PaperlayerView exact revision on arXiv

Source-linked AI summary

Accelerating the Adoption of Residential Solar Power Systems: Policy Analysis using a Dynamic Structural Model

Sebastián Souyris, Jason A. Duan, Anantaram Balakrishnan, Varun Rai

arXiv:2608.23796v1econ.EMcs.LGstat.AP

TL;DR

Residential PV policy requires models that account for household adoption drivers while respecting limited budgets. This paper estimates a dynamic structural model with forward-looking households, economic returns, and neighborhood influence, then simulates rebate counterfactuals. It finds that indefinitely extending rebates is not optimal for diffusion or budgets because households postpone adoption.

  • Problem

    Limited budgets require principled policies that evaluate how incentives affect residential PV adoption and carbon savings.

  • Method

    The paper estimates a dynamic structural model of household PV adoption using forward-looking decisions, economic returns, neighborhood effects, and household heterogeneity.

  • Results

    The model reduces out-of-sample adoption prediction error relative to benchmark models, while simulations show indefinite rebate extension is not optimal for diffusion or budgets.

  • Takeaways & Limitations

    Forward-looking households postpone adoption under indefinite rebates, making limited-duration incentive design relevant for accelerating diffusion within budget constraints.

  • Takeaways & Limitations

    The authors identify additional household information and validation of the PV model as directions for improving unobserved heterogeneity and model fit.

Abstract

from arXiv · show

Problem definition: Solar electricity generation is a strategic component of energy portfolios designed to meet growing demand and reduce carbon emissions. Governments and municipalities encourage household photovoltaic (PV) adoption through upfront rebates and tax credits. Limited budgets require principled, data-driven policies that account for the drivers of adoption and the effects of incentives on adoption rates. Methodology/results: We develop a dynamic structural model of residential PV diffusion based on adoption decisions by forward-looking households that weigh the economic trade-offs between installing now and later. Adoption depends on return on investment and influence from neighboring adopters. The model segments households by home value and urbanization level, incorporates unobserved heterogeneity, and captures spatiotemporal installation dynamics. We estimate the model using Bayesian methods and detailed household-level data from Austin, Texas. In out-of-sample tests, it predicts installations more accurately than contemporary alternatives. We simulate counterfactual policies within the dynamic equilibrium of PV diffusion to evaluate rebate designs. The framework can also be adapted to study the adoption of other durable technologies. Managerial implications: A rebate offered for a limited period generates more adoption and emissions reductions than a prolonged, costlier program. This counterintuitive result arises from forward-looking behavior, neighbor influence, and accelerated adoption before the rebate expires. We also evaluate phased reductions and rebates differentiated by household segment. A two-step reduction outperforms multiple small reductions. Geographic differentiation improves policy performance, whereas differentiation by home value offers little advantage over a uniform rebate.

1. Introduction

The paper develops a dynamic structural framework to evaluate residential PV incentives under budget constraints, modeling household economics, forward-looking timing, and neighborhood influence. Simulations show that limited-duration rebates can outperform longer programs, while policy design affects adoption, emissions savings, and spending.

  • Motivation: Limited budgets motivate data-driven incentive policies that maximize residential PV adoption and carbon savings.Upfront rebates and tax credits are important policy levers for promoting household PV adoption.
  • Modeling challenge: Household adoption reflects the long-term economic benefits of installation and the evolving influence of neighboring adopters.The economic component depends on property-specific investment costs, rebates, and future electricity savings or credits.
  • Method: The model uses forward-looking household decisions, comparing adoption now with deferral to a future period.It combines a time-varying net-present-value component with a neighborhood effect based on nearby prior installations.
  • Method: A dynamic structural model estimated from detailed Austin household data supports counterfactual comparisons of incentive amounts and durations.Treating households as the unit of analysis also accommodates household-level heterogeneity and neighborhood effects.
  • Results: The model with forward-looking households and neighborhood effects outperforms benchmark models in out-of-sample adoption prediction.The estimated results also indicate that peer influence within a one-mile radius is especially salient and wealthier households adopt earlier.
  • Policy simulations: Adoption accelerates near rebate expiration, so a longer and costlier rebate does not necessarily increase adoption or carbon savings.Higher initial rebates increase adoption and carbon savings but require greater budget spending.
  • Broader applications: The framework can extend to durable technologies with substantial upfront investment, policy-dependent incentives, peer influence, and forward-looking timing.The paper specifically discusses extensions to electric vehicles and heat pumps.

2. Literature Review

Prior PV-diffusion research spans reduced-form, structural, experimental, agent-based, and analytical approaches. This paper positions its household-level dynamic structural model as combining forward-looking decisions, unobserved heterogeneity, and local neighborhood effects for policy counterfactuals.

