Households with Solar Installations Are Ideologically Diverse and More Politically Active Than Their Neighbours

Articles https://doi.org/10.1038/s41560-019-0498-8

Households with solar installations are ideologically diverse and more politically active than their neighbours

Matto Mildenberger 1,3*, Peter D. Howe2,3 and Chris Miljanich1

Climate risk mitigation requires rapid decarbonization of energy infrastructure, a task that will need political support from mass publics. Here, we use a combination of satellite imagery and voter file data to examine the political identities of US households with residential solar installations. We find that solar households are slightly more likely to be Democratic; however, this imbalance stems primarily from between-neighbourhood differences in partisan composition rather than within-neighbourhood differences in the rate of partisan solar uptake. Crucially, we still find that many solar households are Republican. We also find that solar households are substantially more likely to be politically active than their neighbours, and that these differences in political participation cannot be fully explained by demographic and socioeconomic factors. Our results demonstrate that individuals across the ideological spectrum are participating in the US energy transition, despite extreme ideological polarization around climate change.

nergy transitions are fundamentally political processes1,2. Even and behaviours are conditional on state-level policy support as renewable energy prices decline and technologies improve, have important implications for energy reforms. If early renewable fossil fuel incumbents mobilize to stymie and delay the energy energy adopters are mostly individuals who are already aligned with E 3–5 transition, often successfully . For instance, carbon-intensive proclimate political parties, then their political voice may only shape utilities have succeeded in repealing or retrenching numerous state- policymaking debates under limited circumstances. By contrast, if level renewable energy policies6. By contrast, emerging clean energy solar energy adopters come from across the political spectrum, then interest groups have sometimes counterbalanced these opponents7,8. a broad-based pool of potential citizen activists may enjoy a louder Members of the public are also important participants in the political voice. Moreover, if US solar households come from across energy transition. Policy advocates hope that individuals who have the political and ideological spectrum, we can likely expect similar installed solar panels may help protect existing clean energy poli- levels of cross-ideological solar adoption in countries where climate cies, or become advocates for policy expansion. These individu- and energy beliefs are not structured as strongly along a left–right als could also be mobilized by clean energy businesses as part of ideological spectrum16. broader lobbying campaigns. Already, solar energy households have A priori, we have reasons to expect either distribution. Energy mobilized to block retrenchment of net metering laws in some US policy has become increasingly polarized in the United States, mir- states, though with uneven success6. roring trends in climate and environmental policy17. Conservatives This theory of change is described by policy feedback theory, are less likely to believe that human-caused climate change is a framework that emphasizes how technologies and policies can happening or to support climate policies18,19. More conservative bring into existence new constituencies that reshape subsequent individuals are less likely to adopt an energy-efficiency measure, political debates9–11. Theoretically, widespread diffusion of energy particularly when framed as an environmental effort20. By contrast, generation capacity could nurture such feedback effects. There are Democratic control of state legislatures predicts more ambitious already about 2 million solar installations in the United States12. At clean energy policymaking21,22. scale, renewable energy infrastructure could involve millions of At the same time, the energy space also continues to be the site additional citizens with home energy installations. For instance, of bipartisan advocacy and legislation23. Many Republican states technical estimates suggest that 57% of US residential buildings, or passed renewable energy support policies in the 1990s and early over 67 million buildings, could support a solar photovoltaic (PV) 2000s24, even though entrenched fossil fuel and conservative lob- installation13. bies have since worked to retrench or repeal these reforms6,8,23. However, any effort to explore energy policy feedback requires In Georgia and Florida, the ‘Green Tea Party’ brought conservatives detailed knowledge of political identities and behaviours among together with environmental groups to campaign for net metering. solar households. Climate beliefs have become increasingly polar- In Arizona, prominent conservative Barry Goldwater Jr acted as ized along ideological lines in several countries, particularly in the spokesperson for a campaign to protect renewable energy laws the United States5,14,15. Understanding whether US households from utility attacks6. Moreover, there are persistent pockets of sup- with existing solar installations are mostly drawn from the politi- port for renewable energy policies among Republicans25. And, all cal left or from across the political spectrum, whether solar house- else equal, we should not expect that financial incentives to install holds are more likely to participate in politics than their non-solar renewable energy differ at the household level on the basis of neighbours, and whether the distributions of these political identities partisan affiliation.

1Department of Political Science, University of California, Santa Barbara, CA, USA. 2Department of Environment and Society, Utah State University, Logan, UT, USA. 3These authors contributed equally: M. Mildenberger, P. D. Howe. *e-mail: [email protected]

Nature Energy | www.nature.com/natureenergy Articles NaTURe EneRgy

Sampled households by county Table 1 | Comparison of solar-adopting households against benchmark neighbouring households Solar Control Statistic N Mean N Mean

