use more local knowledge, which indicates that technologies with a more significant local nature have a higher dependence on localized learning.
Another important contribution of this essay is to add the international dimension to localization learning literature. I argue that geographic proximity is unlikely to be the whole story of promoting innovations and I indeed find that international knowledge significantly contributes to improving patents’ quality, which emphasizes the value of knowledge diversity in technological innovation. Compared to acquiring local knowledge and building local networks, acquiring non- local knowledge/networks requires extra efforts, costs, and more complicated processes. Further, the difficulties in knowledge acquisition are expected to be more serious for start-ups, firms with low absorptive capacity, and firms located in peripheral regions. Thus, for policymakers, it might be valuable to examine whether local firms face difficulties in accessing non-local knowledge and, if so, to help local firms bridge non-local networks with local ones. For example, depending upon
10 the technological characteristics, it would be useful for local governments or industry associations to organize more formal or informal activities that facilitate international collaboration and knowledge exchange. In sum, the second essay incorporates both local and non-local aspects of technological learning/innovation process, thus providing for a more integrated understanding of the trade-offs between tailoring approaches for cultivating local markets with local strategies and leveraging global learning networks.
Essay three ---- Although technological innovation is intrinsically important, it is often of more direct interest to understand its impact on cost reductions for technologies and services offered on the market. Previous literature on solar costs has focused on how market structure, installer scale, demographics, ownership, policy, and system components influence solar PV installation prices (Gillingham et al., 2016b; Nemet et al., 2016). However, the unexplained variation is still large (>70%) after controlling for these factors, which suggests the need for further work to identify the potential for cost reductions in solar PV installations – a key requirement for much broader deployment of PV at the global scale. To address this research opportunity, my third essay focuses on the role of technological innovation, knowledge spillovers, and upstream- downstream networks (i.e., installer-inverter/panel manufacturer networks) in reducing costs of solar PV installations. My analytical strategy is to match the PV BOS patent database with an extraordinarily rich dataset of system-level PV installations in the U.S., allowing me to build models to explore how innovation and network factors impact installation prices, as well as the embeddedness of innovation and networks in geography.
The second contribution of this essay is to separate three major learning mechanisms through which deployment policy is expected to impact cost reductions. The direct impact of deployment policy is to increase firms’ installation experience. When an installer increases its installation experience, the installer can learn from its repeated working on technologically similar installations to improve its productivity and thus reduce its installation costs (learning by doing).
11 Increasing installation experience can also motivate new working processes, installation methods, or related technology innovations and stimulate firms to invest more in research and development activities, thereby potentially leading to cost reductions (learning by searching). Installers can also learn from its upstream suppliers (i.e., panel manufacturers and inverter manufacturers) regarding new products, strategies, technology trends, and processes, which could also lead to cost reductions (learning by interacting). By incorporating experience-related variables, innovation-related variables, and network-related variables, this essay carefully separates these three mechanisms that convert experience accumulation to cost reductions.
12 Chapter 2: Relevant literature
TECHNOLOGICAL INNOVATION AND INNOVATION POLICY
In 1912, Schumpeter published “the theory of economic development” and put forward “innovation theory,” In Schumpeter’s work, he defined innovation as five types of activities: 1) launch of a new product or a new species of already known product, 2) application of new methods or process of production, 3) acquiring of new sources of supply of raw material or semi-finished goods, 4) opening of a new market, and 5) new industry structure and new forms of organization. As for technological innovation, it mainly refers to the first three activities (i.e., new product, new methods/process, and new materials). These activities can be further distinguished between innovation activities and diffusion activities. Here I mainly focus on innovation activities, and thus it can be defined as “the discontinuity in the technical characteristics or in the use of a new product or new material or process or method” and it includes both radical innovation and incremental innovation (Lundvall, 2005). Schumpeter also proposed that innovation is an important factor in economic growth in 1912. In the following decades, scholars began to contribute to the theory of technology innovation. These neo-Schumpeterian economists including Freeman, Rosenberg, and Nelson emphasized the role of innovation in economic growth. As economists began to more widely recognize that technology innovation is an engine of economic growth, interest in studying the drivers of technology innovation grew.
Hicks was the first to propose the hypothesis that innovation could be induced by “a change in the relative prices of the factors of production”, for example, rising wages can induce labor- saving innovation (Hicks, 1932). Applied to the field of energy, the rising prices of energy could induce energy-saving innovations. Inspired by this idea, environmental economists investigated the effect of rising energy prices or environmental regulations on environment-friendly technology innovation. Many empirical studies demonstrate that rising energy prices or external environmental regulations could significantly stimulate technology innovation (Altwies and
13 Nemet, 2013; Hayami and Ruttan, 1985; Jaffe and Palmer, 2009; Lanjouw and Mody, 1996; Newell et al., 1999; Parry et al., 2003; Popp et al., 2011a, 2010; Taylor, 2008).
Another branch of induced innovation hypothesis, proposed by Schmookler and Griliches, emphasizes that the power of induced innovation lies in market demand (Griliches, 1957; Schmookler, 1962). By studying the technology spillover of hybrid corn, Griliches found that technology spillover responds to market size (Griliches, 1957). Schmookler further concluded that demand not only pulls technology spillover, but also pulls technological innovation. Utterback further pointed out that market factors are the primary influence on innovation, and 60% to 80% of important innovations have been in response to demand (Utterback, 1974). Many recent empirical studies also provide strong evidence to support the demand-induced innovation hypothesis that innovation could be induced by demand (Jaffe and Trajtenberg, 2002; Lanjouw and Mody, 1996; Lin, 2010; Gregory F Nemet, 2009; Popp, 2006a).
The basic model of demand-pull innovation is a linear model that indicates firms’ innovation is triggered by market demand. Roy Rothwell (1983), Fusfeld and Kahlish (1985) and von Hippel (1986) were not satisfied with the linear innovation model, so they examined the innovation process as a process of interactions between different elements, such as between firms, between users and firms, and between government and firms (Fusfeld and Haklisch, 1985; Rothwell, 1983; von Hippel, 1986). Later in 1986, Kline and Rosenberg (1986) proposed the chain-linked model, which is similar to the interaction model (Kline and Rosenberg, 1986). Contrary to the linear model of innovation, the chain-link model studies the complex relationship between different stages in the process of innovation. Each stage has effects on other stages and there are complex feedback loops among all the stages. In this dissertation, I also emphasize and incorporate the interaction view in studying innovation process.
14 Inspired by the interaction idea, scholars began to study technology innovation from a systems perspective, including studies in the “national innovation system” framework. In the 1980s, economies of the major Western industrial countries declined drastically, while the economies of Japan and several emerging countries in East Asia prospered, which was an exciting topic for researchers of technological innovation. Freeman, interested in factors that contributed to the economic growth in Japan, studied the role of governments (especially the Ministry of international trade and industry) and the Japanese strategy of “building the country by technology” in Japan’s technological innovation mechanism, pointing out that the national innovation system contributed significantly to Japan’s high-speed development. Since then, national innovation system has become an increasingly important conceptual framework. The national innovation system framework emphases that innovation system includes various interdependent elements, including firms, industries, governments, universities, complementary infrastructure, non- government organizations, other institutions, and their environments. This framework successfully involves all kinds of institutions, their behaviors, and more importantly, their interactions into innovation analysis and contributes a new rationale to identify technological gaps and economic gaps within an industry, within a country, and across different countries (Godin, 2002; Lundvall, 1992; Lundvall et al., 2002; Motohashi, 2005; Mowery and Sampat, 2006; Mowery et al., 1993; Mowery and Rosenberg, 1993; OECD, 2002).
From the perspective of the micro drivers of innovation, the extant literature provided insightful thoughts of understanding firms’ innovation behaviors, such as why firms innovate and how firms innovate. In order to develop competitive advantages and ensure future competitiveness, improve absorptive capacity (absorb external information/knowledge/technology), increase productivity, reduce costs, and increase profits, firms often invest in innovative activities (Baldwin et al., 2000; Baldwin and Johnson, 1999; Cohen and Levinthal, 1990; Crepon Bruno et al., 1998; Cuervo-Cazurra and Annique Un, 2010; Helfat, 2000; Schumpeter, 1942; Teece et al., 1997). Firms innovate through various activities, including formal investment in research and
15 development (R&D), incremental changes on products and process, imitative activities, and exploiting existing knowledge/technologies in new ways (Evangelista et al., 2002; Grimpe and Sofka, 2009; Hansen, 1997; Huang et al., 2010; Kim and Nelson, 2000; Nascia and Perani, 2002; Veugelers and Cassiman, 1999). Although firms have incentives to innovate, why is it still important to involve government and public policy in the innovation process? Technology policy/innovation policy is a set of public policies that aims at influencing or changing speed, direction, and scale of technological innovation. Next, I summarize and discuss three primary reasons to justify policy interventions on technological innovation.
First, technological innovation policies are necessary to correct market failures. Knowledge is a public good, which has positive externalities. Since firms cannot gain all the benefits from their R&D activities and cannot get paid for knowledge spillover, the firms’ R&D investment will be lower than socially efficient levels of R&D investment. Such conclusions have been agreed upon by many researchers (Arrow, 1962a; David et al., 2000; Griliches, 2013; Nelson, 1959). Another important type of market failure is associated with environmental externalities. Knowledge externality discussed above is a type of positive externality, but environmental externality is a type of negative externality. The private benefits of developing or adopting environmental-friendly technologies to reduce environmental externality are very likely to be lower than social benefits. Therefore, without appropriate policy interventions, firms would not have sufficient incentive to reduce emissions of local or global pollutants. In terms of policy instrument, to put a price on pollutants or emissions through tax or emission trading system is widely recognized as the most efficient policy (Baumol, 1972; Pigou, 1920; Coase, 2013; Stavins, 1995). In a second-best world, government R&D investments and demand-side support for technology innovation are economically justified on the basis of externalities, while it is also recognized that different policy instruments have varying effects on different technologies and those effects may even vary by country (Dechezleprêtre and Glachant, 2014; Hassett and Metcalf, 1995; Molas-Gallart and Davies, 2006; National Research Council, 2001; Peters et al., 2012).
16 Technology/innovation policy may also be more political feasible compared to tax or trading system, because the benefits are typically more concentrated, but the costs are dispersed (Jaffe et al., 2004)1.
Second, government action is needed to correct system failure. As discussed above, innovation is widely recognized as a process of interactions among different elements and the innovation system approach emphasizes that the success of innovation depends on various complementary elements in the system, including conditions for generating and diffusing knowledge, financial conditions, demand for innovation, innovation capacity, and the degree of cooperation (Braun, 2008; Edler and Georghiou, 2007; Fagerberg, 2017; Kline and Rosenberg, 1986; Lundvall, 1992, 1988; OECD, 2002; Smits et al., 2013). The insufficient provision of these complementary elements is recognized as a system failure and the existence of these system failures would block technological changes, and thus policy interventions are needed to identify and correct system failures (Braun, 2008; Edquist, 2011; Fagerberg, 2017; Hoppmann et al., 2014; Klein Woolthuis et al., 2005; Negro et al., 2012; Wieczorek and Hekkert, 2012). Third, even in the absence of the market failures from R&D spillovers or externalities, governments may still want to subsidize innovation in areas that support the national interest (e.g., defense, climate change mitigation and adaptation, security).
There are mainly three types of innovation policy: 1) supply-side policy includes providing technological support and financial support; 2) demand-side policy mainly refers to expanding market size, including government procurement; 3) environmental-side policy includes patent policy, environmental regulation, worker health regulation, and so on (Rothwell, 1982). Although
1 It is worthy to note that it is possible to have socially excessive, inefficient investment in research and development. The “patent races” literature shows that non-colluding firms invest excessively in research and development activities based on the assumption that the winners can capture the largest or all the price and the others receive nothing (Reinganum, 1989; Doraszelski, 2008) . From the perspective of competition and leadership, literature also shows that the incentive of a monopolist to invest in new technology is to preempt entry by competitors, so a monopolist may also invest excessively to be ahead in the competition (Grossman and Shapiro, 1985).
17 public policy is important and necessary in technological learning and innovation, complicated innovation processes limit the effectiveness of public policy tools, thus highlighting the importance of studying technological learning processes that can provide direct insights for improving public policies (Rothwell and Zegveld, 1988). Rothwell and Zegveld (1988) also noted that innovation process is geographically and temporally dynamic. Innovation process can be different in different countries, in different industries and in different sectors of an industry. Therefore, public policies should be flexible enough to adapt to different situations.
As noted above, the main focus of my dissertation is demand-side policy (i.e., deployment policy). Accordingly, below I summarize the policy justification for and the insights from key theoretical and empirical research on the impacts of demand-side policy on technological change. In addition to two types of market failures I discussed in the previous paragraphs, Jaffe et al. (2004) pointed out a third market failure: adoption externalities. Adoption externalities refer to the possibility that private benefits of adopting a new technology might depend on the overall scale of technology adoption (Jaffe et al., 2004). Particularly, the positive benefits of demand-side policy may be acquired from various mechanisms including learning by doing, learning by using, learning by interacting, and learning by searching. These learning mechanisms induced by market demand significantly contribute to cost reductions of technologies. I will provide an in-depth discussion regarding these learning mechanisms in section 2.2.
Many studies focus on demand-pull policies specifically. Learning and innovation often result from an interactive process, which benefits from proximity, thus experience-based innovations are often characterized by localization and geographical agglomeration. The importance of local contexts, such as local culture, local regulation, local business environment, and local supply chain, in the innovation process has been recognized. Local firms have comparative advantages resulting from a better understanding of local context. The comparative advantages of local firms are stronger for the business-to-consumer sectors, as they focus more on
18 end-users and demand-induced innovation (Brem and Voigt, 2009). Demand-pull policy can expand local markets and help local firms gain more experience. However, policy-makers are often more interested in seeing that public support for expanding local markets contributes to local technological innovation and economic competitiveness, rather than benefiting other jurisdictions more heavily (Lewis and Wiser, 2007; Maguire et al., 2012).
Recent empirical work has begun to distinguish the effect of demand-pull policy on local technology innovation and foreign technology innovation, but the results of empirical studies are not consistent. Popp (2006) found that innovations for sulfur dioxide and nitrogen oxide control at coal-fired power plants primarily responds to domestic regulation instead of foreign regulation (Popp, 2006b). Dechezleprêtre and Glachant (2014) analyzed that wind technological improvements across OECD countries respond to both domestic and foreign demand-pull policies, they also suggest that the marginal effect of domestic policies is twelve times greater than foreign policies (Dechezleprêtre and Glachant, 2014). Peters et al. (2012) found that both domestic demand-pull policies and foreign demand-pull policies could spur innovations in solar PV panels, but the magnitude of the effect of foreign demand-pull policies is smaller than that of domestic policies (Peters et al., 2012). On the contrary, Lanjouw and Mody (1996) examined vehicle emission regulations in the U.S. and found that the U.S.’s vehicle emission regulations stimulated innovations in Japan and Germany (Lanjouw and Mody, 1996). Popp et al. (2011) observed that technology innovation in the pulp and paper industry responds to both domestic and foreign regulation (Popp et al., 2011). A possible explanation of these apparently inconsistent results could be whether an industry faces a global market (Popp, 2006b). If an industry’s products focus on foreign markets, it is reasonable to expect that technology innovation in this industry responds to foreign regulation. Another possible explanation is whether the local context is significantly important to technology innovation. PV BOS technologies, the focus of my research, need to interface with and respond to various local contextual factors such as climate, standards and codes,
19 supply chain, and customer preferences. Therefore, based on the existing literature, I expect that the innovations in PV BOS should primarily respond to local demand-pull policies.
TECHNOLOGICAL LEARNING
Technological learning can occur at both the individual level and the firm level. There is a long list of different definitions for technological learning: Carayannis and Kassicieh (1997) defines technological learning as “the process by which a technology-driven firm creates, renews, and upgrades its latent and enacted capabilities based on its stock of explicit and tacit resources” (Carayannis and Kassicieh, 1997); Dodgson (1991) defines technological learning as “the ways firms build and supplement their knowledge-bases about technologies, products and processes, and develop and improve the use of the broad skills of their workforces” (Dodgson, 1991); and Figueiredo (2002) defines technological learning from the perspective of capability, and says “the term technological ‘learning’ is usually understood in two alternative senses. The first sense refers to the trajectory or path along which the accumulation of technological capability proceeds; the second sense refers to the various processes by which knowledge is acquired by individuals and converted into the organizational level” (Figueiredo, 2002). Some other definitions emphasize the interactive nature of technological learning. For example, Bathelt et al. (2004) noted that “Innovation, knowledge creation and learning are all best understood if seen as the result of interactive processes where actors possessing different types of knowledge and competencies come together and exchange information with the aim to solve some – technical, organizational, commercial or intellectual – problems” (Bathelt et al., 2004).
Most learning is not a spontaneous process, but rather a conscious, costly and targeted process (Malerba, 1992). At the firm level, learning can occur at almost every step, including R&D, designing, manufacturing, production, marketing, engineering and deployment (Malerba, 1992). Knowledge and innovation are the most important outcomes of firms’ efforts in learning (Fiol, 1996; Hitt et al., 2000; Oliveira, 1999; Wolfe, 1994).
20
A common approach to studying technological learning in economics is to use a learning curve model, which builds the relationship between learning and experience. Learning curve model describes the phenomenon of learning by doing. Many empirical studies consistently show the robust relationships between firms’ production experience and cost reduction/productivity growth (Arrow, 1962b; Bollinger and Gillingham, 2014; Nemet, 2011; Qiu and Anadon, 2012). Learning curves can be used to estimate future working time and number of laborers. There are two main types of learning curves: the traditional learning curve only includes one factor, cumulative experience; newer models includes two factors and try to separate the effect of R&D and cumulative experience on technological change (Qiu and Anadon, 2012; Söderholm and Klaassen, 2007; Söderholm and Sundqvist, 2007).
Firms can learn from both internal sources or external sources. Figueiredo (2002) classified learning processes into knowledge acquisition and conversion using four categories: 1) external knowledge acquisition; 2) internal knowledge acquisition; 3) knowledge socialization (converting knowledge from individual learning to organizational learning); 4) knowledge codification (converting tacit knowledge to codified knowledge). Researches also identified different types of learning channels. These learning channels can be understood as sub-processes of the above four learning processes. For example, internal knowledge acquisition includes learning by doing, learning by using, learning by searching, and learning by training; external knowledge acquisition includes learning by doing, learning from advances in science and technology, learning from inter- industry spillover, learning by interacting, learning by hiring, learning by alliance, learning by imitating, and learning by training (Arrow, 1962a; Enos, 2008; Hitt et al., 2000; Kline and Rosenberg, 1986; Malerba, 1992; Taylor et al., 2006). Some learning processes (or channels) will be discussed in more detail in Section 2.3 on the theories of geography of innovation, where local networks and learning channels play an important role.
21 Additional factors that can influence a firms’ learning that are explored in the literature include location characteristics, industry characteristics and firm characteristics, such as size, structure, managerial ability, culture, R&D investment, human resources, stock of knowledge, absorption ability, industry structure, external technological information availability, and collaboration (Cohen and Levinthal, 2006; Henderson, 1994; Hitt et al., 2000; Sen and Egelhoff, 2000; Teece, 1993).
WHY AND HOW GEOGRAPHY MATTERS IN INNOVATION
Researchers have recognized that innovation activity and knowledge flows have a tendency to be geographically localized. Although firms are exposed to convenient transportation tools and instant communication tools, geography is still an important element in technological learning and innovation. Over a century ago, Marshall (1895) proposed three reasons why location is important: specialized skilled labor pools, an infrastructure of complementary industries, and proximity to a strong knowledge base. Firms can significantly benefit from “being in such locations”.
Geographic proximity is important for facilitating the spillover of tacit knowledge through face-to-face communication, deep interaction, and even local gossip. The difficulty in codifying tacit knowledge prevents its diffusion to other locations, thus tacit knowledge accumulates in a certain location benefitting the local pool of firms. Tacit knowledge also tends to be contextual, which means that local tacit knowledge is more useful for local firms that are in the same context (institutional context, climate context, etc.) and where personalization to meet user demands are more likely to be similar such that user-producer interaction becomes an important source of technological learning and innovation (Brem and Voigt, 2009). For example, Bathelt and Zeng (2012) studied the chemical industry in Shanghai and found that user-producer interaction significantly contributes to firms’ innovation.
22 Localized learning can also shape local networks, as evidenced by the positive relationship between geographical proximity and closed networks. Geographic proximity facilitates trust- building and enables close collaboration, which is beneficial to localized technological learning. Networks ties developed through interaction with local universities, suppliers, utilities, manufacturers and end-users are all useful to firms’ technological learning and innovation (Grübler, 2003; Helfat, 2006; Junginger et al., 2005; Salter and Laursen, 2006; Schneider et al., 2008). Strupeit and Neij (2017) studied mounting technologies and cost dynamics in Germany, and found that the learning process of mounting technologies has a significant local nature, because local rooftop structure, local building codes and the needs of local installers are significant elements, which means interactions between mounting equipment producers and installers are key in technological learning. Laursen et al., (2012) understood the localized network and norms as social capital and found a positive relationship between geographically bound social capital and firms’ innovative ability.
Geographical proximity also facilitates local firms in forming their practice community. “Community of practice” provides another useful perspective to study localized learning. The literature of “community of practice” is built on collective and interactive learning (Wenger, 1998). Firms working in a similar industry in certain geography can form a community where they share and communicate their knowledge, information, and activities. The connection and interaction enhance firms’ stock of information and knowledge, thus improving firms’ practice (Brown and Duguid, 1991; Wenger, 1998). Another related perspective is from the competition view, where firms can compare their performance to other local firms and reassess, improve or change their practice or performance accordingly (Bathelt et al., 2004).
The view of “community of practice” is consistent with the some key observed features in the solar mounting/racking market. Morrris et al., (2013) compared mounting/racking process between Germany and the U.S. from pre-installation to racking attachment, to module installation,
23 to on-roof electrical and to off-roof electrical and found that practices in the U.S. are different from practices in Germany. For example, in the process of pre-installation, Germany has no off-site preparation that can improve productivity, but in the U.S., off-site preparation is a necessary routine. This indicates that each geographic area has formed their community of practice.
From the perspective of public policy, regional innovation systems are important to facilitate localized learning. Maguire et al. (2012) summarized two policy trends that are inspired by localized learning. The first policy trend focuses on mobilizing local resources and local assets and designing local strategies to promote local innovation and economic growth. The second policy trend is that national innovation strategy began to recognize the importance of regional innovation systems, especially the role of using local resources, local assets and enhancing national innovation policy impacts. Studying regional innovation systems (RIS), especially the nature and characteristics of RIS, includes a focus on firm-level technological learning and networks (Martin and Sunley, 2003; Maskell, 2001).
Although localized knowledge spillover, facilitated by geographical proximity, is efficient for tacit knowledge, localized knowledge has limitations as well. In this vein, a contrasting stream of research contends that non-localized knowledge should also be an important part of technological learning processes. Localized knowledge tends to be more homogenous, thus if we recognize diversity as useful in technological learning and innovation, non-localized knowledge should be expected to be a valuable source for external and diverse inputs (Jacobs, 1969). Asheim and Isaksen (2002) found that too much focus on localized knowledge may lead to path dependence. Geographic proximity makes learning networks fixed and dependent. An effective way to break path dependence is to absorb non-localized knowledge, which tends to be more novel and comprehensive than localized knowledge. Owen-Smith and Powell (2004) studied knowledge network in the Boston Biotechnology Community and found that interaction with “pipeline” (non- localized knowledge) was the source of remarkable innovation. Burt (1992) and Grabher (2002)
24 thought that localized knowledge is highly homogenous and highly stable, thus, extra-local sources have an important role in innovation. Bathelt et al. (2004) built a theoretical framework to illustrate that each firm can benefit from network ties outside local clusters and that information or knowledge obtained from the “global pipeline” can also benefit local clusters through “local buzz”.
Localized learning and non-localized learning have different searching processes. Localized learning is effective by just “being there,” because it is relatively easy to build trust and interact with others in the same geography. As Bathelt et al. (2004) said “localized learning can largely take care of itself”. On the other hand, trust, collaboration, and networks are relatively difficult to establish over long distances, and searching extra-local information and extra-local knowledge requires extra effort, cost, and more complicated processes (Bathelt et al., 2004). However, as noted before, policy trends largely focus on regional innovation systems, such as how to mobilize local resources and help local firms establish collaboration with each other. Therefore, an important policy implication of localized learning versus non-localized learning is to emphasize the necessity of establishing “global pipelines” for local firms and providing more institutional support for non-localized learning.
Some descriptive studies show that choices of learning sources depend on the nature of innovation, the maturity level of industries, and knowledge types. From Schumpeter’s combination view of innovation, researchers believe that diverse knowledge, especially knowledge from external industries and geographically faraway places, is necessary for disruptive innovation (Ahuja and Lampert, 2001; Arthur, 2007; Fleming, 2001; Nemet, 2012a). But incremental innovation may need more specialized knowledge, which may be readily provided by localized learning. Knowledge that is suitable for non-localized learning is codified, because codified knowledge is easy to commute over long distances. Tacit knowledge is difficult to diffuse over long distances, but tacit knowledge can be obtained just by “being there” (geographic proximity), which means that the predominant way of transferring tacit knowledge is through localized
25 learning (Bathelt et al., 2004; Lorenzen, 2007; Storper and Venables, 2004). The maturity level of industries also plays a role in different learning sources. Tacit knowledge is important for industries that are at an early stage. When an industry becomes mature, standardized products and mass production make knowledge easier to diffuse globally (Grossman and Helpman, 1991).
In sum, research on the geography of innovation aims to emphasize the importance of local areas and how local conditions relate to technological innovation. However, from the opposite perspective, densely connected localized clusters may also be treated as barriers for large-scale knowledge spillover.
SOFT COSTS REDUCTIONS AND TECHNOLOGICAL INNOVATIONS IN SOLAR ENERGY
The center of cost reduction literature and technological innovation literature in the field of solar PV has been hardware, panels and modules, in particular. Only very recently have researchers begun to study the dynamics of soft (i.e., non-hardware) costs and cost-reduction opportunities they provide. One motivation for this change in focus is the increasingly important role of soft costs to overall PV system costs. Generally, soft costs include marketing and customer acquisition, system design, installation and engineering costs, permitting and inspection costs, and installer margins. Although solar PV system price has decreased by about 69% between 2000 to 2017 (Barbose et al., 2018), solar PV installed prices are still not competitive with other electricity supply options. Compared to Germany, China, Italy, Japan and Australia, residential and non- residential PV system prices are higher in the U.S., and the disparity in soft costs is the primary reason (Barbose et al., 2015). As it is shown in Figure 4, in the U.S., soft costs account for about 60%-70% of solar PV system prices (Barbose et al., 2017). As a major driver of PV system prices, studying solar soft costs has attracted increased attention to historical trends in soft costs and the factors that influence soft cost reductions, but this is still a small body of literature.
26
Figure 4: Residential Solar PV System Pricing. Source: SEIA/Wood Mackenzie Power & Renewables U.S. Solar Market Insight
Nemet et al., (2016) studied factors that influence the prices of solar PV systems by comparing low-price (LP) systems and non-low-price (non-LP) systems, finding that competition, installer scale, markets, demographics, ownership, policy, and system components are all statistically significant factors. Gillingham et al. (2016) explored factors that influence the variation of solar PV system prices across states in the U.S. and found that federal policies, differences in market structure (i.e., competition), new construction or retrofit, installer experience (learning by doing), inclusion of battery storage and tracking equipment all contribute to price variation in solar PV systems. They also found that the unexplained variation remains large after controlling for all of these variables, which indicates more study is needed to understand price discrepancy that could lead to cost reduction. Both these papers point to the role of experience in reducing soft costs. However, while more experience can help improve labor productivity, it can also lead to technological innovation. Moreover, the process of accumulating experience is also a process of learning from other collaborators. These papers do not separate the effect of labor productivity improvement (i.e., streamlining the installation process), the effect of technological innovation, and the effect of learning from other collaborators.
27 The disparity in soft costs being the primary reason for significantly higher solar PV system prices2 in the U.S. has driven some recent research to study the source of soft cost variation between the U.S. and Germany. Seel et al. (2013) summarized the following factors that may contribute to significantly higher soft costs in the U.S.: smaller market size, higher inverter cost, higher customer acquisition costs, lower customer success rates, higher marketing and advertising costs, longer installation process, higher wages, higher labor hour requirements for permitting, interconnection and inspection (PII), higher fees for PII, and higher sales taxes. Seel et al. (2013) conducted two surveys in the U.S. and Germany and found that the total difference of non- hardware costs for residential PV between Germany and the U.S. is about $2.70/W. Specifically, customer acquisition costs, installation labor costs (including working time and working wage), PII processes, and sales/value-added tax can explain $1.38/W difference. The remaining $1.32/W cannot be explained by the above factors. Strupeit and Neij (2017) pointed out future research should study other factors, especially the role of knowledge, learning process and institutional development.
Comparison studies between the U.S. and Germany support the above research direction. Market growth is more correlated with soft cost reductions in Germany. Country-level cumulative installation can explain 91% of the variation of soft costs in Germany, but only 48% in the U.S (Seel et al., 2013). The learning rate of soft costs in Germany is about 15%, but only 7% in the U.S. (Seel et al., 2013; Strupeit and Neij, 2017). Both are lower than the learning rate for PV modules, which is about 21% on average for global scale studies (de La Tour et al., 2013). Strupeit and Neij (2017) implied the importance of studying and supporting deployment processes after they studied the cost dynamics of solar PV deployment in Germany. The lower learning rate of soft costs and the lower effect of market growth on soft costs also indicate that studies of the
2 The median installed price of customer-owned PV systems that is smaller than10 kW in the U.S. for California Solar Initiative (CSI) was about $5.52/W in 2012Q3, but the system price in Germany is only $2.51/W in 2012Q3 (Seel’s presentation on page 10).
28 learning process relevant to soft costs are important. Such studies could inform policies for supporting learning environments for soft cost reductions.
Studies of soft costs reductions are still a small body of literature. Current literature addresses the role of market size (or market growth) on soft costs reductions to some extent but does not address the separate roles of labor productivity and technological innovation in soft costs reductions. Morrris et al. (2013) explored each step in the mounting/racking process, including pre-installation, racking attachment, module installation, on-roof electrical, and off-roof electrical. They found that German installers can finish each of these steps two to four times quicker and better. Institutional and organizational arrangements, different building structures, and installation practice all contribute to differences in installation efficiency between the U.S. and Germany. The comparison between compositions of soft cost and mounting practices between the U.S. and Germany relates to the geography of innovation addressed Section 2.3, because knowledge about mounting practices and soft costs in Germany (non-localized knowledge) can provide an important learning channel for the U.S. The importance of non-localized knowledge suggests that focusing mainly on localized learning channels or the lack of non-local learning channels may be a barrier to industry development in the U.S. to some extent, and research is needed to resolve the role of non-localized knowledge in the technological learning process.
PATENT DATA AND CITATION ANALYSIS
Patent data is a valuable and available resource for analyzing the process of technological change. In the field of technological change, patent data has been widely used, with various limitations, to evaluate study innovation, its drivers and impacts, including to predict the future development direction of technologies, and to trace knowledge spillover among patents.
Patent counts are widely used to measure the innovative output and technological capacity of a firm, a region, an industry, or a country (Almeida et al., 2002; Griliches, 1990; Hausman et
29 al., 1984; Hottenrott and Lopes-Bento, 2014; Jaffe and Trajtenberg, 2002; Lanjouw and Mody, 1996; Patel and Pavitt, 1987; Scherer, 1983; Schmookler, 1962). However, critics argue that patent counts can only reflect the quantity of technological innovation and may not fully reflect the value or quality of technological innovation. Therefore, deeper indicators, beyond just patent counts, need to be developed from patent data or additional data are needed to reflect the quality of innovation.
Tong and Frame (1994) proposed to use numbers of patent claims as an indicator of technological capacity. They pointed out each claim in a patent can be treated as a separated invention, thus, more claims in a patent indicate a higher value. Patent renewal data (i.e., year of renewal) and patent “family” size can also be used to reflect the value of patents. Patent holders are more likely to pay renewal fees when the economic benefits of patents can cover patent costs (Schankerman and Pakes, 1986). The willingness to file a patent in more countries also indicates that the patent is more valuable (Harhoff et al., 2003; Putnam, 1997). Lanjouw (1998) and Schankerman and Pakes (1985) use patent renewal to build a stochastic renewal model or model of decision to renew that can estimate the rate of innovation. The number of forward citations a patent receives is another widely used indicator for the quality of a patent (Hall, 1999; Trajtenberg, 1990). Forward citations are the number of citations received by a patent. Reitzig (2004) suggested another indicator from the procedure of filing a patent: filing an accelerated examination request during the filing process can indicate the value of a patent.
Griliches et al. (1986) emphasized that forward citations and family size reflect the value of a patent more than any other indicator by comparing all these indicators. Reitzig (2004) summarized and analyzed nine indicators that reflect the economic value of a patent, including lifetime of a patent, patent-breadth, technical non-obviousness, difficulty of “inventing around” a patent, “science linkage” for backward citations, disclosure, portfolio position, uses or functions
30 of patents, and bargaining chips. Which indicators provide the most value depends on the specific industry.
Full-text analysis of patents is increasingly popular, because using the whole abstract or the whole text of a patent is more valuable than simple indicators. Using topic modeling, a computer science technique, can help analyze the full-text of patents by classifying patents, building new measurements for the value of patents, and matching patent treatment groups and patent control groups (Azoulay et al., 2007; Boon and Park, 2005; Kaplan and Keyvan, 2013; Mogoutov and Kahane, 2007; Saiki et al., 2006; UPHAM et al., 2009; Venugopalan and Rai, 2015; Winston Smith and Shah, 2013). Kaplan and Vakili (2013) pointed out that citation analysis can reflect the usefulness of a patent, so they built a new measurement to indicate the novelty of a patent by using topic modeling. Based on Kuhn’s ideas about novelty, patents that introduce new language can be treated as novel, but patents with more forward citations can only be treated as more useful.
Citation data is not only used for determining the value of a patent, but also for tracing knowledge spillover through a patent citation network. Jaffe et al. (1993) is the first paper that uses patent citation to illustrate the importance of localized learning processes. Although Jaffe et al. (1993) faced critiques and questions, using patent citations as a proxy of knowledge flows has been recognized as a valid way to study knowledge spillover. Backward citations of a patent can reflect where knowledge of this patent comes from, while forward citation can be used to trace where the new knowledge goes. Inspired by Jaffe et al. (1993), Caballero and Jaffee (1994), Jaffe and Trajtenberg (1999, 1996) Johnson and Popp (2001), Kelly and Hageman (1999), Sonn and Storper (2003) all use citation analysis to study knowledge spillover. Jaffe and Trajtenberg (1996) and Jaffe and Trajtenberg (1998) focus mainly on citations to university patents and government patents. Kelly and Hageman (1999) studied knowledge clusters by controlling production clusters and found that geographic concentration of knowledge still exists. John et al. (2006) added a time
31 variable into the model and found that knowledge can be spread to larger geographical areas over time.
Current literature about measurements and methodologies show the validation of using patent data and citation analysis. The literature acknowledges the limitations of citation analysis as well. Botolfmaurseth and Verspagen (2002) said, “it should be emphasized that knowledge spillovers are a much broader concept than what is captured by patent citations (US or European)”. For example, citation networks cannot capture tacit knowledge spillover. In addition to patent data and patent citations, researchers also use survey data to study knowledge flows. It is worthy to note that Feldman et al. (2009) verified that geographic proximity is important in knowledge spillover by using survey data. Therefore, if feasible, using both patent data and qualitative methods (i.e., interviews and survey data) is a more robust approach to develop a more holistic, triangulated understanding of knowledge flows.
32 Chapter 3: Local demand-pull policy and energy innovation: Evidence from the solar photovoltaic market in China3
INTRODUCTION
Technological innovation and competitiveness of regional economies are closely connected, driving governments to adopt various policy instruments to support innovation, hoping such support would translate into positive economic effects. A central challenge for governments, though, is to design innovation policies that can effectively influence the speed, direction, and scale of technological innovation. A commonly-used policy tool, especially by sub-national governments, is demand creation. Inspired by Griliches (1957) and Schmookler (1962), the literature started to emphasize that market demand is an important driver for inducing technological innovation. It is now recognized that going from innovation to market demand is not a sequential, linear process, rather it is an interactive, dynamic process whereby the influence (between innovation and demand) can flow both ways.
From a firm’s perspective, the process of innovation involves interactions not only within a firm, but also with other elements outside the firm, such as other firms, consumers, and even the government ((Rothwell, 1989; von Hippel, 1986). Thus, firms’ innovative outcomes cannot be understood well if the environment in which they operate is not taken into account (Antonelli, 1996). This ‘interactive’ view of innovation further raises the question whether local market demand is essential to developing a competitive local industry. Existing literature offers different answers, often at odds with each other, to this question (Dechezleprêtre and Glachant, 2014; Lanjouw and Mody, 1996; Peters et al., 2012; Popp, 2006; Popp et al., 2011).
One area where a strong case for the positive, interactive feedback between geography, innovation, and localized economic effects may be expected involves technologies that have a
3 This chapter is published in Energy Policy. Below is the full citation: Gao, X., & Rai, V. (2019). Local demand- pull policy and energy innovation: Evidence from the solar photovoltaic market in China. Energy Policy, 128, 364- 376. Xue Gao contributed to collecting data, building models, analyzing results, and writing of the manuscript.
