Shared research project · Literature-to-design guide

Where rooftop opportunity becomes solar adoption

Use economics to explain the steps between a technically suitable roof and a completed installation—then match each explanation to evidence that New York data can actually provide.

Frame the problemChoose a designStart the project
Core distinctionPotential is not adoption

A modeled roof is an opportunity, not an installed system or an economic optimum.

Useful unitPlace × time

Begin with county-year data; move to ZIP- or tract-time only when the underlying records permit it.

Claim ceilingDescriptive first

Call a pattern causal only when policy or program variation supplies a credible counterfactual.

The research problem

The project’s umbrella question is:

How do technical potential, household and neighborhood conditions, and policy or market access jointly shape realized distributed-solar deployment in New York?

The outcome is not one jump from sunlight to solar panels. It is a sequence of increasingly demanding conditions.

01Solar resourcesunshine
02Technical potentialroof + layout
03Economic potentialprivate or social NPV
04Market potentialconstraints + access
05Adoptioncompleted installation
06Outcomesgeneration + bills + emissions

Technical-potential studies estimate what suitable roofs could generate if used, independent of financial or social constraints (Gagnon et al. 2016). Google’s modeled roof geometry, sunshine, and panel layouts therefore populate an early rung of this ladder. They do not reveal whether a household installed panels, whether installation would be privately profitable, or whether it would maximize social welfare. See the Project Sunroof methodology and Solar API methodology for the construction of these measures.

Which gap?

The energy-efficiency-gap literature is useful here because it forces the analyst to name the benchmark before diagnosing “too little” adoption (Jaffe and Stavins 1994; Allcott and Greenstone 2012; Gerarden et al. 2017).

Technical benchmarkWhat is physically possible?

Google helps measure this.

Private benchmarkWhat is worthwhile for the decision-maker?

Tariffs, financing, tenure, risk, and hassle are needed.

Observed benchmarkWhat actually happened?

NYSERDA records projects, timing, and capacity.

Gap being described Comparison What it means Can current data identify it?
Deployment gap Technical potential ↔︎ observed deployment Some modeled opportunity is not realized Descriptively, with careful aggregation
Private investment gap Private optimum ↔︎ observed deployment Privately worthwhile investment did not occur No—household-specific costs and benefits are missing
Social welfare gap Social optimum ↔︎ observed deployment Deployment differs from the welfare-maximizing level or location No—marginal grid and environmental values are missing
Important

Interpretation rule: a large technical-potential gap is not, by itself, evidence of irrational households, a market failure, or a welfare-improving subsidy.

The adoption decision

For household or firm i, a useful organizing inequality is:

ADOPTi = 1 if expected bill savings + incentives + other private benefits exceed capital + finance + transaction + perceived-risk costs

Technical output enters expected bill savings, but nearly every other term varies across people, buildings, utilities, programs, and time. Borenstein shows that retail tariff design and tax incentives can materially change private solar value (Borenstein 2017). De Groote and Verboven show why the timing of benefits matters: households in their setting heavily discounted future production subsidies, so equivalent support delivered up front could induce adoption at much lower public cost (De Groote and Verboven 2019).

What the literature teaches

1. Nonadoption has competing explanations

Reviews of the energy-efficiency gap emphasize omitted costs, heterogeneous preferences, uncertainty, market failures, and behavioral mechanisms (Gillingham and Palmer 2014; Gerarden et al. 2017). The same low-adoption place can fit four different stories:

Model / measurement

The projected return is wrong.

Old imagery, a generic tariff, roof-repair needs, or optimistic output can make modeled value too high.

Look for: model–quote or prediction–production errors that vary systematically across places.
Rational heterogeneity

Waiting or declining is privately reasonable.

Moving risk, replacement timing, uncertainty, preferences, or option value can make nonadoption optimal.

Look for: adoption timing and household or building circumstances—not only average potential.
Market failure

A correctable wedge blocks adoption.

Credit constraints, split incentives, imperfect information, or spillovers separate private choices from an appropriate benchmark.

Look for: exogenous changes in finance, tenure, information, or incentives.
Behavioral friction

The decision process is distorted.

Inattention, present bias, inertia, and complexity may suppress action even when returns are understood correctly.

Look for: randomized simplification, reminders, salience, or information treatments.