N N (households in 4,129 4,129 2 sample) 10 20 Partisan model score 3,666 60.625 2,951 58.928 100 200 Prop. reg. Democratic 3,427 0.339 2,648 0.306 400 Prop. reg. Republican 3,427 0.200 2,648 0.218 Prop. registered to vote 3,666 0.813 2,951 0.768 Prop. voted in any 3,666 0.766 2,951 0.700 general election 08–17 Prop. voted in any 3,666 0.467 2,951 0.383 primary election 08–17 Fig. 1 | Location of households in study sample. Number (N) of households Prop. voted in any 3,666 0.198 2,951 0.141 aggregated by county. municipal election 08–17 Household income 3,539 107.774 2,785 96.063 (US$1,000) To date, the political identities and behaviours of solar households Prop. female 3,616 0.517 2,897 0.525 have not been a major focus of academic inquiry. Scholars have pri- Prop. White, non- 3,590 0.782 2,882 0.734 marily examined the social dynamics of technology adoption. They Hispanic 26 have identified strong peer effects in US adoption patterns , with Prop. Black or African 3,590 0.048 2,882 0.060 new solar installations more likely in areas with pre-existing installa- American, non-Hispanic tions27,28, partly independent of income and population density pat- Prop. Asian, non- 3,590 0.044 2,882 0.050 terns28. These peer effects, often described as a form of homophily, Hispanic have also been described across Europe29–31. Further, scholars have documented racial disparities in patterns of US solar energy adop- Prop. Hispanic or 3,590 0.121 2,882 0.152 tion, even when controlling for education and income32. However, Latino/a less work has examined the political dimensions of renewable energy Prop. own home 3,164 0.937 2,410 0.884 adoption, despite the partisan polarization of climate and energy Prop. bachelors or higher 3,350 0.429 2,604 0.394 policy preferences. In one analysis, researchers compare rates of solar Demographic data are only available where voter file data are available for a given address. Sample energy adoption in majority Democratic versus majority Republican sizes describe the total number of households with information on the given attribute. communities in Texas and New York33. They find stronger adoption rates in Republican-majority communities. Likewise, a power sector report found that individuals who donate to either Republican or Democratic causes are both likely to adopt solar34. demographic data, US voter file data, political donation data and This paper investigates the political identities of solar energy address-specific estimates of solar potential and capacity. Our voter adoption at the household level. Understanding these identities file data were prepared by a leading political data company (see helps describe the types of political coalition that may emerge to Methods), and included demographic variables, house ownership accelerate the energy transition. Our work departs from the previ- data, an imputed model of household resident partisanship, party ous literature in several ways. Previous work on energy adoption registration data in states where registration is publicly available often relies on self-reported data from opinion surveys. We join an and voting activity in previous elections. Combined, our unique emerging literature that uses address-level databases (compare ref. 28) dataset, all of which draws from publicly available information, and satellite imagery32,33. We also focus on household-level data, allows us to systematically explore the political and demographic moving to a more granular scale of analysis than studies that rely on characteristics of solar adopters at a more granular level than previ- census-tract or regional adoption rates. This responds to recent calls ous aggregated analyses. for more disaggregated analysis of renewable energy adoption35. Finally, our household approach also allows us to compare house- Demographic attributes of solar and non-solar households. We holds with solar installations with randomly selected neighbours to first evaluate the demographic and political attributes of US house- better understand neighbourhood-level politics of energy adoption. holds with solar installations. Table 1 compares descriptive statis- In brief, we first undertake a stratified random sample of US cen- tics for our sample of solar households against randomly selected, sus tracts that Google’s Project Sunroof models as having at least neighbouring, ‘control’ households without solar installations (see one solar rooftop installation (either solar thermal or solar PV)36. Methods for sampling strategies). We find moderate (but statisti- Project Sunroof identifies the presence of rooftop solar installa- cally significant) differences between solar and neighbour house- tions using remotely-sensed imagery. We stratify on the basis of the holds on a variety of demographic attributes. Solar households have density of solar installations in a given census tract, and manually incomes on average US$12,000 higher than neighbouring house- sample addresses within each tract. In total, this creates an initial holds (t(6,309) = −1.47, P = 0.0002, Cohen’s d = 0.09) and, not sur- sample of 4,145 solar households, distributed across the country as prisingly, are about five percentage points more likely to own their visualized in Fig. 1. home rather than rent (t(4,343) = −6.81, P < 0.0001, d = 0.19). They For every sampled solar address, we also collect a set of ran- are slightly more likely to have a resident with a bachelor’s degree domly selected neighbouring addresses without rooftop solar. We (t(5,579) = −3.13, P = 0.0017, d = 0.08) and five percentage points then merge solar and neighbour addresses with household-level more likely to have residents who identify as ‘White, non-Hispanic’

Nature Energy | www.nature.com/natureenergy NaTURe EneRgy Articles

a b 0.020 Unaffiliated/ Democratic Republican no party 0.5

0.015 0.4

0.3 Density

0.010 0.2

0.1

0.005 Proportion by party registration 0.0 Strong Lean Lean Strong Control Solar Control Solar Control Solar Rep. Rep. Dem. Dem. Household type Partisan score

Household type Control Solar

Fig. 2 | Comparing party affiliation across households of solar adopters and neighbouring control households. a, Distribution of mean partisan scores, plotted as kernel density estimates, for matched households where scores are available in all US states in sample (n = 5,278). Partisan scores based on TargetSmart 2016 National Partisan Model, representing the probability of supporting the Democratic Party. The distribution of partisanship is similar among solar households and neighbouring households without solar. b, Proportion of registered Democrats and Republicans among households of solar adopters and neighbouring control households, in states where party registration data are available (n = 4,432). Error bars represent ±2 standard errors.