33 significant knowledge dependence on local factors such as climate, regulations, building codes, and user preferences. Furthermore, to the extent innovation depends on local knowledge, it may be expected to advantage local firms more, since they presumably have better knowledge of the local context. This possibility should pique the interest of policymakers, who are often interested in seeing that public support for expanding local markets contributes to local technological innovation and economic competitiveness, rather than benefiting other jurisdictions more heavily (Lewis and Wiser, 2007; Maguire et al., 2012). However, empirical evidence in this area is relatively thin. Accordingly, I aim to inform this discussion through robust, empirically-driven analysis. Specifically, using patent data for the distributed solar photovoltaic (PV) market in China from 2005 to 2014, in this chapter I address the following questions: (1) Is there evidence of demand-induced innovation? (2) Does the effect of local demand-pull policy differ from the effect of non-local demand-pull policy on demand-induced innovation? These questions are especially interesting since the essential role of local demand in inducing innovation has been challenged by the fact that China has taken the lead in the PV module manufacturing industry in the absence of a large domestic solar PV market demand when PV manufacturing in China flourished (Huang, 2016; Huenteler et al., 2016; Schmidt and Huenteler, 2016).
This chapter considers solar PV balance-of-system (PV BOS) technologies as a case to address the above research questions. The choice to focus on PV BOS technologies is motivated by two factors: (1) large-scale deployment of distributed solar PV technologies is widely considered to be an important piece in addressing the environmental impacts of the electricity sector, making PV an important technology to study; furthermore, there is a general consensus that the biggest barrier for a wider deployment of PV is the high price of non-module components, including PV BOS (Barbose et al., 2015; Goodrich et al., 2012; Seel et al., 2013; Smith and M. J. Shiao, 2012; Yu et al., 2015), and (2) technological and operational aspects of non-module PV- system components make local conditions (climate, regulation, learning networks, etc.) salient, thus affording a rich empirical setting to explore the geography-innovation-economy interactions
34 at the local level. Although innovations in PV BOS may be generally considered as incremental innovations, they are directly related to solar PV adopters’ experience by making PV installations more user-friendly, safe, and reliable, while also reducing installation costs.
Using original and unique data on innovation in solar PV balance-of-system (PV BOS) technologies in the Chinese distributed solar PV market, this chapter estimates the effect of local demand and non-local demand on technology innovations. This chapter offers two main contributions to the existing literature. First, leveraging a clear demand-pull policy context and careful empirical design, I provide new empirical evidence to reinforce the explanatory ability of the demand-induced hypothesis. A big limitation in previous literature is that many empirical studies do not separate the shift in the demand curve from changes in other conditions (Chidamber, S. und Kon, 1994; Mowery and Rosenberg, 1979; Gregory F. Nemet, 2009), thus confounding the impact of demand on innovation. In this empirical case, the market for distributed PV in China has largely been created by policies, meaning that these market formation policies have led to a significant shift in the demand curve. The clear existence of a policy-driven demand shift can help effectively addresses a key limitation in prior literature. Moreover, strong policy signals, as in this empirical case, mean that installers and manufactures are more readily able to identify and respond to market demand, thus creating the basis for how demand is linked to innovation.
Second, I distinguish between the roles that local demand-pull policy and non-local demand-pull policy play in innovation, thereby highlighting the importance of local demand in driving locally-generated innovation. In particular, I exploit variation in province-level PV BOS patenting activities in China to address the influence of locational aspects of demand on innovation. Most empirical studies that have explored this locational aspect have focused at the national level (i.e., domestic vs. foreign), thus my use of subnational policy and demand variation adds a new perspective to the literature on this topic. I also exploit differences in technological characteristics within PV BOS to explore how the linkage between location of demand and innovation depends
35 upon the characteristics of the involved technologies. Thus my analysis also offers insights into the potential for tailoring market-creation policies based on technological characteristics.
BACKGROUND AND RELATED LITERATURE
Related literature
Scholars from a range of disciplines have increasingly emphasized the role of technological innovation as an engine of economic growth (Schumpeter, 1934). Thus understanding the drivers of and barriers to technological innovation has excited much interest across a range of disciplines for several decades.
The induced innovation hypothesis, proposed by Schmookler and Griliches decades ago, emphasizes that the power of induced innovation lies in market demand (Griliches, 1957; Schmookler, 1962). Utterback further pointed out that market factors have a primary influence on innovation and that 60% to 80% of “important innovations” are demand-induced in a large number of fields (Utterback, 1974). Many recent empirical studies also provide strong evidence to support the demand-induced innovation hypothesis, namely that innovation can be induced by demand
(Lanjouw and Mody, 1996; Lin, 2010; Gregory F. Nemet, 2009; Peters et al., 2012; Popp, 2006b). It is noteworthy that demand-induced innovation is not a simple linear model. Feedback loops and interactions between different actors – between firms, between users and firms, and between governments and firms – are also important elements of technological innovation in the context of demand-induced innovation (Kline and Rosenberg, 1986; Rothwell, 1989; Taylor, 2008; von Hippel, 1986).
A parallel stream of literature has explored the linkages between interactive learning and innovation (Christensen and Stoerring, 2011; Lundvall, 2012; Rothwell, 1989; von Hippel, 1986). Learning and innovation often result from an interactive process, which benefits from proximity, thus experience-based innovation is often characterized by localization and geographical
36 agglomeration. The importance of local context factors (such as demand, interaction, culture, regulation, business environment, and supply chain) in the innovation process has been recognized (Asheim et al., 2007, 2011; Fabrizio and Thomas, 2012; Jaffe and Trajtenberg, 2002; Marshall, 1895; Nemet, 2012; Porter and Stern, 2001). Firms with more experience in a certain geographic area have comparative advantages because of their better understanding of the local context (Beise and Rammer, 2006; Fabrizio and Thomas, 2012; Kogut, 1995) and are stronger for business-to- consumer sectors, as they focus more on end-users and demand-induced innovation in those geographic areas (Brem and Voigt, 2009). Thus, it is plausible that demand-pull policy could expand local markets and help firms gain more localized experience, translating into (demand- induced) innovation. However, policymakers often prefer to see that public support for expanding local markets contributes to local technological innovation and economic competitiveness, rather than benefiting outside jurisdictions (Lewis and Wiser, 2007; OECD, 2011).
Recently some empirical work has begun to distinguish between the effects of local demand-pull policy and non-local demand-pull policy on technology innovation, suggesting substantial variation in these effects depending on the industry/sector of study (See Table 1). For example, Popp (2006) found that innovations for sulfur dioxide and nitrogen oxide control at coal- fired power plants primarily respond to domestic regulation instead of foreign regulation. Dechezleprêtre and Glachant (2014) found that technological improvements in wind power across OECD countries respond to both domestic and foreign demand-pull policies, although they also suggested that the marginal effect of domestic policies is twelve times greater than that of foreign policies. In contrast, Peters et al. (2012) found that both domestic demand-pull policies and foreign demand-pull policies could spur innovation in the solar PV (modules) sector, with similar magnitudes of effect for both foreign and domestic demand-pull policies. On the other hand, Lanjouw and Mody (1996) found that vehicle emissions regulation in the U.S. stimulated innovation in Japan and Germany. Popp et al. (2011) observed that technology innovation in the pulp and paper industry responded to both domestic and foreign regulations.
37
Table 1: Summary of relevant empirical studies to test whether the location of demand matters for inducing innovation.
Industry/Technology Domestic Policy Foreign Literature Policy Sulfur dioxide and nitrogen √ Popp (2006) oxide control Wind technological √ √ Dechezleprêtre and improvements Twelve times greater Glachant (2014) √ Solar PV manufacturing √ Peters et al. (2012) Similar magnitude
Vehicle emissions √ Lanjouw and Mody (1996) Pulp and paper industry √ √ Popp et al. (2011) Note：A tick mark denotes if a study found domestic and/or foreign demand-pull policy to significantly induce (domestic) innovation. The text in the ‘Domestic Policy’ column denotes the relative impact on innovation of domestic vs. foreign demand pull.
The variation in these results regarding the impact of policy locus on demand-induced innovation may depend on whether an industry faces a global market (Popp et al., 2011). For example, the U.S. is a major market for Japanese vehicle makers; thus Japanese technology development is appropriately responsive to emissions regulation in the U.S. The seemingly inconsistent results may also depend on whether local context is significantly important to technology innovations (Peters et al., 2012). If technology localization and producer-user interaction are important in an innovation process, such experience-based innovation might depend on a stable and sizable local market (Qiu and Anadon, 2012). This chapter focuses on PV BOS technologies, for which local factors are known to be more significant than they are to PV modules. Innovations in PV BOS are a product of, among other things, local experience and local learning, as well as local climate, local rooftop characteristics, and local complementary industries (Neij et al., 2017; Venugopalan and Rai, 2015). Therefore, based on existing literature, I expect that PV BOS innovations should primarily respond to local demand-pull policies.
38 Background on distributed PV policies in China
Since 2005 China’s solar PV manufacturing industry has developed rapidly. In 2007 Chinese solar PV manufacturing surpassed Japan, and China became the world's largest producer of PV modules. However, in 2011 the Chinese PV industry slowed down significantly due to shrinking foreign markets, Europe in particular. To help the Chinese PV industry digest overcapacity and get through the downturn, Chinese governments at various level expended much effort to expand the domestic market.
Since 2011 the Chinese national government has established several demand-pull policies to expand the distributed PV market, including a grid-connection policy, Feed-in-Tariff policy, generation subsidy policy, and financial services policy (National Development Reform Commission, 2013; National Grid Company, 2012; State Council, 2013; Zhang, 2016). The deployment of distributed PV in China is also through the following programs: the Golden Sun Demonstration Projects jointly led by China’s Minister of Finance, China's Ministry of Science & Technology and National Energy Agent (Ministry of Finance, 2009); the Green Building Plan led by China’s Ministry of Housing and Urban-Rural Construction (Ministry of Housing and Urban- Rural Construction, 2013; National Development Reform Commission, 2013); the distributed- generation PV application and demonstration projects led by National Energy Agent, and individual self-consumption distributed-generation PV.
39
5000 4500 4000 3500 3000 2500 (MW) 2000 1500 1000
lnstalled capacity of distributed PV PV distributed of capacity lnstalled 500 0 2011 2012 2013 2014
Figure 5: Installed capacity (MW) of distributed PV4 in China, 2011-2014. Data source: NEA (2014), NEA (2015) and Zhang (2016).
At the subnational levels, provincial, city, and local governments made great efforts to expand local market sizes. By the end of 2014, 11 provinces issued feed-in tariff policy and other supplementary subsidies to support solar PV deployment. For example, Jiangxi Province offered the feed-in tariff level at 0.2 yuan/kWh and provincial-specific funding is at 4 yuan/W and 3 yuan/W with different eligibility conditions. Some feed-in tariff policies were combined with local content requirement policies. For example, Shangluo City in Shaanxi province provided 0.1 yuan/W and 0.05 yuan/kWh for solar PV projects using local PV products. It also provided favorable tax policies, specifically a 5% local tax refund to locally registered balance-of-system firms. City-level governments also issued demand-pull policies. Hefei City at the Anhui Province provided additional city-level 0.25 yuan/kWh subsidy for roofs and building integrated solar PV for 15 years and 2 yuan/W for residential PV projects. For most provinces, only one or two cities
4 Distributed PV systems include projects with electricity generation located near users with total installed capacity no greater than 6 MW; the electricity is mainly consumed by end users, but excess electricity can be fed into the public grid (National Development Reform Commission, 2013; National Energy Agency, 2014; National Grid Company, 2012).
40 provided additional subsidies at the city level. But in Zhejiang province, twelve local governments offered feed-in tariffs to local markets, and these city-level subsidies can be combined with national and provincial subsidies.
In addition to feed-in tariffs and other subsidies, provincial and local governments also made great efforts in helping find proper rooftops, clarify rooftop ownership, finance, and install and interconnect distributed PV systems, several provinces, especially the Eastern provinces, provide detailed grid-connection services and financing services to local distributed-generation PV projects. For example, the provincial government in Jiangsu assesses all available and suitable rooftop resources for distributed PV and helps address the problem of unclear rooftop ownership. The provincial government in Jiangsu also requests Grid Company to provide specific services to help costumers connect distributed PV electricity to the grid, so that consumers can sell their excess electricity back to Grid Company. Owing to such coordinated government support, distributed PV has developed rapidly in the Eastern provinces of China. Literature shows that these demand-side policies in China are the major divers of solar PV deployment (Corwin and Johnson, 2019; He and Liu, 2014; Hou et al., 2019; Urban et al., 2018; Zhang et al., 2015; Zhang, 2016; Zhao et al., 2015). Therefore, these demand-side policies and developments yield helpful variation (for my analysis) in the level of demand-pull policies and distributed PV deployment across the Chinese provinces.
Increasing installation of distributed PV in China has significantly expanded the practical experience of PV installers with installing and maintaining distributed PV systems. As discussed above, incremental innovations, such as those in PV BOS, are expected to be closely related to local interactions, local customer preferences, local regulation, local business environment, and local supply chain (Bollinger and Gillingham, 2014; Fabrizio and Thomas, 2012; Maskell, 1999; Neij et al., 2017; Nemet, 2012b; Porter and Stern, 2001). In turn, local firms have comparative advantages that result from a better understanding of the local context (Beise and Rammer, 2006). With expanding markets, installers have more opportunity to learn about the nature of local
41 demand, local roof structures, and local codes and regulations. Such practical experiences provide increased opportunities for installers to improve and update their PV BOS related products and services.
DATA AND METHODOLOGY
Data
To explore the effect of local demand on PV BOS innovation, I build an original database of PV BOS patents filed between 2005 and 2014 related to distributed PV based on patent information from the State Intellectual Property Office (SIPO) website. SIPO contains invention announcements, invention authorization, utility models, and design patents. After SIPO receives an application for an invention, it conducts a preliminary review. If the application passes the preliminary examination, the invention is typically announced 18 months after the filing date. After an invention application is announced, SIPO starts a substantive examination if requested by the applicant. If the application passes the substantive examination, it is authorized to be an invention. There is no substantive examination for utility models and design patents. A utility model is similar to an invention, but standards of novelty and inventiveness for utility models are not as high as those of inventions. A design model refers to new designs for a product’s appearance, shape, and color.
China patent system was established in 1985, marked by the passage of the first patent law. Then Chinese patent law experienced four amendments in 1992, in 2000, in 2008, and in 2016, separately. The biggest change in the Chinese patent system happened in 2000. As China prepared to join the World Trade Organization (WTO), they were required to fulfill the Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPS). While my descriptive analysis covers the Chinese PV BOS patents during 2005 and 2014, there are only few patents before 2008. Thus, the change of China patent system in 2008 should not have a significant impact on my descriptive analysis. Furthermore, my regression analysis covers the PV BOS patents from 2012
42 to 2014, during which there was no significant change in the Chinese patent system. Thus, patent data should be a stable indicator during my study period in regression analysis.
For identifying PV BOS patents for each province in China, I adapt the methodology developed by Venugopalan and Rai (2015), which focused on PV BOS patents in the U.S. The first step is to identify specific items and keywords for PV BOS. Following Venugopalan and Rai (2015), this chapter classifies BOS into four categories: inverter, mounting system, monitoring, and site assessment. Inverter is used to convert power generated by solar PV cells/modules from direct current (DC) to alternating current (AC). Mounting structures are to physically support solar PV modules and/or arrays. Monitoring includes methods and devices that monitor health, connectivity and output of solar PV systems. Site assessment includes methods or mechanisms of appraising the photovoltaic output potential of a land parcel, determining the ideal placement of PV systems on potential sites, determining the extent to which obstructions or shade may impede the output potential of PV sites (Venugopalan and Rai, 2015). All these non-module components are used in the solar PV deployment process. In other words, solar panel/module is used to convert solar energy to DC; however, in order to complete a whole deployment process, non-module components (i.e., PV BOS) are needed for identifying the placement of PV modules, attaching PV modules to rooftops, converting DC to AC, and monitoring the operation and performance of PV modules.
Specific items I include for each category can be found in the appendix A. I use the same search terms as Venugopalan and Rai (2015) in English and translate these English terms into Chinese. Note that one keyword in English could be translated into several relevant terms in Chinese. Table 2 lists keywords I used for searching relevant patents of PV BOS.
43 Table 2: Search terms for PV BOS patents in Chinese and English.
BOS category Search Terms Translation in English
Inverter “逆变器” and “光伏”or “太阳能” "inverter" and "photovoltaic" or "solar"
“支架” or “框架”or “齿条”or“屋顶”or "mounting" or "bracket" or "frame" or Mounting “安装”or “围栏”or “轨道” and “光伏”or "roof" or "installation" or "rail" and “太阳能” "photovoltaic" or "solar"
“跟踪” or “评价” or “诊断” or “维持” or "monitoring" or "evaluation" or Monitoring “保持” or “监视” or “监视器” or “维护” "diagnosis" or "maintain" and or “监控” and “光伏”or “太阳能” "photovoltaic" or "solar" “卫星” or “地图” or “阴影” or “遮蔽” or "satellite" or "map" or "shade" or “估计” or “预测” or “支持” or "estimation" or "projection" or "support" Site “安装”and “光伏”or “太阳能” or "installation" and "solar" or Assessment "photovoltaic"
The second step is to use keywords for each category to search patents’ full text on the SIPO website. This step gives me a large sample of potentially relevant patents, but not all patents I extract from keyword searching are actually relevant to PV BOS. After the second step, my search process yields a total of 7626 potentially relevant patents. The third step is the key step, wherein I review each patent’s claims and identify whether the invention embodied in a patent is directly related to one of the four PV BOS categories. Note that a patent may have been an application in earlier years, but went on to become a granted patent in the following years. I remove these overlapping patents. I also drop the patents that are related only to utility-scale PV, since the distributed and utility-scale markets are largely distinct and innovations in utility-scale PV are not expected to be induced by the demand for distributed PV. In any case, the “utility-scale only” patents that I remove account for a very small proportion of total potentially relevant patents.
After the review in the third step, I retain 2122 patents filed between 2005 to 2014 as being actually relevant to PV BOS innovation, and this set forms the basis of the analysis reported below. The fourth step is to extract the date and the location (province) of each patent. A patent has three
44 important dates: filed date, application announcement date, and invention authorized date. Filing date is the better choice for measuring when an innovation actually took place (Griliches, 1990). I use location of the patentee to measure where an innovation took place. In this study, I include all 31 provinces, municipalities, and autonomous regions in China.
Measurement
I use the number of patents in PV BOS as the dependent variable to measure technology innovation in PV BOS. Though with limitations, patenting activity is frequently used as a meaningful proxy for measuring innovative activity. Numerous studies recognize that patent count is an appropriate and useful measure of innovative activity (Almeida et al., 2002; Choi et al., 2011; Griliches, 1990; Hausman et al., 1984; Jaffe and Trajtenberg, 2002; Lanjouw and Mody, 1996; Schmookler, 1962; Venugopalan and Rai, 2015). Patent data is also widely used in the literature as a measure of innovation in China (Choi et al., 2011; Choi and Williams, 2014; Dang and Motohashi, 2015; Sun and Du, 2010; Wang et al., 2015).
I acknowledge limitations of using patents as a measure of innovation. The number of patents does not map directly on to innovation (Botolfmaurseth and Verspagen, 2002). For example, implicit knowledge cannot be captured by patents. Likewise, not all firms are willing to protect their innovations by applying for patents, because they may not be willing to publicly disclose their ideas (Cohen et al., 2000). While there is no perfect measurement for innovation, given that patent data offer the most systematically compiled, spatially and temporally nuanced data about inventive activity in China (Choi et al., 2011) and the steps I take to check robustness of the results, I believe that patenting activity is an appropriate measure for the purposes of this study.
It is problematic to define the scope of demand-pull policy and categorize specific innovation-support policy as demand-pull policy or not (Kemp and Pontoglio, 2011; Taylor, 2008).
45 Furthermore, the major demand-pull policies (i.e., feed-in tariff policy and subsidies to initial investments) vary a lot across provinces and cities with respect to level, duration, eligibility, and other features, and it is very difficult to aggregate different ways of city-level demand-side subsidies with different level, duration, and eligibility at the provincial level (see details on the background section). Local governments’ efforts in finding proper rooftops, clarifying rooftop ownership, financing, and installing and interconnecting distributed PV systems are even harder to quantify. A commonly-used measure of demand-pull policy in the PV industry is PV installed capacity (Dechezleprêtre and Glachant, 2014; Klaassen et al., 2005; Peters et al., 2012; Rai et al., 2013). In the case of distributed PV market in China, literature shows that market demand in China has mainly been created by government policies, especially at the early stage of solar PV deployment (Corwin and Johnson, 2019; He and Liu, 2014; Hou et al., 2019; Urban et al., 2018; Zhang et al., 2015; Zhang, 2016; Zhao et al., 2015). Accordingly, I believe it is appropriate to use province-level distributed PV installed capacity as a proxy for demand-pull policy. The limitation of this measurement choice is that only three years of data are available in China, because the Chinese government began to promote the distributed PV industry and collect relevant data only since 2011. The limited availability of installed capacity data constrains my modeling choices. I discuss those limitations and how I address them in more detail below. Specific description of the measurement for each independent variable is provided in Section 3.3.
Since innovative activity is influenced by provincial-level innovation capacity and innovation environment (Cohen and Levinthal, 1990; Cooke et al., 1997; Lundvall, 2012; OECD, 2011; Porter and Stern, 2001; Prajogo and Ahmed, 2006), these factors need to be accounted for in my analysis. Provincial-level innovation capacity in China varies a lot (Sun and Liu, 2010; Yang and Lin, 2012). Innovative provinces might be more likely to engage in PV BOS innovation activities; thus, I use total number of patents in each province as a control for overall innovative capacity. The incentives for patenting likely vary across provinces, as provinces may implement various pro-patent policies and/or have variations in accessing the patenting system (Li, 2011). A
46 province with more favorable patent promotion policies is likely to have more patenting activity. Thus the general propensity to patent may also be controlled by the total number of all patents in a province. In addition, the development of complementary sectors could also influence PV BOS innovation activities (Choi and Anadón, 2014; Pisano and Teece, 2007). I use earnings of four sectors including manufacturing, utility, wholesale & retail, and scientific & technical services as controls for development of complementary sectors. Table A4 in appendix A summarizes descriptive statistics of key variables and Table A5 shows the correlation table between the independent variables.
Empirical models
Since the number of patents is a type of count data, two appropriate count models are Poisson regression and negative binominal regression. Additionally, the number of patents for each province in this chapter has over-dispersion. Since the mean and variance of a Poisson regression are the same, Poisson regression cannot be used for data with over-dispersion. However, a negative binominal distribution has two parameters, one of which is an over-dispersion parameter, thus negative binominal regression is a better choice for count data with over-dispersion. To check the robustness of the regression results, I also include the regression results of using Poisson regressions in appendix A.
In my database, variation of provincial level installed distributed PV within province over time only accounts for about 9% of the total variation in provincial level demand, and variation of provincial level installed distributed PV between provinces accounts for 91% of the total variation. Variation of annual added installed capacity within province over three years (2012-2014) may not be enough for an efficient two-way fixed effect model. When variation across provinces is great but variation within each province over time is little, fixed effect estimates would be imprecise and standard errors are likely to be too large to tolerate (Allison, 2009). In this case, a random effect model that takes advantage of both within-province variation and between-province
47 variation may be a better choice. Therefore, I use the Hausman Test to compare a random effect model and a fixed effect model. I estimated the following model by using both fixed effect estimation and random effect estimation:
������ = (u + � ) + β ln ����� + β ln �������� + β ln � + � , (1) ������ = u + β ln ����� + β ln �������� + β ln � + (� + � ), (2) ������ = (u + � ) + β ln ����� + β ln �������� + β ln � + � , (3)
where the dependent variable ������ is the number of PV BOS patents for province i in year t; ln ����� is logged local installed capacity (i.e., for a given province) of distributed PV for province i in year t; ln �������� is logged non-local installed capacity (i.e., all of China minus a given province) of distributed PV for province i in year t, and ln � is a vector of log of time variant control variables (fully listed in table A4 of appendix A) for province i in year t. u is time fixed effects, which could be different for each year. α is the unobserved time-invariant individual effect, which only varies among provinces. � is an error term. This error term is different for each province in each year. For the first fixed effect model (Equation 1), α and u are considered as a part of the intercept, so it allows correlation between α and the predictors. In a random effect model (Equation 2), α is considered part of the error term, so it does not allow any correlation between α and the predictors.
I expect that differences in learning processes exist between subcategories of PV BOS technologies (inverter, monitoring, mounting, and site assessment). Inverter technologies are expected to have the least local nature, compared to the other three subcategories, because inverters are closer to globally-traded products that are relatively easily adaptable to the local context. Therefore, I split the PV BOS patents into inverter and non-inverter (to include monitoring, mounting, and site assessment) patents, meaning that I use the total number of inverter and non- inverter patents as two separate dependent variables to run the regression models. The differentiation between inverter and non-inverter patents allows me to explore the technological
48 dependence of whether or not local demand is a prerequisite for inducing localized innovation. Huenteler et al. (2016), Schmidt and Huenteler (2016), and Quitzow et al. (2017) found that whether or not firms in developing countries without a large home market are able to catch up with developed economies in certain sectors depends on the nature of associated learning processes, suggesting the criticality of the location of demand in the development of comparative advantage. This provides the basis for my inverter/non-inverter split models. I expect that the effect of non- local demand on inverter patents will be much larger than the effect on non-inverter patents.
The inverter/non-inverter splitting serves another important purpose. While I use the total number of patents in a province to control for potential differences in propensity to patent across provinces, the potential differences in propensity to patent across BOS categories should also be taken into account. In principle, there are two reasons (Basberg, 1987; Fontana et al., 2009) that potentially cause differences in patenting propensity across BOS categories. First, it may be hard to codify an innovation as a patent. In the case of PV BOS, inverter technologies are most likely to be codified. But non-inverter technologies, such as installation methods, mounting skills, and personalized design, contain more tacit knowledge. As such they are relatively harder, and thus also less likely, to be codified. As a result, an inverter-related innovation is more likely to be filed for a patent. Second, inventors may believe that their innovations are not innovative enough to be granted patent, leading them to forgo attempting to patent, given the non-trivial effort and costs associated with the patenting process. In the case of PV BOS, non-inverter technologies are more likely to be incremental innovations, compared to inverter technologies, and thus are less likely to be filed as patents. Considering these two reasons, in order to account for the potential differences in propensity to patent across BOS categories, it is appropriate to split patents into inverter and non-inverter patents.
To check whether the regression results are sensitive to the quality of patent data, I conduct robustness checks by including only ‘inventions’, which are known to be of the highest quality
49 among different types of patents in China. In other words, I redefine the dependent variable as the annual number of PV BOS inventions for each province from 2012 to 2014, thus excluding utility models and design patents when constructing the dependent variable. Another purpose of this robustness check is to compare the strength of the effects of local and non-local demand on all patents versus only inventions (high-quality patents).
Finally, in order to further address the limitation posed by the relatively small sample size (n = 90) and the small variation in annual added installed capacity within the provinces over three years (2012-2014), I also develop an alternative model. Since the two-way fixed effect model including province fixed effect and year fixed effect will discard information (variation) across provinces and thus lead to large standard errors, in the alternative model I group provinces and estimate a two-way fixed effect model including year fixed effect and group fixed effect (Equation
3). In Equation 3, α is the unobserved time-invariant group effect. I use three ways to group provinces. The first approach is based on geography, where I group provinces into seven groups: Eastern provinces, Northern provinces, Western provinces, Southern provinces, Central provinces, Northeastern provinces, and Southwestern provinces. The provinces in the same geographic areas share similar climate, culture, business environment, and economic growth conditions. The second approach is based on Gross Domestic Product (GDP). I rank the provinces according to provincial GDP in 2014 and group provinces into six groups according to the ranking: top 5 richest provinces, top 6-10 provinces, top 11-15 provinces, top 16-20 provinces, top 21-25 provinces, and top 26-31 provinces. The third approach is based on a comprehensive competitiveness index according to the province ranking in the “Blue book of China’s provincial competitiveness” (Li et al., 2015). I also group provinces into six groups according to the ranking. Figure A1 to Figure A3 in the appendix A illustrate the three ways to group 31 provinces in China.
50 Omitted variables and potential endogeneity
Omitted variable bias potentially exists, because location-specific variables and institution- relevant variables, such as social value, culture, norms, customs, and institutional structures, are potential confounding variables. However, these confounding variables typically change rather slowly, so I do not expect omitting them to impact the results from the models, which focus only on a three-year period. As such, since these confounding variables may be treated as time-invariant during my study period, they would be adequately controlled for through the province-level fixed- effects in my models.
Reverse causality – PVBOS innovations driving demand over the three-year study period – is not a major concern in this study, because the Chinese government’s support for expanding the domestic PV market since 2011 is driven by a shrinking foreign market, Europe in particular. The effort in creating domestic market demand is to help the Chinese PV industry digest overcapacity and get through the downturn. Put differently, market demand (the independent variable) in this case is genuinely exogenous variation, which is not a result of the development of innovations in the field of PV BOS (the dependent variable) in China.
RESULTS
Descriptive analysis of PV BOS patents in China
The Chinese PV BOS patent dataset illustrates trends for PV BOS innovation activities in China from 2005 to 2014. Figure 6 shows the number of PV BOS patents in China over time. The number of PV BOS patents increased from 2 in 2005 to 567 in 2014. A big jump in the number of PV BOS patents occurred in 2011, increasing by over 100% from 155 in 2010 to 326 in 2011, which is the year the Chinese government began to change policy direction from large-scale PV to distributed PV deployment. Under the Chinese government’s strong support, installations of distributed PV in China increased from 0.6GW in 2011 to 2.3GW in 2012. Utility patents account for a higher proportion of the total number of patents than inventions, which indicates that PV
51 BOS patenting activity in China has tilted more toward incremental improvements rather than disruptive innovation.
Design 600 Utility 500 Invention
400 Application
300
200
100
Annual nmuber of patents (counts) 0 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014
Figure 6: Annual distribution of PV BOS applications, inventions, utility and design (counts).
Figure 7 shows distribution of annual number of patents by category. The overall trend for all four categories is increasing from 2005 to 2014. The relatively flat number of inverter patents since 2012 mainly results from the decreasing number of design patents filed (which refer to new designs for a product’s appearance, shape, and color). The number of monitoring patents and site assessment patents have increased slowly but steadily since 2011. Patenting activity related to mounting is more relevant to the expansion of local markets, since better understanding of local context, such as buildings and rooftops, provides comparative advantage for local firms.
52 350 5000 Inverter 4500 300 Monitoring 4000 Mounting 250 3500 Site Assessment 3000 200 2500 150 2000 100 1500 1000 50 500 Annual number of patents 0 0 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 Distributed PV installed capacity (MW)
Figure 7: Annual number of PV BOS patents including invention application, invention, utility and design (count), by category and installed capacity of installed distributed PV (MW). Note: The left axis is the number of PV BOS patents (count) and the right axis is installed capacity of installed distributed PV (MW).
The geographic distribution of PV BOS patenting activity correlates with the geographic distribution of distributed PV installations in China. Figure 8 and figure 9 show the geographic distribution of both distributed PV installations and number of PV BOS patents. Nine provinces in China have few PV BOS patents during 2012-2014, which coincides with almost zero distributed PV installation in these provinces; these provinces mainly focus on utility-scale PV systems. Five provinces have more than 100 PV BOS patents in China during 2012-2014, three of which (Jiangsu, Guangdong, and Zhejiang) are the top three provinces with highest cumulative distributed PV installations. These richer provinces suffer from frequent electricity shortages, but land-intensive utility-scale PV conflicts with their scarce and expensive land. As a result, distributed PV is a reasonable and necessary choice for these rich Eastern provinces.
53
Figure 8: Geographic distribution of the provinces with less than or equal to 10 PV BOS patents from 2012 to 2014 (blue) and the provinces with more than or equal to 60 PV BOS patents from 2012 to 2014 (red).
Figure 9: Geographic distribution of the provinces with nearly zero MW installed capacity of distributed PV from 2012 to 2014 (blue) and the provinces with more than or equal to 500 MW installed capacity of distributed PV (red).
54 Regression model results
Regression results for Equations 1 and 2 are presented in Table 3. Column 1 shows the regression results of the two-way fixed effect model. I am interested in the effect of local demand and non-local demand on PV BOS innovative activity. The results show that local demand has a positive relationship with local PV BOS patenting activity, but that non-local demand has a negative relationship with local PV BOS innovation. Neither of these two coefficients is statistically significant, which is consistent with my previous expectation that only 9% of variation within provinces over three years in my database could lead to large standard errors. Column 2 shows the regression results of the random effect model. The random effect model takes advantage of variation between provinces and within provinces over time. The result of the Hausman test shows that the coefficients of the two-way fixed effect model (model 1) and the random effect model (model 2) are consistent, but estimates of the random effect model (model 2) are more efficient, which means the random effect model is a more appropriate model. The results of the random effect model show that the effect of local demand for distributed PV in stimulating local PV BOS innovation is positive and statistically significant.
The regression results shown in column 3 of Table 3 try to follow the approach used in Peters et al. (2012). The methodology used in Peters et al. (2012) includes an aggregated time trend variable to reflect three development phases of PV in Europe. For Europe, Peters et al. (2012) defines 1974-1985 as the first time-trend, 1986 to 1994 as the second time-trend, and 1995 to 2009 as the third time-trend, and thus use time-trend fixed effect instead of year fixed effect, which would be much more granular (as is the case with model 1 and model 2 in Table 3). In my case, since (for modeling purposes) I have data for only three years (2012-2014) and all these three years may be taken to be part of the first boom of distributed PV development in China, I take this to mean that I have only one time-trend. Thus, in effect, this means that I do not need a dummy variable for this time trend. Therefore, model 3 includes only the province fixed effect. The results of model 3 are similar to those of model 2. Both models show that local demand (i.e., within a
55 province) can positively stimulate local PV BOS innovations. As expected, the total number of patents in each province is also statistically significant. The total number of patents reflects the overall innovative capacity of each province, which is positively related to the number of PV BOS patents.
Table 3: Regression results for Equations 1 and 2 using different estimation methods.
Model 1 Model 2 Model 3 Fixed effect Random effect Fixed effect 0.121 0.208*** 0.244** Log local demand (MW) (0.127) (0.0770) (0.0987) -3.096 -2.426 -0.893 Log non-local demand (MW) (2.069) (1.879) (0.578) 0.536* 0.505*** 0.453* Log total patent (counts) (0.275) (0.113) (0.257) 0.868 -0.420 0.466 ￥ Log manufacture salary ( ) (2.223) (1.420) (2.107) 0.770 -0.116 0.748 Log utility salary (￥) (0.785) (0.568) (0.800) 2.435* 1.615 2.041 ￥ Log wholesale & Retail salary ( ) (1.336) (1.077) (1.344) Log scientific & technical services -0.657 0.332 -0.957 salary (￥) (0.933) (0.630) (0.944) -14.33 0.454 -19.67* Constant (23.54) (15.64) (11.95) Year fixed effect Yes Yes No Province Fixed effect Yes No Yes Observations 90 90 90 *** p<0.01, ** p<0.05, * p<0.1
Table 4 shows the regression results of the province grouping approach to address the problem that variation within province over time only accounts for 9% of total variation in province-level annual installed capacity. As discussed in data and methodology section, I apply three ways to group provinces and include both year fixed effect and group fixed effect in these three models (Equation 3). Models 4, 5, and 6 (Table 4) show regression results for the model grouping provinces based on province ranking in the “Blue Book of China’s Provincial Competitiveness (2015),” geography, and GDP, respectively. The coefficients of local demand in all three models are positive and statistically significant, ranging from 0.17 to 0.26. The magnitude
56 and standard error of the coefficients of local demand in model 4 and model 6 are consistent with the results in model 2 and model 3. All these models show that only the local demand for distributed PV significantly promotes local PV BOS innovation, and that non-local demand for distributed PV does not affect local PV BOS innovation.
Table 4: Regression results for different grouping methods (Equation 3), based on competitiveness index according to the city ranking in the “Blue book of China’s provincial competitiveness,” geography, and GDP.
Model 4 Model 5 Model 6 Fixed Effects Fixed Effects Fixed Effects Group by index Group by geography Group by GDP 0.257** 0.165** 0.241** Log local demand (MW) (0.107) (0.0757) (0.0952) 1.010 -2.703 -2.847 Log non-local demand (MW) (3.189) (2.379) (3.367) 0.626*** 0.748*** 0.666*** Log total patent (Counts) (0.156) (0.0963) (0.164) 0.271 -1.829* -0.432 Log manufacture salary (￥) (1.400) (1.028) (1.515) -0.161 -0.103 0.00903 Log utility salary (￥) (0.634) (0.610) (0.732) 0.224 0.391 0.00388 Log wholesale & Retail salary (￥) (1.309) (0.976) (1.511) 0.784 1.039* 1.207 Log scientific, & technical services salary (￥) (1.007) (0.553) (0.766) -24.47 18.87 7.743 Constant (24.07) (20.20) (29.10) Year fixed effect Yes Yes Yes
Group Fixed effect Yes Yes Yes
Observations 90 90 90 *** p<0.01, ** p<0.05, * p<0.1.