The weatherization experiment by Fowlie, Greenstone, and Wolfram is the cautionary benchmark: realized savings were far below engineering projections, so apparently attractive modeled investments did not deliver the forecast returns (Fowlie et al. 2018). This does not prove that Google Solar overstates rooftop potential. It tells us to validate model-based production or value before treating a prediction as an economic return.

Note

The unsigned “Energy Efficiency Gap” course-note excerpt in the reference folder is useful for teaching behavioral policy debates, but it is a synthesis without identifiable publication metadata. Use the original studies—not the excerpt—as research citations.

2. Policy changes both adoption and incidence

Solar incentives can increase installations, but an adoption effect alone does not tell us whether a policy is well targeted or socially efficient.

Rebates

Hughes & Podolefsky

A California rebate increase raised installations by about 10%; their broader model also implies sizable inframarginal transfers (Hughes and Podolefsky 2015).

NY use: exploit discrete NY-Sun incentive changes, with a comparison market and technical-opportunity controls.
Financing

Kirkpatrick & Bennear

Property-assessed clean-energy financing increased solar investment in the studied California setting (Kirkpatrick and Bennear 2014).

NY use: compare places before and after financing access, if rollout dates and eligibility are recoverable.
Timing

De Groote & Verboven

A dynamic model finds strong discounting of future benefits and much greater cost-effectiveness from up-front support (De Groote and Verboven 2019).

NY use: model adoption timing around anticipated incentive-block changes; do not infer discounting from dates alone.
Tariffs

Borenstein

California’s tiered tariffs created a private solar incentive nearly as large as the federal tax credit in the study period (Borenstein 2017).

NY use: add utility territory, historical retail rates, and compensation rules before interpreting private returns.

3. Diffusion requires people, information, and suppliers

Spatial clustering is consistent with peer learning, but it can also reflect common housing, sorting, shared policies, or the same installers. Bollinger and Gillingham estimate peer effects using adoption histories and a design intended to separate prior exposure from shared neighborhood demand (Bollinger and Gillingham 2012). Solarize campaigns caused additional installations and lower prices, consistent with social learning and lower customer-acquisition costs, but the study does not isolate a pure information channel (Gillingham and Bollinger 2021).

Information also need not cross the entire adoption funnel. In a randomized experiment in India, an information tool improved knowledge and increased strong intent to adopt, yet the effect on actual adoption was statistically insignificant (Mahadevan et al. 2023). In U.S. platform data, prospective customers in lower-income tracts received fewer installer quotes conditional on measured factors; the observational design cannot reveal installer targeting independently of who enters the platform (O’Shaughnessy et al. 2021).

Opportunitysuitable roof
Awarenessknows the option
Accessquote + program
Intentwants to proceed
Applicationfinance + permit
Installationinterconnected system

NYSERDA observes the end of this funnel for recorded projects. Google primarily observes the first stage. Neither source directly observes awareness, quotes, credit approval, intent, or decision authority.

4. Opportunity, ability, and access shape equity

Adopters remain higher-income than the overall population, although participation has broadened over time (Forrester et al. 2024). Tenure adds a separate barrier because a renter may live under a suitable roof without holding the investment decision (Best et al. 2023). Installer attention can make access unequal even after a prospective customer seeks a quote (O’Shaughnessy et al. 2021).

Sunter, Castellanos, and Kammen reported large racial and ethnic deployment disparities using Project Sunroof and ACS tract data (Sunter et al. 2019). Dokshin and Thiede’s reconstruction produced different national magnitudes and emphasized filtering, normalization, state heterogeneity, urban sample bias, and ecological inference (Dokshin and Thiede 2023). Treat this as a methodological replication dispute—not as proof that the broader inequity question is settled in either direction.

Warning

Ecological-inference rule: a tract’s median income, renter share, or racial composition describes the area. It does not identify the income, tenure, or race of the household that adopted solar.

5. Private value and social value are different objects

Private value depends on avoided retail bills and transfers such as rebates or tax credits. Social value depends on real resource costs, displaced marginal generation, local pollution, climate damages, grid effects, and learning spillovers. These values vary over time and place (Borenstein 2012). Sexton and coauthors show that avoided pollution benefits can be poorly aligned with subsidy levels and estimate large gains from better spatial targeting (Sexton et al. 2021).

