全球变化知识资源中心

Addressing climate change in the United States requires enactment of national, state and local mitigation and adaptation policies. The success of these initiatives depends on public opinion, policy support and behaviours at appropriate scales. Public opinion, however, is typically measured with national surveys that obscure geographic variability across regions, states and localities. Here we present independently validated high-resolution opinion estimates using a multilevel regression and poststratification model. The model accurately predicts climate change beliefs, risk perceptions and policy preferences at the state, congressional district, metropolitan and county levels, using a concise set of demographic and geographic predictors. The analysis finds substantial variation in public opinion across the nation. Nationally, 63% of Americans believe global warming is happening, but county-level estimates range from 43 to 80%, leading to a diversity of political environments for climate policy. These estimates provide an important new source of information for policymakers, educators and scientists to more effectively address the challenges of climate change.

Decision makers need locally relevant information about the physical impacts of climate change to inform mitigation and adaptation efforts. In response, climate scientists have developed a variety of methods to downscale climate change projections from global models to finer regional and local scales. Mitigation and adaptation initiatives also depend heavily on social factors such as levels of public awareness, risk perceptions, policy support and knowledge of appropriate behavioural responses^{1, 2, 3, 4, 5, 6}. However, while these critical social data are often available at the national scale (for example, national surveys), they rarely exist at the sub-national levels required by scientists and policy makers. To comprehensively assess climate change risks and the prospects for mitigation and adaptation initiatives, it is necessary to have accurate local-scale data on public climate change beliefs, risk perceptions, policy preferences and behaviour. To address this need, this study uses national-scale data to ‘downscale’ estimates of public responses to climate change to sub-national scales, providing locally relevant information about public opinion for scientists and national, state and local decision makers.

Previous research has found that public climate change policy support and behaviour are greatly influenced by public beliefs, attitudes and risk perceptions^{7, 8, 9}. In turn, these critical variables are influenced by other factors, including knowledge, emotion, ideology, demographics and personal experience^{2, 10, 11, 12, 13}. Climate change mitigation and adaptation decisions are made at multiple spatial scales, including households, cities, counties and states^{14, 15, 16, 17}, so understanding how beliefs, attitudes and risk perceptions shape public responses to climate change requires information on these factors at the appropriate level of decision-making. However, at present we know little about how public beliefs, attitudes and risk perceptions vary geographically or how they relate to policy outcomes at these critical sub-national scales.

Perceptions of climate change are likely to vary geographically as a function of demographics and of cultural and ideological factors, because people with similar demographic, cultural and ideological characteristics tend to cluster together^{2, 6, 12, 18}. In addition, climate change perceptions possibly exhibit geographic patterns due to differences in personal experience with extreme weather events and climate variability, since local weather is known to influence climate change perceptions^{13, 19, 20, 21, 22, 23, 24, 25}. Sociodemographic and ideological characteristics may also affect how personal experiences with climatic phenomena are translated into perceptions and beliefs^{24, 26, 27}.

National surveys in the US consistently find that a majority of Americans believe global warming is happening (63%), while fewer believe that it is human-caused (47%) or that most scientists think it is happening (42%; ref 28). However, national-level statistics obscure large variations in opinions between states. For example, Californians are more likely than Ohioans to believe that global warming is happening and that most scientists think global warming is happening²⁹. While there have been several state and local surveys of public climate change opinion, there are at present no comprehensive assessments of the geographic variation in public climate change beliefs, attitudes and behaviours across the United States at the state, congressional district, county, and other sub-national scales. Conducting a comprehensive set of surveys in all 50 states and the District of Columbia, the 435 congressional districts, or the 3,143 counties across the United States would be prohibitively expensive, and pooling existing survey data from diverse sources is problematic due to often incompatible item wordings and different times of data collection.

Two primary methods have been used to address the problem of sparse public opinion data: national-level disaggregation^{30, 31} and Bayesian approaches such as multilevel regression and poststratification^{32, 33, 34}. Disaggregation involves compiling a large set of nationally representative survey data, then pooling the responses of all respondents located in each unit of the geographic level of interest³⁰—for instance, each state or congressional district. This approach, however, requires a large number of survey respondents to meet the minimal sample sizes necessary to obtain reliable estimates of public opinion. Disaggregating even very large data sets typically provides insufficient sample sizes to produce accurate estimates, especially in small population areas (for example, Wyoming). In addition, disaggregation often requires the compilation of polling data over multiple years and therefore cannot account for changes in public opinion over time.