Table 5 shows the regression results for models where I split the patents into inverter and non-inverter patents. The dependent variables in models 7, 8, and 9 are, respectively, the annual number of all BOS patents at the province level, the annual number of inverter patents at the
57 province level, and the annual number of non-inverter patents at the province level. The coefficients of local demand are consistent with the results of model 1 to model 6 (Tables 3 and 4), further confirming that local demand has an essential role in spurring local innovative activity in PV BOS. While the coefficients of non-local demand are not significant, likely due to limited variation in the non-local demand variable, the opposite signs of the coefficient estimates on inverter and non-inverter patents offer interesting insights. The coefficient for non-local demand on inverter patents is positive, while it is negative on non-inverter patents. This indicates, as expected, that innovations in inverters – a technology that is more globally traded than any of the other BOS categories considered here – could potentially learn from non-local markets. This also helps explain why non-local demand seems to have little or no impact on BOS innovations but positively impact innovations in PV modules. Essentially, previous literature suggests that innovation in PV modules is a phenomenon of global learning, but innovation in PV BOS, on the other hand, is strongly driven by local learning (Neij et al., 2017; Shum and Watanabe, 2008; Wene, 2000; Wiser et al., 2007). Since local learning is predominantly driven by local demand, local learning appears to be the linkage between the positive impact of local demand on PV BOS technologies. In other words, these results suggest that whether or not a local market is a precondition for inducing local innovative activity depends on technology characteristics: local market has an essential role for technologies that have a significant local nature and, thus, depend on local learning for innovations in them.
58 Table 5: Regression results for differentiation between inverter and non-inverter patents. Note that Model 7 is the same as Model 2.
Model 7 Model 8 Model 9
Dependent Variable All BOS patents Inverter patents Non-inverter inventions
Log local demand (MW) 0.208*** 0.267** 0.217** (0.0770) (0.105) (0.0910) Log non-local demand (MW) -2.426 0.998 -3.091 (1.879) (2.826) (2.104) Log total patent (counts) 0.505*** 0.727*** 0.445*** (0.113) (0.147) (0.127) Log manufacture salary (￥) -0.420 1.854 -2.724 (1.420) (1.904) (1.684) Log utility salary (￥) -0.116 -1.444** 0.945 (0.568) (0.732) (0.672) Log wholesale & Retail salary (￥) 1.615 -0.755 2.904** (1.077) (1.414) (1.360) Log scientific, & technical services 0.332 0.260 0.871 salary (￥) (0.630) (0.875) (0.739) 0.452 -14.21 -0.0751 Constant (15.64) (22.61) (17.61) Year fixed effect yes yes yes Observations 90 90 90 *** p<0.01, ** p<0.05, * p<0.1.
Table 6 shows the regression results for models where the dependent variable includes only patent inventions, which are known to be of the highest quality among different types of patents in China. To be specific, the dependent variable in models 10, 11, and 12 are the annual number of all inventions at the province-level, the annual number of inverter inventions at the province- level, and the annual number of non-inverter inventions at the province-level, respectively. These regression results are consistent with previous models (Tables 4 and 5). As before, only local demand has a significant and positive effect on local innovations. Results from these models also indicate that the strength of the effect of local demand on BOS inventions (high-quality patents) (model 10) is somewhat larger than the effect on all BOS patents (model 7). Most interestingly, as
59 expected, the coefficient of local demand on non-inverter inventions (model 12) is much larger compared to inverter inventions (model 11), which may indicate that local demand has a stronger positive effect on non-inverter inventions compared to inverter inventions. This finding needs to be tempered, though, since the coefficient of local demand on non-inverter inventions (model 12) is significant only at the 10% level. Overall, these results suggest that local demand might play a critical role in inducing high-quality local patents in PV BOS. This intuitively makes sense, since larger demand offers greater opportunity to benefit from innovations – and the magnitude of those benefits, in turn, will be larger for higher-quality patents, thus providing a multiplicative incentive.
Table 6: Regression results for including only inventions in the dependent variable.
Model 10 Model 11 Model 12
All BOS Non-inverter Dependent Variable Inverter inventions inventions inventions 0.249** 0.274** 0.391* Log local demand (MW) (0.120) (0.136) (0.221) 1.183 5.704 -3.591 Log non-local demand (MW) (3.289) (3.906) (4.253) 0.700*** 0.907*** 0.564** Log total patent (counts) (0.162) (0.191) (0.236) -0.0484 2.729 -2.804 Log manufacture salary (￥) (2.114) (2.173) (3.556) -0.607 -2.161*** 1.157 Log utility salary (￥) (0.835) (0.749) (1.360) 0.812 -0.657 2.229 Log wholesale & Retail salary (￥) (1.729) (1.728) (2.604) Log scientific, & technical services 0.131 -0.518 1.053 salary (￥) (1.002) (0.863) (1.299) -19.41 -48.18 18.78 Constant (26.28) (30.52) (1012.7) Year fixed effect yes yes yes Observations 90 90 90 *** p<0.01, ** p<0.05, * p<0.1.
60 DISCUSSION AND LIMITATIONS
The results show that for PV BOS in China only local demand significantly induces local PV BOS innovation, while non-local demand has no effect. This emphasizes that local demand could translate into local innovative activity. But why should local PV BOS innovation be spurred only by local market experience (i.e., demand)? I suggest that learning by doing and learning from geographical agglomeration are the two main mechanisms through which such a translation might happen.
On one hand, with expanding local markets, installers have more opportunity to learn about local factors relevant to the design and installation of a PV system, such as local roof structures, codes and standards, supply-chain, and the nature of consumer demand – which, for novel technologies such as PV is expected to have market-specific heterogeneity, especially in early phases of the market. Such local installation experiences provide increased opportunities for installers to improve and update their PV BOS related products and services. On the other hand, with increasing local demand, knowledge gained by each firm operating in the local market contributes to the overall local pool of knowledge. A stronger local knowledge base, in turn, also improves local firms’ innovative potential. Such knowledge is often described as if it were in the air, and local firms benefit from local knowledge just because of “being there” (i.e., without having to expend much effort). Moreover, from the view of competition, in order to compete better in the local market, local competition pushes firms in the local cluster to improve their practices and innovations (Bathelt et al., 2004).
However, the perspective of geographic agglomeration also highlights an inherent limitation of this chapter: my analysis does not fully capture the benefits associated with tacit knowledge that are gained from local clusters, because inventive activities that are captured by patent data mainly embody codified knowledge. Geographic agglomeration is expected to be more important for tacit knowledge, as the spillover of tacit knowledge significantly benefits from face-
61 to-face communication, deep interaction, and even local gossip, and the difficulty in codifying tacit knowledge prevents its diffusion to other locations (Gertler, 2003; Haldin‐Herrgard, 2000; Venkitachalam and Busch, 2012). Thus, using only patenting activity as a measure of innovative outcome in this chapter means that this chapter likely underestimate the effect of local demand on local innovations in PV BOS.
It is worth noting three other limitations of my analysis and how I have tried to address them. First, as noted before, patenting activity is not necessarily a good measure of technological innovation. Previous literature particularly points to concerns about the quality of Chinese patents compared to that of U.S. or European patents. In this chapter, because the timeframe of my analysis is very recent (2012-2014) to effectively detect forward citations, I am unable to weight patent count by using forward citations as a way of addressing the quality of patents. Instead, I conduct a robustness check by excluding utility models and design patents – known to be lower quality patents in China – from my models in order to check whether my regression results are sensitive to the quality of patent data. Dropping lower quality patents does not change the main findings. A second noteworthy limitation of my analysis is that fixed-effects models have limited ability to infer causal relationships, compared to conducting random experiments. Previous literature shows that confounding variables, especially institutional factors and various patenting incentives across provinces, should be carefully controlled for in studies using China patent data. In this chapter, I do not include data on provincial-level institutional factors. However, I believe that these institutional factors, such as social value, culture, norms, customs, and institutional structures, are slow-changing variables. As such, these factors may be expected to not change significantly within three years (2012-2014, the focus of my analysis). Thus, I include province fixed-effects in my models to control for these (time invariant) variables. Thirdly, PV installed capacity is a commonly-used proxy measure for demand-pull policy in the PV industry (Dechezleprêtre and Glachant, 2014; Peters et al., 2012; Rai et al., 2013). This is partly because it is problematic to define the scope of demand-pull policy and categorize specific innovation-support policies as
62 demand-pull policy or not (Kemp and Pontoglio, 2011; Taylor, 2008). Moreover, the major demand-pull policies, namely feed-in tariff policy and subsidy to the initial investment, vary a lot across provinces and cities with different level, duration, and eligibility. It is even harder to quantify local governments’ efforts in finding proper rooftops, clarifying rooftop ownership, financing, and installing and interconnecting distributed PV systems. Yet, I acknowledge that using distributed PV installed capacity at the province level as a proxy for local demand pull is not the most direct way to measure demand-pull policies. Future work could try to address this limitation by leveraging the variation of specific demand-pull policies province-by-province to improve measurement of the local demand-pull variable.
CONCLUSION AND POLICY IMPLICATIONS
In this chapter I have addressed the following question: is local demand essential in inducing local innovation? my empirical design exploits the subnational variation in (policy- created) demand to study the impact of geographically-differential (local vs. non-local) demand on local innovation. Specifically, I use a new patent dataset for photovoltaic balance-of-system technologies at the province level in China to study the effect of local demand and non-local demand on province-level PV BOS innovation.
Geographic distribution of distributed PV installations is unequal across the 31 provinces in China, but it is strongly correlated with geographic distribution of PV BOS patents. The consistency between the explosion of PV BOS patents and local demand is especially telling in the case of mounting technologies: to enable adequate physical support over the life of the PV system (20-30 years), mounting equipment need to be designed for local climatic conditions and variations therein, rooftop structures, and framing and roofing materials. Thus, greater local demand is expected to strongly drive innovations in mounting technologies to cater to that local demand, as I indeed find to be the case.
63 I find that only local demand for distributed PV significantly contributes to local PV BOS innovation in China. This suggests that local distributed PV markets created by demand-pull policies (of both provincial and local governments) significantly promotes (only) local innovative activity in PV BOS. The effect of non-local demand on patenting activity in inverters, while insignificant, is large and positive, but is negative on patenting activity in non-inverter BOS technologies. This suggests, consistent with expectations, that innovations in inverters may benefit from non-local markets, given the more fungible nature of inverter technologies. As expected, overall innovative capacity at the provincial level – as measured by all (i.e., not just PV BOS) patenting activity within each province – significantly and positively affects PV BOS innovation. The significantly different effects of local demand and non-local demand indicate that local context (including local knowledge and local experience) is a key element in the learning and innovation process for PV BOS. The importance of local knowledge and local context in demand-induced hypothesis has a valuable market implication: to the extent innovation depends on local knowledge, it advantages local firms more, since they presumably have better knowledge of the local context. In such cases local demand-pull policies may be more politically feasible, since policymakers would view the potential economic benefits from market creation to be more appropriable by local actors.
Furthermore, my findings serve to inform that drivers of innovation depend on market and technology characteristics. PV BOS helps emphasize that within the same technology (an integrated PV system, here) the technological trajectories of different subcomponents (modules, inverters, etc.) potentially interact differently with local market conditions. Previous studies have suggested that different learning processes are likely at play for PV cell/module versus PV BOS (Neij et al., 2017; Shum and Watanabe, 2008; Wene, 2000; Wiser et al., 2007). PV module innovation is a phenomenon of global learning, and there is no clear boundary between local PV module learning and global PV module learning. However, PV BOS innovation is strongly driven by local learning, which comes from accumulated learning of installations in specific, local
64 markets (Neij et al., 2017). The differences in technology characteristics and the associated learning processes help explain the apparent inconsistencies of my findings with the results of Peters et al. (2012). Peters et al. (2012) analyzed innovation that is directly related to PV cells and modules. They noted that user feedback and producer-user interaction are weak for PV cells and modules – a situation that does not hold for PV BOS technologies, the focus of my study.
Moreover, differences in learning processes are not limited to just between PV cell/module and PV BOS technologies; rather, those differences exist even between subcategories of PV BOS technologies (inverter, monitoring, mounting, and site assessment). Indeed, I find that for the highest-quality patents (i.e., inventions in the Chinese patent system) local demand has a stronger positive effect in inducing innovation in non-inverter BOS technologies compared to inverters. This is consistent with the fact that mounting structures and site assessment technologies have the most prominent local nature. In the case of inverters, on the other hand, although local demand is still a significant factor in spurring innovation, I also find weak support that inverter innovations in China may have been driven by non-local markets. Overall, these findings suggest that whether local demand is an essential precondition to local industry development (i.e., innovations) depends on technological characteristics and the nature of associated learning processes, highlighting the need to open the “technology black box”. In other words, a better understanding of the impact of market-creation policies on the geography innovation necessitates accounting for technology characteristics and in particular how those characteristics accentuate certain learning processes over others.
My findings also insightfully inform the discussion on ‘green industry development’. The ultimate goal of green industrial policies is to integrate multiple policy goals including, but not limited to, environmental protection, energy security, technological innovation, and economic competitiveness (Rodrik, 2014; Schmidt and Huenteler, 2016). My results suggest that demand- pull policies for the distributed PV market in China have significantly shaped local innovative
65 activities. However, the important role of local demand-pull policy in inducing innovation has been challenged by the fact that China has taken the lead in the PV module manufacturing industry even in the absence of a large domestic solar PV market demand when PV manufacturing in China flourished (Huang, 2016; Huenteler et al., 2016; Schmidt and Huenteler, 2016). These arguments have raised a critical question to the literature on green industrial policy: is local demand essential in the development of green industry? My study suggests that the answer to this question depends, at least in part, on technology-specific factors. Specifically, by comparing technology subcategories within PV BOS as well as comparing PV BOS technologies with PV module, I find that technology characteristics are critical factors in explaining the role of local demand in green industry development. If learning and innovation processes depend on local context, then local demand is a prerequisite for developing a local competitive industry. Moreover, most empirical studies that have explored this locational aspect have focused at the national level (i.e., domestic vs. foreign), thus the use of sub-national policy and demand variation adds a new perspective to the literature on this topic. Importantly, my study hints that the arguments typically made in the discussion regarding green industry development may also be applicable to sub-national levels of policy-making. Demand-pull policy could be an important strategic element in the local policy mix of developing green industry, particularly, for technologies that have a significant local nature. By the same token, it also might be difficult for lagging provinces/states to close the technology innovation gap without developing a stable local market demand. This perspective of looking at the sub-national level is especially important for large countries, such as China, but it may be less relevant to small countries.
The success of subnational governments’ efforts in expanding market demand and translating it into local innovative activities underscores the potential viability of a bottom-up approach to renewable energy governance, at least for some technologies. The localized economic gains from expanding local market demand could mobilize a large number of subnational actors in support of advancing certain renewable energy technologies locally. Thus, if the central/federal
66 government is slow to move on environmental policies, or only focuses on supporting export-led growth, subnational policymakers could still take voluntary, interest-driven actions to expand the market demand for renewable energy, because they would view the potential economic benefits from market creation as more appropriable by local actors. These subnational actions could be voluntary and interest-driven, thus potentially underlying a viable mechanism for the bottom-up approach, even if not coordinated.
Finally, one fruitful direction for further work could be development of a framework to explain the conditions under which local demand is a prerequisite for developing a competitive local industry. These conditions include, but are not limited to, technology characteristics, the degree of market maturity, the development stage of technology, local policies, and local institutional ecosystem. This broader framework can help generalize the results of solar PV BOS to other technologies. A broader framework that situates the role of local demand in inducing local innovation could further help explain relevant mechanisms of localized learning.
67 Chapter 4: Does Geography Matter? A Study of Technological Learning and Innovation in the Solar Photovoltaic Balance-of-Systems (PV BOS) Industry
INTRODUCTION
The importance of geography in technological learning and innovation is an important component of endogenous growth theory, innovation theory, and industrial economics. However, the importance of geography as an element of technological learning continues to be debated, within the underlying context that modern transportation and communications infrastructure – key enablers of deepening globalization – make long distance no longer a barrier in technological learning and innovation. One situation where there is presumably a strong case for the positive, interactive feedback between geography and innovation involves technologies whose future development trajectory significantly depends on local factors (e.g., climate, regulations, administrative codes, user preferences, etc.). Technologies in the deployment process are an example of such a situation, as the features of such technologies are associated with the locations in which they are applied and deployed. In such cases Schaeffer et al. (2004) pointed out that “effective learning only takes place when both deployment is increased and the process of how deployment leads to learning is better understood”. Previous literature only offers a limited understanding of the linkages between the geography of learning and innovation outcomes for technologies in the deployment process – a key research gap in the literature (Schaeffer et al., 2004; Strupeit and Neij, 2017). The salience of this topic is ever greater today, when governments across the world are struggling to balance the opposing pulls of inward-looking, protective approaches to economic growth and outward-looking, collaborative, and open approaches to economic growth. Speaking to these theoretical, empirical and policy opportunities, this chapter focuses on learning and innovation processes of green technologies in the deployment process.
The subject of my empirical inquiry is distributed solar photovoltaic (PV) systems. Large- scale deployment of distributed solar PV technologies is widely considered to be an important
68 piece in addressing the environmental impacts of the electricity sector, making PV an important technology to study. Over the last few years solar PV has been the fastest growing electricity generation technology globally, accounting for about 47% of all newly added renewable energy installations in 2016 (REN 21, 2017). However, PV installed costs need to come down further as solar PV enters the next phase of much broader deployment. The hardware for a solar PV system is typically clubbed in two broad categories: modules – a collection of cells that convert solar irradiation on direct current (DC) and non-module or “balance-of-system” hardware – all hardware in a PV system other than the PV modules. Most of the extant literature about solar energy learning and innovation focus on PV modules, and only several research has begun to pay attention to non- module hardware (i.e., PV balance-of-system) innovation and non-module costs (Neij et al., 2017; Nemet et al., 2016; Shum and Watanabe, 2008; Strupeit and Neij, 2017).
This small body of literature emphasizes the technological learning and innovation processes of non-module hardware greatly differ from innovations of PV module. Shum and Watanabe (2008) identify significant differences, driven primarily by the associated geography of learning processes, when comparing the technological innovation processes of PV module and non-module (balance-of-system) hardware. Technological learning in PV modules is widely considered to be a global phenomenon. However, PV balance-of-system (PV BOS) technologies, which enable the deployment of PV modules in a specific local context (say, a rooftop), typically embody local characteristics (Shum and Watanabe, 2008; Strupeit and Neij, 2017) associated with the local deployment context (e.g., local climate, local rooftop structure, and local regulations and codes, etc.). Thus, localized learning may be expected to be critical for innovations in PV BOS technologies. However, there is very little empirical literature exploring the extent to which localized learning is relevant for PV BOS innovations and the comparative contribution of local vs. non-local knowledge in value added to PV BOS innovations. As a result, studying the technological learning and innovation process of PV BOS contributes to completing the whole picture of technological innovations in the solar PV industry.
69
To address this research gap, in this chapter I empirically study the role of geography in the technological innovation of non-module solar energy technologies (PV BOS) using patent citation analysis on a unique database of PV BOS patents in the U.S. issued between 2000 to 2014. My analysis of PV BOS technologies addresses three questions: 1) how does geographic distance influence knowledge spillover, 2) how technological characteristics influence knowledge acquisition choices as regards local knowledge vs. non-local knowledge? and 3) how do local and non-local knowledge affect the value (or quality) of innovation? The remaining discussion is structured as follows: Section 2 provides theoretical background on geography of innovation and technological learning literature; Section 3 describes data collection and methodology used in this chapter; Section 4 presents results and relevant discussions; and Section 5 summarizes the main insights that emerge from my results, describes the limitations of this chapter, and discusses relevant policy implications.
THEORETICAL BACKGROUND
Studies of geography of innovation have recognized that innovation activity and knowledge flows are geographically localized. Although firms have access to convenient transportation and instant communication tools, geography is still an important element in technological learning and innovation. Because of specialized skilled labor pools, an infrastructure of complementary industries, and proximity to a strong knowledge base that facilitates knowledge exchange, firms can significantly benefit from locating in certain geographic regions (Marshall, 1890).
The role of geography is also an important, but understudied, topic for technological learning and innovation in PV BOS. Shum and Watanabe (2008) raised several key questions about whether technological learning and innovation of PV BOS is local or global in character. Their findings support the notion that PV BOS learning (innovation) spillover is local. Neij et al. (2017)
70 reinforced these finding and identified the need for local learning in solar energy technology deployment. They also described what is learned locally, who learns and how, and how spatial proximity supports learning. However, these descriptive studies cannot quantify that to what extent the localized learning is important to PV BOS learning and innovation.
Geographic proximity facilitates knowledge spillover and network-building, and can also shape local networks (Bottazzi et al., 2002; Gertler, 2003; Jaffe et al., 1993), as evidenced by the positive relationship between geographical proximity and closed networks. Trust and close collaboration are easier to build in the same geographic area. Networks developed through interactions with local universities, suppliers, utilities, manufacturers and end-users are useful to firms’ technological learning and innovation (Grübler, 2003; Helfat, 2006; Junginger et al., 2005; Salter and Laursen, 2006; Schneider et al., 2008). Laursen et al., (2012) understood the localized network and norms as social capital and found a positive relationship between geographically bound social capital and firms’ innovative ability. In the field of PV BOS, Strupeit and Neij (2017) studied mounting technologies and cost dynamics in Germany, finding that mounting technology learning processes are highly local in nature. The authors argue that local rooftop structure, local building codes and the needs of local residents are significant explanatory factors behind the highly local nature of learning processes, and that interactions between mounting equipment producers and installers are key in technological learning. Based on prior literature, I expect that PV BOS innovation, especially mounting and site assessment innovation, will absorb more knowledge from the local geographic area. Alternatively, this hypothesis indicates that geographic distance might be a barrier to knowledge spillover and firms at peripheral regions might have insufficient local knowledge, which may negatively impact local firms’ innovation activities. Relatedly, it is of interest to explore whether accessing non-local knowledge is able to compensate insufficient local knowledge. These illustrations of the limited, but emerging, literature around how geography impacts learning and innovation in PV BOS highlight the need to better understand how local and
71 non-local knowledge influence technological innovation, and whether non-local knowledge can compensate for insufficient local knowledge.
Geographical proximity also facilitates local firms in forming their practice community. Collective and interactive learning are key concepts in the “community of practice” literature (Wenger, 1998). Firms working in the same geographic area and similar industries can form a community where they share knowledge, information, and activities. These industry connections and interaction enhance the stock of information and knowledge firms possess, improving the firms’ practice (Brown and Duguid, 1991; Wenger, 1998). The view of “community of practice” literature is consistent with the operations in the solar mounting/racking market. Morrris et al., (2013) compared the mounting/racking process between Germany and the U.S., including pre- installation, racking attachment, module installation, on-roof electrical and off-roof electrical, and found that practices in the U.S. are different from practices in Germany. For example, during the pre-installation process, U.S. firms routinely conduct off-site preparation that can improve productivity, while Germany firms do not conduct off-site preparation. This process illustrates how geographically rooted communities of practice have formed in relation to PV BOS technologies.
Localized learning and the significance of geography indicate the importance of local knowledge in innovative activities. Local knowledge tends to be more closely aligned with local firms’ technological demand (Arza, 2010) and easier to understand and absorb (Qiu et al., 2017), because they are created and applied in the same context. Generally, local knowledge tends to be more homogenous (Jacobs, 1969). Asheim and Isaksen (2002) found that too much focus on local knowledge may lead to path dependence, and geographic proximity makes learning networks fixed and dependent. An effective way to break path dependence is to absorb non-local knowledge, which tends to be more novel and comprehensive than local knowledge (Asheim and Isaksen, 2002). Thus, knowledge at different geographic levels should have different strengths in
72 supporting innovation. Following Schumpeter, innovation researchers believe that diverse knowledge, especially knowledge from external industries and geographically distant places, is a valuable source for innovation, especially disruptive innovation (Jacobs, 1969; Nelson and Winter, 1982; Ahuja and Lampert, 2001; Arthur, 2007; Fleming, 2001; Nemet, 2012). Owen-Smith and Powell (2004) studied knowledge network in the Boston Biotechnology Community and found that interaction with the “pipeline” (non-local knowledge) was the source of remarkable innovation. Burt (1992) and Grabher (2002) pointed out that local knowledge is highly homogenous and highly stable, thus extra-local sources have an important role in innovation. Bathelt et al. (2004) built a theoretical framework to illustrate that each firm can benefit from network ties outside local clusters and information or knowledge obtained from the “global pipeline” can also benefit local clusters through “local buzz”.
As noted above, technologies associated with solar PV deployment embody various locational characteristics, emphasizing the importance of localized learning in developing valuable and relevant innovations. Thus, I hypothesize that the knowledge acquisition when conducting PV BOS innovations has a significant local dependence, which means that the degree of knowledge acquisition decreases with geographic distance. Within the field of PV BOS, technologies regarding mounting and site assessment are hypothesized to have a higher degree of local dependence when acquiring knowledge and conducting innovation, because compared to technologies regarding inverter and monitoring, those technologies are more related with local context. Although local knowledge is expected to be more relevant in conducting PV BOS innovation, due to different advantages and disadvantages of knowledge in various geographic areas and the importance of knowledge diversity on innovation, I hypothesize that non-local knowledge will be an important driver of quality in PV BOS innovations.
The searching processes for localized learning and non-localized learning differ, which has valuable policy implications. Since it is relatively easy to build trust and interact with others in the
73 same geographic area, localized learning results just by “being there.” Bathelt et al. (2004) go so far as to state “localized learning can largely take care of itself”. On the other hand, trust, collaboration, and networks are relatively difficult to establish over long distances, and searching extra-local information and knowledge requires more effort, cost, and complicated processes than local information and knowledge (Bathelt et al., 2004). Moreover, absorbing non-local knowledge also requires firms to have high absorptive capacity (Qiu et al., 2017). Absorbing non-local knowledge might be a more serious impediment for firms that locate outside the technological core areas, have low absorptive capacity, and are in industries that face greater technological uncertainty (Levy and Murnane, 2004; Sonn and Storper, 2003).
Overall, the discussion above implies that it is necessary and valuable to understand different roles of local and non-local knowledge in technological innovation when policymakers design and prioritize innovation policies. If local knowledge and networks are essential for technological innovation, in order to promote technological innovations, policymakers may want to focus on how to mobilize local resources and help local firms establish collaboration with each other; if non-local knowledge and networks also play important roles in technological innovation, policymakers could further help local firms establish “global pipelines” and provide more institutional support for non-localized learning.
DATA AND METHODOLOGY
Data
I used keyword search strings to develop an original database of PV BOS patents in the distributed PV market from the United States Patent and Trademark Office (USPTO) website, registered between 2000 and 2014. I include four technological areas in PV BOS, including inverter, monitoring, mounting, and site-assessment. Inverters convert power generated by solar PV cells and modules from direct current (DC) to alternating current (AC). Mounting is used to physically support solar PV modules and/or arrays. Monitoring includes methods and devices that
74 monitor health, connectivity and output of solar PV systems. Site assessment includes methods and mechanisms for appraising the photovoltaic output potential of a land parcel, determining the ideal placement of PV systems on potential sites, and the extent to which obstructions or shade may impede the output potential of PV sites. The specific technologies included within each of the four main BOS components can be found in the Supplemental information (SI).
The data collection steps are as followings. The first step is to identify specific items and keywords for PV BOS (See table 7). The second step is to use keywords for each category to search patents’ full text on the USPTO website. This step gives me a large sample of potentially relevant patents, but not all patents I extract from keyword searching are actually relevant to PV BOS. Next, the third step is to review all these patents’ claims and identify whether the invention embodied in a patent is directly related to one of the four PV BOS categories. Further, the same patent may have been an application in earlier years but went on to become a granted patent in the following 2 to 3 years. I remove these overlapping patents for my descriptive analysis. The fourth step is to extract the assignee, the date and the location of each patent. The data collection steps are the same as Gao and Rai (2019).
Table 7: The specific items and keywords for PV BOS patent data collection BOS category Key words Inverter "inverter" and "photovoltaic" or "solar" Mounting "mounting” and "photovoltaic" or "solar" Monitoring "monitoring" or "monitor" and "photovoltaic" or "solar" "satellite" or "map" or "shade" or "estimation" or "obstruction" or "potential" Site Assessment or "site" or "terrain" or "assess" or "azimuth" or "survey" and "solar" or "photovoltaic"
I further used USPTO application programming interface (API) to collect backward citations and forward citations for each PV BOS patent in the database, and for each backward citation and forward citation, my database includes the application year, the location of an assignee, and the longitude/latitude of the location of an assignee. Citations indicate that there are intellectual
75 property claims the new patent yields to. Backward citations are patents that are cited by a patent, while forward citations are patents that cite the focal patent. Taking “patent A” as an example, backward citations of the patent A are patents cited by the patent A, and forward citations of patent A are patents that cite the patent A in the future. By exploring the knowledge flows from backward citations to the focal patent then to the forward citations (the whole knowledge spillover chain), this chapter is able to study firms’ technological learning and information acquisition behaviors. The link between a focal patent and each backward citation can also be understood in the context of the underlying citation network; as such, I assess the effect of geographic distance on the citation networks (i.e., any pairs between PV BOS patents and backward citations).
Methodology
Gravity model
Jaffe et al. (1993) was the first paper to use patent citations to illustrate the importance of localized learning processes. They compare an original database to a control database using t-tests, finding a significant localization effect, that patents are more likely to cite patents that come from the same country and state. Following Jaffe et al. (1993), using patent citations as a proxy for knowledge flows has been recognized as a valid way to study knowledge spillover (Caballero and Jaffee, 1994; Jaffe and Trajtenberg, 1999, 1996; Johnson and Popp, 2001; Kelly and Hageman, 1999; Sonn and Storper, 2003; Johnson et al., 2006; Nemet and Johnson, 2012; Nemet 2012). Backward citations are used to indicate the knowledge sources of the focal patent, and the number of forward citations indicates the value (or quality) of the focal patent (Abrams and Popadak, 2013; Gambardella et al., 2008; Lanjouw and Schankerman, 2004; Schoenmakers and Duysters, 2010).
The first research question in this study asks how geographic distance influences knowledge spillover. I use patent citations to capture knowledge spillover and gravity model to quantify the impact of geographic distance on knowledge spillover. Equation (1) and (2) are the
76 corresponding empirical models. The dependent variable is the number of linkages between the location (the location is a state in the U.S.) of the assignee of a U.S. PV BOS patent and the location of the assignee of all the backward citations in the U.S. PV BOS patent database at a certain year.
In other words, the dependent variable � is the number of citations that the “location a” cites the “location b” at year t, which indicates the scale (number of citation linkages) of knowledge flows from location b to location a. I built all possible pairs between locations of U.S. PV BOS patents and locations of all the backward citations of U.S. PV BOS patents, thus the dependent variable is count data with “excess zero”. Therefore, I apply zero-inflated negative binominal regression model (ZINB) to estimate the model.
Equation (1) is derived from the gravity model, which is one of the most successful empirical models in international economics (Frankel and Rose, 2002). The gravity model is widely used to predict international trade flows that is proportional to economic sizes (or populations) of importer and exporter, and is inversely proportional to the distance between importer and exporter. In the citation-based network, firms in a “location a” can be treated as knowledge importers and firms in a “location b” can be treated as knowledge exporters. Following Morescalchi et al. (2015), the total number of linkages that are attached to a “location a” indicates the size of “location a”, the total number of networks that are attached to a “location b” indicates the size of “location b”. The variable “�������� ” indicates the geographic distance between location a and location b; the variable “state” is a dummy variable, and if state equals 1 means that two regions of this network are in the same state in the U.S.; and the variable “border” is also a dummy variable that indicates whether two regions of this network are both in the U.S. I also include year dummies to control for time trends, and include dummies that indicate the country of “location b” to control for unobserved characteristics of cited countries, especially any unobserved variables that affect the underlying propensity of being a location of any given backward citation. In the equation 2, I include one more control variable, which is used to control for a set of alternative potential cited locations in the knowledge acquisition process using a proxy index
77 called “remoteness”. Head (2003) provides the formula to calculate the index of “remoteness” that measures a location’s average weighted distance from all the potential knowledge spillover partners.
Equation (1)
� = � + � ∗ log(���� ) + � ∗ log(���� ) + � ∗ log(�������� ) + � ∗ ����� + � ∗
������ + ∑ � ∗ ���� + ∑ � ∗ ������� + �
Equation (2)
� = � + � ∗ log(���� ) + � ∗ log(���� ) + � ∗ log(�������� ) + � ∗ ����� + � ∗
������ + � ∗ ���������� + ∑ � ∗ ���� + ∑ � ∗ ������� + �
Where ���������� = ∑ /
T-tests
I used t-tests to assess how the degree of importance of localized learning depends on technological characteristics. Therefore, I calculate the proportion of backward citations that are in the same state or country for the following technological categories: inverter, mounting, monitoring, site assessment, inverter and monitoring, and mounting and site assessment. I expect that technologies with more local nature use more local knowledge (local knowledge is defined as knowledge from the same state or the same country). Since mounting and site assessment technologies tend to be more rooted in the local context (Shum and Watanable, 2008; Neij et al., 2017), I expect that mounting and site assessment technological innovations use more local knowledge compared with monitoring patents and site assessment patents.
Backward citations are sometimes added by examiners during the patent filing process rather than the applicant. It is examiners’ job to add all relevant previous work. Alcácer and
78 Gittelman (2006) and Cotropia et al. (2013) argued that backward citations added by patent holders better reflect knowledge spillovers. A caveat of this inference, however, is that patent holders may avoid citing some prior art, such as prior patents that could block their invention, thus it is also reasonable to argue that including all the backward citations added by both patents holders and examiners better reflects the whole body of prior knowledge (Atal and Bar, 2010; Lampe, 2012). In order to reduce the noise of patent citation data, I use two types of backward citations, total backward citations and patent holder added backward citations, to conduct t-tests.
Regression model
The third research question asks: how do local and non-local knowledge effect the value (or quality) of technology innovation? Equation (3) is the corresponding empirical model, where
� is the number of forward citations received by patent i within 5 years of patent filing. Self- citations, where the focal patent and the forward citation have the same assignee, are excluded from the forward citations. In order to check robustness, this chapter also uses the number of forward citations received by patent i within 4 years and 6 years of patent filing as independent variables.
Because the dependent variable is count data, and the variance is much larger than the mean, I use the negative binomial regression model. I also include the regression results of Poisson regression models in appendix C to check whether the results are sensitive to the underlying distribution. In equation (3), SameState is the number of backward citations that are in the same state as patent i; SameCountry is the number of backward citations that are in the same country as patent i; DiffCountry is the number of backward citations that are in different countries than patent i; �����_�� is a dummy variable indicating whether patent i’s assignee is a startup or established firm; �� is a dummy variable indicating that whether patent i’s assignee is located in
California; ���ℎArea indicates whether patent i refers to inverter, mounting, monitoring, or site assessment technology; ������������ is solar PV installations (MW) in the assignee's state j;
79 ������� is mean of backward citation lags for patent i; ��������� is a dummy variable indicating whether patent i is associated with a government contract; � is the fixed effect of application year.
The coefficients of SameState , SameCountry , and DiffCountry can tell me how knowledge from different geographic levels impact the patent’s value. The variation of the number of same state citations, same country citations, and different countries citations can be treated as exogenous variations, because cited patents are just all previous relevant knowledge for the innovative parts of the focal patent. Compared to the backward citations from different country, the same country backward citations are local knowledge, and same state backward citations are local knowledge compared to the backward citations from different states but in the same country.
Following Keijl et al. (2016), I built another indicator that measures the geographic diversity of each patent’s backward citations. If a backward citation comes from the same state as the focal patent, it is weighted with “1”; if a backward citation comes from a different state as the focal patent, but in the same country, it is weighted with “2”; If a backward citation comes from a different country as the focal patent, it is weighted with “3”. Equation (4) is the empirical model that uses the “geographic diversity” to replace the count of different level backward citations. In the Equation (4), all the control variables are the same, except that a new control variable, the total number of backward citations, is also added.
Equation (3) Y = β + β ∗ SameState + β ∗ SameCountry + β ∗ DiffCountry + β ∗ CA + β ∗ Start_up + β ∗ ���ℎArea + β ∗ ������������ + β ∗ ������� + β ∗ ��������� + � + �
Equation (4) Y = β + β ∗ Geo + β ∗ TotalCiations + β ∗ CA + β ∗ Start_up + β ∗ ���ℎArea + β ∗ ������������ + β ∗ ������� + β ∗ ��������� + � + �
80 In addition, this chapter uses longitude/latitude data to define geographic distances directly. The specific information of longitude and latitude allows me to give a fine grain approach to define geographic distance, rather than dividing distances into different political divisions, such as state and country. The geographic longitude and geographic latitude can determine a point on the earth, and the shortest distance between two points on the surface of the earth, that is great-circle distance, can be calculated by using the following formula (known as Haversine formula).
where ∅ and ∅ are geographic latitude of two points; Δ∅ is the difference between these two geographic latitudes; � and � are geographic latitude of two points; Δ� is the difference between geographic longitudes of these two points. Δs is the central angel between these two points, and the distance (the arc length) between two points is the product between Δs and the radius of the Earth (r approximately equals to 6371 km). I classify the geographic distances into six ranges, including less than 1000km, between 1000km and 3000km, between 3000km and 5000km, between 5000km and 7000km, between 7000km and 9000km, and more than 9000km. Then I separately count the number of backward citations within each distance category for each PV BOS patent. The estimation model that incorporates these variables can be found in the equation (5).
81 �������� (5)
Y = β + β ∗ (citations < 1000km) + β ∗ (1000�� < ��������� <= 3000��) + β ∗ (3000�� < ��������� <= 5000��) + β ∗ (5000�� < ��������� <= 7000��) + β ∗ (7000�� < ��������� <= 9000��) + β ∗ (citations > 9000km) + β ∗ CA + β ∗ Start_up + β ∗ ���ℎArea + β ∗ ������������ + β ∗ ������� +
β ∗ ��������� + � + � where citations < 1000km, 1000km < citations <= 3000km, 3000km < citations <= 5000km, 5000km < citations <= 7000km, 7000km < citations <= 9000km, citations>9000km indicate the number of backward citations in each distance category.