Private NPV

bill savings + export compensation + incentives + resilience/preferences − private costs

Social NPV

avoided generation + climate/local pollution + system effects − real resource and integration costs

NYSERDA plus Google can describe deployment and modeled annual production. They cannot, by themselves, value marginal emissions, congestion, capacity, or cost shifting. A welfare analysis is a later-stage project requiring grid and emissions data.

Evidence map: design, result, and transfer

The table below treats each paper as design → estimand → limitation → project use rather than as a free-floating conclusion.

Study Context and design What it credibly contributes Transfer to the NY project
Fowlie, Greenstone & Wolfram (Fowlie et al. 2018) Michigan weatherization; randomized encouragement plus quasi-experimental evidence Validation can overturn engineering-return projections Seek independent production, quotes, or roof assessments; do not transfer the weatherization return estimate to solar
Borenstein (Borenstein 2017) California billing, adopter, tariff, and incentive accounting Tariff design and incentives strongly shape private value; adopter composition matters Add utility territory, tariffs, consumption proxies, and incentive rules
Hughes & Podolefsky (Hughes and Podolefsky 2015) California rebate changes; quasi-experimental analysis and model counterfactual Local adoption response plus evidence of inframarginal transfers Separate a local policy response from a modeled no-program counterfactual
De Groote & Verboven (De Groote and Verboven 2019) Flanders; dynamic structural adoption model Timing and expected future benefits matter for subsidy design Use program schedules and expectations; do not label timing patterns “present bias” without a model
Bollinger & Gillingham (Bollinger and Gillingham 2012) California adoption histories; peer-effect design Prior nearby installations can affect later adoption Use ZIP-time histories, spatial lags, and tests against shared shocks; county-year data are too coarse for replication
Gillingham & Bollinger (Gillingham and Bollinger 2021) Solarize campaigns; program evaluation Campaigns raised installations and lowered prices Recover treated municipalities, dates, and comparison places before making causal claims
O’Shaughnessy et al. (O’Shaughnessy et al. 2021) EnergySage inquiries and quotes; observational platform data Supply-side access differs by area income, conditional on measured factors Developer concentration is a useful proxy, but NYSERDA cannot reveal rejected or missing quotes
Mahadevan, Meeks & Yamano (Mahadevan et al. 2023) Cluster-randomized information intervention in India Knowledge and intent can improve without a detectable installation response If outreach is tested in NY, preregister installation—not only intent—as the primary outcome
Sunter et al.; Dokshin & Thiede (Sunter et al. 2019; Dokshin and Thiede 2023) Project Sunroof + ACS tract comparisons and replication Equity estimates are sensitive to sample, normalization, and regional heterogeneity Report area-level disparities, weights, exclusions, and sensitivity analyses
Dokshin, Gherghina & Thiede (Dokshin et al. 2024) Nearly all NY residential incentive installations, 2010–2020; tract panels NY disparities changed over time and differed sharply by region Establishes the closest baseline; new work should add later years, technical opportunity, market mechanisms, or a sharper design

What is already known in New York

Two studies are especially close to this project. Araújo, Boucher, and Aphale analyze early clean-energy adopters in New York and find that income and home value are important correlates, while local patterns are more nuanced than a single statewide story (Araújo et al. 2019). Dokshin, Gherghina, and Thiede use geocoded NYSERDA records for 100,124 residential installations initiated from 2010–2020—95.6% of the state’s residential installations through 2020 in their data—to study tract-level disparities (Dokshin et al. 2024). They find that racial, income, and rural–urban gaps evolved differently over time and that New York City/Westchester, Long Island, and Upstate followed distinct trajectories.

That evidence raises the bar for a new contribution. A new project should not merely show that high-income places have more solar. It should add at least one of the following:

Update

Extend the panel

Use public NYSERDA records through 2026 and test whether earlier regional or income patterns persist.

Denominator

Add technical opportunity

Use vetted Google measures to distinguish low deployment from low sampled roof suitability—without calling the sample a statewide census.

Mechanism

Add policy or market access

Bring in incentive blocks, utility tariffs, Solarize timing, developers, or permitting to test a specific channel.

Credibility

Audit sensitivity

Show how geography, coverage, imagery vintage, eligible housing, and outcome normalization change the result.