  • Empirical approaches: Reduced-form models describe observed adoption relationships but are not amenable to counterfactual analysis outside observed driver ranges.This limits their use for evaluating incentive policies beyond the data variation used for estimation.
  • Research landscape: PV-diffusion research examines household and regional adoption using empirical, agent-based, analytical, and structural approaches.The literature addresses incentives, prices, peer influence, tariffs, market structure, and social welfare.
  • Structural approaches: Existing dynamic structural PV models incorporate forward-looking consumers or intertemporal preferences but omit peer effects or use limited heterogeneity.De Groote and Verboven omit peer effects and model heterogeneity regionally, while related work studies intertemporal preferences and subsidies over time.
  • Peer effects: Related studies investigate peer influence through experiments, surveys, aggregate forecasting, and single-period econometric models.Findings distinguish active and passive peer influence across diffusion settings and examine word-of-mouth and social diffusion.
  • Positioning of this paper: The paper extends prior work with disaggregated household-level adoption data and local neighborhood effects embedded in dynamic decision-making.It estimates adoption utility and uses the estimated parameters to simulate policy counterfactuals.
  • Policy models: Analytical policy studies evaluate subsidy recipients, suppliers, tariffs, utility incentives, generation subsidies, and third-party ownership.These models focus on specific strategic or welfare questions rather than the paper’s combined household-level diffusion framework.

3. Residential PV Systems Data

The study combines Austin Energy and property records to track residential PV adoption, costs, incentives, and household segmentation from 2004 to 2013.

  • The dataset covers all adopters and non-adopters in Austin Energy’s service area, containing about 170,000 single-family residential households.
  • 2,724 households adopted and installed PV systems from 2004 to 2013, while the housing stock grew approximately 2% annually.
  • 42,384 households were identified as potential adopters using positive perceived behavioral control based on property and site characteristics.
  • Household records include property characteristics, location, market value, PV eligibility, installation timing, system size, costs, and rebates.
  • The average lead time from rebate application to installation was roughly one quarter, or 139 days.
  • Households are classified by home value and urbanization, with most homes below $300,000 and suburban homes nearly twice as numerous as urban homes.
  • High-value homeowners adopted earlier, with 21.4% adopting on average compared with lower economic segments.

4. Dynamic Discrete Choice Model

The paper develops a dynamic discrete-choice model in which households weigh PV economics, timing, and neighborhood influence when deciding whether to adopt.

  • The model represents households as forward-looking decision makers balancing adoption now against deferral under changing costs and incentives.
  • Households are segmented economically by home value and geographically by population density, allowing preferences and neighborhood effects to vary across segments.
  • A household’s PV system is installed in the period after adoption and then generates electricity over its lifespan, typically 20 years.
  • The model’s net installation cost is c_t = (p_t − r_t)(1 − ITC_t), combining PV cost, the local rebate, and the federal investment tax credit.
  • Neighborhood influence is measured using installed PV systems in concentric radial distance rings and modeled as a function of installation counts.
  • The one-period adoption-to-installation lag helps resolve the reflection problem, while the linear neighborhood specification provides the best out-of-sample performance.
  • Adoption utility includes a baseline, discounted net present value, neighborhood effects, household heterogeneity, and idiosyncratic shocks.

5. Estimation Results

The estimation results compare alternative dynamic structural specifications and identify a parsimonious model with neighborhood effects, two economic segments, and household heterogeneity. The selected model performs best in out-of-sample prediction and shows that neighborhood influence and forward-looking household behavior are important for forecasting adoption.

  • Model specifications: The model comparison varied economic segmentation, neighborhood radial rings, and whether unobserved household heterogeneity was included.Specifications included alternative market-value segmentations, radial-ring definitions, and random-effects structures.
  • Estimation and validation: 33 quarters of data were used to estimate the structural parameters, followed by a five-quarter out-of-sample test.The training period spans 2004 Q1–2012 Q1, and the test period spans 2012 Q2–2013 Q2.
  • Model selection: 31.85 to 85.51 adopters: MAE across the 12 specifications, with Model (4)H achieving the lowest error.Model (4)H combines one radial ring, two economic segments, neighborhood effects, and unobserved household heterogeneity.
  • Model selection: 31.85–49.14 adopters versus 67.73–85.51: models with neighborhood effects had lower MAE than models without them.The results provide evidence that nearby PV installations influence household adoption decisions.
  • Model selection: Less than 2 adopters: the out-of-sample error difference between one-mile and half-mile radial rings.The one-mile ring was selected as a parsimonious specification, and estimates remained stable across alternative ring choices.
  • Parameter estimates: The estimates were stable under expanding-window re-estimation, and the parameter distributions showed consistent convergence.Household adoption propensity reflected baseline utility, home-value differences, geographic variation in neighborhood effects, and unobserved heterogeneity.
  • Comparison with benchmark models: 31.85 adopters: Model (4)H had the lowest MAE and outperformed all four benchmark models in out-of-sample prediction.The benchmarks do not jointly capture the forward-looking structure, individual household drivers, and neighborhood effects represented in the selected model.