The patent database is truncated because patents issued in different years vary in the number of potential years in which forward citations can occur. For example, a patent that is filed in 2000 has 15 years to receive forward citations until 2015, but a patent that is filed in 2010 only has 5 years to receive forward citations until 2000. Therefore, this chapter forces a fixed 5-year window for each patent to receive forward citations, which is justified by previous studies (Gilsing et al., 2008; Noailly and Shestalova, 2017). In other words, this chapter only counts the number of forward citations that are received by the focal patent within 5 years. A longer forward citation window might be better (a 10-year window is also frequently used), but the number of patents is very limited in the early 2000s. If I force a 10-year window, the sample size is too small to conduct an effective regression analysis. In order to check the robustness, this chapter also conducts regression analyses for a 4-year window and for a 6-year window of receiving forward citations.
RESULTS AND DISCUSSION
Descriptive analysis
My patent citations database includes the focal patents, their backward citations, and their forward citations. Backward citations are patents that are cited by the focal patent. The backward citations of a patent can indicate the knowledge source (previous knowledge). In other words, backward citations refer to where knowledge comes from when conducting patents. Forward
82 citations are patents that cite the focal patent in the future, which indicates where knowledge goes to (knowledge diffusion). There are 526 U.S. patents between 2000 to 2014 as being relevant to PV BOS innovation, and this set forms the basis of the analysis reported below. Figure 10 illustrates the trend of PV BOS patents in the U.S. by technological category over time. Interestingly, figure 10 shows a fall in PV BOS patenting activities since 2009. It is mainly due to the decrease of monitoring patents. The fall in inverter and site assessment patenting activities occurred around 2010 and 2011, while the number of site assessment patents were always very limited. But the trend of mounting patents is relatively steady over time. It indicates that the decreasing trend is more important for globally-traded products (i.e., inverters and monitoring), which supports the conclusion in Carvalho et al. (2017), namely the decrease is driven by severe market competition from China. I further explored the reason behind the trend of patenting activities by examining the number of assignees. Figure 11 shows that the annual number of assignees by technology. Solar inverter and monitoring have experienced a decreasing trend in the number of assignees since 2009 and 2011, respectively, but the general trend of the number of assignees in mounting is increasing. The different trends of these two types of technologies (inverter and monitoring vs. mounting and site assessment) support that many producers of globally-traded products have exited the market which might be driven by competition from China. But the number of market players who heavily depend on local market scale is increasing with the expansions of local market size.
83 90 80 70 60 50 40 30 20 10 0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014
inverter monitoring mounting site assessment
Figure 10: Annual distribution of PV BOS patents by technological category (counts).
30
25
20
15
10
5
0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014
Solar Inverter Site Assessment Solar Monitoring Solar Mounting/Rack
Figure 11 Annual distribution of assignee numbers by technological category (counts).
84 Table 8 and table 9 show the results of knowledge sources (backward citations) and knowledge diffusion (forward citations) of U.S. PV BOS patents. Table 8 shows the percentage of backward and forward citations for U.S. PV BOS patents located in different countries. Table 9 shows the percentage of backward citations and forward citations located in different states within the U.S.
Table 8: Comparison between the percentage of knowledge source and percentage of knowledge diffusion of U.S. PV BOS patents across different countries Percentage of Percentage of forward Country backward citations citations (knowledge source) (knowledge diffusion) U.S. 63% 66% Japan 20% 2% Germany 6% 2% France 3% 2% Canada 2% 0.4% China Taiwan 1% 2% U.K. 1% 1% South Korea 1% 1% Switzerland 1% 0.4% Netherlands 1% 0.3% Israel 0.2% 20% Other 3% 3%
85 Table 9: Comparison between the percentage of knowledge source and percentage of knowledge diffusion of U.S. PV BOS patents across different states in the U.S. Percentage of Percentage of forward State in the U.S. backward citations citations (knowledge source) (knowledge diffusion) California 24% 43% New York 7% 8% Michigan 7% 2% Illinois 7% 2% Ohio 5% 5% Massachusetts 5% 3% Colorado 4% 10% Pennsylvania 4% 2% New Jersey 4% 1% Texas 3% 6% Washington 3% 2% New Mexico 3% 1% Arizona 2% 2% North Carolina 2% 1% Connecticut 1% 2% Kansas 1% 1% Delaware 1% 1% Virginia 1% 1% Rhode Island 1% 1% Vermont 0.2% 2% Indiana 0.1% 1% Other 15% 3%
The majority of both backward and forward citations for U.S. PV BOS patents come from U.S. patents (63% and 66%, respectively). This suggests that these patents also mainly contribute to subsequent U.S. patents. Japan was the foreign country with the largest share of knowledge inflows to U.S. PV BOS patents. Twenty percent of U.S. PV BOS patent backward citations are from Japan, while Japanese patents among all the forward citations only cite 2% of these U.S. PV BOS patents. Among all foreign countries, Israel has the largest share of knowledge inflows from the U.S. Only 0.2% of U.S. PV BOS patents backward citations come from Israel, but Israel
86 accounts for about 20% of these U.S. PV BOS patents’ forward citations, which indicates that Israel contributes little knowledge to U.S. PV BOS patents, but absorbs a lot of knowledge from U.S. PV BOS patents.
Within the U.S., California provides the largest share of knowledge sources (24%) for U.S. PV BOS patents, while California also absorbs the largest share of knowledge (43%) from U.S. PV BOS patents. In other words, compared to the knowledge source (where the knowledge comes from), U.S. PV BOS patents contribute more to subsequent innovation in California. Colorado, Texas, and New York also have positive knowledge inflows, which means that these states absorb a larger share of knowledge from these PV BOS patents compared to the share of knowledge that they contribute to these PV BOS patents. Most of the states have positive knowledge outflows, which means the knowledge they contribute to the U.S. PV BOS patents is more than the knowledge they absorb from these PV BOS patents.
Results and discussions
Distance and border effect on citation networks
The regression results of gravity models are shown in table 10. Model 1 is the standard gravity model that only includes the sizes of two regions and the geographic distance between these two regions. The model 2 include two more dummy variables. The variable “within the same state” indicates that both “location a” and “location b” are in the same state in the U.S. (i.e., within same-state citations), and the variable “within the U.S.” indicates that both two locations are in the U.S. (i.e., within same-country citations). In model 3, in order to control for unobserved characteristics of cited locations, especially any unobserved variables that affect the propensity to be a location of any backward citation, I include all the dummies that indicate each cited region. In model 4, I use the proxy index “remoteness” to control for a set of alternative locations of knowledge source. In model 5, I add both cited-country fixed effect (as in model 3) and the index
87 “remoteness” (as in model 4) to double-control for the issue of citation propensity. While the self- citations in the backward citations only accounts for 1.4%, and the correlation between total links and the links without self-citations is as high as 99.16%. I still check whether the results are robust to self-citations. Therefore, in model 6, I exclude self-citations from backward citations, meaning that the dependent variable in the model 6 measures the number of linkages between the location of the focal patent and the location of the backward citations without self cites.
In general, the results are consistent with my hypotheses that the number of (citation) linkages between two regions is proportional to the knowledge size of each region and inversely proportional to the geographic distance between these two regions. The coefficient of “distance” shows that the number of citation-related linkages decreases with distance. The coefficient of “within the same state” shows that there are significantly more citation-related linkages between regions that are in the same state, implying a localization effect at the subnational level. However, the coefficient of “within the same state” becomes insignificant when the self-citations are excluded from backward citations, which indicates that the state-level localization effect may be driven by self-citations. But the coefficient of “distance” is still positive and significant, indicating that firms acquire more knowledge from the locations that closer to their own locations. In terms of the national border, I do not find any evidence supporting that national border has an impact on knowledge acquisitions.
88 Table 10: Regression results of the gravity model
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Methods ZINB ZINB ZINB ZINB ZINB ZINB 1.122*** 1.114*** 1.115*** 1.112*** 1.112*** 1.118*** Log(size_a) (0.0147) (0.0147) (0.0146) (0.0121) (0.0121) (0.0122) 0.893*** 0.874*** 0.894*** 0.901*** 0.911*** 0.919*** Log(size_b) (0.0226) (0.0221) (0.0200) (0.0178) (0.0187) (0.0187) -0.0912*** -0.0776*** -0.0972*** -0.0511** -0.0603** -0.0710*** Log(distance) (0.0251) (0.0246) (0.0295) (0.0211) (0.0235) (0.0233) 0.895*** 0.877*** 0.892*** 0.883*** 0.125 Within the same state (0.127) (0.129) (0.129) (0.130) (0.116) 0.00706 -0.102 0.0334 -0.0226 -0.0436 Within the U.S. (0.0448) (0.0722) (0.0442) (0.0642) (0.0642) 4.78e-09*** 4.72e-09*** 4.77e-09*** Remoteness (6.12e-10) (6.05e-10) (6.07e-10) -4.201*** -4.310*** -4.065*** -4.568*** -4.447*** -4.478*** Constant (0.316) (0.316) (0.358) (0.308) (0.329) (0.311) Year fixed effect yes yes yes yes yes yes Cited region fixed effect yes yes yes inflate part
-0.0544 -0.0723 -0.0828* -0.0820* -0.0873* -0.0848* Log(size_a) (0.0448) (0.0447) (0.0450) (0.0454) (0.0460) (0.0468) -25.04*** -22.89*** -23.15*** -24.60*** -25.52*** -24.96*** Log(size_b) (1.102) (2.326) (2.918) (1.086) (1.131) (1.084) -0.213*** -0.256*** -0.232*** -0.234*** -0.221** -0.226*** Log(distance) (0.0785) (0.0850) (0.0857) (0.0858) (0.0869) (0.0869)
-18.07*** -18.04*** -18.05*** -20.50*** -19.43*** Within the same state (6.366) (5.443) (5.537) (5.232) (1.723)
-0.886*** -0.937*** -0.873*** -0.897*** -0.914*** Within the U.S. (0.188) (0.194) (0.195) (0.199) (0.200) 3.132*** 3.740*** 3.591*** 3.501*** 3.419*** 3.435*** Constant (0.639) (0.695) (0.699) (0.706) (0.711) (0.711) Num of observation 68791 68791 68791 68791 68791 68791 *p<0.1; **p<0.05; ***p<0.01
89 T-test results
Although I find the evidence of localized learning in the PV BOS innovations, I argue that the degree of the importance of localized learning may depend on technological characteristics. I use t-test to assess the different degrees of localized learning based on technological characteristics. Tables 11 to 14 show t-test results for the proportions of backward citations that are in the same state and in the same country (the U.S.) for the following categories: inverter, mounting, monitoring, site assessment, inverter and monitoring, and mounting and site assessment. All these categories are implemented as binary variables. Group 1 means that the value of these variables is
1, and group 0 means that the value of these variables is 0. The mean for group 1 for each category is the mean of the proportion of backward citations where the focal patent and the cited patent are in the same state (table 12 and table 14) or in the same country (table 11 and table 13).
The results in table 11 and table 12 show that, compared to non-inverter patents (group 0 column in the ‘inverter’ row in tables 11 and 12), patents related to inverters use significantly less U.S. knowledge, and less knowledge from the same state. Note that the non-inverter group primarily includes ‘mounting’ and ‘site assessment’ patents (as well as ‘monitoring’ patents). Compared to non-mounting patents, patents related to mounting use significantly more U.S. knowledge, but there’s no significant difference in the use of same-state knowledge. The same holds true for site assessment patents. Furthermore, I combined inverter and monitoring patents into one group and mounting and site assessment as the balancing group. Compared to inverter and monitoring patents (group 0 column in the ‘inverter and monitoring’ row in tables 11 and 12), patents related to inverter and monitoring use significantly less U.S. knowledge, and less knowledge from the same state. In other words, compared to inverter and monitoring patents, patents related to mounting and site assessment use significantly more U.S. knowledge and same- state knowledge. Overall, these results indicate that technologies that are more associated with the local context (mounting and site assessment) absorb more knowledge from local areas. It indicates
90 that localized learning is more important for technologies that are more associated with the local context.
Because backward citations are added by both patent holders and examiners, there are disagreements about whether all backward citations (added by both patent holders and examiners) or only backward citations that are added by patent holders better reflect the relevant previous knowledge (Alcácer and Gittelman, 2006; Cotropia et al., 2013; Thompson, 2006). Therefore, I also conduct t-tests for the variables that only include backward citations added by the patent holders (see table 13 and table 14). Most of the results are consistent with the results in tables 11 and 12. Compared to mounting and site assessment patents, inverter and monitoring patents use less same-state and same-country knowledge. Startups use less same-state knowledge, and firms located in California use significantly more knowledge from within California.
Table 11: T-test results of the proportion of total backward citations (added by both applicants and examiners) that are in the same country as the focal patents for different categories Same country citations Mean for group 1 Mean for group 0 t-statistic p-value Inverter 0.58 0.69 4.34 0.00*** Mounting 0.69 0.62 -2.80 0.01*** Monitoring 0.69 0.64 -1.22 0.23 Site Assessment 0.74 0.65 -1.79 0.10* Inverter and Monitoring 0.61 0.70 3.28 0.00***
*p<0.1; **p<0.05; ***p<0.01
Table 12: T-test results of the proportion of total backward citations (added by both applicants and examiners) that are in the same state as the focal patents for different categories Same state citations Mean for group 1 Mean for group 0 t-statistic p-value Inverter 0.16 0.23 2.31 0.02** Mounting 0.23 0.19 -1.39 0.17 Monitoring 0.23 0.20 -0.69 0.49 Site Assessment 0.25 0.20 -0.76 0.47 Inverter and Monitoring 0.13 0.20 2.32 0.02**
*p<0.1; **p<0.05; ***p<0.01
91 Table 13: T-test results of the proportion of backward citations (only added by applicants) that are in the same country as the focal patents for different categories Same country citations Mean for group 1 Mean for group 0 t-statistic p-value added by applicants Inverter 0.49 0.55 1.28 0.20 Mounting 0.61 0.47 -3.60 0.00*** Monitoring 0.42 0.55 2.14 0.04** Site Assessment 0.42 0.53 1.11 0.29 Inverter and Monitoring 0.47 0.60 3.09 0.00***
*p<0.1; **p<0.05; ***p<0.01
Table 14: T-tests results of the proportion of the backward citations (only added by applicants) that are in the same state as the focal patents for different categories Same State Citations Mean for group 1 Mean for group 0 t-statistic p-value added by applicants Inverter 0.11 0.19 2.65 0.01*** Mounting 0.20 0.13 -2.26 0.02** Monitoring 0.17 0.16 -0.12 0.90 Site Assessment 0.17 0.16 -0.16 0.87 Inverter and Monitoring 0.13 0.20 2.32 0.02**
*p<0.1; **p<0.05; ***p<0.01
The effect of local/nonlocal knowledge on patent quality
Regression results for equation 3, 4, and 5 are presented in table 15. Model 1, model 3, and model 5 are results for equation 3, which uses the count of backward citations that are at different geographic levels as the key independent variables. Model 2, model 4 and model 6 are results for equation 4, which uses geographic diversity to measure the geographic proximity between the focal patent and backward citations, instead of the count of backward citations that are at different geographic levels. Model 7 to model 9 in table 15 are results for equation 5, which uses the direct distances as independent variables to define local and non-local knowledge sources.
92 In order to check the robustness of setting a fixed window for receiving forward citations, I force a 4-year window for receiving forward citations in model 1 and model 2, and a 5-year window for receiving forward citations in model 3 and model 4, and a 6-year window for receiving forward citations in model 5 and model 6. The results in table 15 show that the number of same- state backward citations is negatively associated with the number of forward citations. The results are consistent across the different time-windows. If a patent cites one more backward citation from the same state, this patent will receive 0.047 less forward citations within 4 years, 0.046 less forward citations within 5 years, and 0.045 less forward citations within 6 years. This means that absorbing more knowledge from the same state is significantly associated with lower value (or subsequent consequences) of this patent. Note that the count of forward citations in table 15 does not include self-citations (I present results that include self-citations in the count of forward citation in the Appendix C for checking whether the results are sensitive to self-citations. I find that the results are not sensitive to self-citations.)
The coefficients of “different countries” are positive and significant, and the results are consistent across different windows of receiving forward citations. This result points out that if a patent cites more backward citations that are from a different country, then more forward citations are received by this patent. The different coefficients for the 4-year, 5-year, and 6-year results are reasonable, as patents receive more forward citations in a longer period. The opposite directions of the coefficients of “same state” and “different countries” demonstrate the importance of knowledge diversity in PV BOS innovation in terms of geographic levels.
Models 2, 4, and 6 use geographic diversity instead of the number of backward citations from different geographic areas, and control for the total number of backward citations. The results are also consistent across different time windows for receiving forward citations. The coefficients of “geographic diversity” are positive and significant, which shows the importance of absorbing diverse knowledge from different geographic areas, and the results are consistent with the results
93 in models 1, 3, and 5. The results in table 16 are consistent with model 1 to model 6 in table 15, showing that the number of short-distance citations (particularly within 5000km) is significantly negative associated with the patent's value, while the number of long-distance citations (particularly more than 9000km) is significantly positive associated with the patent's value.
The results of key independent variables are consistent across 4-year, 5-year and 6-year window, which shows that the results are not sensitive to the number of years over which a patent receives forward citations. Moreover, that the results are consistent across the three types of distance measurements shows that the results are not sensitive to how the distance metric is specifically defined.
94 Table 15: Regression results of equation 3 and equation 4 using different windows for receiving forward citations (exclude self-citations)
Dependent variable
Forward Citations within 4 Forward Citations within Forward Citations within years 5 years 6 years Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 0.001 -0.004 -0.017 Same Country (0.012) (0.013) (0.015) 0.028* 0.044** 0.060*** Different Countries (0.017) (0.017) (0.020) -0.047** -0.046* -0.045 Same State (0.023) (0.024) (0.029)
Geographic 0.038*** 0.047*** 0.058*** Diversity (0.014) (0.013) (0.016) 0.158 0.145 -0.038 -0.038 0.247 0.244 Inverter (0.400) (0.399) (0.417) (0.417) (0.468) (0.467) 0.322 0.321 0.141 0.140 0.294 0.284 Mounting (0.370) (0.370) (0.402) (0.402) (0.458) (0.458) -0.100 -0.099 -0.396 -0.398 -0.189 -0.201 Monitoring (0.413) (0.413) (0.436) (0.436) (0.485) (0.485) 0.484*** 0.485*** 0.297* 0.297* 0.065 0.058 Startup (0.165) (0.165) (0.164) (0.164) (0.182) (0.182) -0.190 -0.195 -0.428 -0.426 -0.227 -0.190 California State (0.297) (0.296) (0.443) (0.441) (0.450) (0.447) 0.007 0.007 0.002 0.002 0.0004 0.0001 Num of Claims (0.007) (0.007) (0.007) (0.007) (0.007) (0.007) 0.051 0.059 0.132 0.132 0.205 0.188 Govint Contract (0.303) (0.302) (0.294) (0.293) (0.314) (0.313)
Total Backward -0.081** -0.098*** -0.123*** Citations (0.032) (0.031) (0.036) 0.001 0.001 0.004* 0.004* 0.003 0.003 Installation (0.001) (0.001) (0.002) (0.002) (0.003) (0.003) Mean Lag of -0.008 -0.007 -0.017 -0.017 -0.030 -0.033 Backward Citations (0.018) (0.018) (0.019) (0.019) (0.021) (0.021) 2.318* 2.340* 3.346*** 3.345*** 3.464*** 3.447*** Constant (1.354) (1.352) (1.199) (1.197) (1.199) (1.198) Year fixed effect yes yes yes yes yes yes Observations 351 351 261 261 201 201 *p<0.1; **p<0.05; ***p<0.01
95 Table 16: Regression results of equation 5 using different windows for receiving forward citations (exclude self-citations)
Dependent variable
Forward Citations Forward Citations Forward Citations in 4 years in 5 years in 6 years
Model 7 Model 8 Model 9 -0.036** -0.026 -0.022 citations < 1000km (0.017) (0.017) (0.020) -0.023** -0.026* -0.022 1000km < citations <= 3000km (0.012) (0.013) (0.014) 0.004 -0.045* -0.094*** 3000km < citations <= 5000km (0.026) (0.026) (0.033) 0.115 0.013 0.118 5000km < citations <= 7000km (0.078) (0.074) (0.085) 0.025 0.036 0.055** 7000km < citations <= 9000km (0.023) (0.022) (0.026) 0.091*** 0.099*** 0.085*** citations > 9000km (0.027) (0.028) (0.031) -0.001 -0.011 -0.010 Mean Lag of Backward Citations (0.019) (0.019) (0.022) -0.038 -0.061 0.372 Inverter (0.406) (0.419) (0.474) 0.282 0.265 0.488 Mounting (0.377) (0.403) (0.464) -0.368 -0.355 -0.365 Monitoring (0.424) (0.439) (0.497) 0.469*** 0.306* 0.001 Start-up (0.171) (0.166) (0.188) -0.166 -0.509 -0.319 California State (0.315) (0.440) (0.463) 0.005 -0.0002 0.0004 Num of Claims (0.007) (0.007) (0.007) 0.111 0.146 0.169 Govint Contract (0.307) (0.292) (0.315) 0.001 0.004* 0.005* Installation (0.001) (0.002) (0.003) 2.469* 3.490*** 3.246*** Constant (1.362) (1.189) (1.194)
Observations 346 261 196 *p<0.1; **p<0.05; ***p<0.01
96 CONCLUSION AND POLICY IMPLICATIONS
Recent literature has started to emphasize the importance of local experience, local markets, local policy, and other local contexts in solar PV deployment. It has also started to emphasize the role of the local nature in PV balance-of-system technologies. However, there is no systematic, quantitative study addressing the extent to which localized learning is happening and its importance in PV deployment. At the same time, most studies usually omit the value of non-local knowledge and, thus, fail to evaluate empirically and directly the role of non-local knowledge in PV BOS innovation. In other words, it remains unclear which factors influence the degree of localized learning, the extent to which local and non-local knowledge contribute to innovation, and how geography influences firms’ knowledge acquisition. To address this research gap, in this chapter I investigated firms’ learning and innovation processes in relation to geography.
One main contribution of this chapter is that the results of the gravity model help answer “how local is local” in the localized learning of U.S. PV BOS industry. Previous descriptive studies have discussed the local nature of technological learning. Strupeit and Neij (2017) studied mounting technologies and cost dynamics in Germany, finding that mounting technology learning processes are highly local in nature. Morrris et al., (2013) compared the mounting/racking process between Germany and the U.S., and found that these technological practices are geographically bounded. My results qualify these important early findings. I find that citation-related network has a localization effect at state level, while the state-level localization effect may be driven by self- citations. The country border has no effect on citation-related networks. I also find that the number of citations-based linkages decreases with distance. This result provides a direct evidence that the technological learning of PV BOS is a local phenomenon, and the localized learning is at state level in the U.S. Furthermore, this chapter provides empirical evidence that knowledge acquisition is associated with technology characteristics. Technologies that are more related to local context use more local knowledge, which indicates that technologies with a more significant local nature have a higher dependence on localized learning.
97
The above results are consistent with the local search literature. The basic conclusion of the local search literature is that knowledge searching behaviors are localized geographically, technologically, and organizationally (Jaffe et al. 1993, Stuart and Podolny 1996, Rosenkopf and Nerkar, 2003). The community of practice literature (Brown and Duguid, 1991; Wenger, 1998; Bathelt et al., 2004) also supports this view. This chapter empirically demonstrates that, at least in the case of PV BOS, knowledge searching behaviors are geographically bounded. Local search behavior is understood to be primarily driven by two sets of factors – relational and organizational. First, trust, collaboration, and knowledge share are relatively easy to build in the same geographic location, which is evidenced by positive relationship between geographical proximity and closed networks (Grübler, 2003; Helfat, 2006; Junginger et al., 2005; Salter and Laursen, 2006; Schneider et al., 2008). Second, local searching behavior is reinforced by organizational routines, which means that because of organizational routines, easily established localized learning becomes continuing and persistent organizational behaviors (Nelson and Winter, 1982). Besides relational and organizational reasoning, this chapter points to a third possible driver for local search: technology characteristics. Technologies that are more rooted in the local context (i.e., local climate, local building code, and local regulations, etc.) tend to have a higher degree of dependence on localized learning for innovation. Because local knowledge is created and applied in the same context, this knowledge tends to be more closely aligned with local firms’ technological demand and easier to understand and absorb (Arza, 2010, Qiu et al., 2017).
Moreover, firms’ persistent localized searching behavior reflects the value of local knowledge, which is easy to access and absorb, and firms’ innovative activities benefit from local knowledge. Consistent with the territorial agglomerations (clustering) literature, I find that innovation is a geographically embedded process, suggesting the vital role of geographical regions in enhancing both firm-level innovative outcomes and potentially even region-specific advantages. This leads to the following question: how to build a strong local knowledge base and a cohesive
98 local knowledge network to develop local green industries? For technologies that have a significant local dependence and interactive learning processes, such as PV BOS, learning by doing is the key mechanism to build the initial knowledge base, and persistent learning by doing will provide a strong knowledge base for the local industrial cluster. The realization of significant learning by doing effect requires a stable and large size of local demand (Bathelt et al., 2004, Gao and Rai, 2018, Neij et al., 2017), meaning that expanding local demand is one way to build local knowledge. In turn, the solid pool of local knowledge improves local firms’ innovative potential, and aggregative local innovative behaviors can translate to region-specific advantages.
Another important contribution of this chapter is to add the international dimension to the literature regarding on localized learning. Previous studies that focus on localized learning tend to overemphasize the role of geographic proximity against the international dimension. I argue that geographic proximity is unlikely to be the full story in understanding both the generation and the eventual value of innovation. I find that international knowledge significantly contributes to high- quality patents, emphasizing the value of knowledge diversity in technological innovation and rejecting the notion that local knowledge in PV BOS innovation is of exclusive importance. The importance of searching for knowledge from distant areas may because knowledge from a broader set of technological areas and geographic areas can help fill in some gaps in current innovations (Rosenkopf and Nerkar, 2003).
However, trust, collaboration, and relational networks are relatively difficult to establish over long distances, and searching extra-local information and knowledge requires more effort, cost, and complicated processes than local information and knowledge (Bathelt et al., 2004). Furthermore, the skills and routines required to recombine knowledge from distant areas are also different, which challenges firms’ absorptive capacity. The knowledge spillover barriers may be expected to be more serious for startups firms and firms located in peripheral regions, as reaching non-local resources might be more challenging for these firms. One key policy implication is that
99 an essential element in building local innovative capacity is to help local firms go beyond localization, and reach out for broad, distant, and unique knowledge. This need is certainly present for PV BOS, for which there is strong localization at the state-level, yet non-local knowledge is essential for creating higher-quality innovation. Policymakers may want to pay attention to whether local firms face difficulties in accessing non-local knowledge, and help local firms bridge non-local networks with local networks. For example, local governments or industry associations can organize more formal or informal activities that facilitate knowledge spillover between local and non-local firms. Moreover, building a prosperous local industry cluster is a long and difficult process. The value of non-local knowledge indicates an alternative way that local policymakers can start by helping local firms build networks with firms in other areas – this can partially compensate for insufficient local knowledge, especially in early phases of cluster development.
Patent data and citation analysis have several limitations in reflecting knowledge spillovers. In particular, knowledge spillovers that can be captured by citation analysis of patent data mainly represent codified knowledge. The knowledge-absorbing behavior may be different between tacit knowledge and codified knowledge, and the degree of the importance of knowledge diversity for innovation quality may differ as well. Therefore, future studies could more directly focus on knowledge flows and learning networks associate with tacit knowledge. Key questions of interest are: 1) whether, how often, and how local installers use local and non-local tacit knowledge; 2) how installers evaluate the value of local and non-local tacit knowledge in their innovation efforts; and 3) whether firms that locate in non-core regions suffer from insufficient local tacit knowledge and how that influences firms’ knowledge acquisition behavior and innovation outcomes.
100 Chapter 5: Separating multiple learning mechanisms that facilitate cost reductions: Evidence from the U.S. solar PV industry
INTRODUCTION
The large-scale deployment of distributed solar photovoltaic (PV) technologies is widely considered to be an important piece in addressing the environmental impacts of the electricity sector, while increasing resilience of the electricity system and energy access. However, the prices of solar PV technologies are not the most competitive compared to other energy resources, therefore, a big challenge in achieving a wider deployment of solar PV technologies in the future is to reduce the installation prices. Although solar costs have come down dramatically over the last decade, costs need to come down even more to realize the full promise of solar. There is a general consensus that the next big tranche of cost reductions in solar will come not from hardware costs (cells, module, etc.), but rather from “non-hardware costs”, including installation, racking, permitting, financing, overhead, marketing, and other non-hardware costs associated with installing solar systems (SEIA, 2018). In the case of the U.S. distributed solar PV industry, non- hardware costs account for about 70% of installed costs, indicating that lowering solar PV system prices to facilitate solar energy deployment hinges on non-hardware cost reductions. Compared to
Germany, China, Italy, Japan, and Australia, residential and non-residential PV system prices are higher in the U.S., and the disparity in non-hardware costs is the primary reason (Barbose et al., 2015).
While studying non-hardware cost reductions has triggered scholars’ interests, so far this important area is understudied. Previous literature on solar costs has focused on how market structure, installer scale, demographics, ownership, policy, and system components influence solar PV installation price (Gillingham et al., 2014; Nemet et al., 2016). However, the unexplained variation is still large (>60%) after controlling for these factors, which suggests the need for further work to identify the potential for cost reductions of solar PV installations. To address this research opportunity, this essay aims to further explore the role of technological innovation, knowledge
101 spillover, and upstream-downstream collaboration in reducing non-hardware costs of solar PV installations with a focus on separating different mechanisms through which deployment policy potentially facilitates cost reductions.
The direct impact of deployment policy (i.e., renewable portfolio standards, feed-in tariff, tax credits, rebates) on the market is to expand market demand, which also increases firms’ installation experience. The ultimate goal is to induce technological changes and cost reductions along the learning curves (Hoppmann et al., 2013; Jaffe et al., 2002; Popp, 2019). For example, cost reduction goals could be achieved through the learning by doing mechanism. Learning by doing refers to the phenomenon that accumulation of experience leads to cost reductions (Arrow, 1962b; Nemet, 2012b; Qiu and Anadon, 2012). Although cost reductions through experience accumulation is an expected outcome, the level of cost reductions may vary across different firms given a certain volume of experience accumulation. What factors, other than just learning by doing, can explain diverse levels of cost reductions resulting from experience accumulation?
This essay focuses on separating different learning mechanisms through which deployment policy can facilitate cost reductions. In the case of solar PV installers, learning by doing, learning by searching, and learning by interacting are three major learning mechanisms. When an installer increases its installation experience, the installer can learn from its repeated working on technologically similar installations to improve its productivity and thus reduce its installation costs. This is the learning by doing mechanism. Installation experience can also motivate new working processes, installation methods, or related technology innovations and stimulate private research and development investment, and thus lead to cost reductions. This learning mechanism is learning by searching. Furthermore, as installation-service providers, installers have to work, for each solar PV system, with products from upstream suppliers (i.e., panel manufacturers and inverter manufacturers) and assemble and install these manufacturers’ products (i.e., panel, inverters, and other hardware) on the rooftops. The upstream-downstream (panel/inverter
102 suppliers-installers) interactions are not a simple seller-buyer relationship. These upstream suppliers can provide installers with training and knowledge regarding installation technique, product information, new technologies, and other useful information about the solar PV industry. Therefore, the process of accumulating experience is also a process of building networks and learning from these upstream suppliers, and this learning mechanism is learning by interacting. Most of the extant studies mainly focus on verifying the learning-curve relationship for cost reductions in manufacturing. This study contributes to explore whether the argument that the accumulation of experience leads to cost reductions can be applied to service industries. As illustrated above, upstream suppliers are important learning sources for service industries, and thus through studying a service industry, this chapter contributes to incorporate supply chain collaboration theories in the learning-curve relationship.
Specifically, this study aims to address the following two research questions: 1) what are the characteristics of inverter manufacturer-installer and module manufacturer-installer networks? 2) how do installers’ accumulative experience, technology innovation, upstream-downstream network (i.e., knowledge spillover from upstream inverter/panel manufacturers to downstream installer) contribute to non-hardware price reductions of solar PV installations? The second research question serves to separate different learning mechanisms (learning by doing, learning by search, and learning by interacting) through which deployment policy can facilitate cost reductions. To address these questions, I develop a new database of solar PV balance-of-system (BOS) patents in the distributed PV market from the USPTO database between 2000 and 2014 by using Boolean keywords searching. Then I matched the patent database with the dataset of small-scale PV installations in the U.S., allowing me to build regression models with non-hardware installation price as the dependent variable and experience, innovation, and network as independent variables, plus other controls (explained in detail below).
103 This study sheds light on the drivers of non-hardware cost reductions of solar PV installations, with a focus on the role of technology innovation and upstream-downstream networks. Studying technological innovations in the solar PV industry is of intrinsic interest, but it is more important to understand how technological innovations can translate to cost reductions of technologies or services offered in the market. To the best of my knowledge, I am not aware of any studies that investigate the relationship between relevant technological innovations and non- hardware costs reductions, which may be related to the difficulties in collecting the relevant data for technological innovations in a specific industry (i.e., patent data). Besides technological innovations, firms can also achieve competitiveness through their collaborative partners, especially their upstream suppliers. This essay sheds lights on the extent to which networks with upstream suppliers can help firms achieve competitiveness in terms of cost reductions, and how the effectiveness of network structures associated with cost reductions may vary based on the specific type of upstream suppliers. In addition, because of the coexistence of various learning sources and learning mechanisms in the process of accumulating experience, it is important to separate these learning effects. This chapter quantifies how much of the learning effect of accumulating experience is due to technology innovation and upstream-downstream networks, separately. In other words, this chapter contributes to exploring how learning by searching and learning by interacting mediate the relationship between experience accumulation and cost reduction.
THEORETICAL BACKGROUND
Learning curve and learning mechanism
The learning curve was proposed and formalized by Arrow (1962) and nowadays it is a widely used approach for studying technological change in economics. It builds the relationship between learning and the accumulation of experience. There are two main types of learning curves: the traditional learning curve is a One-Factor-Learning Curve, where the one factor is cumulative experience; a second type of newer models include two factors that aim to separate the effect of
104 cumulative experience (learning by doing) and R&D investment (learning by searching) on technological change. Learning by doing refers to learning from cumulative experience leading to cost reductions and learning by searching refers to technology innovation leading to cost reductions (Qiu and Anadon, 2012; Söderholm and Klaassen, 2007; Söderholm and Sundqvist, 2007).
One-Factor-Learning-Curve hypothesizes that cost reductions are only determined by accumulative experience. However, the cost reductions may be due to multiple factors, such as R&D investment, economies of scale, knowledge spillover, and so on. A major criticism of this model is that only including one variable fails to separate other cost drivers from the accumulative experience (Nemet, 2006; Wene, 2000). In order to address this limitation, the Two-Factor- Learning-Curve incorporates another cost driver – R&D investment – which yields both the learning by doing rate and learning by searching rate. The R&D investment measurement is a type of input in the process of technological change; however, there would be gaps or uncertainties between innovation inputs and innovation outcomes. Therefore, Popp et al. (2013) suggest that using innovation outcomes (e.g., patent counts) rather than R&D investment should be a closer indication of innovation. In this study, I collected a unique dataset including all the patents associated with non-panel components in the solar PV systems, which allows me to directly use innovation outcomes to quantify the effect of learning by searching in the solar PV industry.
In addition to learning by doing and learning by searching, there may be other important cost drivers, including learning by using and learning by interacting Kahouli-Brahmi (2008). Learning by using refers to learning from users’ experience and feedback (Rosenberg, 1982). Learning by interacting means that firms can learn from their interactions with different market actors and government actors ( Lundvall, 1988). Previous literature mainly focuses on One-Factor- Learning-Curve and Two-Factor-Learning-Curve, because learning by doing and learning by searching are expected to be the two most important cost reduction drivers in manufacturing.
105 However, the focus of this study is the installations in the solar PV industry and thus it is also important to take into account any potential effect of learning by interacting. For service providers in service industries such as installers, they have to use products from their upstream suppliers and knowledge spillover from these upstream suppliers is a valuable learning source. Therefore, this study contributes to exploring whether learning-curve relationship can be successfully applied to service industries and separating the effects of learning by doing, learning by searching, and learning by interacting through including experience-related variables, innovation outcome-related variables, and network-related variables.
Supply chain collaboration
Network relationships, especially tacit and explicit information and knowledge transfer within networks, play a critical role in reducing costs, enhancing product quality, and improving technological innovations (Habermeier, 1990; Liker and Choi, 2004). Among different types of interactions and collaborations, supply chain collaborations are recognized as valuable and effective inputs to benefit both upstream firms and downstream firms (Cachon and Fisher, 2000; Eisenhardt and Tabrizi, 1995; Modi and Mabert, 2007; Saeed et al., 2005), providing unique competencies for both the upstream firms and downstream firms (Tidd et al., 2005). For example, involving supply networks can accelerate the development of new products (Eisenhardt and Tabrizi, 1995; Hirotaka and Nonaka, 1986). Information and knowledge sharing in supply chain collaborations can enhance firms’ performance and improve process efficiency (Cachon and Fisher, 2000; Saeed et al., 2005).