What each dataset can—and cannot—measure

Source Economic role Useful measures Appropriate join Cannot establish
NYSERDA Statewide Distributed Solar Projects Realized deployment project timing, ZIP/county, capacity, estimated production, utility/developer fields in the full source ZIP- or county-time, depending on extract Household identity, rejected applications, technical eligibility, actual metered generation
Class county-year file Beginner-ready deployment panel projects, installed kWdc, estimated annual kWh county-year Building matching, peer exposure, household mechanisms
Google Solar faculty snapshot Sampled technical opportunity match quality, imagery date, roof area/segments, sunshine, modeled layouts and production aggregate vetted records to ZIP/county Installed panels, adoption date, household NPV, representative statewide rates
American Community Survey Area conditions and denominator housing units, owner occupancy, income, tenure, structure type, demographics tract/ZIP/county with compatible years Adopter characteristics or individual behavior
TIGER/Line Geographic crosswalk and spatial structure boundaries, GEOIDs, adjacency, land/water area stable geographic identifiers Economic mechanism or causal effect
Policy, tariff, and market records Mechanism and comparison incentive blocks, rates, compensation, program dates, installer presence, permits utility/place-time Credible causality unless timing and comparison assumptions are defensible

Current Google snapshot: research constraints

The Lecture 3 audit reports 10,000 candidate points, 8,311 API returns, 4,829 valid matches within 30 meters, and 4,791 valid matches with potential fields. New York City supplies 67.45% of valid matches; 531 valid returned buildings lie outside New York; imagery dates span 2012–2024. Those facts imply four rules:

  1. Filter geography explicitly. A successful API return is not necessarily in New York.
  2. Use the vetted parent table. Child panel/configuration rows must be semi-joined to valid parent input_id values.
  3. Model coverage and vintage. Missing, distant, or old-image matches are not zero-potential roofs.
  4. Do not estimate a statewide adoption rate. The metro-quota snapshot is not a census or documented probability sample.
Tip

Stock with stock, flow with flow. Compare cumulative installed capacity through a common date with a technical-capacity stock. Use annual project counts for policy timing or diffusion—not as the numerator of a cross-sectional “realized share.”

A research agenda for New York

Hypotheses and discriminating evidence

Hypothesis Observable implication Rival explanation Evidence that would move the claim forward
H1 · Technical opportunity Places with greater suitable sampled capacity have more cumulative deployment Google coverage and urban sampling A representative building frame, coverage weights, and sensitivity to imagery/match quality
H2 · Private value Deployment differs across utilities or tariff regimes at comparable potential Income, demand, housing, and local policy Historical tariffs, net-metering rules, consumption proxies, and boundary/time variation
H3 · Liquidity / tenure Lower deployment in low-income or renter-heavy areas at comparable opportunity Building type, roof condition, or installer targeting Financing eligibility/rollout, parcel tenure, and program take-up or application data
H4 · Peer learning Prior nearby installations predict later adoption Sorting, shared shocks, policy, and installers Fine time/space histories plus an instrument, boundary, campaign, or other source of exogenous exposure
H5 · Installer access Low deployment where few developers operate or quotes arrive Low underlying demand Quote/lead data, installer entry, travel costs, or market-boundary changes
H6 · Policy timing Installations bunch or change around incentive blocks Anticipation, seasonality, prices, and administrative delay Exact rule dates, eligibility, application/completion timing, and untreated comparison places

Claim ladder

Level 1

Describe

Where and when did projects, capacity, and sampled potential occur?

NYSERDA + vetted Google + transparent denominators
Level 2

Condition

How do patterns differ with income, tenure, housing, region, or installer presence?

Add ACS, geography, coverage, and uncertainty
Level 3

Identify

What changed adoption because of a policy, information intervention, or market-access shift?

Add treatment timing, a counterfactual, and design-specific assumptions
Level 4

Value

Did the change improve private or social welfare, and for whom?

Add prices, real costs, transfers, grid effects, emissions, and distribution

Move upward only when the data support the next claim. A sophisticated model does not substitute for a credible comparison.

Minimum viable project

The first project cycle can remain entirely within open data while the restricted Google snapshot is reviewed.

1

Reproduce the NY baseline

Audit county-year rows, units, missingness, time coverage, and cumulative versus annual outcomes.

2

Choose a defensible denominator

Use owner-occupied or suitable housing units when the claim requires an opportunity set; explain why it matches the outcome.