6. Counterfactual Policy Analysis

The counterfactual analysis evaluates rebate amount, duration, phased reductions, and segment differentiation under a limited budget. Results favor time-limited, intermediate designs that accelerate adoption and carbon savings, while geographic differentiation performs better than differentiation by home value.

  • One-Step Rebate Switching Policies: Households accelerate adoption as a pre-announced rebate deadline approaches, then adoption drops after expiration, creating a kink in the diffusion curve.Lower initial net costs and neighbor influence drive diffusion, while deadline anticipation raises the adoption rate near expiration.
  • One-Step Rebate Switching Policies: Indefinite rebates are dominated because households wait longer, weakening neighborhood effects and reducing adoption over time.Although RSP(∞) offers the lowest net cost over time, its rebates never expire and waiting becomes more attractive.
  • One-Step Rebate Switching Policies: RSP(36) maximizes carbon savings, while RSP(48) achieves the highest installed capacity with a lower rebate budget than RSP(52).Carbon savings depend on cumulative generation over time, so earlier adoption can outweigh slightly greater final capacity.
  • One-Step Rebate Switching Policies: Intermediate switch periods improve budget efficiency: delaying the switch beyond the interior optimum raises spending while lowering cumulative carbon savings.The policy implications persist across evaluation horizons, with installed capacity showing diminishing returns to budget.
  • Rebate Amount and Duration: For a $35 million budget, moderate initial rebates with an RSP around 36 maximize carbon savings, while an RSP between 40 and 44 is most effective for an initial net cost of $1.34.The preferred duration and rebate depend on the policy objective and budget constraint.

7. Conclusions

The paper develops and estimates a dynamic household-level model of residential PV adoption, then uses it to evaluate incentive policies and broader durable-technology applications. Its results emphasize peer influence, forward-looking timing, and policy designs that balance adoption, emissions savings, and limited budgets.

  • Model contribution: The model combines economic benefits, neighborhood effects, forward-looking behavior, and unobserved heterogeneity in household adoption decisions.It is estimated on a 9-year household-level dataset spanning economic and geographic segments.
  • Empirical performance: Adding peer influence from nearby installations and forward-looking behavior reduces out-of-sample adoption prediction error relative to benchmark models.The estimates and predictive performance support the importance of these behavioral components.
  • Policy implications: Indefinitely extending rebates is not optimal for diffusion or budgets because forward-looking households postpone adoption, weakening neighborhood effects.A limited-period rebate can accelerate adoption as households anticipate expiration.
  • Policy implications: The simulations characterize a trade-off curve for policymakers seeking greater adoption and carbon savings under a limited rebate budget.The framework evaluates policy outcomes using adoption, budget, and carbon-savings considerations.
  • Broader applications: The framework can support incentive-driven adoption studies of electric vehicles, home chargers, heat pumps, emergency generators, and battery storage.These applications share large one-time investments, policy-shaped effective prices, observable peer adoption, and forward-looking decisions.
  • Future research: Future research could add household information, validate the model in other locations, develop seeding strategies, and apply it to other capital-intensive technologies.Additional information such as prior green-technology investment or social-media exposure could reduce unobserved heterogeneity and improve fit.

Dynamic Structural Model

The dynamic structural model represents household PV adoption as a forward-looking decision governed by evolving costs, benefits, neighborhood installations, and household-specific heterogeneity. Its parameters and state transitions are estimated through observed adoption and cost dynamics using nested fixed-point logic.

  • Model structure: Household choices depend on state variables for costs and neighborhood installations, household-specific utility parameters, and net present value.The model includes a household-specific unobserved heterogeneity term.
  • Identification: State transition functions are identified from observed costs and PV installations over time.Choice probabilities are estimated separately across economic and geographic segments.
  • Identification: The model assumes observed conditional choice probabilities are generated by the parametric structural model.This links the empirical choice probabilities to the structural parameters.
  • Identification: The discount factor β is fixed because it is generally not separately identifiable from the structural parameters.Given fixed β and state transition probabilities, the dynamic discrete choice parameters are uniquely identified.
  • Estimation: Estimation iterates between expected-value-function calculations and parameter estimation until the model reaches a fixed point.The final parameter vector includes economic effects, geographic effects, and household-specific heterogeneity.