However, Villena et al. (2011) pointed out that collaborative upstream-downstream networks may also lead to negative outcomes, and their research shows that cognitive, relational, and structural social capitals can positively influence downstream firms’ performance, while an extreme relationship (i.e., too little or too much social capital) between upstream suppliers and downstream buyers could lead to negative performance at the firm level. There is continuing
106 debate in the literature whether a very high level of dependency between upstream suppliers and downstream buyers could hurt downstream firm performance. On one hand, the dependency increases downstream buyers’ product and process innovations and financial performance (Corsten and Felde, 2005). One the other hand, a high-level dependency on upstream suppliers could gradually and eventually make the downstream firms lose innovativeness (Fine and Whitney, 1996; Heide and John, 1988), and it could also increase the upstream suppliers’ opportunistic behaviors (Villena et al., 2011). In addition, some other studies differentiate between different types of dependency. For example, Lawson et al. (2008) find that ‘closeness’ can increase the social capital between upstream suppliers and downstream buyers. Particularly, managerial communication and technical employees’ communication can improve downstream buyers’ performance. Krause et al. (2007) show that different dimensions of social capital may have different impacts on different types of firms’ performance. The dependency between suppliers and buyers (relational capital) is positively related to firms’ performance in terms of cost reductions. If quality and flexibility are considered as firms’ performance, technical employee communication and tacit knowledge (structural capital (information) transfer are more important.
Based on the literature, as summarized above, upstream-downstream collaborations are not guaranteed to result in improving firms’ performance and different types of network structures (e.g., different levels of centrality, dependency, and dispersion) may lead to different impacts on firms’ performance (e.g., cost reductions, quality improvement, innovative capacity). In the case of the solar PV industry, upstream suppliers and downstream installers have multiple opportunities to collaborate closely. In fact, for most solar PV system it installs, an installer has to work with a products and services from upstream suppliers: panels from panel manufacturers and inverters from inverter manufacturers. It is noteworthy that the network between installers and upstream suppliers is a typical bipartite network (or also known as a two-mode network). Bipartite network structure is often used to model interactions between two different sets of nodes and there is no interaction between nodes that belong to the same set. In the case of solar PV, for each solar PV
107 system, there is no direct interaction between installers or between upstream suppliers. The interactions exist only between installers and their upstream suppliers.
As an important learning source, these upstream suppliers can provide installers with training and knowledge regarding installation techniques, product information, new technologies, and other useful information about the solar PV industry. However, different structures of networks between upstream suppliers and installers in the solar PV industry may lead to different outcomes. For example, a high dependency between installers and upstream suppliers, such as working with an exclusive supplier, might help improve the labor productivity because of the familiarity with its products, reduce management/cooperation costs, and obtain more deep and specialized knowledge about products, technology trends, and the industry, which can lead to cost reductions. On the contrary, a high dependency might also result in the lack of competitive pricing and diverse knowledge from various suppliers for the installer, potentially driving installation costs higher. Therefore, it is important to study which condition or type of network structure is beneficial to which type of downstream-firm performance improvement. In this study, I contribute to improving our understanding of the functional role of upstream-downstream networks on cost reductions with a focus on which type of upstream-downstream network structures are associated with non-hardware cost reductions in the U.S. solar PV industry.
DATA AND METHODOLOGY
Data
The empirical analysis in this study uses two novel databases. The first database covers all the photovoltaic balance-of-system (PV BOS) patents in the U.S. from 2000 to 2014. This patent dataset is collected from the United States Patent and Trademark Office (USPTO) website by using Boolean keywords searching. I include four technological areas in PV BOS, including inverter, monitoring, mounting, and site-assessment (Table 17). I use the same methodology for building
108 the PV BOS patent database as in Venugopalan and Rai (2015). The data collection steps are described in Chapter 4.
Table 17: The specific items and keywords for PV BOS patent data collection BOS category Key words Inverter "inverter" and "photovoltaic" or "solar" Mounting "mounting” and "photovoltaic" or "solar" Monitoring "monitoring" or "monitor" and "photovoltaic" or "solar" "satellite" or "map" or "shade" or "estimation" or "obstruction" or Site Assessment "potential" or "site" or "terrain" or "assess" or "azimuth" or "survey" and "solar" or "photovoltaic"
The second database is an extraordinarily rich dataset of system-level PV installations in the U.S. This dataset is collected as part of LBNL’s Tracking the Sun (TTS) report series, and the detailed description of the dataset can be found in the annual TTS report (Barbose et al., 2017). In my analysis, I focus on residential PV installations with system sizes that are less than 15 kW direct current (DC). The TTS dataset is based on each solar PV system, which includes the transaction price of installation, the date of installation, the location of installation, the system size, the characteristics of the system (e.g., building integrated PV or not, thin-film PV or crystalline PV, Chinese made module or not, and micro-inverters or not). I match the patent database with the system-level PV installation database, which allows me to know the number of PV BOS patents owned by each PV installer and inverter manufacturer and thus to build the innovation-related variables (described in more details below).
The dataset also includes information about the installer, the panel manufacturer, and the inverter manufacturer for each installation system, which allows me to identify the upstream- downstream (suppliers-installers) network. An installer might work with different panel manufacturers and inverter manufacturers for different solar PV systems. Since the dataset includes information about the installer, the panel manufacturer, and the inverter manufacturer for each system, I know the entire upstream - downstream network for solar PV installations. In the
109 upstream-downstream (installer-supplier) bipartite network, the interactions are only between installers and panel/inverter manufacturers, and there is no interaction between installers or between panel/inverter manufacturers. The network-related variables are constructed based the entire upstream-downstream collaboration network.
The dependent variable in my analysis is the non-hardware part of the transaction price between the customer and the installer. The original transaction price is the installed price of PV systems, which includes PV panel cost, inverter cost, other hardware costs (wiring, support structure, and meters), non-hardware costs, and installer profits. My analysis focuses on non- hardware price, meaning that I exclude PV panel cost and inverter cost when building the dependent variable. There are three categories of key independent variables, introduced here in broad terms, and explained in more details in the next section. The first category is experience- related variables, which describe learning from the installer’s own experience and other installers’ experience. The second category is innovation-related variables, which describes the installer’s own knowledge stock and the learning from innovative inverter manufacturers. The third category is network-related variables, which describe the network connectedness between installers and their upstream suppliers (i.e., panel manufacturers and inverter manufacturers). I also control for system characteristics, market share, competition, and county-level labor costs and household income. Detailed description of these independent variables and control variables is in the next section.
Econometric models
This section presents the econometric models for analysing the effect of experience-related factors, innovation-related factors, and network-related factors on the non-hardware installation prices of PV system installations. Consistent with recent work in this area (e.g., Davidson and Steinberg, 2013; Gillingham et al., 2016; Haynes and Thompson, 2008; Nemet et al., 2016, 2017; Skidmore et al., 2005; Wiser et al., 2007), this analysis assumes the price at time t is impacted by
110 its intrinsic attributes (or packages of characteristics), which includes potential supply shifters, demand shifters, and exogenous characteristics. Therefore, this study uses a reduced form model of price and includes experience-related variables, innovation-related variables, network-related variables, market structure, user characteristics, and system characteristics as independent variables. Equation (1) is the basic model. This model only includes experience-related variables.
The dependent variable � is the non-hardware transaction prices of PV system i installed by installer j in county c in year t. The two key independent variables are Installation and
Installation . The former is the cumulative installed capacity of the installer j in county c in year t-1, and the latter is the cumulative installed capacity of all the other installers (except installer j) in county c in year t-1.
The endogeneity between prices and installed capacity is a general concern in learning- curve models (Gilingham et al. 2014; Hayashi et al., 2018; Nemet, 2006; Nemet, 2012; Qiu and Anadon, 2012), because the price reductions could be both a cause for and a result of the increased solar adoptions. However, it is not a major concern for this study, because the dependent variable in this study, non-hardware prices, is not the dominant factor that leads to solar PV adoptions in the past. Rather, the dominant driver of the increased installed capacity in the past is price reductions in hardware. Since 2000, the price reductions in solar modules and inverters account for about 43% and 9% of the total installation price reductions, respectively (Barbose et al., 2018). Besides these hardware-related price factors, public inclination towards the adoption of solar PV technology is another major driver. Rising concerns about the environmental, increasing awareness of solar PV technologies (through neighbourhood peer effects and direct marketing by installers), and higher quality products have significantly reduced adoption barriers (Karakaya and Sriwannawit,2015; Rai and Reeves, 2016).
Even with these more dominant drivers of adoption in the past, an impact of non-hardware prices on adoption levels is very likely. As such, I take two measures to address this potential
111 endogeneity issue. In my main models, I use one-year lagged accumulative installation capacity as my experience-related independent variables. The rationale is that installation prices in year t should not have an impact on the cumulative installation capacity in year t-1. In principle, current price should affect flow of new system installations only (new added installations annually). But the independent variable in the equation is a stock at time t, meaning that it measures cumulative sum of flows of new installations from 0 through t-1, so it should not be affected by current price. However, any shock to installation price in t-1 will affect installed capacity in time t and time t-1, so if shock/disturbance to price in time t correlated with shock disturbance to price in time t-1, installed capacity is still endogenous. Therefore, I also introduced an instrumental variable to address potential endogeneity issues. The instrumental variable for the installed capacity is zip- code level annual insolation. This is a valid instrument variable because it is not associated with non-hardware installation prices. However, this is a weak instrumental variable because the correlation between installed capacity and insolation is small. Thus, this valid but weak instrumental variable could lead to biased IV estimates.
I control for installer fixed effect (α ), county fixed effect (γ ), and year fixed effect (� ). It is important to control for installer fixed effect, because heterogeneity among installers could be potentially confounding (Gilingham et al. 2014). Many studies using learning-curve model are unable to control for firm-level fixed effect because normally each firm only has a limited number of repeated observations, meaning that, typically, there is not enough within-firm variation to incorporate firm-level fixed effect. The rich dataset on the solar PV system level provides enough within-installer variations to introduce installer fixed effect. Therefore, the model is able to control for all the unobserved time-invariant heterogeneity among different installers. In addition, the model includes year fixed effect to control for time-varying factors. County fixed effect is also included to control for all the time-invariant characteristics among different counties. A county is political subdivision of a state, and thus when I include county fixed effect, I cannot include any
112 variables that only have variations across different states or counties, such as time-invariant policies at state or county level.
I also control for county-level labor costs, household income, and system characteristics. Previous literature indicates that regional differences in labor costs and demand-side characteristics should be taken into consider in learning-curve models (Gilingham et al. 2014; Nemet, 2006; Nemet, 2012; Qiu and Anadon, 2012). Therefore, I control for county level labor costs and household income. The system characteristics should also be taken into consider, because they are related with the different degrees of difficulties in installing solar PV systems and the total value of the systems, and thus have an impact on the non-hardware costs. The system characteristics include a series of dummy variables to indicate whether a PV system is integrated into roof materials (bipv ), whether a PV system has a battery backup system (battery ), whether the panels were manufactured in China (China ), whether the PV cells are thin film
(thin ), whether the PV system has sun tracking equipment (tracking ), and whether the system uses micro-inverters attached to each panel (microinv ).
Equation (1)
� = � + � ∗ Experience + � ∗ Experience + � ∗ China + � ∗ thin
+ � ∗ bipv + � ∗ battery + � ∗ tracking + � ∗ microinv + �
∗ LaborCost + � ∗ Income + α + γ + � + �
Equation 2 includes innovation-related variables. Previous literature uses cumulative installed capacity to capture a joint learning rate of learning by doing and learning by searching. This chapter uses the depreciated cumulative number of PV BOS patents owned by a certain installer to measure that installer’s knowledge stock, which allows me to separate the effect of learning by doing that is mainly driven by productivity improvement (i.e., repeatedly working on technologically similar installations) and learning by searching that is mainly driven by technology
113 innovation (i.e., changes in working processes or installation methods). Although using firm-level patent data is a more direct way to measure innovation outcomes compared to firm-level R&D expenditures, the limitation of the patent data is that it can only measure codified knowledge but does not capture tacit knowledge that lies with each installer. Therefore, the coefficient of knowledge stock may underestimate the impact of learning by searching, because the impact of tacit knowledge stock is not included. However, I am able to measure tacit knowledge spillover from innovative inverter manufacturers to installers. An inverter manufacturer that has patents in the field of inverter technologies is defined as an innovative inverter manufacturer. To complete the installation of a solar PV system, an installer has to work with an inverter manufacturer. The knowledge spillover from innovative inverter manufacturers to an installer is measured by the installer’s cumulative installed capacity that uses inverters from all innovative inverter manufacturers. In equation 2, �������������� is the cumulative installed capacity with innovative inverter manufacturers for installer j in year t-1; Knowledge is the depreciated cumulative sum of patents owned by installer j in year t, and the depreciation rate, γ, is 0.8. All the other control variables are the same as in equation 1.
Equation (2)
� = � + � ∗ Experience + � ∗ Experience + � ∗ Experience���� + �
∗ Knowledge + � ∗ China + � ∗ thin + � ∗ bipv + � ∗ battery
+ � ∗ tracking + � ∗ microinv + � ∗ LaborCost + � ∗ Income
+ α + γ + � + �
Where Knowledge = (1 − γ)Knowledge ( ) + Patent
Equation (3) includes network-related variables. When an installer increases its installation experiences, it also expands its network with panel manufacturers and inverter manufacturers through collaborations with these upstream suppliers. In other words, only including experience related variables cannot separate the effects of learning by doing and learning by interacting.
114 Therefore, including network-related variables contributes to separating the effect of learning by interaction from the overall effect of increasing installation experience.
In equation 3, CentralPanel is the value of centrality for installer j in year t-1 in the installer-panel manufacturer bipartite network; and CentralInv is the value of centrality for installer j in year t-1 in the installer-inverter manufacturer bipartite network. I use centrality to measure both the direct ties and indirect ties. Direct ties refer to the number of links between an installer and panel/inverter manufacturers. Although installers do not work with each other in the installation of solar PV systems, they could be linked by their common upstream suppliers. The indirect ties refer to the number of links among different installers through collaborations with their common upstream suppliers. For example, if an installer collaborates with three inverter manufacturers, and each of these three inverter manufacturers works with four different installers, then this installer has twelve indirect ties, meaning that the value of centrality is twelve.
FCPanel and FCInv are the first-order couplings between installer j and its panel manufacturer collaborators in year t-1 and between installer j and its inverter manufacturer collaborators in year t-1, respectively. First-order coupling captures the degree of dependency between an installer and its upstream collaborators. The first-order coupling is calculated by squaring the proportion of an installer’s installed capacity that results from collaboration with each manufacturer and then summing the resulting numbers. Put differently, the first-order coupling is
defined as �� = ∑ � , where � is the percentage of installer i's total installed capacity that comes from manufacturer k and � is the number of manufacturers that installer j works with. A high value on first-order coupling indicates a high degree of dependency between an installer and its upstream suppliers, because a high value on first-order coupling means that the installer j concentrates its solar PV installation projects with a limited number of panel/inverter manufacturers. SCPanel and SCInv are second-order couplings between installer j and its panel manufacturer collaborators in year t-1 and between installer j and its inverter manufacturer
115 collaborators in year t-1, respectively. The second-order coupling is calculated in two steps. The first step is to calculate the first-order coupling for each inverter manufacturer and each panel manufacturer. The second step is to sum the value of first-order coupling for each of the inverter (or panel) manufacturers in the installer’s network, and then divide by the total number of inverter (or panel) manufacturer collaborators in this installer’s network. Put differently, the first step is
to calculate � = ∑ � , where � is percentage of manufacturer k's market output worked with the installer j and � is the total number of installer that manufacturer k works with. The ∑ second step is to calculate �� = , where � is the number of manufacturers that installer j work with. A high value of second-order coupling indicates that an installer’s collaborators have a concentrated network with other installers. Previous literature shows that the relationship between the second-order coupling and the dependent variable is quadratic (Uzzi, 1996), and, thus,
I also include the squared terms of SCPanel and SCInv .
Equation (3)
� = � + � ∗ Experience + � ∗ Experience + � ∗ Experience���� + �
∗ Knowledge + � ∗ CentralPanel + � ∗ FCPanel + � ∗ SCPanel + � ∗ ������� + � ∗ CentralInv + � ∗ FCInv + � ∗ SCInv + � ∗ ����� + � ∗ China + � ∗ thin + � ∗ bipv + �
∗ battery + � ∗ tracking + � ∗ microinv + � ∗ LaborCost + �
∗ Income + α + γ + � + �
In equation 4, I use degree to replace centrality and species specificity index (SSI) to replace first-order coupling. DegreePanel is the number of links with panel manufacturers for installer j in year t-1, which only counts the number of direct ties between an installer and panel manufacturers. DegreeInv is the number of links with inverter manufacturers for installer j in year t-1. Another similar measurement as the first-order coupling is SSI, which also measures the degree of concentration between an installer and its collaborators. The SSI measures the coefficient of variation of interactions between an installer and its collaborators, which is the ratio of the standard deviation to the mean. A high value of SSI indicates a concentrated network
116 between an installer and its upstream collaborators, which means that an installer only works with a few upstream suppliers. A low value of SSI indicates a diverse network between an installer and its upstream suppliers. In equation 4, SSIPanel and SSIInv are the values of SSI between an installer j and the panel manufacturers in its network in year t-1, and between an installer j and the inverter manufacturers in its network in year t-1, respectively.
Equation (4)
� = � + � ∗ Experience + � ∗ Experience + � ∗ Experience���� + �
∗ Knowledge + � ∗ DegreePanel + � ∗ SSIPanel + � ∗ SCPanel + � ∗ ������� + � ∗ DegreeInv + � ∗ SSIInv + � ∗ SCInv + � ∗ ����� + � ∗ China + � ∗ thin + � ∗ bipv + �
∗ battery + � ∗ tracking + � ∗ microinv + � ∗ LaborCost + �
∗ Income + α + γ + � + �
RESULTS AND DISCUSSION
Network characteristics
Figure 12 shows the annual number of total installers in the system-level PV installation database and the annual number of installers with PV BOS patents from 2000 to 2014. I match the PV BOS patent database with the system-level PV installation database, which allows me to know the number of PV BOS patents owned by each PV installer annually. The left y-axis and the blue bar show the total number of installers annually, and the right y-axis and orange line show the annual number of installers that have at least one PV BOS patent. The number of total installers in the U.S. increases steadily after 2000, but it has a decreasing trend after 2011. The number of installers that have patents in the field of PV BOS is very limited: at most 1% of the total number of installers annually.
117 3000 20 18 2500 (count) 16 14 2000 12 1500 10 8 1000 6 4 The Numer of The InstallersNumer (count) 500 2
0 0 of The InstallersNumber with Patents 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014
num of installers num of installers with patents
Figure 12: The annual number of total installers and the annual number of installers with PV BOS patents from 2000 to 2014
The upstream-downstream network is a typical bipartite network. Installers and upstream suppliers are two different sets of nodes, and there are no interactions between nodes that belong to the same set. For each solar PV system, an installer has to work with a panel manufacturer and an inverter manufacturer, but these is no interaction between just the upstream suppliers and between just the installers. Table 18 shows the annual number of installers, inverter manufacturers, and panel manufacturers during 2001 and 2014. The bipartite network between installers and inverter/panel manufacturers is asymmetric, as the number of installers is consistently much higher than the number of upstream suppliers. Figure 13 shows an example network between two installers (i.e., Sungevity and Verengo) and their upstream suppliers in year 2013. Figure 13 (a) is the installers-inverter manufacturers network and figure 13 (b) is the installers-panel manufacturers network. Sungevity worked with seven inverter manufacturers in 2013 and mainly worked with two inverter manufacturers. 77% of Sungevity’s projects were distributed to two
118 inverter manufacturers, Power-One and ABB. Verengo also worked with seven inverter manufacturers and mainly worked with the same two inverter manufacturers, Power-One and ABB. There are more options for installers when choosing panel collaborators. Compared with inverter manufacturers, installers worked with a more diverse set of panel manufacturers. Sungevity worked with eleven panel manufacturers and concentrated 59% of its projects with one panel manufacturer, ET. Verengo worked with seven panel manufacturers and concentrated 79% of its projects with one panel manufacturer, Yingli, in 2013. Overall, the installers distribute their PV projects with several upstream suppliers but concentrate their projects with one or two major suppliers.
Table 18: The annual number of installers, panel manufacturers, and inverter manufacturers Number of Number of panel Number of inverter Year installers manufacturers manufacturers (count) (count) (count) 2001 46 4 5 2002 167 5 4 2003 370 12 7 2004 457 11 10 2005 422 16 12 2006 390 18 12 2007 647 27 19 2008 809 35 24 2009 1519 66 23 2010 1951 98 29 2011 2035 141 38 2012 1891 145 46 2013 1700 134 45 2014 1077 97 38
119
Figure 13 (a) : An example network structure between two installers and their inverter manufacturers
Figure 13 (b) : An example network structure between two installers and their panel manufacturers
120
Figure 14 shows the cumulative distribution of all the system-level collaborations between installers and their upstream suppliers (panel manufacturer and inverter manufacturer). Figure 14 (a) shows how installers distribute their projects to panel manufacturers and inverter manufacturers in the market, and figure 14 (b) presents how installers distribute the projects to their principal panel/inverter manufacturer. I distinguish innovative installers from non-innovative installers. An innovative installer is defined as an installer that has at least one patent in the field of PV BOS. The blue, red, purple, and green lines in figure 14 present the network between innovative installers and inverter manufacturers, non-innovative installers and inverter manufacturers, innovative installers and panel manufacturers, and non-innovative installers and panel manufacturers, respectively.
The only difference between figure 14 (a) and figure 14 (b) is the x-axis. The x-axis in figure 14 (a) is the percentage of the installed capacity associated with each panel/inverter manufacturer, and the x-axis in figure 14 (b) is the percentage of the installed capacity associated with the largest panel/inverter partner (the principal collaborator). For example, if installer A works with two panel manufacturers and distributes the projects equally, then the percentage of the installed capacity per collaboration is 50% and 50%, and the percentage of the principal collaboration is also 50%. The installer B works with three panel manufacturers and distributes
30%, 30%, and 40% of its projects to panel manufacturer a, b, and c, respectively. In the case of the installer B, the percentage of the installed capacity per collaboration is 30%, 30%, and 40%, and the percentage of the principal collaboration is 40%.
The y-axis in figure 14 (a) is the cumulative percentage of all the collaborations between installers and panel/inverter manufacturers, and the y-axis in figure 14 (b) is the cumulative percentage of collaborations between installers and their largest panel/inverter manufacturers partner (the principal collaborators). Take the same example as above, the installer A works with
121 two different panel manufacturers and the installer B works with three different panel manufacturers. Thus, there are five collaborations in the market. Assume that installer A and installer B occupy all the market, and the market share for each installer is 50%. For the case of taking into account all five collaborations, there are three possible values on the x-axis, namely 30%, 40%, and 50%. When the value on the x-axis is 30%, the corresponding y value is 30% (30% *50% *2=30%). It means that 30% of all the installations in the market is for 30% of an installer’s total installation. When the value on the x-axis is 40%, the corresponding y value is 50% (40%*50% + 30% =50%). When the value on the x-axis is 50%, then the corresponding y value is 100% (50%*50%*2 + 50%= 100%).
I use the same calculation methodology and expand the above simple example to all the collaborations between installers and panel/inverter manufacturers in the dataset. The pattern of four lines in figure 14 (a) shows that about 20%-25% of all collaborations between installers and panel/inverter manufacturers are for 50% of an installer’s total installed capacity, and about 40%- 65% of all collaborations between installers and panel/inverter manufacturers are for 100% of an installer’s total installed capacity. About 65% of all collaborations from non-innovative installers to panel/inverter manufacturers are for 100% of an installer’s total installed capacity. In other words, for 65% of capacity from non-innovative installers, the installer only works with one panel and one inverter manufacturer. The figure 14 (b) shows the pattern of the principal collaboration is more concentrated. About 14%-20% of installers distribute 80% of their projects to one principal panel/inverter manufacturer. About 60%-70% of installers distribute all their projects to one panel/inverter manufacturer. These two figures show that the network between installers and upstream suppliers are concentrated, and there are no significant differences regarding the degree of dependency among innovative installers and non-innovative installers.
122
About 40%-65% of upstream- downstream collaborations are for 100% of an installer’s total installed capacity
About 20%-25% of upstream- downstream collaborations are for 50% of an installer’s total installed capacity
Figure 14 (a) The distribution of installers’ all projects to panel manufacturers and inverter manufacturers in the market.
123 About 60%-70% of collaborations through a principle supplier are for 100% of an installer’s total installed capacity
About 14%-20% of collaborations through a principle supplier are for 80% of an installer’s total installed capacity
Figure 14 (b) The distribution of installers’ principle projects to panel manufacturers and inverter manufacturers in the market.
Regression results
Table 19 shows the regression results of equation 1 to equation 4. The results show that all three categories of independent variables – experience, innovation, and network – can explain variations in the non-hardware part of the solar PV installation prices. The coefficients of experience-related variables show that an installer’s own installed capacity in a certain county can significantly contribute to reducing the non-hardware prices; however, the other installers’ installed capacity at the same county increases that installer’s non-hardware prices. Given the same level of an installer’s installed capacity in a certain county, the higher the installed capacity of other installers in the county, the higher the total installed capacity in the county. The total installed capacity in a county is a proxy for the level of public preference towards solar PV in that county. Thus a potential explanation for this counterintuitive result is that public preference towards solar
124 PV technologies might be positively associated with non-hardware prices (Gillingham et al., 2014). The relevant tests and more discussions about this hypothesis can be found in the texts between table 19 and table 20.
The innovation-related variables show that an installer’s technological innovations can significantly reduce the non-hardware installation prices and working with innovative inverter manufacturers can also significantly reduce this installer’s non-hardware installation prices. These variables suggest that installers’ learning from their own technological innovations and their innovative upstream suppliers can help them reduce prices. After adding innovation-related variables, the coefficient of an installer’s own experience is reduced by about 15%, which indicates that the effect of learning by searching, which is mainly driven by technology innovation (i.e., changes of working process or installation methods), accounts for about 15% of the overall effect of cumulative experiences. It is worthy to note that I may overestimate the proportion of learning by searching effect to the overall effect of cumulative experience. That is because patent data measures innovations outcomes, but these innovation outcomes may be motivated not just by cumulative experience, but also by other investments. Furthermore, the study may underestimate the effect of learning by searching on price reduction, since the number of patents can only capture codified knowledge.
The network-related variables can also significantly explain the variations in the non- hardware part of installation prices. The coefficients of degree and centrality show that collaborating with more panel manufacturers and inverter manufacturers is associated with lower non-hardware installation prices. However, controlling for degree or centrality, the coefficients of first-order coupling/species specificity index show that a concentrated network with panel manufacturers and inverter manufacturers is associated with low non-hardware installation prices. The network results suggest that the installers who work with diverse upstream suppliers but concentrate their projects with several main upstream suppliers within their networks have low
125 non-hardware installation prices. The advantage of this type of network is that installers can learn diverse knowledge from their diverse network, but they can also learn deeply and intensively with their major network partners. The signs of the coefficients of second-order coupling are different between installers-panel manufacturers network and installers-inverter manufacturers networks. The installers who partner with panel manufacturers who have moderately concentrated networks are associated with lower non-hardware installation prices. However, the installers who partner with inverter manufacturers who have very concentrated networks are associated with lower non- hardware installation prices. This result may be driven by the different degrees of maturity between the panel market and the inverter market. The panel market is a much more matured market than the inverter market, and thus a mutually concentrated structure may have less value in the installers-panel manufacturers network.
After adding the network-related variables, the coefficient of an installer’s own installed capacity is reduced by 28%. This indicates that the effect of learning by interacting that is driven by upstream-downstream collaborations accounts for about 28% of the overall effect of learning by doing (i.e., cumulative experiences). Adding the innovation-related variables and the network- related variables reduces the coefficient of an installer’s own installed capacity by about 43%. When an installer increases its installed capacity, part of the learning effect comes from the increased installation experience (i.e., by repeated working); however, at the same time, the installer can also significantly learn from its own technology innovations and collaborations with its upstream suppliers. Results above suggest that, in addition to cost reductions due to sheer learning from repeated installation experience, technological innovations and learning from upstream suppliers can also significantly help reduce installation prices.
It is worth discussing why learning by interacting accounts for a larger share of the overall effect of cumulative experience. One explanation is that learning by interacting is a more important learning mechanism for non-hardware price reductions, compared to learning by searching.
126 Literature shows that learning by interacting is a critical and major learning way in a service industry (Cachon and Fisher, 2000; Eisenhardt and Tabrizi, 1995; Modi and Mabert, 2007; Saeed et al., 2005). Contrary to manufacturing, service providers such as solar PV installers typically do not produce any products and their cost reductions are not driven by direct product innovations. This argument is supported by the number of installers that have patents in BOS. The US BOS patent dataset shows that at most only 1% of installers have BOS patents. However, the training and knowledge regarding installation techniques, product information, new technologies, and other useful information about the solar PV industry from upstream suppliers (learning by interacting) can significantly contribute to their everyday service and help reduce non-hardware costs. Another explanation is that learning by searching might be as important as learning by interacting, but the number of BOS patents only captures a small proportion of the outcomes of learning by searching. Particularly, technological innovations associated with cumulative experience are more likely to be process innovations or tacit knowledge. The former is less likely to be projected through patenting and patent data cannot capture the latter one. Thus, it is possible that this study underestimates the proportion of learning by searching in the overall effect of cumulative experience. The third explanation is that learning by searching effects could be achieved through other channels, such as research and development investment. Put differently, learning by searching through cumulative experience is only one channel through which technological innovations can be achieved. But learning by interacting is mainly captured through upstream-downstream networks in the case of solar PV installers. Therefore, a smaller proportion of learning by searching in the overall effect of cumulative experience cannot be interpreted as a smaller impact of learning by searching in non-hardware price reductions, compared to learning by interacting.
127 Table 19: The regression results of the effect of learning by doing, learning by searching, and learning by interacting
Model 1 Model 2 Model 3 Model 4
-1.18e-08*** -1.00e-08*** -6.95e-09*** -6.63e-09*** Experience (watt) (1.43e-09) (1.54e-09) (1.55e-09) (1.54e-09) Other installers' experience 2.85e-09*** 2.76e-09*** 2.37e-09*** 3.26e-09*** (watt) (3.39e-10) (3.39e-10) (3.39e-10) (3.35e-10) -0.0365*** -0.0232*** -0.0247*** Knowledge stock (0.00382) (0.00423) (0.00400) Experience with innovative -3.38e-09*** -2.94e-09*** -2.25e-09*** inverter manuf (watt) (4.81e-10) (5.10e-10) (5.10e-10) -0.00340 Degree_Panel (0.00231) -0.0588*** Degree_Inverter (0.00371) -0.0000271*** Centrality_Panel (0.00000371) -0.0000302*** Centrality_Inverter (0.00000484) 0.000000679 FC_Panel (0.00000240) -0.00000718*** FC_Inverter (0.00000263) -0.0510 SSI_Panel (0.0342) -0.271*** SSI_Inverter (0.0372) -0.000255*** -0.000279*** SC_Panel (0.0000219) (0.0000221) 0.000233*** 0.000259*** SC_Inverter (0.0000206) (0.0000207) 4.02e-08*** 4.56e-08*** SC_Panel_Sqaure (5.42e-09) (5.50e-09) -2.01e-08*** -2.54e-08*** SC_Inverter_Sqaure (4.37e-09) (4.39e-09) 0.00000111** 0.00000114** 0.00000105** 0.000000961* Labor Cost (0.000000527) (0.000000527) (0.000000525) (0.000000524) 0.0166 0.0165 0.0131 0.0174 Household Income (0.0233) (0.0233) (0.0232) (0.0231)
128 -0.319*** -0.318*** -0.301*** -0.299*** China Panel (0.00916) (0.00916) (0.00914) (0.00910) 0.0350 0.0413 0.0149 0.00849 Thin Film (0.0708) (0.0708) (0.0710) (0.0709) 1.520*** 1.491*** 1.550*** 1.549*** BIPV (0.0367) (0.0374) (0.0375) (0.0379) 3.280*** 3.276*** 3.278*** 3.261*** Battery (0.572) (0.572) (0.572) (0.574) 1.437*** 1.441*** 1.450*** 1.437*** Tracking (0.207) (0.207) (0.206) (0.204) -0.0639*** -0.0642*** -0.0773*** -0.0838*** Micro Inverter (0.00986) (0.00987) (0.00991) (0.00990) -0.0662*** -0.0663*** -0.0659*** -0.0658*** System Size (0.00127) (0.00127) (0.00127) (0.00127) 6.080*** 6.143*** 6.479*** 6.687*** Constant (0.0341) (0.0346) (0.0502) (0.0574) Num of Obs 126904 126851 126851 126692 Installer fixed effect yes yes yes yes County fixed effect yes yes yes yes Year fixed effect yes yes yes yes * p<0.10 ** p<0.05 *** p<0.01
The positive coefficient of other installers’ cumulative installed capacity at the same county in table 19 is counterintuitive, so I run two new models to test the hypothesis regarding the potential explanations. A potential explanation is that the level of public preference towards solar PV technologies in a county might be positively associated with non-hardware prices. I use the total installed capacity in a county as a proxy for the public preference towards solar PV in the county (Gillingham et al., 2014). Given the same level of an installer’s installed capacity in a certain county, the higher the installed capacity of other installers in the county, the higher the total installed capacity in the county. Thus, the positive coefficient of other installers’ capacity may result from a high preference towards the solar PV in the county. To test this hypothesis, I include the total installed capacity in a county in the model 5 in table 20. The coefficient of the total installed capacity is positive, which supports the hypothesis that a higher preference towards solar PV technologies could lead to higher prices.
129
The degree of competition and market share are two potentially confounding variables in the above hypothesis. Therefore, I add competition and market share in the model 6 in table 20 to further test the above hypothesis. The variable competition is defined as the number of installers in a county, and the market share refers to the proportion of the installer i’s installed capacity in a county to the total installed capacity in the county. The coefficient of competition shows that the number of installers in a county is negatively associated with non-hardware prices, indicating that competition can help bring down the non-hardware prices. The coefficient of market share shows that the non-hardware prices are higher for installers who have a higher market share, which may result from the market power of installers who have a large market size. After controlling for competition and market share, the coefficient of the total installed capacity in a county is still positive and statistically significant. Therefore, the results support the hypothesis that the preference towards solar PV technologies is positively associated non-hardware prices.
130 Table 20: The regression results of including total installed capacity, market share and competition
Model 5 Mode 6 -9.33e-09*** -1.31e-08*** Experience (watt) (1.61e-09) (1.77e-09) 2.37e-09*** 2.67e-09*** Total experience (watt) (3.39e-10) (3.44e-10) -0.0000336*** Competition (0.00000399) 0.434*** Market share (0.0818) -0.0232*** -0.0214*** Knowledge stock (0.00423) (0.00422) Experience with innovative inverter -2.94e-09*** -2.43e-09*** manuf (watt) (5.10e-10) (5.16e-10) -0.0000271*** -0.0000272*** Centrality_Panel (0.00000371) (0.00000372) -0.0000302*** -0.0000337*** Centrality_Inverter (0.00000484) (0.00000486) 0.000000679 0.000000635 FC_Panel (0.00000240) (0.00000240) -0.00000718*** -0.00000803*** FC_Inverter (0.00000263) (0.00000263) -0.000255*** -0.000257*** SC_Panel (0.0000219) (0.0000219) 0.000233*** 0.000231*** SC_Inverter (0.0000206) (0.0000206) 4.02e-08*** 4.05e-08*** SC_Panel_Sqaure (5.42e-09) (5.43e-09) -2.01e-08*** -2.03e-08*** SC_Inverter_Sqaure (4.37e-09) (4.36e-09) 0.00000105** 0.00000105** Labor Cost (0.000000525) (0.000000525) 0.0131 0.0123 Household Income (0.0232) (0.0232) -0.301*** -0.300*** China Panel (0.00914) (0.00913) 0.0149 0.0134 Thin Film (0.0710) (0.0708) BIPV 1.550*** 1.545***
131 (0.0375) (0.0376) 3.278*** 3.278*** Battery (0.572) (0.574) 1.450*** 1.443*** Tracking (0.206) (0.206) -0.0773*** -0.0760*** Micro Inverter (0.00991) (0.00990) -0.0659*** -0.0658*** System Size (0.00127) (0.00127) Constant -9.33e-09*** -1.31e-08*** (1.61e-09) (1.77e-09) Num of Obs 126898 126898 Installer fixed effect yes yes County fixed effect yes yes Year fixed effect yes yes * p<0.10 ** p<0.05 *** p<0.01
CONCLUSION AND POLICY IMPLICATIONS
There is a general consensus that the next big barrier of cost reductions in the solar PV deployment will come not from hardware (e.g., panel, inverter, etc.), but rather from “non- hardware costs”, including installation, racking, permitting, financing, overhead, marketing, and other non-hardware costs associated with installing solar systems. In this essay I studied how experience accumulation, technological innovations in the field of solar PV BOS, and networks between installers and their upstream suppliers (e.g., panel manufacturer, inverter manufacturer) contribute to non-hardware cost reductions. Most of the extant studies mainly focus on verifying the learning-curve relationship for cost reductions in manufacturing. This study contributes to explore whether the argument that accumulation of experience leads to cost reductions can be applied to service industries. Different from manufacturing industries, upstream suppliers are important learning sources for service industries, and thus through studying a service industry, this chapter also contributes to incorporate supply chain collaboration theories in the learning-curve relationship. More importantly, this chapter contributes to separate the effect on prices due to
132 learning by doing, learning by searching, and learning by interacting in the process of accumulating experience.
I use a unique database of solar PV balance-of-system (BOS) patents in the distributed PV market in the U.S. between 2000 and 2014, matched with a system-level dataset of small-scale PV installations in the U.S. Matching these two datasets allows me to know the number of patents owned by each installer and inverter manufacturer and thus build innovation-related variables. The database also includes information about the installer, the panel manufacturer, and the inverter manufacturer for each solar PV system, thus allowing me to know the entire upstream-downstream (suppliers-installers) network and build network-related variables.