(t(5,947) = −4.68, P < 0.0001, d = 0.12). Broadly, these findings partisanship scores. The partisan distribution of households in these replicate demographic disparities in solar adoption that have states is illustrated in Fig. 2b. Among matched households, the propor- previously been shown in aggregate32, though here at the house- tion of registered Democrats per household is slightly higher among hold level. Additional descriptive data that compare solar house- solar households and control households (mean difference = −0.021, holds with benchmark neighbouring households are provided in t(2,215) = −1.97, P = 0.048, d = 0.05). The proportion of registered Supplementary Tables 1–3. Note that our sample reflects the extant Republicans per household is not significantly different between distribution of US solar installations; more than 99% of our sample solar households and control households (mean difference = 0.018, is located in counties that contain at least one metro area with more t(2,215) = 1.89, P = 0.058, d = 0.05). Again, we find that party differ- than 250,000 residents. Our results may not generalize to future ences are small to non-existent and that solar installations are simi- solar households among the 15% of the US population that live in larly distributed across Democratic and Republican households. non-metropolitan counties. Differences in the aggregate number of solar households with Democratic versus Republican party affiliation could either stem Political affiliation among solar and neighbouring non-solar from within-neighbourhood differences in solar uptake between households. Despite these demographic differences, we find only Democrats and Republicans or between-neighbourhoods differ- small differences between solar and neighbouring non-solar house- ences in partisan composition (for example, neighbourhoods with holds with respect to partisan affiliation. When relying on official solar adopters are more likely to have Democrats in them). The party registration, 34% of members of solar-adopting households are results in Fig. 2 suggest the absence of within-neighbourhood dif- registered Democrats as compared with 31% in control households. ferences. We further investigate this possibility in Fig. 3 where we In solar-adopting households, 20% are registered Republicans, as show that, across different levels of geography, solar households compared with 22% in control households. While households with have similar partisan scores compared with neighbouring house- solar installations are slightly more likely to be Democratic than holds. Correspondingly, small imbalances in the aggregate number Republican, households with solar installations exist across the of Democratic versus Republican solar-adopting households most political spectrum. probably stem from higher solar uptake in neighbourhoods that In Fig. 2, we compare the partisan distribution of solar house- have more Democrats in them. holds with neighbouring non-solar households. Figure 2a includes all pairs of sampled households across the United States for which Political participation among solar and non-solar households. partisan scores are available, including in states that do not make While there are no substantial differences between the partisan individual-level party registration public. Here, we use an imputed affiliation of solar households and their neighbours, there are sub- partisanship score prepared by a leading voter file company stantial differences in political participation (Fig. 4). Solar-adopting (Methods). This score models the probability of being a Democrat households have a higher proportion of residents who voted in a or Republican for each US adult resident on a 1–100 scale. When previous general election between 2008 and 2017 (M = 0.77 in comparing only matched pairs of households, the mean parti- solar houses versus M = 0.70 in control houses; t(2,638) = −6.76, san score is similar in solar households versus control households P = 0.000, d = 0.18). Similarly, solar-adopting households are (mean difference = −1.19; t(2,638) = −1.48, P = 0.14, d = 0.03). more likely to have residents who voted in a previous primary The probability distributions of partisan scores are also similar election (M = 0.47 in solar houses versus M = 0.38 in control across solar and control households (Kolmogorov–Smirnov test: houses; t(2,638) = −7.66, P = 0.000, d = 0.20) or municipal elec- D = 0.03, P = 0.31). tion (M = 0.20 in solar houses versus M = 0.14 in control houses; We extend this analysis to the subset of sampled households t(2,638) = −6.87, P = 0.000, d = 0.16). Thus, solar households are living in states that publish party registration data. For these states, more politically active than adjacent non-solar households, and we directly compare partisanship without relying on imputed these differences in political participation are more substantial

Nature Energy | www.nature.com/natureenergy Articles NaTURe EneRgy

State level County level Table 2 | Linear probability model predicting presence of 100 household solar installation using demographics, partisan affiliation and political participation indicators 75 Household type: solar

50 Prop. White, non-Hispanic 0.098*** (0.023) Prop. female −0.032 (0.027) 25 Prop. bachelors or higher 0.011 (0.019) Own home 0.132*** (0.028) Mean HH income (log) 0.037** (0.012) ZIP code level Census-tract level 100 Prop. registered to vote 0.002 (0.038) Mean partisan model score 0.001*** (0.0002) 75 Prop. voted general election −0.004 (0.037) Prop. voted primary election 0.086*** (0.022) Partisan score for control households (mean) 50 N 4,916 R2 0.062 25 Residual std error 0.496 (d.f. = 4,568) F-statistic 0.865 (d.f. = 347; 4,568)

0 25 50 75 100 0 25 50 75 100 Model includes county fixed effects. Standard errors in parentheses. *P < 0.05; **P < 0.01; Partisan score for solar households (mean) ***P < 0.001.