3

Build the literature mechanism map

For each proposed variable, name the theory, predicted sign or comparison, and at least one rival mechanism.

4

Audit the Google sample

Filter valid New York parents, quantify coverage and imagery vintage, and aggregate without exposing coordinates or identifiers.

5

Join at the coarsest valid geography

Start with county or ZIP summaries; preserve a table of every crosswalk and unmatched unit.

6

Stress-test the headline

Re-estimate under alternative years, denominators, coverage thresholds, regional exclusions, and outcome definitions.

Contribution paths

Data audit

Make the denominator trustworthy

Document rows, units, missingness, geography, time, duplicates, and stock/flow construction.

Literature

Own one mechanism

Trace one claim—tariffs, finance, peers, installer access, tenure, or equity—from theory to estimand and limitation.

Policy

Build the NY chronology

Record incentive blocks, rate changes, Solarize campaigns, and program rules with dates and geographic eligibility.

Spatial

Map opportunity and deployment

Test how maps change with normalization, coverage, region, and the Modifiable Areal Unit Problem.

Replication

Reproduce a published pattern

Translate a paper’s unit, outcome, sample restriction, and comparison into a public-data approximation.

Governance

Design the safe research layer

Separate public tables from restricted coordinates, credentials, raw responses, and derivative exports.

Evidence memo template

For every result, record:

  1. Claim: the one sentence the figure or model supports;
  2. Unit and sample: what one row represents and which places/times are included;
  3. Outcome and denominator: including stock/flow and measured/estimated distinctions;
  4. Comparison: what variation identifies the result;
  5. Mechanism: which theory predicts it;
  6. Rivals: at least two alternative explanations;
  7. Sensitivity: which cleaning, geography, or model choice changes it; and
  8. Claim ceiling: descriptive association, causal effect, or welfare statement.
CautionData governance before expansion
  • Never place API keys, raw coordinates, or restricted Google responses in the course repository.
  • Students should not make billable API calls. The first class projects can use the open NYSERDA county-year file.
  • Before sharing or publishing Google-derived records, complete institutional/legal review of the current Maps Platform terms and Solar API policies.
  • Publish only approved, privacy-preserving aggregates with a provenance note and explicit coverage limitations.

Decision rule for the project

Research principle Policy follows the diagnosed mechanism—not the size of the raw technical-potential gap.

Begin with a transparent descriptive baseline. Add data only when they distinguish an economic explanation or unlock a more credible comparison.