Through the descriptive analysis of innovation-related and network-related indicators, I find that 1) the number of installers that have patents in the field of PV BOS is very limited, which accounts for at most 1% of the total number of installers annually; 2) the network between installers and two types of upstream suppliers, panel manufacturers and inverter manufacturers, are concentrated, which means that installers mainly concentrate their installation projects with a few panel/inverter manufacturers. The concentrated network structure is a double-edged sword. On one hand, keeping a concentrated network might be beneficial for installers to learn intensively and deeply with their upstream suppliers; a concentrated network is also easier to manage, which may help reduce the costs associated with coordination and management. On the other hand, a concentrated network may result in the lack of competitive pricing, and the lack of diverse knowledge from various suppliers, and thus lead to higher costs. Therefore, I further conduct regression analysis to study the functional role of network structures on non-hardware solar PV installation prices.
The regression results show that all the three categories of independent variables – experience, innovation, and network help explain the variations of the non-hardware part of the
133 solar PV installation prices. Technological innovations and concentrated networks with upstream suppliers significantly contribute to reducing the non-hardware installation prices. The installers who concentrate their projects with their major partners in the networks with their upstream suppliers are associated with lower non-hardware installation prices. A mutually concentrated network structure between installers and their suppliers is more important for installers-inverter manufacturers networks. Collaborating with panel manufacturers that have a moderately concentrated network is associated with lower non-hardware installation prices, while collaborating with inverter manufacturers that have a very concentrated network is associated with lower non-hardware installation prices. This result may be driven by the different degrees of maturity between the panel market and the inverter market. The panel market is a much more matured market than the inverter market, and thus a mutually concentrated structure may be more valuable for installers-panel manufacturers networks.
Furthermore, the effects of innovations and networks on non-hardware costs are part of the effect of accumulating installation experience. After adding innovation-related variables, the coefficient of an installer’s own installations is reduced by about 15%, and after adding the network-related variables, the coefficient of an installer’s own installed capacity is reduced by another 28%. This indicates that the effect of learning by searching, which is mainly driven by technology innovation (i.e., changes in working processes or installation methods), accounts for about 15% of the overall effect of accumulative experiences. The effect of learning by interacting that is driven by upstream-downstream collaborations accounts for about 28% of the overall effect of accumulative experiences. In other words, the results suggest that the relationship between experience accumulation and cost reductions are mediated by learning by searching and learning by interacting.
It is worth to note several limitations of modeling and the interpretation of the results. I may be overestimating the proportion of learning by searching effect in the overall effect of
134 cumulative experience and the true proportion may be lower than 15%. That is because innovation outcomes may be driven by factors other than just installation experience. In any case, including the innovation-related variables lowers the installation-experience related coefficient by 15%; just that my analysis cannot tell if all that 15% effect is taken fully away from the experience variable or potentially comes from other channels. At the same time, it is also possible that I underestimate the proportion of learning by searching effect to the overall effect of cumulative experience, because I only include codified knowledge stock of installers in the models and the conversion from experience accumulation to tacit knowledge stock might be even higher.
In addition, endogeneity between installed capacity and non-hardware price may exist in the main model, because the price reductions could be both a cause and a result of increased solar adoptions. I employed two strategies to address this potential endogeneity issue. In the main models, I use one year lagged accumulative installation capacity as the experience-related independent variables. The second way is to use zip-code level annual insolation as an instrumental variable for the installed capacity (See supplementary information). This is a valid but weak instrumental variable (IV) because it is not associated with non-hardware installation prices, but the correlation between installed capacity and insolation is small. The innovation-related and network-related variables are almost identical as the main results, but the sign of installer’s own experience is changed and not statistically significant. The coefficients of experience-related variable with IV may be biased because of the weak IV. Future work may try to address this limitation by using a strong IV.
In sum, the direct impact of deployment policy is to increase firms’ installation experience, but I care more about the how increased deployment operating through higher installer experience facilitates cost reductions. The results show that learning by doing, learning by searching, and learning by interacting are three important learning mechanisms that can contribute to reducing non-hardware installation prices. When a firm increases its experience, it can trigger different
135 learning mechanisms, including the effects of learning by doing, learning by searching, and learning by interacting. The positive relationship between experience accumulation and cost reductions as typically explained as learning by doing are mediated by other learning mechanisms, such as learning by searching and learning by interacting.
Policy implications of this study are two folds. First, this indicates that deployment policies that aim to enlarge market demand could help firms to reduce installation costs through experience accumulation. Secondly, this indicates that a deliberate policy-mix design is needed to reduce the solar PV deployment barrier in terms of installation cost reductions, because deployment policies could potentially interact with policies that facilitate network-building and technological innovations. A combination of deployment policies, innovation-support policies, and network facilitating policies could potentially lead to a more desired policy outcome through achieving higher joint learning rate from firms’ cumulative experiences developed in a more integrated production and deployment ecosystem.
136 Chapter 6: Conclusion
CONTRIBUTIONS TO THE LITERATURE
First of all, this dissertation contributes to demand-induced innovation literature by taking into account how technological characteristics interact with local context to shape technological change. Although many empirical studies support that demand-pull policy can induce technological innovation, the role of location of demand-pull policy in inducing technological innovations has been debated, with results pointing in opposite directions. More broadly, a detailed understanding of how local and non-local market demand differentially impact technological innovations has been lacking. Furthermore, we also lack knowledge about to what extent local demand-pull policy plays an indispensable role in spurring local technological innovations. In essay one of this dissertation results show that whether local demand-pull policy is essential in promoting local innovation depends on technological characteristics and the nature of associated learning processes. For technologies that have a significant local nature, the major benefits, as reflected in patenting patterns, of local demand-pull policy are appropriated by local firms. But technological innovations of standard and globally traded technologies can potentially be stimulated by non-local markets as well. These findings lend support for previous studies that the technological learning process in solar PV panels is a global phenomenon, while adding entirely novel empirical evidence that the technological learning process for non-module components in PV is a local phenomenon. Furthermore, findings from essay one help reconcile apparently inconsistent results from prior literature on whether the location of demand matters for inducing innovation. Therefore, essay one contributes to providing a better understanding that the impact of demand-pull policies on the geography of innovation necessitates accounting for technological characteristics and, in particular, how those characteristics accentuate certain learning processes over others.
137 Second, this dissertation sheds lights on bottom-up approach to renewable energy governance. The difficulties of supporting the local green industry development lie in the necessity of achieving multiple policy goals. These policy goals include, but are not limited to, environmental protection, energy security, technological innovation, and economic competitiveness. The success of subnational governments’ efforts in expanding market demand and translating it into local innovative activities and economic benefits underscores the potential viability of a bottom-up approach to renewable energy governance. The localized economic gains from expanding local market demand could mobilize a large number of subnational actors in support of advancing certain renewable energy technologies locally. These subnational actions could be voluntary and interest-driven, thus potentially underlying a viable mechanism for the bottom-up approach.
Third, this dissertation contributes to integrating localized learning literature and global learning literature. The localized learning literature mainly focuses on the local aspect of technological learning and the global learning literature mainly emphasizes the role of global pipeline and international collaborations in technological innovations. Essay two in this dissertation takes into account both the local and non-local aspects of technological learning at the same time. Results show that in firms’ knowledge acquisition process, firms acquire more knowledge from locations closer to their own locations. This finding suggests that knowledge searching behaviors for PV BOS are geographically localized. Furthermore, the results indicate that knowledge acquisition behaviors are associated with technology characteristics. I find that technologies that are more related to the local context use more local knowledge, which indicates that technologies with a more significant local nature have a higher dependence on localized learning. However, localized learning is not the whole story of technological learning. Owing to their different characteristics, local knowledge and non-local knowledge play different roles in technological innovation. In essay two I find that the geographic diversity of knowledge acquisition, especially absorbing international knowledge, has a significantly positive impact on
138 innovation quality. Incorporating both the local and non-local aspects of technological learning provides a better understanding of how to balance the protective approach (i.e., develop local green industries with local strategies) and the collaborative approach (i.e., involved in the global learning network).
Fourth, this dissertation contributes to separating the effects of three major learning mechanisms (learning by doing, learning by searching, and learning by interacting) through which demand-pull policy can facilitate cost reductions. The direct impact of demand-pull policy is to increase market demand and many studies support that the accumulation of experience as directly engendered through market expansion is positively associated with cost reductions. However, because of the existence of multiple learning sources and learning mechanisms in the experience accumulation process, the level of cost reductions may vary across different firms given a certain volume of experience accumulation. Essay three of this dissertation contributes a concrete understanding of different learning sources and learning mechanisms in the experience- accumulation process. The results of essay three show that the relationship between experience accumulation and cost reduction is mediated by learning by searching and learning by interacting. Importantly, essay three is the first study to explore the drivers of non-hardware installation costs in the solar PV industry by including experience-related factors, innovation-related factors, and network-related factors. Through studying a service industry, this essay also contributes to incorporating supply-chain collaboration theories into the learning-curve relationship. Specifically, I find that firms in the service industry can achieve cost advantages through their collaborative partners, especially their upstream suppliers, while the effectiveness of network structures associated with cost reductions may vary based on the specific type of upstream suppliers.
POLICY IMPLICATIONS
A critical aspect of policy designs in supporting local green industry is to understand how technological learning and innovation happen locally. In the case of demand-pull policy, although
139 demand-pull policies are widely recognized as an effective way to induce technological innovation, the benefits of local demand-pull policy are not necessarily only captured by local firms. It is possible that local demand-pull policy can help generate competitive advantages for non-local firms. Therefore, when considering and designing local demand-pull policy, policymakers may mainly focus on supporting technologies that have significant local nature.
The success of subnational governments’ efforts in expanding market demand and translating it into local innovative activities in PV BOS emphasizes the potential feasibility of a bottom-up approach to renewable energy governance. Thus, if the central/federal government is slow to move on environmental policies, or only focuses on supporting export-led growth, it behoove subnational policymakers to recognize the potential economic benefits from market creation may be more appropriable by local actors and consider taking voluntary, interest-driven actions to expand local market demand for renewable energy.
Another essential element in building an innovative local green industry is to engage with global learning networks. Policymakers should realize the critical role of international knowledge in local innovations and the difficulties of accessing and absorbing international knowledge faced by local firms. To engage productively with the global learning network might be more difficult for firms that locate outside the technological core areas, have a low absorptive capacity, and are in industries that face greater technological uncertainty. Therefore, policymakers may consider helping local firms go beyond localization and reach out for broad, distant, and unique knowledge. For example, local governments or industry associations can organize more formal or informal activities that facilitate knowledge spillover between local and non-local firms. Moreover, building a prosperous local industry cluster is a long and difficult process. The value of non-local knowledge indicates an alternative way that local policymakers can start by helping local firms build networks with firms in other areas – this can partially compensate for insufficient local knowledge, especially in early phases of cluster development.
140
Last but not the least, it is well recognized that the key challenge of emerging renewable energy sources is the relatively higher cost compared to incumbent energy sources. Therefore, an ultimate goal of demand-pull policy should be cost reductions. In order to achieve cost reductions through demand-pull policy, policymakers should understand that demand-pull policy may interact with innovation-supporting policies and network-facilitating policies, because multiple learning sources and mechanisms exist in the experience accumulation process. In order to achieve a greater learning rate from the experience accumulation, policymakers need to consider the interactions among different policy instruments and designs of policy mix.
FUTURE DIRECTIONS
First, although installed capacity is a commonly-used proxy measure for demand-pull policy, it is not a direct way to measure demand-pull policy. Using capacity as a proxy is necessitated partly because it is problematic to define the scope of demand-pull policy and categorize specific innovation-support policies as demand-pull policy or not (Kemp and Pontoglio, 2011; Taylor, 2008). Future work could try to address this limitation by developing a clear typology of demand-pull policy and leveraging the variation of specific demand-pull policies state- by-state to improve measurement of the local demand-pull variable. Based on a clear policy classification, future work could further explore how local policies and institutions are involved in innovation ecosystems, and their roles in spurring local innovations, building local innovative communities, and creating a region-specific advantage.
Second, while the main topic of this dissertation was the impact of deployment/demand- pull policy on technological innovations and its conversion to cost reductions in the solar PV industry, a further related question is to what extent policymakers should support solar PV technology deployment, considering there are various types of new energy technologies. Even within the solar PV industry, similar questions arise: to what extent policymakers should support
141 different types of solar modules (crystalline vs. thin film) and different types of solar PV projects (utility-scale vs, distributed PV). The high-level related question is how policymakers should choose between technology-neutral or technology-specific deployment policy. Future works could explore the motivations of policymakers’ decision between technology-neutral and technology- specific policy and compare the cost-effectiveness of technology-neutral and technology-specific policy under different economic, political, and social contexts.
Third, this dissertation emphasizes the value of international knowledge, which suggests the role of international collaboration on promoting technological change. A fruitful future direct might be to explore whether and how policy can help improve collaboration and coordination among countries. My dissertation only focuses on the impact of geographic distance in technological innovation, but it would be valuable to further explore the impact of social, cultural, institutional and technical distances and barriers on international collaborations. Furthermore, to limit global warming requires responses and actions from all the counties around the world. International collaborations also play an essential role in facilitating developing countries’ catch- up on renewable energy deployment. Therefore, an important direction to explore in future is how to take advantage of international collaborations to overcome the technical and institutional barriers faced by developing countries.
142 Appendices
APPENDIX A: SUPPLEMENTAL INFORMATION FOR CHAPTER 3
Robustness check
In the first robustness check, I add global demand into my empirical models and check whether including global demand would confound my estimate on local demand. The growth trends of global demand and non-local demand in China are almost identical. The correlation between global demand and nonlocal demand in China is as high as 0.9887 (See Table B2: Correlation table for independent variables), including both two variables in the model could raise the issue of multicollinearity. Consequently, I only include local demand and global demand in this robustness check.
Table A1 shows the regression results that include global demand. The dependent variables (DV) in the model 1, model 2, and model 3 are the annual number of all inventions at the province- level, the annual number of inverter inventions at the province-level, and the annual number of non-inverter inventions at the province-level, respectively. As can be seen, these results are very similar to the regression results in the main text. Local demand has a significant and positive effect on local innovations. The global demand has a positive effect on local inverter inventions, but it has negative effects on both all inventions and non-inverter inventions. It suggests the similar implication in the main text, which is the local demand has an essential role in spurring local innovative activities, and the role of global demand in local innovations depends on technology characteristics.
143 Table A1: Regression results of including global demand
Model 1 Model 2 Model 3
All BOS Inverter Non-inverter Dependent Variable inventions inventions inventions Log local demand (MW) 0.229** 0.154* 0.560*** (0.0916) (0.0977) (0.185) Log global demand (MW) -0.538 1.021 -6.097 (3.333) (3.908) (4.727) Log total patent (counts) 0.677*** 0.759*** 0.588** (0.151) (0.163) (0.235) Log manufacture salary (￥) -0.000506 3.164 -3.349 (2.132) (2.285) (3.532) Log utility salary (￥) -0.628 -2.147*** 1.207 (0.834) (0.800) (1.351) Log wholesale & Retail salary (￥) 0.609 -1.166 2.816 (1.709) (1.760) (2.565) Log scientific, & technical services 0.242 -0.165 0.406 salary (￥) (0.954) (0.846) (1.256) Constant -7.682 -8.656 14.19 (8.196) (9.989) (1153.8) Year fixed effect yes yes yes Observations 90 90 90 * p < 0.10, ** p < 0.05, *** p < 0.01
In the second robustness check, I exclude the independent variable log manufacture salary. The correlation table (See Table A2) shows that the manufacture salary is highly correlated with utility salary, wholesale & retail salary, and scientific, & technical services salary. As we can see in the Table A2, manufacture salary is only highly correlated with control variables, and the variables of interest (local demand, nonlocal demand, and global demand) do not highly correlate with manufacture salary, and thus it only influences the standard errors of these control variables. More important is that including these control variables can keep the performance of the control variables as controls, and thus I keep all these control variables in the models in my main text. However, it is valuable to test whether including manufacture salary can have an impact on the variables of my interest.
144
Table A2 shows the regression results that exclude the independent variable log manufacture salary. These results are very similar as those in my main text. The dependent variables (DV) in the model 4, model 5, and model 6 are the annual number of all patents at the province-level, the annual number of inverter patents at the province-level, and the annual number of non-inverter patents at the province-level, respectively. The dependent variables (DV) in the model 7, model 8, and model 9 are the annual number of all inventions at the province-level, the annual number of inverter inventions at the province-level, and the annual number of non-inverter inventions at the province-level, respectively. The results show that local demand has a significant and positive effect on local innovations. The non-local demand has a positive effect on local inverter innovations, but it has negative effects on both all innovations and non-inverter innovations.
145 Table A2: Regression results of excluding colinear independent variable
Model 4 Model 5 Model 6 Model 7 Model 8 Model 9
Dependent All BOS Inverter Non-inverter All BOS Inverter Non-inverter Variable patents patents inventions inventions inventions inventions
Log local demand 0.206*** 0.280*** 0.175** 0.249** 0.317** 0.364* (MW) (0.0766) (0.104) (0.0890) (0.118) -0.148 (0.218) Log non-local -2.419 1.002 -3.511 1.177 6 -3.375 demand (MW) (1.876) (2.887) (2.182) (3.280) -4.126 (4.260) Log total patent 0.505*** 0.713*** 0.441*** 0.700*** 0.841*** 0.580** (counts) (0.113) (0.148) (0.128) (0.161) -0.2 (0.240) Log utility salary -0.206 -0.983* 0.380 -0.619 -1.278* 0.475 (￥) (0.481) (0.569) (0.571) (0.634) -0.752 (1.089) Log wholesale & 1.380* 0.329 1.244 0.782 1.269 0.564 Retail salary (￥) (0.731) (0.870) (0.917) (1.095) -1.204 (1.535) Log scientific, & technical services 0.299 0.416 0.746 0.128 -0.521 0.909 salary (￥) (0.620) (0.858) (0.748) (0.996) -1.073 (1.318) Constant 1.250 -1.076 2.023 -0.906 -4.13 1.632 (1.304) (2.014) (1.523) (2.294) -2.903 (2.959) Year fixed effect yes yes yes yes yes yes Observations 90 90 90 90 90 90 * p < 0.10, ** p < 0.05, *** p < 0.01
146 In the third robustness check, I include utility-scale installed capacity as an additional independent variable to check whether the spillover between utility-scale solar PV and distributed solar PV exists and whether my main results are sensitive to any potential spillover between these two markets. The new independent variable is �� ��������� , which is logged utility-scale installed capacity for province i in year t. I add this independent variable to the six main models. The Table A3 is the regression results of these six models including utility-scale installed capacity. The dependent variable in the models 1 to 6 are, respectively, the annual number of all BOS patents at the province level, the annual number of inverter patents at the province level, the annual number of non-inverter patents at the province level, the annual number of all inventions at the province- level, the annual number of inverter inventions at the province level, and the annual number of non-inverter inventions at the province level.
The regression results show that the coefficients of utility-scale installed capacity are not significant across all six models, and other coefficients are very similar to the results in the main text. The insignificant results of the coefficients of utility-scale installed capacity are consistent with my expectations, as I stated in the main text that utility-scale solar PV and distributed solar PV are two separated markets at province-level in China and PV BOS are not identical in these two markets.
Firstly, each province in China mainly focuses on either utility-scale installations or distributed installations. Among all the provinces in China, three provinces only have installed capacity in distributed solar and no installed capacity in utility-scale solar, four provinces only have installed capacity in utility-scale solar and no installed capacity in distributed solar, nine provinces have more than 80% of their total installed capacity in distributed solar, and four provinces have more than 80% of their total installed capacity in utility-scale solar. These descriptive data indicate that utility-scale solar PV and distributed solar PV are two separated
147 markets in China, meaning that there would be very low spillover between these two separated markets at the province level.
Secondly, non-inverter technologies are different in utility-scale market and distributed markets. Even for inverters, they are not identical in utility-scale solar PV and distributed solar PV as well. For example, central inverters are commonly used in utility-scale PV because of economies of scale and high value placed on utility interactive controls. But distributed solar projects are increasingly moving away from such central inverter design. In my data collection process, I also dropped the patents that are related only to utility-scale PV. Therefore, I believe it is reasonable to find that the utility-scale installed capacity has no effect on inverter innovations – which are measured in the variables to closely mirror innovations in the distributed PV market.
148 Table A3: Regression results of including utility-scale installed capacity
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Non- Non- All BOS Inverter All BOS Inverter Dependent Variable inverter inverter patents patents inventions inventions patents inventions Log local demand 0.203*** 0.259** 0.205** 0.250** 0.280** 0.417* (MW) (0.0748) (0.103) (0.0878) (0.120) (0.137) (0.224) Log non-local demand -2.580 1.086 -3.400 1.180 5.797 -3.364 (MW) (1.864) (2.777) (2.112) (3.291) (3.878) (4.278) Log local utility-scale 0.00547 0.00751 0.00808 -0.000489 0.00843 -0.0125 capacity (MW) (0.00550) (0.00682) (0.00678) (0.00815) (0.00840) (0.0119) Log total patent 0.527*** 0.749*** 0.464*** 0.699*** 0.884*** 0.585** (counts) (0.110) (0.142) (0.121) (0.162) (0.181) (0.242) Log manufacture -0.171 2.298 -2.213 -0.0811 3.126 -3.375 salary (￥) (1.419) (1.897) (1.707) (2.185) (2.160) (3.586) -0.316 -1.706** 0.634 -0.589 -2.437*** 1.624 Log utility salary (￥) (0.588) (0.751) (0.693) (0.884) (0.788) (1.405) Log wholesale & 1.588 -0.820 2.775** 0.823 -0.699 1.967 Retail salary (￥) (1.080) (1.412) (1.365) (1.738) (1.713) (2.618) Log scientific, & 0.271 0.184 0.762 0.132 -0.513 1.236 technical services salary (￥) (0.614) (0.880) (0.716) (1.003) (0.826) (1.297) 1.848 -15.40 2.530 -19.36 -49.39 17.29 Constant (15.43) (22.26) (17.49) (26.30) (30.53) (749.6) Year fixed effect yes yes yes yes yes yes Observations 90 90 90 90 90 90 * p < 0.10, ** p < 0.05, *** p < 0.01
149 Table A4, table A5, and table A6 are regression results using Poisson regression. Table A4 shows regression results of equation 1 and 2 and use all the BOS patents in China. Table A5 shows regression results of equation 3 and use all the BOS patents. Table A6 uses only inventions in the BOS patent dataset. Model 8 and model 9 in the table A6 also differentiate inverter patents and non-inverter patents. The regression results in table A4, table A5, and table A6 using Poisson regression are similar to the results in the main text using negative binominal regressions.
Table A4: Regression results of equation 1 and 2 using Poisson regression
Model 1 Model 2 Model 3 Fixed effect Random effect Fixed effect 0.0776 0.164** 0.166** Log local demand (MW) (0.0937) (0.0758) (0.0711) -2.610 0.270 -1.994 Log non-local demand (MW) (1.757) (0.499) (1.335) 0.741* 0.264 0.578*** Log total patent (counts) (0.411) (0.330) (0.106) -2.742 -2.840* -1.353 Log manufacture salary (￥) (1.749) (1.605) (1.168) 1.277** 1.225** 0.569 Log utility salary (￥) (0.593) (0.586) (0.502)
Log wholesale & Retail salary 2.543** 2.144** 2.226** (￥) (1.053) (1.011) (0.885)
Log scientific & technical -1.234* -1.424** -0.423 services salary (￥) (0.636) (0.625) (0.508) -0.865*** Constant (0.308)
Year fixed effect Yes Yes No
Province Fixed effect Yes No Yes
Observations 90 90 90 * p < 0.10, ** p < 0.05, *** p < 0.01
150 Table A5: Regression results of equation 3 using Poisson regression Model 4 Model 5 Model 6 Fixed Effects Fixed Effects Fixed Effects Group by Group by index Group by GDP geography 0.260*** 0.0899 0.334*** Log local demand (MW) (0.0793) (0.114) (0.103) 1.947 -3.155 -1.509 Log non-local demand (MW) (2.660) (2.523) (2.686) 0.798*** 0.957*** 1.184*** Log total patent (Counts) (0.109) (0.128) (0.174) 2.322* 0.352 -0.789 Log manufacture salary (￥) (1.382) (1.436) (1.393) -0.216 -0.316 -0.0949 Log utility salary (￥) (0.374) (0.643) (0.296)
Log wholesale & Retail -1.270 -1.398 0.359 salary (￥) (0.914) (1.158) (1.023)
Log scientific, & technical -0.438 0.630 0.112 services salary (￥) (0.615) (0.501) (0.370) -25.90 22.72 3.363 Constant (19.70) (20.23) (21.79)
Year fixed effect Yes Yes Yes
Group Fixed effect Yes Yes Yes
Observations 90 90 90 * p < 0.10, ** p < 0.05, *** p < 0.01
151 Table A6: Regression results for including only inventions in the dependent variable and differentiating inverters and non-inverters using Poisson regressions Model 7 Model 8 Model 9 All BOS Inverter Non-inverter Dependent Variable inventions inventions inventions Log local demand (MW) 0.298*** 0.344** 0.391* (0.110) (0.134) (0.221) Log global demand (MW) 1.912 6.943* -3.590 (2.687) (3.755) (4.252) Log total patent (counts) 0.797*** 0.876*** 0.564** (0.153) (0.182) (0.236) Log manufacture salary (￥) 0.0988 0.967 -2.803 (1.877) (2.249) (3.556) Log utility salary (￥) -0.129 -0.952 1.156 (0.805) (1.006) (1.360) Log wholesale & Retail salary (￥) 1.174 0.924 2.229 (1.476) (1.804) (2.603) Log scientific, & technical services salary -1.115 -1.659* 1.052 (￥) (0.847) (0.985) (1.299) Constant -23.65 -56.21* 3.015 (22.05) (30.45) (34.36) Year fixed effect yes yes yes Observations 90 90 90 * p < 0.10, ** p < 0.05, *** p < 0.01
152 Summary statistics
Table A7: Summary statistics for key province-level variables. Data source: Local demand and non-local demand are from China NEA (2014), NEA (2015), and Zhang (2016). Total patents for each province are from the China Statistical Yearbook. The salaries for four complementary industries are from China Statistical Yearbook and China Labor Statistical Yearbook.
Variables Mean SD Min Max
Local demand (MW) 103 146 0 850
Non-local demand (MW) 3257 971 2040 4680 Total number of BOS patents 16 30 0 182 (Counts per year) Total number of patents (Counts 67545 98565 170 504500 per year) Manufacture salary 30150 6057 17849 50154 (Chinese Yuan) Utility salary (Chinese Yuan) 29700 6891 10058 46725 Wholesale & retail salary 28392 5941 19201 47109 (Chinese Yuan) Scientific, & technical services 35901 9053 19670 70425 salary (Chinese Yuan)
153 Table A8: Correlation table for independent variables Scientific, & Non- technical Local local Global services Wholesale Utility Total demand demand demand salary & Retail salary Manufacture patent (MW) (MW) (MW) (￥) salary (￥) (￥) salary (￥) (counts) Local demand (MW) 1 Non-local demand (MW) -0.0614 1 Global demand (MW) 0.046 0.9887 1 Scientific, & technical services salary (￥) 0.3278 0.3895 0.4306 1 Wholesale & Retail salary (￥) 0.2772 0.4302 0.4674 0.3396 1 Utility salary (￥) 0.0972 0.4364 0.4545 0.4683 0.5244 1 Manufacture salary (￥) 0.3326 0.473 0.5138 0.787 0.8865 0.7211 1 Total patent (counts) 0.5452 0.0006 0.0821 0.4628 0.4314 0.2285 0.4123 1
154 Three ways of grouping the 31 provinces in China
Central East NorthEast North South SouthWest West
Figure A1: The geographic distribution of province groups based on geography. Note: I grouped the provinces into seven groups including Eastern provinces, Northern provinces, Western provinces, Southern provinces, Central provinces, Northeastern provinces, and Southwestern provinces.
Group1 Group2 Group3 Group4 Group5 Group6
Figure A2: The geographic distribution of province groups based on competitiveness index according to the city ranking in the “Blue book of provincial competitiveness”. Note: I grouped provinces into six groups according to ranking: top 5 provinces, 6-10 ranked provinces, 11-15 ranked provinces, 16-20 ranked provinces, 21-25 ranked provinces, and 26-31 ranked provinces.
155 GDP1 GDP2 GDP3 GDP4 GDP5 GDP6
Figure A3: The geographic distribution of province groups based on GDP. Note: I ranked the provinces according to provincial GDP in 2014 and grouped provinces into six groups according to ranking: top 5 provinces, 6-10 ranked provinces, 11-15 ranked provinces, 16-20 ranked provinces, 21-25 ranked provinces, and 26-31 ranked provinces.
156 Specific items for PV BOS classification
Table A9: Scope of the technologies included within each of the four main BOS components analyzed in the main chapter. Source: Venugopalan, S. and Rai, V., 2015. Topic based classification and pattern identification in patents. Technol. Forecast. Soc. Change 94, 236–250. doi:10.1016/j.techfore.2014.10.006
Broad Category Sub categories Power conditioning equipment and methods 1) Solar Inverter Power conversion equipment and methods Voltage detection Electrical safety Maximum power point tracking Electrical connection and management Energy Storage Mounting hardware 2) Solar Mounting/Racking Structural equipment, connections and placement Solar tracking equipment Installation methods Rails Hardware Maintenance/ Repair 3) PV System Monitoring Power generation continuity management Physical security and maintenance Remote diagnostics and alerts System operation and control Graphical interface Remote geographic placement 4) Site Assessment Shade detection/ estimation Estimation of PV output potential Use of satellite imagery for placement
157 APPENDIX B: AN ANALYSIS OF US BOS INNOVATIONS
The U.S. is another leader of global solar energy development. Due to rich solar resources and plenty of available rooftops, distributed PV is a suitable choice for the U.S., especially for customers who have concerns about climate change and energy transition. Both the federal government and state governments in the U.S. issued several kinds of policies to create and expand distributed solar installations. To study the different effects of local demands and non-local demands on PV BOS innovations in the U.S., I built a new database of PV BOS patents related to distributed-generation PV based on patent information from both the United States Patent and Trademark Office (USPTO) between 2000 and 2015. There are three types of patents in the U.S., including utility patents, design patents and plant patents. The USPTO defines utility patents as “any new and useful process, machine, manufacture, or composition of matter, or a new and useful improvement thereof”. All the patents in my U.S. patent database are utility patents, which is consistent with the USPTO statement, “most patent applications filed at the USPTO are utility applications”. A design patent is defined as “a new, original, and ornamental design for an article of manufacture”, and a plant patent is defined as “any distinct and new variety of plant”. These two patent types are not relevant to this analysis.
Figure B1 shows the U.S. PV BOS patents by technology category. The mounting technologies account for the largest share of PV patents in the U.S. (about 55%). Inverters are the second large category, which accounts for about 26% of all the BOS patents. Monitoring and site assessment are 15% and 4% of all the BOS patents, respectively. An interesting phenomenon is that the total number of PV BOS patenting activity decreased since 2011. The decrease in patenting activities has been found in the literature (Carvalho et al., 2017), and the fall in patenting activities since 2011 has occurred in production equipment, silicon, and cells. Figure B1 shows that the decreases also occurred in inverters and monitoring. But the patenting activities in mounting and site assessment was relatively steady. It indicates that the decreasing trend is more important for globally-traded products, which supports the conclusion in Carvalho et al. (2017), namely the
158 decrease is driven by severe market competition from China. I further explored the reason behind the trend of patenting activities by examining the number of assignees. Figure B2 shows that the annual number of assignees by technology. Both solar inverter and monitoring have experienced a decreasing trend in the number of assignees since 2011 and 2012, but the general trend of the number of assignees in mounting and site assessment is increasing. The different trends of these two types of technologies (inverter and monitoring vs. mounting and site assessment) support that many producers of globally-traded products have exited the market which might be driven by competition from China. But the number of market players who heavily depend on local market scale is increasing with the expansion of local market size.
180 160 140 120 100 80 60 40 20 0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 Inverter Mounting Monitoring Site Assessment
Figure B1: Annual U.S. PV BOS patents by technology category (count)
159 60
50
40
30
20
10
0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015
Solar Inverter Site Assessment Solar Monitoring Solar Mounting/Rack
Figure B2: Annual U.S. PV BOS assignees by technology category (count)
Figure B3 shows the distribution of US BOS patents by state. 40% of the BOS patents belong to California state. The shares of all the other states are all below 10%. Six states only owe the BOS patents less than 1%. Eight states have no patents filed in the field of BOS components.
160 Oregon Tennessee 2% 1% California New York Colorado Texas Arizona California 2% 40% Ohio Massachusetts Pennsylvania New Mexico Florida New Jersey Illinois Michigan 2% 2% Washington Florida Washington New Jersey Arizona 2% Oregon Delaware Michigan North Carolina Maryland 2% Utah Virginia Vermont Connecticut Illinois 2% Georgia Rhode Island Minnesota Tennessee New Mexico 3% Indiana Wisconsin Louisiana Hawaii Pennsylvania Missouri DC 3% Mississippi New Hampshire Massachusetts Nevada South Carolina 3% Iowa Idaho New York Ohio Kansas Maine 5% Texas 8% Colorado Montana 5% 8%
Figure B3: The share of U.S. PV BOS patents by state
Table B1 shows the data source and description of dependent variables. The independent variable is the number of BOS patents for each state in each year. I also include year fixed effect and state fixed effect in the models. Table B2 shows the regression results of equation 1 in the main text. Because there might be time-lag between market demand and innovation outputs. Therefore, Model 1 has no lag of the demand variable, model 2 to model 6 has one-year to five- year lag of the demand variable. The regression results are consistent with the results of China’s solar PV industry. All the models show that local demand (i.e., within a state) can positively stimulate local PV BOS innovations, but non-local demand has no effect on promoting local innovations.
161 Table B1: Data sources and descriptions of dependent variables Variable Description Data source
Local demand Distributed PV installation in a state or a Solar Market Trend Report province Non-local demand Distributed PV installation in all the other Solar Market Trend Report states or provinces within a country
Median house price Median house price Federal Housing Finance Agency Complementary Sectors: Computer and Peripheral Equipment United States Patent and Trademark sector patents (NAICS code: 3341), Electrical Equipment, Office Appliances, and Components (NAICS code: 335), Fabricated Metal Products (NAICS code: 332), and Nonmetallic Mineral Products (NAICS code: 327) Gas retail price Motor Gasoline Sales to End Users Prices Energy Information Administration (EIA) Complementary Sectors: Utilities (NAICS code: 22), Statistics of U.S. Businesses sector payroll Manufacturing (NAICS code: 31-33), Professional, scientific, & technical services (NAICS code: 54), and Wholesale trade (NAICS code: 42). Residential electricity Residential electricity rate Energy Information Administration rate (EIA) Commercial Commercial electricity rate Energy Information Administration electricity rate (EIA) Household income Household income U.S. Census Bureau Industry scale Number of firms, number of establishments, Statistics of U.S. Businesses number of employments, and annual payroll.
162 Table B2: Regression results for the equation1 using different year lags
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 no lag 1-year lag 2-year lag 3-year lag 4-year lag 5-year lag Log local demand 0.139** 0.242*** 0.182*** 0.122* 0.151** 0.273*** (MW) (0.0696) (0.0719) (0.0677) (0.0656) (0.0646) (0.0740) Log non-local -0.106 0.350 0.322 -0.0929 -0.420* 0.244 demand (MW) (0.231) (0.267) (0.224) (0.241) (0.232) (0.218) 0.436 0.710 0.933 0.745 0.859 1.216* Log house price (0.634) (0.762) (0.674) (0.643) (0.602) (0.659) 1.585 1.575 1.780 2.574 1.487 2.238 Log median income (1.728) (1.707) (1.793) (1.821) (1.801) (1.902) Log manufacture -0.0236 -0.0287 -0.0330 -0.0269 -0.0127 -0.00419 patent (0.0227) (0.0230) (0.0228) (0.0241) (0.0243) (0.0254) 0.00631 -0.0136 -0.0127 0.0227 -0.00614 -0.0416 Log utility patent (0.0558) (0.0559) (0.0572) (0.0712) (0.0717) (0.0673) Log scientific & 0.772** 0.730** 0.601* 0.616 0.501 0.356 technical services patent (0.331) (0.345) (0.346) (0.379) (0.386) (0.383) Log wholesale & 0.0115 -0.00296 -0.000525 0.0242 0.00923 0.0129 Retail patent (0.0585) (0.0591) (0.0621) (0.0799) (0.0830) (0.0805) -4.022 4.805 1.454 -15.80 -13.28 -20.35 Log number of firms (16.58) (16.77) (16.70) (17.62) (18.82) (18.25) Log number of -1.163 -0.0518 0.368 -1.860 -2.900 -3.319 employees (3.557) (3.639) (3.662) (3.876) (4.584) (4.118) Log number of 6.830 -3.109 1.868 20.51 14.80 20.55 establishments (17.35) (17.90) (17.55) (18.64) (19.76) (18.97) -1.491 -1.470 -3.176 -2.463 1.724 3.325 Log payroll (3.190) (4.064) (3.356) (3.857) (6.115) (4.322) Log commercial -1.369 -0.812 -0.660 0.559 1.839 -0.434 electricity (1.198) (1.198) (1.285) (1.304) (1.352) (1.302) Log residential 1.613 0.792 0.962 0.969 0.633 2.671* electricity (1.521) (1.507) (1.577) (1.568) (1.512) (1.558) -17.05 -23.75 -22.54 -26.71 -39.47 -64.74** Constant (18.78) (22.14) (20.11) (20.80) (30.56) (25.87) Num of observations 269 247 216 193 174 151 State fixed effect yes yes yes yes yes yes Year fixed effect yes yes yes yes yes yes *p<0.1; **p<0.05; ***p<0.01
163 APPENDIX C: SUPPLEMENTAL INFORMATION FOR CHAPTER 4
Robustness check
Regression results for equation 3 and 4 are presented in table C1. Model 1, model 3, and model 5 are results for equation 3, which uses the count of backward citations that are at different geographic levels as the key independent variables. Model 2, model 4 and model 6 are results for equation 4, which uses geographic diversity to measure the geographic proximity between the focal patent and backward citations, instead of the count of backward citations that are at different geographic levels. Model 7 to model 9 in table C2 are results for equation 5, which uses the direct distances as independent variables to define local and non-local knowledge sources.