Fig. 3 | Comparisons of mean partisan score for solar households and neighbouring control households across different geographic levels. are robust to alternative specifications, including a multilevel logis- Mean scores for geographic areas with at least five pairs of households are tic regression (Supplementary Table 4) or a spatially lagged model shown at the state, county, ZIP code and census-tract level. The diagonal (Supplementary Table 5). reference line indicates a 1:1 correspondence between partisan scores. We also explore other potential factors that might shape house- While points illustrate the wide range of partisanship across households hold-level political behaviours. First, we assess whether differences from different neighbourhoods, the lack of clustering above or below the between solar households and non-solar neighbours are a function reference line supports the absence of within-neighbourhood partisan bias of differentiated solar potential or capacity. We find no evidence for in uptake of solar technologies. this. There are no significant differences between solar households and neighbouring non-solar households with respect to usable sunlight hours per year (difference in means: 32.5 h, t(577) = −1.51, P = 0.13, d = 0.12), total roof area (in square feet) available for solar panels (difference in means: 180 square feet, t(577) 1.55, P 0.12, General Municipal Primary = = d = 0.12) and estimated net savings for the property projected over 0.8 20 years (difference in means: US$250, t(546) = −0.42, P = 0.68, d = 0.03). We then test whether there are differences in the partisan com- 0.6 position of solar energy adopters as a result of state-level policy environments. We find no differences in the partisan score distri- Household bution of solar adopters if we subset our analysis to states that have type 0.4 adopted a Renewable Portfolio Standard (D = 0.032, P = 0.206) or to Control states that have passed legislation to support residential net metering (D = 0.031, P = 0.246). Proportion voted Solar We also explore the distribution of political donations among 0.2

in any election (2008–2017) solar-adopting households, using political donation records (see Methods for data collection details). Of the Democrats living at sampled solar addresses, 0.9% had made a political donation. By 0.0 contrast, only 0.3% of our Republican sample had made politi- Control Solar Control Solar Control Solar cal donations over this period. The median donation size among Household type Democrats was US$500. The median donation size among the sample Republicans was US$657. Of course, most individuals did not Fig. 4 | Voting behaviour comparisons. Proportion of residents who voted donate to either party. in at least one general, primary or municipal election from 2008–2017, among households of solar adopters and neighbouring control households Comparing neighbourhoods with varying rooftop solar density. Finally, we explore whether early solar energy adopters have system- (n = 5,278). Error bars represent ±2 standard errors. atically different political profiles from later adopters. Per our sampling strategy (Methods), our sampled census tracts were divided than cross-group differences in partisanship. We find that these into five quintiles on the basis of the aggregate density of solar instal- large differences in primary election participation persist even after lations across all Project Sunroof census tracts. We can thus compare controlling for household-level demographic and socioeconomic households who have the only solar installation in their census tract attributes using a linear probability model (Table 2). These findings with households who have one of many installations in their census

Nature Energy | www.nature.com/natureenergy NaTURe EneRgy Articles

Unaffiliated/ General Municipal Primary Democratic Republican no party

0.75 Household 0.6 type 0.50 Household Control type 0.4 Solar Control (2008–2017) 0.25 Proportion voted

Proportion Solar 0.2

by party registration 0.00 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 0.0 Quintile of census tracts by density of solar installations 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 Quintile of census tracts by density of solar installations Fig. 6 | Voting behaviour by density of solar distributions. Distribution of proportion of residents who voted in general, municipal and primary Fig. 5 | Party registration by density of solar installations. Distribution elections from 2008–2017, by density of solar installations at census-tract of Democratic and Republican party registration among households of level. Error bars represent ± 2 standard errors (n = 5,278). solar adopters and neighbouring control households, by density of solar I installations at census-tract level (n = 4,502). Observations are divided by quintile of solar installation density. Error bars represent holds in the same neighbourhoods, we find similar proportions ±2 standard errors. of Democrats and Republicans living in households with a solar installation. This pattern is consistent with theories of solar uptake that emphasize material incentives to deploy solar, for instance tract. A priori, we might expect our results to be conditioned by as driven by solar leasing companies. It suggests that ideological this installation density. For instance, first adopters in a low-density polarization on climate and energy issues is not yet a binding con- tract may have invested in solar as a result of proenvironmental straint on solar technology uptake. Of course, despite this polariza- convictions. By contrast, households in neighbourhoods with more tion, recent survey research indicates that large pools of potentially solar may have stronger financial incentives to adopt solar (on the proclimate Republicans still exist in the United States. Our data basis of the revealed behaviour of area residents). suggest that these data translate into real economic decisions, since Empirically, we do not find evidence for such differences. Instead, there are sizable numbers of Republicans with a household-level we find that the distribution of party registration is similar between material stake in the energy transition. solar households and control households irrespective of tract-level Moreover, we do not see a gradient in the partisan affiliations of installation density (Fig. 5). This finding also addresses potential solar household across neighbourhood-level installation density. This concerns about homophily in patterns of solar adoption27,28,31,37. suggests that early solar adopters are not more ideologically-driven If neighbourhood-level clustering were shaping our results, then this than their immediate neighbours, nor are later solar adopters. Of would generate differences between solar and neighbouring house- course, we cannot easily compare these individuals with members of holds across density quintiles. However, we do not find any such evi- either party in other census tracts. The distribution of solar instal- dence. We also examine the distributions of political participation lations exhibits geographic patterns at the neighbourhood scale and in each quintile. In Fig. 6, we confirm our previous finding of asym- above due to peer effects and policy differences26. This is particularly metry. Across four of the five quintiles, we find that solar households true in the context of known racial and socioeconomic disparities are substantially more likely to also be households with high political between census tracts with solar energy deployment and those census participation, as indicated by voting in primary elections. tracts that do not have much deployment32,37. However, there are reasons to believe that our findings are not structured by partisan sorting Discussion into neighbourhoods, as such sorting is empirically rare38: individuals Using a dataset combining satellite imagery and individual-level do not move to neighbourhoods on the basis of their political prefer- voter file data, we compared individuals living in households hav- ences or their desire to live amidst like-minded partisans. ing rooftop solar installations with their neighbours. This analysis Future research should evaluate whether and how these house- allowed us to adjust for geographic differences in solar adoption, an holds will mobilize to contest public policy or defend existing important consideration given that solar adoption varies between policies against retrenchment efforts (compare ref. 6). We do find neighbourhoods, cities and states27,28,31. Our household-level strong differences in political participation when comparing solar analysis provides information to understand how people living in households and benchmark control households. US residents of households with and without solar panels may differ (or not). All solar households are more likely to vote in general, primary and things being equal, when comparing a house having solar panels municipal elections, often substantially more so. This would suggest with its neighbours, our analysis suggests that the people living in that solar households may enjoy a greater and more reliable politi- the house with solar panels are more likely to own their home (by cal voice than their non-solar neighbours. This gradient in political 13 percentage points), identify as White, non-Hispanic (by 10 per- participation is not entirely surprising: solar households also have centage points) and participate in elections (by 9 percentage points). higher incomes, a demographic attribute known to correlate with Our analysis indicates that US households with solar instal- voting behaviour. However, the association between solar installations include both Democrats and Republicans. While solar lation presence and political participation—as indicated by a sub- household residents are slightly more likely to be Democrats, stantially greater likelihood of voting in primary elections—is still the difference is smaller than other demographic factors: house- present after controlling for a full range of demographic attributes. holds composed of Democrats are only about 4 percentage points As the pace of the energy transition accelerates, solar adopters may more likely to have rooftop solar than households composed of become a growing constituency of relevance for policymakers on Republicans. Further, this effect is a function of neighbourhood both sides of the ideological aisle. composition, not differential partisan uptake of the technology More broadly, the energy transition will require accelerating within a given neighbourhood. When comparing matched house- deployment of clean energy resources and, probably, increased