References

Allcott, Hunt, and Michael Greenstone. 2012. “Is There an Energy Efficiency Gap?” Journal of Economic Perspectives 26 (1): 3–28. https://doi.org/10.1257/jep.26.1.3.
Araújo, Kathleen, Jean Léon Boucher, and Omkar Aphale. 2019. “A Clean Energy Assessment of Early Adopters in Electric Vehicle and Solar Photovoltaic Technology.” Journal of Cleaner Production 216: 99–116. https://doi.org/10.1016/j.jclepro.2018.12.208.
Best, Rohan, Andrea Chareunsy, and Madeline Taylor. 2023. “Emerging Inequality in Solar Panel Access Among Australian Renters.” Technological Forecasting and Social Change 194: 122749. https://doi.org/10.1016/j.techfore.2023.122749.
Bollinger, Bryan, and Kenneth Gillingham. 2012. “Peer Effects in the Diffusion of Solar Photovoltaic Panels.” Marketing Science 31 (6): 900–912. https://doi.org/10.1287/mksc.1120.0727.
Borenstein, Severin. 2012. “The Private and Public Economics of Renewable Electricity Generation.” Journal of Economic Perspectives 26 (1): 67–92. https://doi.org/10.1257/jep.26.1.67.
Borenstein, Severin. 2017. “Private Net Benefits of Residential Solar PV: The Role of Electricity Tariffs, Tax Incentives, and Rebates.” Journal of the Association of Environmental and Resource Economists 4 (S1): S85–122. https://doi.org/10.1086/691978.
De Groote, Olivier, and Frank Verboven. 2019. “Subsidies and Time Discounting in New Technology Adoption: Evidence from Solar Photovoltaic Systems.” American Economic Review 109 (6): 2137–72. https://doi.org/10.1257/aer.20161343.
Dokshin, Fedor A., Mircea Gherghina, and Brian C. Thiede. 2024. “Closing the Green Gap? Changing Disparities in Residential Solar Installation and the Importance of Regional Heterogeneity.” Energy Research & Social Science 107: 103338. https://doi.org/10.1016/j.erss.2023.103338.
Dokshin, Fedor A., and Brian C. Thiede. 2023. “Revised Estimates of Racial and Ethnic Disparities in Rooftop Photovoltaic Deployment in the United States.” Nature Sustainability 6: 752–55. https://doi.org/10.1038/s41893-023-01134-4.
Forrester, Sydney, Galen Barbose, Eric O’Shaughnessy, and Naím Darghouth. 2024. Residential Solar-Adopter Income and Demographic Trends: 2024 Update. Lawrence Berkeley National Laboratory. https://emp.lbl.gov/publications/residential-solar-adopter-income-3.
Fowlie, Meredith, Michael Greenstone, and Catherine Wolfram. 2018. “Do Energy Efficiency Investments Deliver? Evidence from the Weatherization Assistance Program.” Quarterly Journal of Economics 133 (3): 1597–644. https://doi.org/10.1093/qje/qjy005.
Gagnon, Pieter, Robert Margolis, Jennifer Melius, Caleb Phillips, and Ryan Elmore. 2016. Rooftop Solar Photovoltaic Technical Potential in the United States: A Detailed Assessment. NREL/TP-6A20-65298. National Renewable Energy Laboratory. https://doi.org/10.2172/1236153.
Gerarden, Todd D., Richard G. Newell, and Robert N. Stavins. 2017. “Assessing the Energy-Efficiency Gap.” Journal of Economic Literature 55 (4): 1486–525. https://doi.org/10.1257/jel.20161360.
Gillingham, Kenneth, and Bryan Bollinger. 2021. “Social Learning and Solar Photovoltaic Adoption.” Management Science 67 (11): 7091–112. https://doi.org/10.1287/mnsc.2020.3840.
Gillingham, Kenneth, and Karen Palmer. 2014. “Bridging the Energy Efficiency Gap: Policy Insights from Economic Theory and Empirical Evidence.” Review of Environmental Economics and Policy 8 (1): 18–38. https://doi.org/10.1093/reep/ret021.
Hughes, Jonathan E., and Molly Podolefsky. 2015. “Getting Green with Solar Subsidies: Evidence from the California Solar Initiative.” Journal of the Association of Environmental and Resource Economists 2 (2): 235–75. https://doi.org/10.1086/681131.
Jaffe, Adam B., and Robert N. Stavins. 1994. “The Energy-Efficiency Gap: What Does It Mean?” Energy Policy 22 (10): 804–10. https://doi.org/10.1016/0301-4215(94)90138-4.
Kirkpatrick, A. Justin, and Lori S. Bennear. 2014. “Promoting Clean Energy Investment: An Empirical Analysis of Property Assessed Clean Energy.” Journal of Environmental Economics and Management 68 (2): 357–75. https://doi.org/10.1016/j.jeem.2014.05.001.
Mahadevan, Meera, Robyn Meeks, and Takashi Yamano. 2023. “Reducing Information Barriers to Solar Adoption: Experimental Evidence from India.” Energy Economics 120: 106600. https://doi.org/10.1016/j.eneco.2023.106600.
O’Shaughnessy, Eric, Galen Barbose, Ryan Wiser, and Sydney Forrester. 2021. “Income-Targeted Marketing as a Supply-Side Barrier to Low-Income Solar Adoption.” iScience 24 (10): 103137. https://doi.org/10.1016/j.isci.2021.103137.
Sexton, Steven, A. Justin Kirkpatrick, Robert I. Harris, and Nicholas Z. Muller. 2021. “Heterogeneous Solar Capacity Benefits, Appropriability, and the Costs of Suboptimal Siting.” Journal of the Association of Environmental and Resource Economists 8 (6): 1209–44. https://doi.org/10.1086/714970.
Sunter, Deborah A., Sergio Castellanos, and Daniel M. Kammen. 2019. “Disparities in Rooftop Photovoltaics Deployment in the United States by Race and Ethnicity.” Nature Sustainability 2: 71–76. https://doi.org/10.1038/s41893-018-0204-z.
Back to top