All the dependent variables in model 1 to model 9 include self-citations citations in the count of forward citation. The results in table C1 are almost identical to the results in table 15 and the results in table C2 are also similar to the results in table 16. It shows that the regression results are not sensitive to including or excluding self-citations.
Model 10 to model 19 use Poisson regressions to check whether the results are sensitive to estimation methods. Poisson regression is another widely used method to estimate count data. The results in table C3 are almost identical to the results in table 15 and the results in table C4 are also similar to the results in table 16. It shows that the regression results are not sensitive to the estimation method used.
164 Table C1: Regression results of equation 3 and equation 4 using different windows for receiving forward citations (include self-citations)
Dependent variable (include self-citations) Forward Citations in 4 Forward Citations in 5 Forward Citations in 6 years years years Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
-0.002 -0.007 -0.017 Same Country (0.012) (0.013) (0.015) 0.033** 0.047*** 0.060*** Different Country (0.017) (0.017) (0.020) -0.049** -0.042* -0.045 Same State (0.023) (0.024) (0.029)
Geographic 0.043*** 0.047*** 0.058*** Diversity (0.014) (0.014) (0.016) -0.019 -0.029 0.096 0.094 0.247 0.244 Inverter (0.412) (0.411) (0.395) (0.394) (0.468) (0.467) 0.208 0.208 0.245 0.238 0.294 0.284 Mounting (0.392) (0.392) (0.386) (0.386) (0.458) (0.458) -0.257 -0.257 -0.274 -0.283 -0.189 -0.201 Monitoring (0.426) (0.426) (0.415) (0.415) (0.485) (0.485) 0.455*** 0.455*** 0.305* 0.303* 0.065 0.058 Startup (0.165) (0.165) (0.164) (0.164) (0.182) (0.182) -0.030 -0.033 -0.449 -0.439 -0.227 -0.190 California State (0.304) (0.303) (0.440) (0.438) (0.450) (0.447) 0.006 0.006 0.003 0.002 0.0004 0.0001 Num of Claims (0.007) (0.007) (0.007) (0.007) (0.007) (0.007) -0.087 -0.083 0.104 0.098 0.205 0.188 Govint Contract (0.294) (0.293) (0.291) (0.290) (0.314) (0.313)
Total Backward -0.091*** -0.099*** -0.123*** Citations (0.032) (0.031) (0.036) 0.001 0.001 0.004* 0.004* 0.003 0.003 Installation (0.001) (0.001) (0.002) (0.002) (0.003) (0.003) Mean Lag of -0.008 -0.007 -0.010 -0.011 -0.030 -0.033 Backward Citations (0.019) (0.019) (0.018) (0.018) (0.021) (0.021) 2.533* 2.551* 3.197*** 3.191*** 3.464*** 3.447*** Constant (1.342) (1.340) (1.197) (1.196) (1.199) (1.198) Observations 339 339 268 268 201 201 *p<0.1; **p<0.05; ***p<0.01
165 Table C2: Regression results of equation 5 using different windows for receiving forward citations (include self-citations)
Dependent variable (include self-citations) Forward Citations Forward Citations Forward Citations in 4 years in 5 years in 6 years Model 7 Model 8 Model 9 -0.034** -0.027 -0.024 citations < 1000km (0.017) (0.016) (0.019) -0.023** -0.023* -0.023 1000km < citations <= 3000km (0.011) (0.013) (0.014) 0.004 -0.038 -0.094*** 3000km < citations <= 5000km (0.025) (0.026) (0.032) 0.082 -0.002 0.099 5000km < citations <= 7000km (0.076) (0.073) (0.082) 0.021 0.031 0.056** 7000km < citations <= 9000km (0.022) (0.022) (0.025) 0.092*** 0.097*** 0.087*** citations > 9000km (0.026) (0.028) (0.029) -0.001 -0.016 -0.017 Mean Lag of Backward Citations (0.018) (0.019) (0.021) 0.079 -0.113 0.145 Inverter (0.393) (0.415) (0.449) 0.311 0.151 0.232 Mounting (0.364) (0.399) (0.440) -0.329 -0.450 -0.655 Monitoring (0.409) (0.434) (0.471) 0.407** 0.262 -0.016 Start-up (0.165) (0.165) (0.180) -0.125 -0.441 -0.243 California State (0.303) (0.437) (0.445) 0.005 0.001 0.002 Num of Claims (0.007) (0.007) (0.007) 0.059 0.104 0.155 Govint Contract (0.296) (0.29) (0.302) 0.001 0.004* 0.004* Installation (0.001) (0.002) (0.003) 2.364* 3.537*** 3.505*** Constant (1.319) (1.183) (1.143) Observations 346 261 196 *p<0.1; **p<0.05; ***p<0.01
166 Table C3: Regression results of equation 3 and 4 using Poisson regression
Model 10 Model 11 Model 12 Model 13 Model 14 Model 15 Forward Citations within Forward Citations within Forward Citations within 4 years 5 years 6 years -0.0232 -0.0255 -0.0311 Same State (0.0259) (0.0257) (0.0284) -0.0235 -0.0249 -0.0292 Same Country (0.0162) (0.0160) (0.0184) 0.0563** 0.0615** 0.0705*** Different Countries (0.0243) (0.0245) (0.0272) 0.0489* 0.0530* 0.0627* Geographic Diversity (0.0279) (0.0298) (0.0338) 0.743 0.787 0.619* 0.648* 0.963*** 1.002*** Inverter (0.641) (0.662) (0.328) (0.344) (0.337) (0.336) 0.657 0.680 0.474 0.489 0.789** 0.813** Mounting (0.653) (0.666) (0.336) (0.348) (0.362) (0.362) 0.0976 0.128 0.0886 0.0985 0.235 0.244 Monitoring (0.553) (0.584) (0.292) (0.316) (0.378) (0.378) 0.465** 0.436* 0.322 0.297 0.110 0.0782 Startup (0.219) (0.225) (0.265) (0.266) (0.301) (0.304) -0.440 -0.398 -0.699 -0.675 -0.518 -0.491 California State (0.370) (0.365) (0.451) (0.442) (0.540) (0.533) 0.00604 0.00557 0.00121 0.000694 -0.00162 -0.00235 Num of Claims (0.00731) (0.00731) (0.00832) (0.00812) (0.00993) (0.00959) 0.419** 0.372** 0.526** 0.474** 0.613*** 0.539*** Govint Contract (0.182) (0.169) (0.204) (0.195) (0.192) (0.176) -0.108* -0.116* -0.138* Total Backward Citations (0.0650) (0.0699) (0.0795) 0.00204 0.00230 0.00453* 0.00495* 0.00377 0.00432 Installation (0.00152) (0.00157) (0.00269) (0.00262) (0.00347) (0.00341)
Mean Lag of Backward 0.0248 0.0226 0.0187 0.0151 0.0207 0.0158 Citations (0.0254) (0.0251) (0.0265) (0.0257) (0.0293) (0.0286) 1.690** 1.615** 2.702*** 2.638*** 2.576*** 2.514*** Constant (0.795) (0.819) (0.467) (0.486) (0.412) (0.419) Year fixed effect yes yes yes yes yes yes Observations 346 346 261 261 196 196 *p<0.1; **p<0.05; ***p<0.01
167 Table C4: Regression results of equation 5 using Poisson regression
Model 16 Model 17 Model 18 Forward Citations Forward Citations Forward Citations within 4 years within 5 years within 6 years -0.0308 -0.0358 -0.0308 citations < 1000km (0.0304) (0.0307) (0.0304) -0.0408*** -0.0466*** -0.0408*** 1000km < citations <= 3000km (0.00924) (0.0133) (0.00924) -0.0505** -0.0699*** -0.0505** 3000km < citations <= 5000km (0.0254) (0.0146) (0.0254) 0.0765 0.0308 0.0765 5000km < citations <= 7000km (0.0483) (0.0280) (0.0483) 0.0529*** 0.0596*** 0.0529*** 7000km < citations <= 9000km (0.0143) (0.0144) (0.0143) 0.119*** 0.133*** 0.119*** citations > 9000km (0.0221) (0.0318) (0.0221) 0.458 0.318 0.458 Inverter (0.437) (0.246) (0.437) 0.507 0.377 0.507 Mounting (0.460) (0.233) (0.460) -0.216 -0.250 -0.216 Monitoring (0.357) (0.243) (0.357) 0.385*** 0.223 0.385*** Startup (0.144) (0.185) (0.144) -0.133 -0.401 -0.133 California State (0.302) (0.388) (0.302) 0.00398 -0.00104 0.00398 Num of Claims (0.00525) (0.00650) (0.00525) 0.355** 0.443*** 0.355** Govint Contract (0.138) (0.125) (0.138) 0.00157 0.00432** 0.00157 Installation (0.00113) (0.00191) (0.00113) 0.0207 0.0126 0.0207 Mean Lag of Backward Citations (0.0198) (0.0198) (0.0198) 1.959*** 3.083*** 1.959*** Constant (0.535) (0.297) (0.535) Year fixed effect yes yes yes Observations 346 261 346 *p<0.1; **p<0.05; ***p<0.01
168 Summary statistics
Table C5: Summary statistics for key variables.
Variable Mean Std. Dev. Min Max Number of citation links 0.13 1.96 0.00 214.00 Forward citations in 5 years 10.40 17.50 0.00 134.00 Size_a 21.20 81.90 0.00 784.00 Size_b 3.81 18.48 0.00 327.00 Distance 6399.63 3852.54 3.99 16718.99 Within the same state 0.01 0.07 0.00 1.00 Within the U.S. 0.25 0.43 0.00 1.00 Same country 14.10 19.70 0.00 180.00 Different country 8.50 12.62 0.00 99.00 Same state 2.56 4.58 0.00 34.00 Distance < 1000km 4.47 6.58 0.00 61.00 1000km < Distance <= 3000km 6.71 15.88 0.00 144.00 3000km < Distance <= 5000km 3.23 4.72 0.00 29.00 5000km < Distance <= 7000km 0.45 1.26 0.00 9.00 7000km < Distance <= 9000km 3.52 5.90 0.00 30.00 Distance > 9000km 3.74 7.18 0.00 54.00 Mean lag of backward citations 8.77 4.47 0.00 24.29 Inverter 0.38 0.49 0.00 1.00 Mounting 0.41 0.49 0.00 1.00 Monitoring 0.18 0.38 0.00 1.00 Site assessment 0.04 0.18 0.00 1.00 Start up 2.46 0.50 2.00 3.00 Government contract 0.08 0.27 0.00 1.00 Installation 80.96 94.95 0.00 258.70 Number of claims 20.06 11.62 1.00 129.00 Total backward citations 22.50 31.60 1.00 279.00 California state 0.39 0.49 0.00 1.00
169 APPENDIX D: SUPPLEMENTAL INFORMATION FOR CHAPTER 5
Robustness check: excluding vertically integrated installers
There are several large installer firms that are vertically integrated. I test whether the network structures are sensitive to exclude these vertically integrated installers. I can identify these vertically integrated installers by the names of installers and panel/inverter manufacturers. For vertically integrated installers, the names of their panel/inverter manufacturers are the same as their own names. In the case of inverter manufacturers, there are three vertically integrated installers in the dataset, including ge energy, solectria renewables, and sunpower (see Table D1). These three installers used inverters from their own companies. About 56 kw installed capacity that were installed by ge energy in 2007 used inverters from the ge energy, which accounts for about 23% all the installed capacity finished by the ge energy (240 kw) in 2007. The solectria renewables only used inverters from their own company in 2005. The sunpower increased the percentage of using its own inverters from 6% to 96% from 2006 to 2011, and then decreased the percentage from 78% to 17% from 2012 to 2014.
The vertically integrated installers in term of panels is relatively more common (see Table D2). Seven installers in the dataset used their own panels. The overall trend is that these installers kept using increasingly more panels from their own companies. The solar integrated technologies and the solartech renewables only used their own panels in all their solar PV installation systems. The sunpower was the only installer that used both of its own inverters and panels from 2010 to 2014. The sunpower increased the percentage of using its own panels from 3% in 2001 to 100% in 2014. The bp solar increased the percentage of using its own panels from 82% in 2001 to 100% in 2005 and continued to only use its own panels until 2007. The ge energy increased the percentage of using its own panels from 92% in 2005 to 98% in 2007. The solar power industries increased the percentage from 8% in 2010 to 57% in 2011. The rec solar increased the percentage
170 of using its own panels from 39% in 2009 to 79% in 2012 and then slightly decreased to 68% in 2013 and 2014.
Table D1: Installed capacity and the percentage of vertically integrated installers that use their own inverters
Inverter Installed Total installed Installer Year Percent manufacturer capacity (w) capacity (w) ge energy 2007 ge energy 55968 239868 23% solectria solectria 2005 3080 3080 100% renewables renewables sunpower 2006 sunpower 1512 26944 6% sunpower 2008 sunpower 24443 48869 50% sunpower 2009 sunpower 517440 740497 70% sunpower 2010 sunpower 690230 1150855 60% sunpower 2011 sunpower 2928438 3053617 96% sunpower 2012 sunpower 5663990 7269697 78% sunpower 2013 sunpower 2282114 9249896 25% sunpower 2014 sunpower 1254529 7206442 17%
171 Table D2: Installed capacity and the percentage of vertically integrated installers that use their own panels
Panel Installed Total installed Installer Year Percent manufacturer capacity (w) capacity (w) bp solar 2001 bp solar 11550 14101.2 82% bp solar 2005 bp solar 154170 154170 100% bp solar 2006 bp solar 101080 101080 100% bp solar 2007 bp solar 56000 56000 100% ge energy 2005 ge energy 314520 341400 92% ge energy 2006 ge energy 356024.2 363744.2 98% ge energy 2007 ge energy 236148 239868 98% rec solar 2009 rec solar 2806840 7289608 39% rec solar 2010 rec solar 3454610 7567204 46% rec solar 2011 rec solar 2957360 8414875 35% rec solar 2012 rec solar 8567745 10810175 79% rec solar 2013 rec solar 7205040 10581050 68% rec solar 2014 rec solar 2336535 3418170 68% sunpower 2001 sunpower 1470 54286.1 3% sunpower 2006 sunpower 1512 26944 6% sunpower 2008 sunpower 28857 48869 59% sunpower 2009 sunpower 514834 740497 70% sunpower 2010 sunpower 690230 1150855 60% sunpower 2011 sunpower 2443458 3053617 80% sunpower 2012 sunpower 6583557 7269697 91% sunpower 2013 sunpower 9208321 9249896 100% sunpower 2014 sunpower 7206442 7206442 100% solar power solar power 2010 10400 135800 8% industries industries solar power solar power 2011 31400 55000 57% industries industries solar integrated solar integrated 2006 5952 5952 100% technologies technologies solar integrated solar integrated 2007 11904 11904 100% technologies technologies solar integrated solar integrated 2008 6528 6528 100% technologies technologies solartech solartech 2012 13340 13340 100% renewables renewables
I further test that whether the overall network structure would be changed if I exclude these vertical integrated installers. The average market share of these vertically integrated installers in
172 terms of inverter is 1.2%. If I only consider the installers’ installed capacity that use their own inverters, the average market share is only 0.5%. As for panels, the average market share of these vertically integrated installers is about 1.4%. Similarly, if I only consider the installers’ installed capacity that use their own panels, the average market share is only 0.9%. As there are only eight vertically integrated installers and their market shares are around 1%, so I expect that excluding these vertical integrated installers won’t impact the original network structure.
Figure D1 shows the distribution of installers’ projects to panel manufacturers and inverter manufacturers in the market. Figure D1 (a) includes all the installers and figure D1 (b) excludes the vertically integrated installers. These two figures are almost identical, showing that excluding these vertical integrated installers doesn’t impact the original network structure. I also plot the same figures (see figure D2) as figure D3 in the chapter 5 using the subset that excludes the vertically integrated installers. Figure 2D distinguishes innovative installers and non-innovative installers. Figure D2 (a) (b) and Figure D3 (a) (b) are also almost identical. Therefore, excluding these vertical integrated installers doesn’t impact the original network structure.
Figure D1 (a) Figure D1 (b)
Figure D1: The distribution of installers’ projects to panel manufacturers and inverter manufacturers in the market. Figure on the left includes all the installers and figure on the right excludes the vertically integrated installers.
173
Figure D2 (a) Figure D2 (b)
Figure D2: The distribution of installers’ projects to panel manufacturers and inverter manufacturers in the market (excluding the vertically integrated installers)
Figure D3 (a) Figure D3 (b)
Figure D3: The distribution of installers’ projects to panel manufacturers and inverter manufacturers in the market (including all the installers) (the same graph in the main text)
174 Robustness check: Instrumental variables
The table D3 shows the regression results including the instrumental variable. I use zip- code level annual insolation as an instrumental variable for the installed capacity. This is a valid instrumental variable (IV) because it is not associated with non-hardware installation prices, but this is a weak IV because the correlation between installed capacity and insolation is small. The innovation-related and network-related variables are almost identical as the results of models in the main texts, but the signs of installer’s own experience are changed and not statistically significant. The coefficients of experience-related variable with IV may be biased because the IV used in these models is weak.
Table D3: Regression results including the instrumental variable
Model 1 Model 2 Model 3 Model 4
9.87e-08 4.61e-08 0.000000104 8.17e-08 Experience (watt) (0.000000106) (6.90e-08) (7.05e-08) (7.41e-08) Other installers' experience 3.50e-10 1.61e-09 9.39e-11 1.49e-09 (watt) (2.39e-09) (1.42e-09) (1.48e-09) (1.50e-09) 0.00000107** 0.00000123** 0.00000122** 0.00000110** Labor Cost (0.000000539) (0.000000540) (0.000000547) (0.000000544) 0.0305 0.0203 0.0201 0.0233 Household Income (0.0277) (0.0241) (0.0242) (0.0241) -0.322*** -0.317*** -0.297*** -0.296*** China Panel (0.00964) (0.00934) (0.00957) (0.00944) 0.00231 0.0360 0.00659 0.00000763 Thin Film (0.0777) (0.0712) (0.0717) (0.0717) 1.784*** 1.615*** 1.804*** 1.756*** BIPV (0.265) (0.164) (0.168) (0.180) 3.279*** 3.273*** 3.272*** 3.253*** Battery (0.574) (0.572) (0.572) (0.574) 1.386*** 1.424*** 1.419*** 1.412*** Tracking (0.214) (0.208) (0.208) (0.206) -0.0458** -0.0563*** -0.0625*** -0.0724*** Micro Inverter (0.0208) (0.0145) (0.0140) (0.0141) -0.0656*** -0.0657*** -0.0646*** -0.0648*** System Size (0.00147) (0.00150) (0.00153) (0.00157) Knowledge stock -0.0339*** -0.0219*** -0.0202***
175 (0.00391) (0.00436) (0.00493) Experience with innovative -1.02e-08 -1.57e-08* -1.27e-08 inverter manuf (watt) (8.64e-09) (8.21e-09) (8.92e-09)
-0.0000364*** Centrality_Panel (0.00000706)
-0.0000226*** Centrality_Inverter (0.00000700)
-0.00000461 FC_Panel (0.00000419)
-0.00000881*** FC_Inverter (0.00000285) 0.000234*** 0.000261*** SC_Inverter (0.0000210) (0.0000211) -0.000318*** -0.000333*** SC_Panel (0.0000460) (0.0000508) -2.20e-08*** -2.69e-08*** SC_Inverter_Sqaure (4.57e-09) (4.62e-09) 4.93e-08*** 5.35e-08*** SC_Panel_Sqaure (8.06e-09) (8.84e-09)
0.00406 Degree_Panel (0.00699)
-0.0662*** Degree_Inverter (0.00740) -0.0281 SSI_Panel (0.0407) -0.326*** SSI_Inverter (0.0593) Total experience (watt)
Market share
Competition
Num of Obs 126904 126851 126851 126692 Installer fixed effect yes yes yes yes County fixed effect yes yes yes yes Year fixed effect yes yes yes yes *p<0.1; **p<0.05; ***p<0.01
176 Summary statistics
Table D4: Summary statistics for key variables. Variable Obs Mean Std. Dev. Min Max Soft cost 189,075 5.62 1.75 1.50 19.99 Installers' own experience 250,377 1486129 3032330 1002 24100000 (watt) Other installers' experience 250,377 27500000 33800000 0 142000000 in a county (watt) All the installers' experience 250,377 29000000 35100000 1351 142000000 in a county (watt) Knowledge stock 250,847 0.91 4.68 0.00 29.36 Experience with innovative 250,377 4841284 8471726 0 48800000 inverter manuf (watt) Market share 209,846 0.05 0.08 0.00 0.93 Competition 250,377 1424.62 1764.74 1.00 7146.00 Centrality_Panel 232,998 3860.01 1794.14 0.00 6867.00 Centrality_Inverter 232,998 5018.73 1678.72 0.00 7515.00 FC_Panel 232,998 5849.08 2637.81 1341.33 10000.00 FC_Inverter 232,998 5649.71 2609.49 1475.39 10000.00 SC_Inverter 232,998 651.66 715.03 107.56 10000.00 SC_Panel 232,998 1031.36 1021.04 139.80 10000.00 Degree_Panel 251031 4.78 3.43 1.00 17.00 Degree_Inverter 251031 4.43 2.37 1.00 11.00 SSI_Panel 226962 0.78 0.18 0.37 1.00 SSI_Inverter 226962 0.74 0.18 0.37 1.00 Labor cost 250,160 59643.02 15949.34 19468.41 199696.50 Household income 248,922 0.34 0.15 0.00 0.92 China panel 202,740 0.30 0.46 0.00 1.00 Thin film 250,847 0.01 0.12 0.00 1.00 BIPV 250,847 0.01 0.09 0.00 1.00 Battery 250,847 0.00 0.01 0.00 1.00 Tracking 250,847 0.00 0.05 0.00 1.00 Micro inverter 207,151 0.26 0.44 0.00 1.00 System size 250,847 5.83 2.80 1.00 15.00
177 References Ahuja, G., Lampert, C.M., 2001. Entrepreneurship in the large corporation: A longitudinal study of how established firms create breakthrough inventions. Strateg. Manag. J. 22, 521–543. Alcácer, J., Gittelman, M., 2006. Patent Citations as a Measure of Knowledge Flows: The Influence of Examiner Citations. Rev. Econ. Stat. 88, 774–779. https://doi.org/10.1162/rest.88.4.774 Allison, P.D., 2009. Fixed Effects Regression Models, SAGE publications. https://doi.org/http://0-dx.doi.org.innopac.up.ac.za/10.4135/9781412993869.d4 Almeida, P., Song, J., Grant, R.M., 2002. Are Firms Superior to Alliances and Markets? An Empirical Test of Cross-Border Knowledge Building. Organ. Sci. 13, 147–161. https://doi.org/10.1287/orsc.13.2.147.534 Altwies, J.E., Nemet, G.F., 2013. Innovation in the U.S. building sector: An assessment of patent citations in building energy control technology. Energy Policy 52, 819– 831. Antonelli, C., 1996. Localized knowledge percolation processes and information networks. J. Evol. Econ. 6, 281–295. https://doi.org/10.1007/bf01193634 Ardani, K., Margolis, R., Feldman, D., Ong, S., Barbose, G., Wiser, R., 2012. Benchmarking Non-Hardware Balance of System (Soft) Costs for U.S. Photovoltaic Systems Using a Data-Driven Analysis from PV Installer Survey Results. Ardani, K., Seif, D., Davidson, C., Morris, J., Truitt, S., Torbert, R., Margolis, R., 2013. Preliminary non-hardware ('soft’) cost-reduction Roadmap for residential and small commercial solar photovoltaics, 2013-2020. Conf. Rec. IEEE Photovolt. Spec. Conf. 3463–3468. https://doi.org/10.1109/PVSC.2013.6745192 Arrow, K.J., 1962a. Economic Welfare and the Allocation of Resources for Invention. Rate Dir. Inven. Act. Econ. Soc. Factors 609, 614–16. https://doi.org/10.1007/978-1- 349-15486-9_13 Arrow, K.J., 1962b. The Economic Implications of Learning by Doing. Rev. Econ. Stud. 29, 155. https://doi.org/10.2307/2295952 Arthur, W.B., 2007. The structure of invention. Res. Policy 36, 274–287. https://doi.org/10.1016/j.respol.2006.11.005 Arza, V., 2010. Channels, benefits and risks of public–private interactions for knowledge transfer: conceptual framework inspired by Latin America. Sci. Public Policy 37, 473–484. https://doi.org/10.3152/030234210X511990 Asheim, B.T., Coenen, L., Vang, J., 2007. Face-to-face, buzz, and knowledge bases: Sociospatial implications for learning, innovation, and innovation policy. Environ. Plan. C Gov. Policy 25, 655–670. https://doi.org/10.1068/c0648 Asheim, B.T., Isaksen, A., 2002. Regional Innovation Systems: The Integration of Local “Sticky” and Global “Ubiquitous” Knowledge. J. Technol. Transf. 22, 77–86. Asheim, B.T., Smith, H.L., Oughton, C., 2011. Regional Innovation Systems: Theory, Empirics and Policy. Reg. Stud. 457, 875–891. https://doi.org/10.1080/00343404.2011.596701 Atal, V., Bar, T., 2010. Prior art: To search or not to search. Int. J. Ind. Organ. 28, 507– 521. https://doi.org/10.1016/j.ijindorg.2009.12.002 178 Azoulay, P., Ding, W., Stuart, T., 2007. The determinants of faculty patenting behavior: Demographics or opportunities? J. Econ. Behav. Organ. 63, 599–623. https://doi.org/10.1016/j.jebo.2006.05.015 Baldwin, J., Hanel, P., Sabourin, D., 2000. The Determinants of Innovation in Canadian Manufacturing Firms, in: Kleinknecht, A., Mohnen, P. (Eds.), Innovation and Firm Performance. https://doi.org/10.2139/ssrn.229792 Baldwin, J.R., Johnson, J., 1999. Entry, Innovation and Firm Growth, in: Are Small Firms Important? Their Role and Impact. Springer US, Boston, MA, pp. 51–77. https://doi.org/10.1007/978-1-4615-5173-7_4 Barbose, G., Darghouth, N., 2018. Tracking the Sun - Installed Price Trends for Distributed Photovoltaic Systems in the United States, LBNL Report. Barbose, G., Darghouth, N., Weaver, S., 2014. Tracking the Sun VI An Historical Summary of the Installed Price Tracking the Sun VI An Historical Summary of the Installed Price of. SunShot - U.S. Dep. Energy 70. Barbose, G.L., Darghouth, N.R., Millstein, D., Spears, M., Wiser, R.H., Buckley, M., Widiss, R., Grue, N., 2015. Tracking the Sun VIII: The Installed Price ef Residential and Non-Resldential Photovoltaic Systems in the United States. Berkeley, CA (United States). Basberg, B.L., 1987. Patents and the measurement of technological change: A survey of the literature. Res. Policy 16, 131–141. https://doi.org/10.1016/0048- 7333(87)90027-8 Bathelt, H., Malmberg, A., Maskell, P., 2004. Clusters and knowledge: local buzz, global pipelines and the process of knowledge creation. Prog. Hum. Geogr. 28, 31–56. https://doi.org/10.1191/0309132504ph469oa Bathelt, H., Zeng, G., 2012. Strong growth in weakly-developed networks: Producer–user interaction and knowledge brokers in the Greater Shanghai chemical industry. Appl. Geogr. 32, 158–170. https://doi.org/10.1016/j.apgeog.2010.11.015 Baumol, W.J., 1972. On Taxation and the Control of Externalities. Am. Econ. Rev. 62, 307–322. Beise, R., Rammer, C., 2006. Local user-producer interaction in innovation and export performance of firms. Small Bus. Econ. 27, 207–222. https://doi.org/10.1007/s11187-006-0013-z Bollinger, B., Gillingham, K., 2014. Learning-by-Doing in Solar Photovoltaic Installations. Work. Pap. 46. Boon, B., Park, Y., 2005. A systematic approach for identifying technology opportunities: Keyword-based morphology analysis. Technol. Forecast. Soc. Change 72, 145–160. https://doi.org/10.1016/j.techfore.2004.08.011 Botolfmaurseth, P., Verspagen, B., 2002. Knowledge Spillovers in Europe: A Patent Citations Analysis. Scand. J. Econ. J Econ. 104, 531–545. https://doi.org/10.1111/1467-9442.00300 Bottazzi, L., Peri, G., Universit, L.B., 2002. Innovation and Spillovers in Regions : Innovation and Spillovers in Regions : Evidence from European Patent Data. Eur. Econ. Rev. 47, 1–40. Braun, D., 2008. Organising political coordination of knowledge and innovation policy. Sci. Public Policy 35, 227–39. 179 Brem, A., Voigt, K.I., 2009. Integration of market pull and technology push in the corporate front end and innovation management-Insights from the German software industry. Technovation 29, 351–367. https://doi.org/10.1016/j.technovation.2008.06.003 Brown, J.S., Duguid, P., 1991. Organizational learning and communities-of-practice: Toward a unified view of working, learning, and innovation. Organ. Sci. 2, 40–57. https://doi.org/10.1287/orsc.2.1.40 Burt, R., 1992. Structural holes: the social structure of competition (Harvard, MA, Harvard University Press). Caballero, R.J., Jaffee, A., 1993. How High are the Giant’s Shoulders: An Empirical Assessment of Knowledge Spillovers and Creative Destruction in a model of Economic Growth. N.B.E.R. Macroecon. Annu. 15074. Cachon, G.P., Fisher, M., 2000. Supply Chain Inventory Management and the Value of Shared Information. Manage. Sci. 46, 1032–1048. https://doi.org/10.1287/mnsc.46.8.1032.12029 Carayannis, E.G., Kassicieh, S., 1997. Higher order technological learning, strategic assets, core capabilities, and market performance in technology-driven firms: An empirical study. Innov. Technol. Manag. - Key to Glob. Leadersh. 71. https://doi.org/Doi 10.1109/Picmet.1997.653254 Carvalho, M.D., Dechezleprêtre, A., Glachant, M., 2017. Understanding the dynamics of global value chains for solar photovoltaic technologies. Chidamber, S. und Kon, H., 1994. A Research Retrospective of Innovation Inception and Success: The Technology-Push Demand-Pull Question. Int. J. Technol. Manag. 53, 1689–1699. https://doi.org/10.1017/CBO9781107415324.004 Choi, H., Anadón, L.D., 2014. The role of the complementary sector and its relationship with network formation and government policies in emerging sectors: The case of solar photovoltaics between 2001 and 2009. Technol. Forecast. Soc. Change 82, 80– 94. https://doi.org/10.1016/j.techfore.2013.06.002 Choi, S.B., Lee, S.H., Williams, C., 2011. Ownership and firm innovation in a transition economy: Evidence from China. Res. Policy 40, 441–452. https://doi.org/10.1016/j.respol.2011.01.004 Choi, S.B., Williams, C., 2014. The impact of innovation intensity, scope, and spillovers on sales growth in Chinese firms. Asia Pacific J. Manag. 31, 25–46. https://doi.org/10.1007/s10490-012-9329-1 Christensen, J.L., Stoerring, D., 2011. Interactive Learning for Innovation: A Key Driver within Clusters and Innovation Systems. Palgrave Macmillan UK, London. https://doi.org/10.1057/9780230362420 Cohen, W., Nelson, R., Walsh, J., 2000. Protecting Their Intellectual Assets: Appropriability Conditions and Why U.S. Manufacturing Firms Patent (or Not) (No. w7552.). Cambridge, MA. https://doi.org/10.3386/w7552 Cohen, W.M., Levinthal, D.A., 2006. Innovation and Learning: The Two Faces of R&D. Econ. J. 99, 569. https://doi.org/10.2307/2233763 Cohen, W.M., Levinthal, D.A., 1990. Absorptive Capacity – A New Perspective on Learning and Innovation. Adm. Sci. Q. 35, 128–152. https://doi.org/10.2307/2393553 180 Cooke, P., Uranga, M.G., Etxebarria, G., 1997. Regional innovation systems: Institutional and organizational dimensions. Res. Policy 26, 475–491. https://doi.org/10.1016/S0048-7333(97)00025-5 Corsten, D., Felde, J., 2005. Exploring the performance effects of key-supplier collaboration: An empirical investigation into Swiss buyer-supplier relationships. Int. J. Phys. Distrib. Logist. Manag. 35, 445–461. https://doi.org/10.1108/09600030510611666 Corwin, S., Johnson, T.L., 2019. The role of local governments in the development of China’s solar photovoltaic industry. Energy Policy 130, 283–293. https://doi.org/10.1016/j.enpol.2019.04.009 Cotropia, C.A., Lemley, M.A., Sampat, B., 2013. Do applicant patent citations matter? Res. Policy 42, 844–854. https://doi.org/10.1016/j.respol.2013.01.003 Crepon Bruno, Duguet, E., Mairesse Jacques, 1998. Research, Innovation and productivity: an econometric analysis at the firm level, NBER No 6696. Cuervo-Cazurra, A., Annique Un, C., 2010. Why some firms never invest in formal R&D. Strateg. Manag. J. 31, 759–779. https://doi.org/10.1002/smj.836 Dang, J., Motohashi, K., 2015. Patent statistics: A good indicator for innovation in China? Patent subsidy program impacts on patent quality. China Econ. Rev. 35, 137–155. https://doi.org/10.1016/j.chieco.2015.03.012 David, P. a., Hall, B.H., Toole, A. a., 2000. Is public R&D a complement or substitute for private R&D? A review of the econometric evidence. Res. Policy 29, 497–529. https://doi.org/10.1016/S0048-7333(99)00087-6 Davidson, C., Steinberg, D., 2013. Evaluating the impact of third-party price reporting and other drivers on residential photovoltaic price estimates. Energy Policy 62, 752– 761. https://doi.org/10.1016/j.enpol.2013.07.112 de La Tour, A., Glachant, M., Ménière, Y., 2013. Predicting the costs of photovoltaic solar modules in 2020 using experience curve models. Energy 62, 341–348. https://doi.org/10.1016/j.energy.2013.09.037 Dechezleprêtre, A., Glachant, M., 2014. Does Foreign Environmental Policy Influence Domestic Innovation? Evidence from the Wind Industry. Environ. Resour. Econ. 58, 391–413. https://doi.org/10.1007/s10640-013-9705-4 Dodgson, M., 1991. The Management of Technological Learning: Lessons from a Biotechnology Company. W. de Gruyter. Doraszelski, U., 2008. Rent dissipation in R&D races. Contrib. to Econ. Anal. https://doi.org/10.1016/S0573-8555(08)00201-0 Edler, J., Georghiou, L., 2007. Public procurement and innovation-Resurrecting the demand side. Res. Policy 36, 949–963. https://doi.org/10.1016/j.respol.2007.03.003 Edquist, C., 2011. Design of innovation policy through diagnostic analysis: Identification of systemic problems (or failures). Ind. Corp. Chang. 20, 1725–1753. Eisenhardt, K.M., Tabrizi, B.N., 1995. Accelerating Adaptive Processes: Product Innovation in the Global Computer Industry. Adm. Sci. Q. 40, 84. https://doi.org/10.2307/2393701 Enos, J.L., 2008. The adoption and diffusion of imported technology: the case of Korea, Routledge. Evangelista, R., Iammarino, S., Mastrostefano, V., Silvani, A., 2002. Looking for 181 regional systems of innovation: Evidence from the Italian innovation survey. Reg. Stud. 36, 173–186. https://doi.org/10.1080/00343400220121963 Fabrizio, K., Thomas, L.G., 2012. The Impact of Local Demand on Innovation in a Global Industry. Strateg. Manag. J. 33, 42–64. https://doi.org/doi: 10.1002/smj.942 Fagerberg, J., 2017. Innovation Policy: Rationales, Lessons and Challenges. J. Econ. Surv. 31, 497–512. https://doi.org/10.1111/joes.12164 Figueiredo, P.N., 2002. Learning processes features and technological capability- accumulation: Explaining inter-firm differences. Technovation 22, 685–698. https://doi.org/10.1016/S0166-4972(01)00068-2 Fiol, C.M., 1996. Squeezing harder doesn’t always work: continuing the search for consistency in innovation research. Acad. Manag. Rev. 21, 1012–1021. https://doi.org/10.5465/AMR.1996.15868543 Fleming, L., 2001. Recombinant Uncertainty in Technological Search. Manage. Sci. 47, 117–132. https://doi.org/10.1287/mnsc.47.1.117.10671 Fontana, R., Nuvolari, A., Verspagen, B., 2009. Mapping technological trajectories as patent citation networks. An application to data communication standards. Econ. Innov. New Technol. 18, 311–336. https://doi.org/10.1080/10438590801969073 Fusfeld, H.I., Haklisch, C.S., 1985. Cooperative R-AND-D for Competitors. Harv. Bus. Rev. 63. Gao, X., Rai, V., 2019. Local demand-pull policy and energy innovation: Evidence from the solar photovoltaic market in China. Energy Policy 128, 364–376. https://doi.org/10.1016/j.enpol.2018.12.056 Gertler, M.S., 2003. Tacit knowledge and the economic geography of context, or The undefinable tacitness of being (there). J. Econ. Geogr. 