Nature Energy | www.nature.com/natureenergy Articles NaTURe EneRgy adoption of distributed solar PV systems. Our results shed light on behaviour data for both our solar roof sample and the matched neighbouring the conditions under which this transition can occur. To date, politi- address sample. This included imputed partisan score data for each household using the TargetSmart 2016 National Partisan Model. For each address in our cal barriers to clean energy adoption remain a subject of considerable dataset, we received the names of all adults living at that address, their partisan debate. At a policy level, ideological polarization of political actors affiliation, their voting history, estimates of their ideological position, whether has led to substantial efforts to retrench clean energy laws, includ- the household rented or owned the property, and basic socioeconomic and ing support for such policies as net metering and solar mandates8. demographic data. In total, our sampling and data collection process resulted However, we show that this elite polarization is not yet reflected in in a dataset of 4,129 housing units with solar installations matched to 4,129 neighbouring housing units. We obtained voter file data for at least one individual patterns of US public uptake. If such polarization is not present in in 3,666 of the solar-adopting housing units and 2,951 of the neighbouring control the United States, the global setting with the most substantial polar- housing units. Records contained missing data for one or more variables, as ization over climate change, our results may place a meaningful indicated in Table 1. If one or both of each pair of solar and control households bound on the ideological gradient of solar adopters in other parts contained missing data for a particular test or figure, that pair of households of the world where climate and energy beliefs are less polarized. was dropped from the test. We used two-sided paired t-tests and a linear probability model for comparisons between solar and matched control households Broadly, they suggest the existence of a cross-ideology public coali- (collinearity diagnostics are provided in Supplementary Table 6 and tests for spatial tion with a growing stake in energy sector decarbonization. autocorrelation in Supplementary Table 7). Alternative regression specifications are provided in in Supplementary Tables 4 and 5. Methods Sampling solar households using aerial imagery. We sampled housing units Additional data sources. A human coder then manually looked up every with and without solar panels using a stratifed spatial random sample clustered individual in solar-adopting households in the Center for Responsive Politics’ by census tract. We frst used a dataset of US existing home solar installations Donor Lookup tool (https://www.opensecrets.org/donor-lookup). This database aggregated at the census-tract level, available from Google’s Project Sunroof36. includes all US Federal Election Commission recorded political donation data. Project Sunroof data are derived from aerial imagery and contain estimates of Generally, only donations above US$200 are publicly available. For each individual, the number of existing solar installations identifed using machine learning, we recorded the amount donated to Republican and Democratic candidates, the available aggregated at the census-tract, ZIP code and county levels. Locations of Republican and Democratic parties generally and political organizations affiliated individual housing units with solar installations were not publicly available but with left- and right-leaning organizations. were identifed by human coders. To do so, we downloaded tract-level Project In addition, a random sample of 600 matched pair addresses was taken, and Sunroof data and stratifed all tracts into quintiles on the basis of the density each address was manually searched within the Project Sunroof database to return of estimated solar installations (installations per housing unit) as calculated solar potential data for each address, including number of usable sunlight hours by Project Sunroof. Te frst quintile of tracts contained only one identifed per year, total roof area (in square feet) available for solar panels, and estimated net household with a solar installation. Te remaining quintiles contained 2–3, 4–8, savings (in US$) for the property projected over 20 years. 9–25 and 26–847 households respectively. We then randomly sampled (without This study was approved as exempt by the University of California Santa replacement) a list of census tracts within each quintile. In each quintile, starting Barbara Office of Research as part of protocol number 11-18-0104. Further with the frst randomly sampled tract, we located roofops with solar installations details are available on request from the corresponding author. Additionally, in within the boundaries of the census tract. To search for solar installations we Supplementary Note 1 we describe our approach to data de-identification. randomly assigned a cardinal direction to each tract, and coders began the search at the corner or side of the tract identifed by this cardinal direction. Coders