An alternative approach, multilevel regression and poststratification (MRP), also involves compiling data from multiple national surveys, but incorporates demographic, geographic and time variables to partially pool information about respondents across sub-groups. The first stage of MRP (multilevel regression) models individual outcome variables (for example, beliefs, attitudes, policy support, and so on) as a function of demographics, state- or region-specific geographic effects, and temporal effects to account for changes in public opinion over time. In the second stage (poststratification), modelled coefficients for each demographic–geographic respondent type are weighted by the proportion of each type within each geographic area. Unlike disaggregation, MRP methods can reliably project opinion in areas with sparse data coverage by partially pooling information from survey responses outside of that local geographic unit. Previous research has demonstrated that MRP methods can greatly improve the accuracy of public opinion estimates and reduce uncertainties compared to disaggregation methods at the state and congressional district level^{35, 36, 37, 38}. Questions remain, however, about its validity with small samples, higher-resolution geographies, and its applicability beyond the narrow range of political opinion variables to which it has previously been applied³⁹.

Here we present validated MRP model estimates of public climate change beliefs, risk perceptions, policy preferences and behaviour based on 12 nationally representative surveys conducted by the Yale Project on Climate Change Communication and George Mason Center for Climate Change Communication between 2008 and 2013 (n = 12,061). The models use individual-level demographic predictors, state-, district- and county-level random effects, random effects based on the year of the survey and the survey mode, and geographic-level covariates. By incorporating random effects for the year of each survey we can account for changes in public opinion over time. Estimates presented here are averages for the year 2013. This paper provides estimates of public climate change beliefs, risk perceptions, policy preferences and behaviours at four geographic levels within the US: each of the 50 states and the District of Columbia, 381 Metropolitan Statistical Areas, 435 congressional districts, and 3,143 counties or county-equivalents.

The data set comprises surveys with dozens of identical questions measuring public responses to climate change, providing an unusually comprehensive source of detailed information on climate change beliefs, risk perceptions, policy preferences and behaviours. This data set contrasts with previous research using MRP, which has focused on only a narrow set of public opinion variables and relied on data sets that collapse differently worded questions from multiple independent surveys into a single latent construct.

The results are validated using two methods: internal cross-validation, and external validation against independently conducted surveys at the state and metropolitan levels. The external validations, using multiple independent surveys across diverse state and metropolitan areas with identical questions, improve on and extend previous MRP research. Both validation methods indicate that the MRP model provides accurate estimates of public responses to climate change at each sub-national scale investigated (typically within 0–5% points). Bootstrap margins of error based on 95% confidence intervals average ±5% points for the state-level models, ±7% points for congressional district-level models, and ±8% points for the county-level models.

We illustrate the approach by describing the distribution of public belief that global warming is happening, the belief that global warming is human-caused, beliefs about scientific agreement regarding global warming, public support for climate policy, and global warming risk perceptions. Model estimates of additional public climate change opinion variables for a range of geographies in 2014 are available at http://environment.yale.edu/poe/v2014

The results demonstrate that public responses to climate change vary substantially within the United States. Figure 1 illustrates model estimates at the state level for the four following beliefs and policies: global warming is happening; if global warming is happening it is caused mostly by human activities; most scientists think global warming is happening; and support for regulating carbon dioxide as a pollutant. The left-hand column of Fig. 1 provides the estimated absolute levels of belief in each state; maps in the right-hand column depict the differences between the estimated opinion in each state and the national average. The model estimates that a majority of adults in all states believe that global warming is happening, ranging from lows of 54% and 55% in West Virginia and Wyoming to highs of 75% and 81% in Hawaii and the District of Columbia respectively. Geographic patterns depend on the particular belief, risk perception, policy, or behaviour in question. For example, although majorities of the public in all states believe that global warming is happening (Fig. 1a) and that carbon dioxide should be regulated as a pollutant (Fig. 1d), many states do not have majorities who believe that global warming, if it is happening, is caused mostly by human activities (Fig. 1b) or that most scientists think that global warming is happening (Fig. 1c). The Supplementary Information includes the full estimates at each geographic level.

a–d, The maps depict the percentage of American adults in each state who believe that global warming is happening (a); believe global warming is mostly human-caused (b); believe that most scientists think global warming is happening (c); somewhat or strongly support the regulation of CO₂ as a pollutant (d). Left-hand panels depict the projected population percentages, whereas right-hand panels depict the relative differences from the national average to facilitate comparisons between states.

The model estimates were verified using two different types of validation analysis: direct external validation using independent representative public opinion surveys and internal cross-validation.

For direct external validation, representative telephone-based surveys were conducted in four states (California, Texas, Ohio and Colorado, n = 800 per state) and two MSAs (Columbus and San Francisco, n = 700 per MSA). The validation surveys were conducted in 2013 using mobile and landline telephones, whereas the majority of survey data used in the model were collected via a nationally representative online panel.