3, 75–99. https://doi.org/10.1093/jeg/3.1.75 Gillingham, K., Deng, H., Wiser, R., Darghouth, N.R., Nemet, G., Barbose, G., Rai, V., Dong, C., 2016. Deconstructing solar photovoltaic pricing: The role of market structure, technology, and policy. Energy J. 37, 231–250. https://doi.org/10.5547/01956574.37.3.kgil Gilsing, V., Nooteboom, B., Vanhaverbeke, W., Duysters, G., Van Den Oord, A., 2008. Network embeddedness and the exploration of novel technologies: Technological distance, betweenness centrality and density. Res. Policy 37, 1717–1731. Godin, B., 2002. Technological gaps : an important episode in the construction of S & T statistics. Technol. Soc. 24, 387–413. Goodrich, A., James, T., Woodhouse, M., 2012. Residential, Commercial, and Utility- Scale Photovoltaic (PV) System Prices in the United States: Current Drivers and Cost-Reduction Opportunities. Tech. Rep. NREL 64. https://doi.org/10.2172/1036048 Grabher, G., 2002. Cool Projects, Boring Institutions: Temporary Collaboration in Social Context. Reg. Stud. 36, 205–214. https://doi.org/10.1080/00343400220122025 Griliches, Z., 2013. The Search for R&D Spillover. Scand. J. Econ. 94, 28–47. https://doi.org/10.2307/3440244 Griliches, Z., 1990. Patent Statistics as Economic Indicators: A Survey (No. w3301.). https://doi.org/10.3386/w3301 Griliches, Z., 1957. Hybrid Corn: An Exploration in the Economics of Technological 182 Change. Econometrica 25, 501. https://doi.org/10.2307/1905380 Griliches, Z., Pakes, A., Hall, B., 1986. The Value of Patents as Indicators of Inventive Activity. Econ. Policy Technol. Perform. https://doi.org/10.3386/w2083 Grimpe, C., Sofka, W., 2009. Search patterns and absorptive capacity: Low-and high- technology sectors in European countries. Res. Policy 38, 495–506. https://doi.org/10.1016/j.respol.2008.10.006 Grossman, G., Helpman, E., 1991. R & D Spillovers and the Geography of Innovation and Production. Production 86, 630–640. https://doi.org/Article Grossman, G.M., Shapiro, C., 1985. Dynamic R&D competition (No. 1674), NBER. Grübler, A., 2003. Technology and Global Change, Cambridge University Press. Habermeier, K.F., 1990. Product use and product improvement. Res. Policy 19, 271–283. https://doi.org/10.1016/0048-7333(90)90040-D Haldin‐Herrgard, T., 2000. Difficulties in diffusion of tacit knowledge in organizations. J. Intellect. Cap. 1, 357–365. https://doi.org/10.1108/14691930010359252 Hall, B.H., 1999. Innovation and market value. NBER Work. Pap. https://doi.org/10.2139/ssrn.151912 Hansen, A. and S., 1997. Will Low Technology Products Disappear ? The Hidden Innovation Processes in Low Technology Industries. Technol. Forecast. Soc. Chang. 191, 179–191. Harhoff, D., Scherer, F.M., Vopel, K., 2003. Citations, family size, opposition and the value of patent rights. Res. Policy 32, 1343–1363. https://doi.org/10.1016/S0048- 7333(02)00124-5 Hassett, K.A., Metcalf, G.E., 1995. Energy tax credits and residential conservation investment: Evidence from panel data. J. Public Econ. 57, 201–217. https://doi.org/10.1016/0047-2727(94)01452-T Hausman, J., Hall, B.H., Griliches, Z., 1984. Econometric Models for Count Data with an Application to the Patents-R&D Relationship. Econometrica 52, 909. https://doi.org/10.2307/1911191 Hayami, Y., Ruttan, V., 1985. Agricultural Development: An Agricultural Perspective. Haynes, M., Thompson, S., 2008. Price, price dispersion and number of sellers at a low entry cost shopbot. Int. J. Ind. Organ. 26, 459–472. https://doi.org/10.1016/j.ijindorg.2007.02.003 He, F., Liu, S., 2014. DG Policy to trigger a peak season for installation. New York. Helfat, C.E., 2006. Open Innovation: The New Imperative for Creating and Profiting from Technology. Acad. Manag. Perspect. 20, 86–88. https://doi.org/10.5465/AMP.2006.20591014 Helfat, C.E., 2000. Guest Editor's Introduction to the Special Issue : the Evolution of Firm Capabilities. Strateg. Manag. J. 959, 955–959. Henderson, R.M., 1994. Managing Innovation in the Information Age. Harv. Bus. Rev. 72, 100–106. Hicks, J.R., 1932. The theory of wages. Springer. Hirotaka, T., Nonaka, I., 1986. The new new product development game. Harv. Bus. Rev. 137–146. Hitt, M.A., Ireland, R.D., Lee, H.-U., 2000. Technological learning, knowledge management, firm growth and performance: an introductory essay. J. Eng. Technol. 183 Manag. 17, 231–246. https://doi.org/10.1016/S0923-4748(00)00024-2 Hoppmann, J., Huenteler, J., Girod, B., 2014. Compulsive policy-making—The evolution of the German feed-in tariff system for solar photovoltaic power. Res. Policy 43, 1422–1441. https://doi.org/10.1016/j.respol.2014.01.014 Hoppmann, J., Peters, M., Schneider, M., Hoffmann, V.H., 2013. The two faces of market support - How deployment policies affect technological exploration and exploitation in the solar photovoltaic industry. Res. Policy 42, 989–1003. https://doi.org/10.1016/j.respol.2013.01.002 Hottenrott, H., Lopes-Bento, C., 2014. (International) R&D collaboration and SMEs: The effectiveness of targeted public R&D support schemes. Res. Policy 43, 1055–1066. https://doi.org/10.1016/j.respol.2014.01.004 Hou, J., Luo, S., Cao, M., 2019. A review on China’s current situation and prospects of poverty alleviation with photovoltaic power generation. J. Renew. Sustain. Energy 11, 13503. https://doi.org/10.1063/1.5048102 Huang, C., Arundel, A., Hollanders, H., 2010. How firms innovate: R&D, non-R&D, and technology adoption Working Paper Series. Huang, H.F., 2016. The Globalization of Clean Energy Technology: Lessons from China by Kelly Sims Gallagher. China Rev. Int. 21, 44–47. https://doi.org/10.1353/cri.2016.0056 Huenteler, J., Schmidt, T.S., Ossenbrink, J., Hoffmann, V.H., 2016. Technology life- cycles in the energy sector - Technological characteristics and the role of deployment for innovation. Technol. Forecast. Soc. Change 104, 102–121. https://doi.org/10.1016/j.techfore.2015.09.022 IEA, 2018. Trends in Photovoltaic Applications 2018. IPCC, 2018. Global warming of 1.5°C. An IPCC Special Report on the impacts of global warming of 1.5°C above pre-industrial levels and related global greenhouse gas emission pathways, in the context of strengthening the global response to the threat of climate change,. IRENA, 2014. REmap 2030 A Renewable Energy Roadmap. Jacobs, J., 1969. The Economy of Cities, Vintage. Jaffe, A.B.., Trajtenberg, M., Henderson, R., 1993. Geographic Localization of Knowledge Spillovers as Evidenced by Patent Citations. Q. J. Econ. 108, 577–598. Jaffe, A.B., Newell, R.G., Stavins, R.N., 2004. Technology Policy for Energy and the Environment. Innov. Policy Econ. 4, 35–68. https://doi.org/10.1086/ipe.4.25056161 Jaffe, A.B., Newell, R.G., Stavins, R.N., 2002. A Tale of Two Market Failures: Technology and Environmental Policy. Forum Am. Bar Assoc. 72–75. Jaffe, A.B., Palmer, K., 2009. Environmental regulation and innovation: a panel data study. Rev. Econ. Stat. 79, 610–619. Jaffe, A.B., Trajtenberg, M., 2002. Patents, Citations, and Innovations: A Window on the Knowledge Economy. MIT Press. Jaffe, A.B., Trajtenberg, M., 1999. International Knowledge Flows: Evidence From Patent Citations. Econ. Innov. New Technol. https://doi.org/10.1080/10438599900000006 Jaffe, A.B., Trajtenberg, M., 1996. Flows of knowledge from universities and federal laboratories: modeling the flow of patent citations over time and across institutional 184 and geographic boundaries. Proc. Natl. Acad. Sci. U. S. A. 93, 12671–12677. https://doi.org/10.1073/pnas.93.23.12671 Joe Tidd, Bessant, J., keith pavitt, 2005. Managing Innovation: Integrating Technological, Market and Organizational Change, third. ed. John Wiley & Sons Inc. https://doi.org/10.1016/S0166-4972(98)80033-3 Johnson, D., Popp, D., 2001. Forced out of the closet: The impact of the American inventors protection act on the timing of patent disclosure. RAND J. Econ. 34, 96– 112. https://doi.org/10.2307/3087445 Junginger, M., Faaij, A., Turkenburg, W.C., 2005. Global experience curves for wind farms. Energy Policy 33, 133–150. https://doi.org/10.1016/S0301-4215(03)00205-2 Kahouli-Brahmi, S., 2008. Technological learning in energy-environment-economy modelling: A survey. Energy Policy 36, 138–162. https://doi.org/10.1016/j.enpol.2007.09.001 Kaplan, S., Keyvan, V., 2013. Novelty vs. usefulness in innovative breakthroughs: A test using topic modeling of nanotechnology patents. Chem. Inf. Model. 53, 1689–1699. https://doi.org/10.1017/CBO9781107415324.004 Keijl, S., Gilsing, V.A., Knoben, J., Duysters, G., 2016. The two faces of inventions: The relationship between recombination and impact in pharmaceutical biotechnology. Res. Policy 45, 1061–1074. https://doi.org/10.1016/j.respol.2016.02.008 Kelly, M., Hageman, A., 1999. Marshallian Externalities in Innovation. J. Econ. Growth 4, 39–54. https://doi.org/10.1023/A:1009874508579 Kemp, R., Pontoglio, S., 2011. The innovation effects of environmental policy instruments — A typical case of the blind men and the elephant? Ecol. Econ. 72, 28–36. https://doi.org/10.1016/j.ecolecon.2011.09.014 Kim, L., Nelson, R.R., 2000. Technology, learning, and innovation: Experiences of Newly Industrializing Economies, Cambridge University Press. Cambridge. Klaassen, G., Miketa, A., Larsen, K., Sundqvist, T., 2005. The impact of R&D on innovation for wind energy in Denmark, Germany and the United Kingdom. Ecol. Econ. 54, 227–240. https://doi.org/10.1016/j.ecolecon.2005.01.008 Klein Woolthuis, R., Lankhuizen, M., Gilsing, V., 2005. A system failure framework for innovation policy design. Technovation 25, 609–619. https://doi.org/10.1016/j.technovation.2003.11.002 Kline, S.J., Rosenberg, N., 1986. An Overview of Innovation. Eur. J. Innov. Manag. 38, 275–305. https://doi.org/10.1108/14601069810368485 Kogut, B., 1995. Country Competitiveness: Technology and the Organizing of Work. Adm. Sci. Q. 40, 181. https://doi.org/10.2307/2393704 Krause, D.R., Handfield, R.B., Tyler, B.B., 2007. The relationships between supplier development, commitment, social capital accumulation and performance improvement. J. Oper. Manag. 25, 528–545. https://doi.org/10.1016/j.jom.2006.05.007 Lampe, R., 2012. Strategic Citation. Rev. Econ. Stat. 94, 320–333. https://doi.org/10.1162/REST_a_00159 Lanjouw, J.O., Mody, A., 1996. Innovation and the international diffusion of environmentally responsive technology. Res. Policy 25, 549–571. https://doi.org/10.1016/0048-7333(95)00853-5 185 Laursen, K., Masciarelli, F., Prencipe, A., 2012. Regions matters: how localized social capital affects innovation and exhernal knowledge acquisition. Organ. Sci. 23, 177– 193. https://doi.org/10.2307/41429024 Lawson, B., Tyler, B.B., Cousins, P.D., 2008. Antecedents and consequences of social capital on buyer performance improvement. J. Oper. Manag. 26, 446–460. https://doi.org/10.1016/j.jom.2007.10.001 Levy, F., Murnane, R.J., 2004. The New Division of Labor: How Computers Are Creating the Next Job Market. Princeton University Press, Princeton. Lewis, J.I., Wiser, R.H., 2007. Fostering a renewable energy technology industry : An international comparison of wind industry policy support mechanisms. Energy Policy 35, 1844–1857. https://doi.org/10.1016/j.enpol.2006.06.005 Li, Jianping, Li, M., Gao, Y., Li, Jianjian, Su, S., Huang, M., 2015. Blue book of China’s provincial competitiveness. Social Science Academic Press. Li, X., 2011. Behind the recent surge of Chinese patenting: An institutional view. Res. Policy 41, 236–249. https://doi.org/10.1016/j.respol.2011.07.003 Liker, J.K., Choi, T.Y., 2004. Building deep supplier relationships. Harv. Bus. Rev. https://doi.org/Article Lin, J.Y., 2010. Education and Innovation Adoption in Agriculture: Evidence from Hybrid Rice in China. Agric. Econ. 73, 713–723. https://doi.org/10.2307/1242823 Lorenzen, M., 2007. Social Capital and Localised Learning: Proximity and Place in Technological and Institutional Dynamics. Urban Stud. 44, 799–817. https://doi.org/10.1080/00420980601184752 Lundvall, B.-Å., 2012. National Systems of Innovation: Toward a theory of innovation and interactive learning, The Anthem Other Canon series. https://doi.org/10.7135/UPO9781843318903 Lundvall, B.-Å., 2005. National Innovation Systems: Analytical Concept and Development Tool, in: DRUID-Conference in Copenhagen. pp. 301–312. https://doi.org/10.1016/j.compstruct.2018.07.072 Lundvall, B.-Å., 1992. National systems of innovationtowards a theory of innovation and interactive learning. Towards a theory of innovation and interactive learning. London, Pinter. Lundvall, B.-Å., 1988. Innovation as an Interactive Process: From User– Producer Interaction to the National Systems of Innovation, in: Technical Change and Economic Theory. https://doi.org/http://dx.doi.org/10.1016/B978-0-7506-7009- 8.50019-7 Lundvall, B.-Å., Johnson, B., Andersen, E.S., Dalum, B., 2002. National systems of production, innovation, and competence-building, in: Research Policy. pp. 213–231. https://doi.org/10.1017/CBO9780511493386.010 Maguire, K., Marsan, G., Nauwelaers, C., Halkier, H., 2012. OECD Reviews of Regional Innovation. Malerba, F., 1992. Learning by Firms and Incremental Technical Change. Econ. J. 102, 845. https://doi.org/10.2307/2234581 Marshall, A., 1895. Principles of Economics. The English Langugage Book Society, London. Martin, R., Sunley, P., 2003. Deconstructing clusters: chaotic concept or policy panacea? 186 J. Econ. Geogr. 3, 5–35. https://doi.org/10.1093/jeg/3.1.5 Maskell, P., 2001. The Concept of the Firm in Economic Geography. Econ. Geogr. 77, 329–344. https://doi.org/10.2307/3594103 Maskell, P., 1999. Localised learning and industrial competitiveness. Cambridge J. Econ. 23, 167–185. https://doi.org/10.1093/cje/23.2.167 Ministry of Finance, 2009. Notice on the implementation of Golden Sun demonstration project (in Chinese). Ministry of Finance. Ministry of Housing and Urban-Rural Construction, 2013. Green Building Action Plan (in Chinese). Ministry of Housing and Urban-Rural Construction. Modi, S.B., Mabert, V.A., 2007. Supplier development: Improving supplier performance through knowledge transfer. J. Oper. Manag. 25, 42–64. https://doi.org/10.1016/j.jom.2006.02.001 Mody, A., Lanjouw, J.O., 1996. Innovation and the international diffusion of environmentally responsive technology. Res. Policy 25, 549–571. Mogoutov, A., Kahane, B., 2007. Data search strategy for science and technology emergence: A scalable and evolutionary query for nanotechnology tracking. Res. Policy 36, 893–903. https://doi.org/10.1016/j.respol.2007.02.005 Molas-Gallart, J., Davies, A., 2006. Toward theory-led evaluation: The experience of European science, technology, and innovation policies. Am. J. Eval. 27, 64–82. https://doi.org/10.1177/1098214005281701 Morescalchi, A., Pammolli, F., Penner, O., Petersen, A.M., Riccaboni, M., 2015. The evolution of networks of innovators within and across borders: Evidence from patent data. Res. Policy 44, 651–658. https://doi.org/10.1016/j.respol.2014.10.015 Morrris, J., Calhoun, K., Goodman, J., Seif, D., 2014. Reducing Solar PV Soft Costs: A focus on installation labor. Photovolt. Spec. Conf. (PVSC), 2014 IEEE 40th. IEEE. Motohashi, K., 2005. University-industry collaborations in Japan: The role of new technology-based firms in transforming the National Innovation System. Res. Policy 34, 583–594. https://doi.org/10.1016/j.respol.2005.03.001 Mowery, D., Rosenberg, N., 1979. The influence of market demand upon innovation: a critical review of some recent empirical studies. Res. Policy 8, 102–153. https://doi.org/10.1016/0048-7333(79)90019-2 Mowery, D., Sampat, B., 2006. Universities in National Innovation Systems. Oxford Handb. Innov. 209–239. https://doi.org/10.1093/oxfordhb/9780199286805.003.0008 Mowery, D.C., Rosenberg, N., 1993. The U.S. National Innovation System, in: National Innovation Systems: A Compartive Analysis. pp. 29–75. Mowery, D.C., Rosenberg, N., Nelson, R.R., Teubal, M., Walker, W., 1993. National Innovation Systems: A Comparative Analysis, in: National Innovation Systems: A Comparative Analysis. pp. 158–191. Nascia, L., Perani, G., 2002. Diversity of innovation in Europe. Int. Rev. Appl. Econ. 16, 277–293. https://doi.org/10.1080/02692170210136118 National Development Reform Commission, 2013. Notice About Playing the Role of the Price Lever to Promote the Healthy Development of Photovoltaic Industry (in Chinese). National Development Reform Commission. National Grid Company, 2012. Notice on Providing Grid Connection Service for DG PV Projects (Provisional) (in Chinese). National Grid Company. 187 National Research Council, 2001. Energy Research at DOE: Was it Worth It?, National Academy Press. https://doi.org/10.17226/10165 Negro, S.O., Alkemade, F., Hekkert, M.P., 2012. Why does renewable energy diffuse so slowly? A review of innovation system problems. Renew. Sustain. Energy Rev. 16, 3836–3846. https://doi.org/10.1016/j.rser.2012.03.043 Neij, L., Heiskanen, E., Strupeit, L., 2017. The deployment of new energy technologies and the need for local learning. Energy Policy 101, 274–283. https://doi.org/10.1016/j.enpol.2016.11.029 Nelson, R., 1959. The Simple Economics of Basic Scientific Research. J. Polit. Econ. https://doi.org/10.1086/258177 Nemet, G.F., Shaughnessy, E., Wiser, R.H., Darghouth, N., Barbose, G.L., Gillingham, K., Rai, V., 2016. Characteristics of Low-Priced Solar Photovoltaic Systems in the United States. Nemet, G.F., 2012a. Inter-technology knowledge spillovers for energy technologies. Energy Econ. 34, 1259–1270. https://doi.org/10.1016/j.eneco.2012.06.002 Nemet, G.F., 2012b. Subsidies for New Technologies and Knowledge Spillovers from Learning by Doing : Subsidies and Learning by Doing. J. Policy Anal. Manag. 31, 601–622. https://doi.org/10.1002/pam.21643 Nemet, G.F., 2011. Knowledge spillovers from learning by doing in wind power. Nemet, G.F., 2009. non-incremental technical change. Res. Policy 38, 700–709. https://doi.org/10.1016/j.respol.2009.01.004 Nemet, G.F., 2009. Demand-pull, technology-push, and government-led incentives for non-incremental technical change. Res. Policy 38, 700–709. https://doi.org/10.1016/j.respol.2009.01.004 Nemet, G.F., 2006. Beyond the learning curve: factors influencing cost reductions in photovoltaics. Energy Policy 34, 3218–3232. https://doi.org/10.1016/j.enpol.2005.06.020 Nemet, G.F., O’Shaughnessy, E., Wiser, R., Darghouth, N.R., Barbose, G., Gillingham, K., Rai, V., 2017. What factors affect the prices of low-priced U.S. solar PV systems? Renew. Energy 114, 1333–1339. https://doi.org/10.1016/j.renene.2017.08.018 Newell, R.G., Jaffe, A.B., Stavins, R.N., 1999. The Induced Innovation Hypothesis and Energy-Saving Technological Change. Q. J. Econ. 114, 941–975. https://doi.org/10.1162/003355399556188 Noailly, J., Shestalova, V., 2017. Knowledge spillovers from renewable energy technologies: Lessons from patent citations. Environ. Innov. Soc. Transitions 22, 1– 14. https://doi.org/10.1016/j.eist.2016.07.004 OECD, 2011. Demand-side Innovation Policies. OECD, 2002. Dynamizing National Innovation Systems. Oliveira, M., 1999. Core competencies and the knowledge of the firm. Dyn. Strateg. Resour. 17–41. Olson Lanjouw, J., 1998. Patent Protection in the Shadow of Infringement: Simulation Estimations of Patent Value. Rev. Econ. Stud. 65, 671–710. https://doi.org/10.1111/1467-937X.00064 Owen-Smith, J., Powell, W.W., 2004. Knowledge Networks as Channels and Conduits: 188 The Effects of Spillovers in the Boston Biotechnology Community. Organ. Sci. 15, 5–21. https://doi.org/10.1287/orsc.1030.0054 Parry, I.W.H., Pizer, W.A., Fischer, C., 2003. How Large are the Welfare Gains from Technological Innovation Induced by Environmental Policies? J. Regul. Econ. 23, 237–255. https://doi.org/10.1023/A:1023321309988 Patel, P., Pavitt, K., 1987. Is Western Europe loosing the technological race? Res. Policy 16, 59–85. Peters, M., Schneider, M., Griesshaber, T., Hoffmann, V.H., 2012. The impact of technology-push and demand-pull policies on technical change - Does the locus of policies matter? Res. Policy 41, 1296–1308. https://doi.org/10.1016/j.respol.2012.02.004 Pigou, A., 1920. The Economics of Welfare Policies. London: Macmillan. https://doi.org/10.2307/2520637 Pisano, G.P., Teece, D.J., 2007. How to Capture Value from Innovation: Shaping Intellectual Property and Industry Architecture. Calif. Manage. Rev. 50, 278–296. https://doi.org/10.2307/41166428 Popp, D., 2019. Environmental policy and innovation: a decade of research, NBER Working paper No. 25631. Popp, D., 2006a. They Don’t Invent Them Like They Used To: An Examination of Energy Patent Citations over Time. Econ. Innov. New Technol. 15, 753–776. https://doi.org/10.1080/10438590500510459 Popp, D., 2006b. International innovation and diffusion of air pollution control technologies: The effects of NOX and SO2 regulation in the US, Japan, and Germany. J. Environ. Econ. Manage. 51, 46–71. https://doi.org/10.1016/j.jeem.2005.04.006 Popp, D., Hafner, T., Johnstone, N., 2011. Environmental policy vs. public pressure: Innovation and diffusion of alternative bleaching technologies in the pulp industry. Res. Policy 40, 1253–1268. https://doi.org/10.1016/j.respol.2011.05.018 Popp, D., Newell, R.G., Jaffe, A.B., 2010. Energy, the environment, and technological change. Handb. Econ. Innov. Popp, D., Santen, N., Fisher-Vanden, K., Webster, M., 2013. Technology variation vs. R&D uncertainty: What matters most for energy patent success? Resour. Energy Econ. 35, 505–533. https://doi.org/10.1016/j.reseneeco.2013.05.002 Porter, M.E., Stern, S., 2001. Innovation: Location matters. MIT Sloan Manag. Rev. 42, 28–36. https://doi.org/10.1016/j.neuron.2009.08.001 Prajogo, D.I., Ahmed, P.K., 2006. Relationships between innovation stimulus, innovation capacity, and innovation performance. R D Manag. 36, 499–515. https://doi.org/10.1111/j.1467-9310.2006.00450.x Putnam, J.D., 1997. The Value of International Patent Rights. J. Int. Bus. Stud. 28, 437– 437. Qiu, S., Liu, X., Gao, T., 2017. Do emerging countries prefer local knowledge or distant knowledge? Spillover effect of university collaborations on local firms. Res. Policy 46, 1299–1311. https://doi.org/10.1016/j.respol.2017.06.001 Qiu, Y., Anadon, L.D., 2012. The price of wind power in China during its expansion: Technology adoption, learning-by-doing, economies of scale, and manufacturing 189 localization. Energy Econ. 34, 772–785. https://doi.org/10.1016/j.eneco.2011.06.008 Quitzow, R., Huenteler, J., Asmussen, H., 2017. Development trajectories in China’s wind and solar energy industries: How technology-related differences shape the dynamics of industry localization and catching up. J. Clean. Prod. 158, 122–133. https://doi.org/10.1016/j.jclepro.2017.04.130 R. H. Coase, 2013. The Problem of Social Cost. J. Law Econ. 56, 837–877. Rai, V., Metteauer, M., Querejazu, D., Wise, R., Hamilton, G., 2013. Demand-Pull and Innovation in the US Solar Market. https://doi.org/10.2139/ssrn.2297140 Reinganum, J.F., 1989. The timing of innovation: Research, development, and diffusion. Handbook of industrial organization, in: Handbook of Industrial Organization. pp. 849–908. Reitzig, M., 2004. Improving patent valuations for management purposes - Validating new indicators by analyzing application rationales. Res. Policy 33, 939–957. https://doi.org/10.1016/j.respol.2004.02.004 REN 21, 2017. Renewables 2017: global status report. https://doi.org/10.1016/j.rser.2016.09.082 REN21, 2018. Renewables 2018 global status report. Rodrik, D., 2014. Green industrial policy. Oxford Rev. Econ. Policy 30, 469–491. https://doi.org/10.1093/oxrep/gru025 Rosenberg, N., 1982. Inside the black box: technology and economics. Rothwell, R., 1989. Smfs inter firm relationships and technological change. Entrep. Reg. Dev. 1, 275–291. https://doi.org/10.1080/08985628900000024 Rothwell, R., 1983. Innovation and firm size: a case for dynamic complementarity or, is small really so beautiful. J. Gen. Manag. 8, 5–25. Rothwell, R., Zegveld, W., 1988. An Assessment of Government Innovation Policies, in: Government Innovation Policy. Palgrave Macmillan UK, London, pp. 19–35. https://doi.org/10.1007/978-1-349-08882-9_2 Saeed, K.A., Malhotra, M.K., Grover, V., 2005. Examining the impact of interorganizational systems on process efficiency and sourcing leverage in buyer- supplier dyads. Decis. Sci. 36, 365–395. https://doi.org/10.1111/j.1540- 5414.2005.00077.x Saiki, T., Akano, Y., Watanabe, C., Tou, Y., 2006. A new dimension of potential resources in innovation - a wider scope of patent claims can lead to new functionality development. Technovation 26, 796–806. Salter, A., Laursen, K., 2006. Open for innovation: the role of openness in explaining innovation performance among U.K. manufacturing firms. Strateg. Manag. J. 27, 131–150. Schaeffer, G.J., Alsema, E., Seebregts, A., Beurskens, L., de Moor, H., van Sark, W., Durstewitz, M., Perrin, M., Boulanger, P., Laukamp, H., Zuccaro, C., 2004. Learning from the Sun. Analysis of the use of experience curves for energy policy purposes. The case of photovoltaic power. Schankerman, M., Pakes, A., 1986. Estimates of the Value of Patent Rights in European Countries During the Post-1950 Period. Econ. J. 96, 1052. https://doi.org/10.2307/2233173 Scherer, F.M., 1983. The propensity to patent. Int. J. Ind. Organ. 1, 107–128. 190 https://doi.org/10.1016/0167-7187(83)90026-7 Schmidt, T.S., Huenteler, J., 2016. Anticipating industry localization effects of clean technology deployment policies in developing countries. Glob. Environ. Chang. 38, 8–20. https://doi.org/10.1016/j.gloenvcha.2016.02.005 Schmookler, J., 1962. Economic Sources of Inventive Activity. J. Econ. Hist. 22, 1–20. https://doi.org/10.1017/S0022050700102311 Schneider, M., Holzer, A., Hoffmann, V.H., 2008. Understanding the CDM’ s contribution to technology transfer. Energy Policy 36, 2930–2938. https://doi.org/10.1016/j.enpol.2008.04.009 Schumpeter, J.A., 1942. Socialism, capitalism and democracy. Routledge. Schumpeter, J.A., 1934. The Fundamental Phenomenon of Economic Development. Harvard University Press, Cambridge, MA. Seel, J., Barbose, G.L., Wiser, R.H., 2013. Why are Residential PV Prices in Germany So Much Lower Than in the United States. Lbnl 59. Seel, J., Barbose, G.L., Wiser, R.H., 2014. An analysis of residential PV system price differences between the United States and Germany. Energy Policy 69, 216–226. https://doi.org/10.1016/j.enpol.2014.02.022 Sen, F., Egelhoff, W., 2000. Innovative capabilities of a firm and the use of technical alliances. IEEE Trans. Eng. Shum, K.L., Watanabe, C., 2008. Towards a local learning (innovation) model of solar photovoltaic deployment. Energy Policy 36, 508–521. https://doi.org/10.1016/j.enpol.2007.09.015 Skidmore, M., Peltier, J., Alm, J., 2005. Do state motor fuel sales-below-cost laws lower prices? J. Urban Econ. 57, 189–211. https://doi.org/10.1016/j.jue.2004.10.004 Smith, S., M. J. Shiao, 2012. Solar PV Balance of System (BOS) Markets: Technologies, Costs and Leading Companies, 2013-2016. Smits, R., Kuhlmann, S., Shapira, Phillip, Shapira, Philip, 2013. Introduction. A Systemic Perspective: The Innovation Policy Dance, in: The Theory and Practice of Innovation Policy. https://doi.org/10.4337/9781849804424.00006 Söderholm, P., Klaassen, G., 2007. Wind power in Europe: A simultaneous innovation- diffusion model. Environ. Resour. Econ. 36, 163–190. https://doi.org/10.1007/s10640-006-9025-z Söderholm, P. Sundqvist, T., 2007. Empirical challenges in the use of learning curves for assessing the economic prospects of renewable energy technologies. Renew. Energy 32, 2559–2578. https://doi.org/10.1016/j.renene.2006.12.007 Sonn, J.W., Storper, M., 2003. The Increasing Importance of Geographical Proximity in Technological Innovation: An Analysis of U.S. Patent Citations, 1975-1977. What Do we Know about Innov. 1975–1997. State Council, 2013. Several Opinions on Promoting the Healthy Development of Photovoltaic Industry (in Chinese). State Council. Stavins, R.N., 1995. Transaction costs and tradeable permits. J. Environ. Econ. Manage. 29, 133–148. Storper, M., Venables, A.J., 2004. Stroper, M., Venables, A.J., 2004. Buzz: face-to-face contact and the urban economy. J. Econ. Geogr. 4, 351–370. J. Econ. Geogr. 4, 351– 370. https://doi.org/10.1093/jnlecg/lbh027 191 Strupeit, L., Neij, L., 2017. Cost dynamics in the deployment of photovoltaics: Insights from the German market for building-sited systems. Renew. Sustain. Energy Rev. 69, 948–960. https://doi.org/http://dx.doi.org/10.1016/j.rser.2016.11.095 Sun, Y., Du, D., 2010. Determinants of industrial innovation in China: Evidence from its recent economic census. Technovation 30, 540–550. https://doi.org/10.1016/j.technovation.2010.05.003 Sun, Y., Liu, F., 2010. A regional perspective on the structural transformation of China’s national innovation system since 1999. Technol. Forecast. Soc. Chang. 77, 1311– 1321. https://doi.org/10.1016/j.techfore.2010.04.012 Taylor, M., 2008. Beyond technology-push and demand-pull: Lessons from California’s solar policy. Energy Econ. 30, 2829–2854. https://doi.org/10.1016/j.eneco.2008.06.004 Taylor, M., Thornton, D., Nemet, G., Colvin, M., 2006. Government actions and innovation in environmental technology for power production: the cases of selective catalytic reduction and wind power in California. Teece, D.J., 1993. Profiting from technological innovation: Implications for integration, collaboration, licensing and public policy. Res. Policy 22, 112–113. Teece, D.J., Pisano, G., Shuen, A., 1997. Dynamic capabilities and strategic management. Strateg. Manag. J. 18, 509–533. Thompson, P., 2006. Patent Citations and the Geography of Knowledge Spillovers: Evidence from Inventor- and Examiner-added Citations. Rev. Econ. Stat. 88, 383– 388. https://doi.org/10.1162/rest.88.2.383 Tong, X., Frame, J.D., 1994. Measuring national technological performance with patent claims data. Res. Policy 23, 133–141. https://doi.org/10.1016/0048-7333(94)90050- 7 Trajtenberg, M., 1990. A Penny for Your Quotes: Patent Citations and the Value of Innovations. RAND J. Econ. 21, 172. https://doi.org/10.2307/2555502 Upham, P., Rosenkopf, L., Ungar, L., 2009. Innovating knowledge communities: An analysis of group collaboration and competition in science and technology. Scientometrics 82, 525–554. https://doi.org/10.5465/AMBPP.2007.26523086 Urban, F., Wang, Y., Geall, S., 2018. Prospects, Politics, and Practices of Solar Energy Innovation in China. J. Environ. Dev. 27, 74–98. https://doi.org/10.1177/1070496517749877 Utterback, J.M., 1974. Innovation in Industry and the Diffusion of Technology. Science (80-. ). 183, 620–626. https://doi.org/10.1126/science.183.4125.620 Uzzi, B., 1996. The Sources and Consequences of Embeddedness for the Economic Performance of Organizations: The Network Effect. Am. Sociol. Rev. 61, 674. https://doi.org/10.2307/2096399 Venkitachalam, K., Busch, P., 2012. Tacit knowledge: review and possible research directions. J. Knowl. Manag. 16, 357–372. https://doi.org/10.1108/13673271211218915 Venugopalan, S., Rai, V., 2015. Topic based classification and pattern identification in patents. Technol. Forecast. Soc. Change 94, 236–250. https://doi.org/10.1016/j.techfore.2014.10.006 Veugelers, R., Cassiman, B., 1999. Make and buy in innovation strategies: evidence from 192 Belgian manufacturing firms. Res. Policy 28, 63–80. Villena, V.H., Revilla, E., Choi, T.Y., 2011. The dark side of buyer-supplier relationships: A social capital perspective. J. Oper. Manag. https://doi.org/10.1016/j.jom.2010.09.001 von Hippel, E., 1986. Lead Users: A Source of Novel Product Concepts. Manage. Sci. 32, 791–805. https://doi.org/10.1287/mnsc.32.7.791 Wang, Y., Sutherland, D., Ning, L., Pan, X., 2015. The evolving nature of China’s regional innovation systems: Insights from an exploration–exploitation approach. Technol. Forecast. Soc. Chang. 100, 140–152. https://doi.org/10.1016/j.techfore.2015.07.010 Wene, C., 2000. Experience Curves for Energy Technology Policy. Wenger, E., 1998. Communities of practice: Learning, meaning, and identity. Wieczorek, A.J., Hekkert, M.P., 2012. Systemic instruments for systemic innovation problems: A framework for policy makers and innovation scholars. Sci. Public Policy 39, 74–87. https://doi.org/10.1093/scipol/scr008 Winston Smith, S., Shah, S.K., 2013. Do innovative users generate more usefulinsights? An analysis of CVC investment in the medical device industry. Strateg. J. 7, 151– 167. Wiser, R., Bolinger, M., Cappers, P., Margolis, R., 2007. Analyzing historical cost trends in California’s market for customer-sited photovoltaics. Prog. Photovoltaics Res. Appl. 15, 69–85. https://doi.org/10.1002/pip.726 Wolfe, A., 1994. Organizational innovation; Review, critique and suggested research directions. J. Manag. Stud. 31, 405–431. Yang, C.H., Lin, H.L., 2012. Openness, absorptive capacity, and regional innovation in China. Environ. Plan. A 44, 333–355. https://doi.org/10.1068/a44182 Yu, H.J.J., Popiolek, N., Geoffron, P., 2015. A comprehensive approach to assessing PV system economics and opportunities for PV policies, in: International Conference on the European Energy Market, EEM. https://doi.org/10.1109/EEM.2015.7216707 Zhang, F., Deng, H., Margolis, R., Su, J., 2015. Analysis of distributed-generation photovoltaic deployment, installation time and cost, market barriers, and policies in China. Energy Policy 81, 43–55. https://doi.org/10.1016/j.enpol.2015.02.010 Zhang, S., 2016. Analysis of DSPV (distributed solar PV) power policy in China. Energy 98, 92–100. https://doi.org/10.1016/j.energy.2016.01.026 Zhao, X.G., Wan, G., Yang, Y., 2015. The turning point of solar photovoltaic industry in China: Will it come? Renew. Sustain. Energy Rev. https://doi.org/10.1016/j.rser.2014.08.045
193