Reporting Summary. Further information on research design is available in the searched in a line from this origin to the point at the opposite boundary of the Nature Research Reporting Summary linked to this article. tract, recording the location of any solar installations along this line. We recorded the locations of up to 10 solar installations within each tract. If no installations Data availability were found, the tract was skipped. In some cases, tracts were estimated by Project De-identified data that support the findings of the study have been deposited in the Sunroof data to contain more solar installations than were visible to human Harvard Dataverse40. coders. In other instances, installations that exceeded the number estimated by Project Sunroof were identifed (Fig. 1). Code availability We then created a database of matched neighbouring addresses for each Replication code to produce the figures and analyses reported in this study have housing unit identified with a solar installation. To identify neighbouring housing been deposited in the Harvard Dataverse 40. units, we used the point coordinates of each solar housing unit and randomly generated nine matching points within a buffer of 200 m around each point. Each randomly generated point was then reverse geocoded using the Google geocoding Received: 6 May 2019; Accepted: 9 October 2019; application programming interface in R 39. Reverse geocoding produced a list Published: xx xx xxxx of postal addresses, which was then cleaned to remove duplicate addresses and those identified by the geocoding application programming interface as ‘range References interpolated’, a range of addresses rather than a single address, which were less 1. Jones, C. F. Routes of Power (Harvard Univ. Press, 2014). likely to represent an actual housing unit. This protocol produced a set of addresses 2. Breetz, H., Mildenberger, M. & Stokes, L. Te political logics of clean energy of housing units identified as likely to have solar PV or solar thermal panels, along transitions. Bus. Polit. 20, 492–522 (2018). with up to nine matched neighbouring addresses. We randomly selected one of the 3. Oreskes, N. & Conway, E. M. Merchants of Doubt: How a Handful of Scientists matched neighbouring addresses for inclusion in the analysis. Obscured the Truth on Issues from Tobacco Smoke to Global Warming We sampled neighbouring households on the basis of spatial distance from the (Bloomsbury, 2011). solar household. This raises a concern that our control households may sometimes 4. Layzer, J. A. Open for Business: Conservatives’ Opposition to Environmental include a household with a solar installation. If this were true of a large number of our Regulation (MIT Press, 2012). control households, we might be concerned this would distort our descriptive results 5. McCright, A. M. & Dunlap, R. E. Te politicization of climate change and by making the control households look more similar to the treated households than is polarization in the American public’s views of global warming, 2001-2010. actually the case. We know from previous research that solar technology adoption is 27,28 Sociol. Q. 52, 155–194 (2011). shaped by peer effects and can often be spatially clustered . We extensively validate 6. Stokes, L. C. Short-circuiting Policy: Interest Groups and the Battle Over Clean our data to ensure that our control group does not have any households with visible Energy and Climate Policy in the American States (Oxford Univ. Press, 2020). solar installations. For each control household, we manually search the address 7. Aklin, M. & Urpelainen, J. Political competition, path dependence, and using Google Maps satellite imagery. Overall, as expected in the presence of spatial the strategy of sustainable energy transitions. Am. J. Polit. Sci. 57, clustering of solar installations, we found that 109 of 2,672 households in our initial 643–658 (2013). control group also had solar installations (just over 4%). These were concentrated in 8. Stokes, L. C. & Breetz, H. L. Politics in the US energy transition: case the quintile with the highest installation density, where 10% of the control group was studies of solar, wind, biofuels and electric vehicles policy. Energy Policy 113, also a solar household. For each of these 109 households we resampled from our list 76–86 (2018). of up to nine neighbouring households, until we found a household without solar 9. Pierson, P. When efect becomes cause: policy feedback and political change. (see below). In total, we were able to find good control matches for 93 of these 109 World Polit. 45, 595–628 (1993). cases. For 16 cases, we did not have a match in our dataset. We dropped these 16 solar 10. Levin, K., Cashore, B., Bernstein, S. & Auld, G. Overcoming the tragedy of households from our analysis. super wicked problems: constraining our future selves to ameliorate global climate change. Policy Sci. 45, 123–152 (2012). Comparing solar and non-solar households using voter file data. We then 11. Schmidt, T. S. & Sewerin, S. Technology as a driver of climate and energy contracted with a leading US voter file vendor to purchase political and voting politics. Nat. Energy 2, 17084 (2017).