The external validation indicates that the model estimates are highly accurate. Figure 4 depicts the independent validation survey results (x-axis) against the model estimates (y-axis) for four states and two metropolitan areas across 11 variables measuring diverse constructs (for example, climate change beliefs, risk perceptions, policy preferences and behaviour). The model estimates and survey results were strongly correlated within each geographic area. Across the 11 variables and accounting for mode differences, the mean absolute difference between model estimates and validation survey results was 2.9 (s.d. = 1.5) percentage points among the four states and 3.6 (s.d. = 2.9) percentage points among the two metropolitan areas, within the margins of error at 95% confidence for the survey results alone. The Supplementary Information also reports results from a comparison of the MRP model results with disaggregated results from a second independent survey data set, the Cooperative Congressional Election Survey (CCES; ref. 40), using a differently worded measure of public belief that global warming is happening. Disaggregated CCES climate opinion and the most similar question in our data set were correlated above the 0.8 level for all geographies.

a, Colorado. b, Texas. c, Ohio. d, California. e, Columbus, Ohio. f, San Francisco, California.

Modelling public climate change beliefs, risk perceptions, policy preferences and behaviours using MRP methods on a large survey data set (for example, more than 10,000 respondents) produces highly accurate results, as verified by both independent validation data (Fig. 4) and cross-validation techniques (Fig. 5). Such high levels of accuracy may seem unexpected given the inherent difficulty in predicting individual-level beliefs, but they are analogous to model projections of long-term climate versus short-term weather. Although climate models cannot accurately project weather conditions in a specific place on a single day, they are able to accurately project long-term average weather conditions. Similarly, it is possible to accurately estimate average opinion among sub-groups of the population even while estimates for a particular individual would be less accurate. The MRP models presented here are not designed to be individual-level predictive models, because the independent variables are limited to those that can be obtained for the entire US at each geographic level of analysis. However, these models take advantage of the hierarchical geographic structure of national survey data sets, combined with geographic covariates, to produce valid estimates for aggregated populations at sub-national scales.

The results demonstrate that, as with previous MRP studies of controversial policy issues, public opinion about global warming exhibits substantial variation between and within regions, states and cities. Geographic patterns in beliefs are often consistent with what one might expect from political patterns, with traditionally ‘blue states’ such as California and New York, for example, showing relatively high concern about climate change, and ‘red states’ such as Wyoming and Oklahoma showing lower concern. However, summarizing perceptions at the state level obscures variability at finer scales. In Teton County, Wyoming, for example, we estimate that 64% of adults believe that global warming is happening, similar to the national average, despite an estimate for the state as a whole of 55%. Likewise, projected belief in global warming is relatively low (55%) in Lewis County, Washington, a blue state, whereas the level of belief in the state as a whole is higher (67%).

Additional patterns showing geographic variation in public opinion are visible in the congressional district- and county-level maps. Southwestern Texas, for example, shows belief in global warming in the 60–70% range, about 10% points higher than other areas of the state, possibly due to the greater proportion of Hispanic/Latino adults there who tend a greater tendency to believe that global warming is happening than whites, on average⁴¹. Similar geographic variations in racial and ethnic composition at the county level also translate to elevated levels of belief that global warming is happening in the majority-Black counties of central Alabama that stand in contrast to the rest of the state. The interplay of demographic and geographic influences on climate change opinions are also reflected in variations between urban and rural areas of the country. Most counties that include the nation’s largest cities, such as New York, Chicago and Los Angeles, show relatively high levels of belief that global warming is happening, whereas proportions in most rural counties are significantly lower. In rural areas, we also find lower levels of belief that global warming is happening in some Midwestern and Western counties with large greenhouse-gas-producing industries, such as coal-fired power plants. The presence of colleges and universities also appears to be a factor associated with high levels of belief that there is a scientific consensus that global warming is happening.

Cross-validation results indicate that the model estimates in low-population areas are likely to be somewhat less accurate than the estimates for areas that have a larger number of respondents in nationally representative survey data sets, although they still far exceed the accuracy of estimates from disaggregation. This uncertainty is more pronounced in county-level models. Although 1,281 of 3,143 counties (40.7%) lack respondents in the national survey data because of their low populations, these counties represent only 6.5% of the total US population. In these cases, estimates are driven by demographic and geographic covariates rather than any endogenous random effects that might be detected if respondents from the area were included in the baseline survey data set, and are thus likely to exhibit less variance than their true values. Additional survey data from low-density areas would probably improve estimates in these areas, and future research should be directed towards validating and refining MRP estimates in low-population areas. Future research should also investigate how time interacts with different geographic subunits, which may improve the model estimates. However, our validation results indicate that the estimates are highly accurate measures of contemporary public opin