A.2. Institutional Background and Exogeneity Evidence

The institutional setting provides a posted, uniform rebate schedule whose changes were administratively tied to installation-cost benchmarks rather than individual household adoption. Additional tests find little evidence that temporal demand shocks materially bias the estimated price response.

  • Institutional setting: Residential PV net cost incorporates the city rebate, institutional tariff, federal investment tax credit, and electricity-generation credits.The rebate rate changed through discrete downward revisions during the observation period.
  • Institutional setting: Austin Energy administered the rebate through a city budget cycle and applied a posted dollars-per-watt rate uniformly to qualifying applicants.Applicants could not negotiate the subsidy or condition it on neighborhood adoption history.
  • Institutional setting: Rebate reductions followed documented installation-cost declines and were infrequent, discrete, and downward-only.The schedule moved from $5.00 per watt initially to about $1.50 in mid-2013.
  • Exogeneity evidence: The exogeneity concern is that annual rebate budgets and rates may anticipate aggregate demand shocks in subsequent periods.The authors test this possibility using temporal fixed effects for unobserved citywide demand shocks.
  • Exogeneity evidence: The estimated NPV coefficients differ by less than one standard error with and without temporal fixed effects.This indicates that the NPV variation identifying price response is orthogonal to citywide temporal shocks.

Appendix B: Proof of Consistency of Using Cluster-Based Sampling

The appendix justifies aggregating similar non-adopting households into clusters to reduce computational cost while preserving consistent maximum-likelihood estimation under large-sample conditions.

  • Choice model: Household characteristics include economic and geographic segments, net present value, and other attributes entering the probabilistic adoption-choice model.The choice probability is represented as p(a|x_it;ϑ) for adoption decisions a = 0,1.
  • Likelihood construction: The full and weighted cluster-based likelihoods account for different sampling distributions, including cluster size and random household selection.A household selected from cluster M_k receives the corresponding inverse cluster-size weighting.
  • Consistency result: Under uniqueness and regularity conditions, the maximum-likelihood estimators from full and cluster-based samples converge almost surely to the same ϑ* as sample size and cluster count grow.The proposition assumes a unique population maximizer and requires M →∞ and K →∞.
  • Cluster construction: Cluster-based sampling selects households from groups of non-adopters with similar characteristics within each census block group.Clusters use home market value, house size, tree cover, and received irradiance as features, and remain fixed over the observation period.
  • Empirical implementation: The empirical clustering produced 1,185 multi-household clusters, with average size 26 and median size 7, and was judged not to introduce appreciable estimation bias.The large number of clusters supports applying the asymptotic consistency result.

Appendix D: Supporting Figures for Bayesian Posterior Sampling and Policy Simulations

The appendix reports Bayesian sampling diagnostics and a counterfactual rebate-switching trajectory, showing stable posterior estimates alongside an adoption surge before rebate expiration.

  • Posterior sampling: The MCMC chains reach their stationary distribution within the first 2,500 iterations, which are discarded as burn-in.Posterior means and 95% credible intervals are computed from retained draws, while trace and density plots support convergence.
  • Mixing diagnostics: Effective sample sizes are in the thousands for heterogeneity and the majority-segment net-present-value coefficient but only tens to low hundreds for several neighborhood effects.Neighborhood-effect coefficients are more strongly autocorrelated, motivating longer runs beyond burn-in.
  • Robustness of sampling: Extending the MCMC run from 4 to 8 hours leaves every posterior mean within roughly a quarter of its posterior standard deviation.This provides a sampling-duration stability check for the reported estimates.
  • Sampling caveat: The urban neighborhood-effect coefficient is weakly mixed, with its mean shifting by about one-third of a standard deviation.This coefficient has a low effective sample size relative to the other reported parameters.
  • Policy simulation: The adoption rate rises as the pre-announced rebate switching period approaches and drops after the rebate ends, creating a kink at period 40.The corresponding cumulative-adopter trajectory is reported for the one-step rebate switching policy RSP(40).

Appendix E: Additional Model Comparison and Selection Results

Alternative specifications yield broadly stable structural estimates, while the one-mile radial ring with two economic segments is selected as a parsimonious, interpretable model with comparable predictive accuracy.