Nature Energy | www.nature.com/natureenergy NaTURe EneRgy Articles

12. Solar Industry Research Data (Solar Energy Industries Association, 2019). 32. Sunter, D., Castellanos, S. & Kammen, D. Disparities in roofop photovoltaics 13. Sigrin, B. O. & Mooney, M. E. Roofop Solar Technical Potential for deployment in the United States by race and ethnicity. Nat. Sustain. 2, Low-to-Moderate Income Households in the United States (National Renewable 71–76 (2019). Energy Laboratory, 2018). 33. Sunter, D. A., Dees, J., Castellanos, S., Callaway, D. & Kammen, D. M. in 2018 14. Tranter, B. & Booth, K. Scepticism in a changing climate: a cross-national IEEE 7th World Conference on Photovoltaic Energy Conversion (WCPEC) study. Glob. Environ. Change 33, 154–164 (2015). (A Joint Conference of 45th IEEE PVSC, 28th PVSEC & 34th EU PVSEC) 15. McCright, A. M., Dunlap, R. E. & Marquart-Pyatt, S. T. Political ideology and 24262429 (IEEE, 2018). views about climate change in the European Union. Environ. Polit. 25, 34. Who Installs More Solar: Democrats or Republicans? (Power Scout, 2018). 338–358 (2016). 35. Tidemann, C., Engerer, N., Markham, F., Doran, B. & Pezzey, J. C. Spatial 16. Lewis, G. B., Palm, R. & Feng, B. Cross-national variation in determinants of disaggregation clarifes the inequity in distributional outcomes of household climate change concern. Environ. Polit. 28, 793–821 (2019). solar PV installation. J. Renew. Sustain. Energy 11, 035901 (2019). 17. Stokes, L. C. & Warshaw, C. Renewable energy policy design and framing 36. Project Sunroof Data Explorer (Google, 2018). infuence public support in the United States. Nat. Energy 2, 17107 (2017). 37. Graziano, M., Fiaschetti, M. & Atkinson-Palombo, C. Peer efects in the 18. McCright, A. M. Political orientation moderates Americans’ beliefs and adoption of solar energy technologies in the United States: an urban case concern about climate change. Clim. Change 104, 243–253 (2011). study. Energy Res. Soc. Sci. 48, 75–84 (2019). 19. McCright, A. & Dunlap, R. Te politicization of climate change and 38. Mummolo, J. & Nall, C. Why partisans do not sort: the constraints on polarization in the American public’s views of global warming. Sociol. Q. 52, political segregation. J. Polit. 79, 45–59 (2017). 155–194 (2011). 39. Kahle, D. & Wickham, H. ggmap: Spatial Visualization with ggplot2. R J 5, 20. Gromet, D. M., Kunreuther, H. & Larrick, R. P. Political ideology afects 144–161 (2013). energy-efciency attitudes and choices. Proc. Natl Acad. Sci. USA 110, 40. Mildenberger, M., Howe, P. & Miljanich, C. Replication data for: Households 9314–9319 (2013). with solar installations are ideologically diverse and more politically active than 21. Lyon, T. P. & Yin, H. Why do states adopt renewable portfolio standards?: their neighbors (Harvard Dataverse, 2019); https://doi.org/10.7910/dvn/4kleou an empirical investigation. Energy J. 31, 133-157 (2010). 22. Huang, M.-Y., Alavalapati, J. R., Carter, D. R. & Langholtz, M. H. Is the choice of renewable portfolio standards random? Energy Policy 35, 5571–5575 (2007). Acknowledgements 23. Hess, D. J., Mai, Q. D. & Brown, K. P. Red states, green laws: ideology and The authors wish to thank K. Goldstein for contributions to the data collection, as well as renewable energy legislation in the United States. Energy Res. Soc. Sci. 11, I. Stadelmann, L. Schaffer, P. Bergquist, participants at the 2019 Coevolution of Politics 19–28 (2016). and Technology workshop, ETH Zurich, and participants at the Comparative Political 24. Rabe, B. G. Statehouse and Greenhouse: the Emerging Politics of American Economy of Energy Transitions workshop, University of Lucerne, for comments on Climate Change Policy (Brookings Institution Press, 2004). earlier drafts of this article. 25. Mildenberger, M., Marlon, J. R., Howe, P. D. & Leiserowitz, A. Te spatial distribution of Republican and Democratic climate opinions at state and local Author contributions scales. Clim. Change 145, 539–548 (2017). M.M. and P.D.H. jointly participated in all stages of this study, including design, data 26. Rai, V., Reeves, D. C. & Margolis, R. Overcoming barriers and uncertainties collection, analysis and writing. C.M. participated in data collection and analysis. in the adoption of residential solar PV. Renew. Energy 89, 498–505 (2016). 27. Bollinger, B. & Gillingham, K. Peer efects in the difusion of solar photovoltaic panels. Mark. Sci. 31, 900–912 (2012). Competing interests 28. Graziano, M. & Gillingham, K. Spatial patterns of solar photovoltaic system The authors declare no competing interests. adoption: the infuence of neighbors and the built environment. J. Econ. Geogr. 15, 815–839 (2014). 29. Rode, J. & Weber, A. Does localized imitation drive technology adoption? A Additional information case study on roofop photovoltaic systems in Germany. J. Environ. Econ. Supplementary information is available for this paper at https://doi.org/10.1038/ Manage. 78, 38–48 (2016). s41560-019-0498-8. 30. Palm, A. Local factors driving the difusion of solar photovoltaics in Sweden: Correspondence and requests for materials should be addressed to M.M. a case study of fve municipalities in an early market. Energy Res. Soc. Sci. 14, Reprints and permissions information is available at www.nature.com/reprints. 1–12 (2016). 31. Dharshing, S. Household dynamics of technology adoption: a spatial Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in econometric analysis of residential solar photovoltaic (PV) systems in published maps and institutional affiliations. Germany. Energy Res. Soc. Sci. 23, 113–124 (2017). © The Author(s), under exclusive licence to Springer Nature Limited 2019

Nature Energy | www.nature.com/natureenergy nature research | reporting summary

Corresponding author(s): Matto Mildenberger

Last updated by author(s): 2019/10/06 Reporting Summary Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency in reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.

