Unconventional Industrial Policy in the US Residential Solar Industry
Are subsidies to consumers effective industrial policy?
This question is important for several active policy debates — perhaps none more so than efforts to decarbonize the US economy, where supporting the adoption of early-stage, clean technologies through demand subsidies is increasingly popular.
Much of the debate surrounding these interventions highlights the potential for industry growth, a goal more often associated with conventional industrial policies such as entry or production subsidies. Proponents of consumer subsidies on these grounds argue that greater demand increases industry learning-by-doing, which in turn reduces costs for potential entrants.
However, economic theory emphasizes that learning spillovers make experience-based cost reductions a public good, which reduces firms’ incentives to expand output and lower costs. The net effect of consumer subsidies on growth in target industries therefore depends on the extent to which the public good nature of learning reduces incentives to lower costs.
To answer this question in the real world, we need to know something about firms’ learning curves and how firms make entry, exit, and output decisions.
My Focus: California Solar Installers
I examine this tradeoff in a policy-relevant setting: consumer subsidies for solar photovoltaics (PV). I study the impact of this form of policy on solar installation firms, which account for a growing share of the final cost to consumers in this industry.
To assess this industry empirically, I focus on the market for residential solar PV in California, which is home to half of all US residential PV systems and has extensive experience subsidizing household solar adoption over the past few decades. One of the largest subsidy programs in the state was the California Solar Initiative (CSI), which ran from 2007 to 2013 and provided $2.2 billion in direct cash rebates to households that adopted solar PV.
Understanding the impact of California’s solar PV subsidies on market size and structure requires knowledge of how firms in this industry learn. The arguments in favor of consumer subsidies as a tool to spur industry growth assume that firms become more efficient as they gain experience.
Do solar installers experience learning-by-doing? What do installers’ learning curves look like?
Solar Installers’ Costs Decline with Experience
Average Costs Dropped 65%, over 2008-2013
Using data on over 50,000 residential solar systems in California for 2008-2013, I develop and estimate a dynamic structural model of the market for solar installations. The model allows me to not only recover solar installers’ learning curves — including the extent to which knowledge spills over across firms — but also simulate firm behavior under counterfactual policies.
I find that solar installers’ costs decline dramatically with the quantity of cumulative production — i.e., “experience.” On average, installers’ marginal costs per unit of installed PV capacity dropped $1.70 over 2008-2013 in California, a 68% reduction.
These estimates of installers’ costs match both the magnitude and time-path of publicly available estimates produced by the National Renewable Energy Laboratory (NREL). As I show above, estimating a static model of installers’ decisions gives vastly different estimates from NREL, which emphasizes the importance of a number of key features of my model.
So solar installers do experience learning-by-doing. But how much of this is firms learning on their own versus benefitting from the knowledge of others?
Observed Cost Reductions Are Mainly Driven by Spillovers
Though the Marginal Effect of Own Experience is 25% Larger
Though I find that on the margin firms’ own experience contributes more to their learning, I estimate that the vast majority of observed cost reductions come from knowledge spillovers across firms.
I test several models of firm spillovers, including constant spillovers, spillovers that differ based on whether rival firms are in the same geographic market (i.e., county), and spillovers that differ based on whether rival firms install the same type of panel. Across all models, cost reductions from rivals’ learning far outweigh those from a firm’s own learning.
Solar installers learn with experience and this knowledge spills over across firms. What role do these large knowledge spillovers play in determining the impact of consumer subsidies in this industry?
CA Consumer Subsidies Increased Market Size
Increased Number of Firms by 9%, Installations by 4% over 2008-2013
I use my model to simulate the industry with and without the California Solar Initiative (CSI)’s consumer subsidies. I find that the number of active firms and the number of solar installations decrease by 9% and 4% without the CSI, respectively. Click “Simulate Industry” below to see what this looks like.
The fact that the CSI’s subsidies result in both demand- and supply-side growth suggests that any disincentive firms face to take advantage of learning-by-doing is outweighed by its benefits. Despite large knowledge spillovers across firms, industry cost reductions due to the additional experience from consumer subsidies outweigh the firms’ strategic incentives to influence the costs of their rivals by producing less.
If policymakers approach to decarbonizing the economy involves targeting growth in specific clean technologies—perhaps as a result of political constraints on first-best policy tools like a carbon tax—consumer subsidies may be an effective means of doing so.
So consumer subsidies can be effective industrial policy. But how do they compare to more conventional forms of industrial policy?
Entry Subsidies Outperform Consumer Subsidies in CA
Lead to Increase in Active Firms and Installations, Lower Costs
I use my model to implement a set of counterfactual policies in which I remove the California Solar Initiative (CSI)’s consumer subsidies and replace them with entry subsidies of varying sizes. These entry subsidies are similar in total magnitude to the CSI, ranging $2.1–8.0 billion in cost.