  • Model selection: 31.85 for Model (4)H versus 31.37 for Model (10)H, with out-of-sample MAE differences below 2% between leading one-mile and half-mile specifications.Because predictive accuracy is nearly equivalent, structural stability and interpretability motivate the one-mile choice.
  • Parameter stability: The heterogeneity scale is statistically indistinguishable across twelve specifications, with 95% credible intervals overlapping near 0.051.The baseline utility remains within [−6.85,−6.64] and is precisely estimated in every specification.
  • Economic segmentation: Higher-value home segments receive larger net-present-value coefficients, although the segmented estimates are not distinguishable from one another at the 95% level.The low-value coefficient is near 3 × 10^-4, while higher-segment estimates overlap around 6.2 × 10^-4 and 8.5 × 10^-4.
  • Geographic segmentation: The neighborhood-effect ordering γ_rural > γ_suburban > γ_urban holds across one-mile-ring specifications.Standalone and inner-ring one-mile estimates overlap, including rural means of 0.121 and 0.135.
  • Selected specification: Model (4)H uses a one-mile radial ring and two economic segments as the parsimonious specification.The selected model provides a well-identified, monotone neighborhood structure.

Appendix F: Estimation with Expanding Data Window

Expanding-window estimation shows that structural parameters stabilize as additional high-adoption periods enter the sample, while neighborhood effects remain more sample-sensitive.

  • Economic coefficients: The two economic-segment net-present-value coefficients cross between windows T = 30 and T = 32 as more high-value adoptions enter the sample.The high-value coefficient rises above the majority-segment coefficient at T = 32 and settles near 6.4 × 10^-4 at T = 33.
  • Interpretation of convergence: The crossover reflects improved identification of the high-value segment’s price response rather than changing preferences.The ordering α_[300K,∞) > α_[0,300K) is recovered and persists from T = 32 onward.
  • Neighborhood effects: Neighborhood-effect estimates are larger and less precise in short early windows before gradually leveling off as the sample expands.Their economic ordering γ_rural > γ_suburban > γ_urban is preserved, and no coefficient changes sign.
  • Policy relevance: The authors conclude that the full estimation window provides the parameter stability needed for counterfactual policy analysis.The expanding-window exercise tracks posterior means and 95% credible intervals from T = 26 through T = 36.

Appendix H: Details of Differentiated Rebate Policy Simulations

The simulations compare geographically and home-value differentiated rebates by adoption, budget, and segment-specific effects. Geographic differentiation can improve adoption beyond budget predictions, whereas home-value differentiation largely follows budget.

  • Geographic differentiation: 27 geographic policies show adoption rising from 7,837 adopters at $38.4 million to 8,480 adopters at $45.0 million.
  • Geographic differentiation: 94.9 adopters per million dollars and R2 = 0.97 summarize the relationship between adoption and budget across geographic policies.
  • Geographic differentiation: 75 additional adopters per 0.1 suburban discount exceed the corresponding effects for urban and rural discounts, at 54 and 31 adopters.
  • Geographic differentiation: A 0.1 urban discount adds about 14 adopters beyond budget predictions, while suburban and rural discounts move policies below the fitted line.
  • Geographic differentiation: At about $42 million, policy (0.8,1.0,1.2) yields 8,212 adopters versus 8,154 under the observed uniform policy.
  • Home-value differentiation: Home-value policy adoption tracks budget closely, with a slope of 95.7 adopters per million dollars and R2 = 0.998.
  • Home-value differentiation: Discounting high-value and low-value households’ net costs adds about 246 and 82 adopters, respectively, while absorbing proportionally more budget.
  • Home-value differentiation: Differentiating rebates by home value does not substantially improve adoption for a fixed budget.

Appendix I: End-Horizon Sensitivity Analysis for Policy Simulations

End-horizon simulations show that installed capacity increases with budget but carbon savings peak at an interior rebate-switch date. The cost-effective switch remains intermediate across horizons.

  • Horizon robustness: At every horizon, installed capacity rises with budget as a concave relationship.
  • Horizon robustness: At every horizon, carbon savings are non-monotone in budget and maximize at an interior switch rather than the most expensive policy.
  • Horizon robustness: RSP(32) maximizes carbon savings at 2 years, while RSP(36) does so from 3 years onward.
  • Horizon robustness: The carbon-maximizing budget is about $30 million at 2 years and about $33 million from 3 years onward, while carbon savings rise from 174.2 to 604.2.
  • Mechanism: Earlier switching front-loads installations, allowing capacity to operate longer and accumulate more carbon savings.
  • Mechanism: At period 56, RSP(48) installs 38.6 MW versus 36.4 MW for RSP(36), yet carbon savings are 552.3 versus 604.2.
  • Horizon robustness: Across 2 to 6 years beyond estimation, an intermediate switch yields the most carbon savings per budget dollar, while later switching raises spending and lowers cumulative carbon.
Loading 2608.23796v1…