Statistics For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section. n/a Confirmed The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly The statistical test(s) used AND whether they are one- or two-sided Only common tests should be described solely by name; describe more complex techniques in the Methods section. A description of all covariates tested A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated

Our web collection on statistics for biologists contains articles on many of the points above. Software and code Policy information about availability of computer code Data collection Data was collected manually from the Google Project Sunroof website (https://www.google.com/get/sunroof)

Data analysis All analysis was conducted within the open-source statistical computing platform, R For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information. Data Policy information about availability of data All manuscripts must include a data availability statement. This statement should provide the following information, where applicable: - Accession codes, unique identifiers, or web links for publicly available datasets - A list of figures that have associated raw data - A description of any restrictions on data availability

Replication code, replication data and supporting information has been deposited within the Harvard Dataverse and can be accessed via the permanent DOI: https://doi.org/10.7910/DVN/4KLEOU October 2018

Field-specific reporting Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection. Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences

1 For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf nature research | reporting summary Behavioural & social sciences study design All studies must disclose on these points even when the disclosure is negative. Study description This study merges address-level data on the presence of solar panels with household-level voting and political outcomes. Population comparisons are then made between solar and neighboring non-solar households.

Research sample We sampled housing units with and without solar panels using a stratified spatial random sample clustered by census tract, using a dataset of U.S. existing home solar installations aggregated at the census tract level, available from Google Project Sunroof.

Sampling strategy We downloaded tract level Project Sunroof data and stratified all tracts into quintiles based on the density of estimated solar installations (installations per housing unit) calculated by Project Sunroof. The first quintile of tracts contained only one identified household with a solar installation. The remaining quintiles contained 2-3, 4-8, 9-25, and 26-847 households respectively. We then randomly sampled (without replacement) a list of census tracts within each quintile. In each quintile, starting with the first randomly sampled tract, we located rooftops with solar installations within the boundaries of the census tract. We then created a database of matched neighboring addresses for each housing unit identified with a solar installation.

Data collection To search for solar installations we randomly assigned a cardinal direction to each tract, and coders began the search at the corner or side of the tract identified by this cardinal direction. Coders searched in a line from this origin to the point at the opposite boundary of the tract, recording the location of any solar installations along this line. We recorded the locations of up to 10 solar installations within each tract. If no installations were found, the tract was skipped. To identify neighboring housing units, we used the point coordinates of each solar housing unit and randomly generated nine matching points within a buffer of 200 meters around each point. Each randomly generated point was then reverse geocoded using the Google geocoding API in R (Kahle and Wickham 2013). Reverse geocoding produced a list of postal addresses, which was then cleaned to remove duplicate addresses and those identified by the geocoding API as 'range interpolated,' a range of addresses rather than a single addresses, which were less likely to represent an actual housing unit.

We then contracted with a leading U.S. voter file vendor to purchase political and voting behavior data for both our solar roof sample and the matched neighbouring address sample. This included imputed partisan score data for each household using the TargetSmart 2016 National Partisan Model. For each address in our dataset, we received the names of all adults living at that address, their partisan affiliation, their voting history, estimates of their ideological position, whether the household rented or owned the property, and basic socioeconomic and demographic data.

A human coder also manually looked up every individual in solar-adopting households in the Center for Responsive Politics' Donor Lookup tool (https://www.opensecrets.org/donor-lookup). This database includes all U.S. Federal Election Commission recorded political donation data. Generally, only donations above $200 are publicly available. For each individual, we recorded the amount donated to Republican and Democratic candidates, the Republican and Democratic parties generally, and political organizations affiliated with left- and right-leaning organizations. In addition, a random sample of 600 matched pair addresses was taken, and each address was manually searched within the Google Sunroof database to return solar potential data for each address, including number of usable sunlight hours per year, total roof area (in square feet) available for solar panels, and estimated net savings (in U.S. dollars) for the property projected over 20 years.

Timing January 2018 to October 2019

Data exclusions n/a

Non-participation n/a

Randomization See details in Data Collection section.

Reporting for specific materials, systems and methods We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. Materials & experimental systems Methods n/a Involved in the study n/a Involved in the study Antibodies ChIP-seq

Eukaryotic cell lines Flow cytometry October 2018 Palaeontology MRI-based neuroimaging Animals and other organisms Human research participants Clinical data

2 Human research participants nature research | reporting summary Policy information about studies involving human research participants Population characteristics The study was reviewed and determined to be exempt by the University of California Santa Barbara Human Subjects Committee as part of as part of protocol number 11-18-0104.

Recruitment n/a

Ethics oversight University of California Santa Barbara Office of Research Note that full information on the approval of the study protocol must also be provided in the manuscript. October 2018