The entry subsidies greatly increase not only the number of active incumbents, but also the number of solar installations, resulting in an increase in aggregate welfare.
Key Takeaways
These findings suggest that consumer subsidies can be an appealing second-best tool for policymakers to achieve decarbonization and industrial policy goals in specific technologies. However, my results indicate that other approaches such as entry subsidies may be more effective on these outcomes.
While the empirical results of this paper are technology- and context-specific, they are likely informative for similar policies in other areas. For example, heat pumps are a closely related clean technology with a similar set of intermediary installers.
I also compare the CSI’s consumer subsidies to a number of alternative consumer rebate designs and climate policies, including a carbon tax. I leave those results for the paper, which you can download here.
Methodology
I develop a dynamic structural model of the market for residential solar PV installations in California based on the theoretical framework for dynamic oligopoly of Ericson and Pakes (1995). The model endogenizes consumer demand for solar PV installations as well as installation firms' entry, exit and quantity-setting decisions. Consumer demand for differentiated solar PV installations is static and follows the random coefficient nested logit model of Brenkers and Verboven (2006). Incumbent installers' installation-specific costs are a function of their own cumulative production as well as the cumulative production of their rivals to allow for learning-by-doing and learning spillovers. In each geographic market, incumbent firms dynamically choose a quantity of installations to provide conditional on their marginal costs and their beliefs about future learning. I model firms' product market decisions as dynamic to capture the incentives to select a production level today based on its impact on own and rival costs in the future. Incumbent firms then choose whether to exit by comparing their expected discounted future profits with an idiosyncratic scrap value while a market-specific pool of potential entrants make one-shot entry decisions based on their expected discounted future profits and an idiosyncratic cost of entry. Firms' strategies lead to a Markov Perfect Equilibrium, which I assume is well-approximated by a Moment-based Markov Equilibrium concept Ifrach and Weintraub (2017).
I estimate the model using detailed, system-level data on prices, rebates, installed capacities, and hardware costs for a substantial share of California residential PV systems from 2008 to 2013. I acquire the bulk of the system-level data from the Lawrence Berkeley National Lab's "Tracking the Sun" database, which includes detailed data on the near universe of PV systems installed from 2000 to 2021 in the US Barbose et al. (2022). The "Tracking the Sun" database includes detailed data on over 2.5 million PV systems installed from 2000 to 2021 which are collected annually from state agencies and utilities that administer PV incentive programs, registration systems for renewable energy credits, or grid interconnection processes. I combine these data with information on hardware costs associated with specific residential installations in California, which I acquire from the California Public Utilities Commission. These data, which are broadly available for California PV systems installed between 2008 and 2013, are important as they allow me to isolate the component of installers' marginal costs in which learning-by-doing could occur, which I otherwise do not observe. Given the importance of these hardware cost data to isolating installer learning, I subset the combined PV system-level data to those California systems installed between 2008 and 2013 before aggregating the data to the market-time-level for all observed installers.
My approach to estimation builds on the family of two-step estimators of dynamic games and their various applications. The main primitives of the model include: the demand system for residential solar PV installations, firms' marginal installation cost function, and the distributions of scrap values and entry costs. In the first stage, I recover the static demand parameters and flexibly estimate the exit policy function and transition process of state variables from the data. Firms' value functions, which approximate expected discounted future profits, are nonlinear in the target production cost parameters based on necessary functional form assumptions. I therefore use the first stage estimates to obtain a flexible approximation of the value function following recent work in the dynamic games literature. In the second stage, I form moments from the model's optimal quantity-setting and exit conditions to first recover the parameters governing production costs (i.e., learning) and exit. I then use these estimates to formulate the likelihood of observed entry decisions to recover the full set of dynamic parameters of interest. Having recovered the main parameters of the model, I evaluate the implications of various counterfactual policy environments, simulating California's solar PV installation industry over time and turning on and off different policies.
I solve the model under counterfactuals using a two-step approach similar to parametric policy iteration. Solving the model involves two steps: first, solving for the new Bellman equation, policy functions, and product market equilibrium in a given period and second simulating the industry forward one period. In each of the counterfactual scenarios that I describe below, I initiate this two-step procedure at the observed data in the first period of my main estimation sample---the first quarter of 2008---and then repeat the two-step procedure until I reach the end of the main estimation sample---the last quarter of 2013. This produces a single industry path over the full counterfactual period. I repeat this process multiple times, averaging market outcomes across a number of distinct, forward-simulated industry paths. In practice, I repeat the process of simulating the model forward over the 6 years in the estimation sample 60 distinct times for each counterfactual scenario and then average key outcomes across all 60 model runs.