1. Introduction
Climate change is frequently framed as a binary issue despite a wide range of opinions and policy support. Americans are more ideologically divided than at any point in recent history (Pew Research Center 2014), and research on the climate discourse has produced a variety of explanations to understand its polarization. Among them, the strongest predictor of climate opinion is political affiliation, which influences the salience of the issue and policy support across the political divide (Dunlap and McCright 2011; Merkley and Stecula 2018; Marquart-Pyatt et al. 2011). Exposure to partisan media coverage (Carmichael et al. 2017; Feldman et al. 2012, 2017), communication of scientific findings (Kahan 2015; Hart and Nisbet 2012), and environmental attitudes (Fielding and Hornsey 2016; Smith et al. 2017) also influence climate change perceptions and beliefs, to varying degrees, but political ideology remains the “elephant in the room” (Hamilton et al. 2016) when it comes to understanding public opinions on this controversial issue.
Despite the partisanship over the climate issue, there are wide-ranging opinions that are generally shared by large segments of the American public (Leiserowitz et al. 2009). For example, Maibach et al. (2011) show that a majority of Americans are either alarmed or concerned about global warming, which means they remain convinced that it is a problem requiring some type of action. In the same study, an estimated 33% of Americans are identified as cautious, disengaged, or doubtful about the issue, indicating different levels of understanding and acceptance of it. Only a small portion of Americans (i.e., 10%) are dismissive of global warming entirely. Even across the political divide, most Democrats and most Republicans believe in anthropogenic climate change (Van Boven et al. 2018). Public opinion polls show that in the most recent 2016 general election, nearly half of Trump voters believed that global warming was happening and a majority of them supported different climate change mitigation policies (Leiserowitz et al. 2017). In cases where partisan identity is strong, targeted communication strategies can make climate change beliefs more malleable (Joslyn and Demnitz 2019). Together, this body of research suggests that many of the barriers to bipartisan public support for climate policy are arbitrary, exaggerated, and psychological in nature (Van Boven et al. 2018).
A major challenge for climate opinion research and policy making is understanding the geographic distribution and magnitude of support for different climate change beliefs. This is because the local- to state-level opinion may differ from the national-level opinion (Howe et al. 2015). Many regions throughout the country exhibit a different psychological, social, cultural, or demographic character (Rentfrow et al. 2008; Leiserowitz and Akerlof 2010). Political decision-making by representatives also may not reflect the public opinion of their constituencies (Stover 2017). In addition, the uneven geographic distribution of extreme weather events and the vulnerability or resilience of residents in different regions may further modify or mask climate change perceptions and beliefs at different scales (Lyons et al. 2018; Marquart-Pyatt et al. 2014; Howe and Leiserowitz 2013). Given this geographic variability, it is particularly relevant to identify commonalities and differences in climate opinion and where they exist geographically.
To better understand geographic patterns of climate change beliefs, studies have examined regional (Hamilton and Keim 2009) to global (Capstick et al. 2015; Lee et al. 2015) climate opinions at different scales. Few of these have utilized geovisual analytics that take advantage of computational and visual environments to explore the multidimensional, often nuanced, character of American climate opinions. In addition, national climate opinion estimates have recently been made available through multilevel regression and poststratification modeling (Howe et al. 2015; Mildenberger et al. 2017). This technique has provided independently validated climate opinions down to the county level, which offers an opportunity for fine-scaled spatial analyses of climate beliefs across the American landscape.
Studies have used various clustering techniques to quantify the different levels of agreement among climate change opinions, including hierarchical cluster analysis (Barnes and Toma 2012; Jones and Song 2014), classification tree models (Lee et al. 2015), and audience segmentation analysis (Maibach et al. 2011; Rolfe-Redding et al. 2011; Detenber et al. 2016). A few of them conducted general opinion and sentiment analysis research using self-organizing maps (SOM) as a spatiotemporal clustering method (Sharma and Dey 2013; Neme et al. 2011; Janetzko et al. 2013), though none of these explicitly focused on climate change opinions. There remains some concern that these and similar methods may contribute to increased local polarization, do little to promote changing values on climate change, and can reflect a particular researcher’s methodological preference (Hine et al. 2014). However, the SOM ability to reduce the complexity of multidimensional data and visualize the most and least common patterns in a geographic context is beneficial to understand the diverse understandings and responses to climate change and how to more effectively communicate their threats. It thus answers the call by Marquart-Pyatt et al. (2011) for research to inform a more nuanced understanding of climate change opinion.
This study uses county-level climate opinion estimates in the exploratory geovisualization environment of a SOM in order to address three primary questions: what is the spatial variability of different types of climate beliefs; which types of climate beliefs are most and least common; and for regions that share similarities or exhibit differences in the character of their beliefs, which climate opinions typically garner the most and least support? The SOM approach, which is suited to data mining, pattern recognition, and knowledge discovery (Koua 2003; Hagenauer and Helbich 2013; Koua et al. 2006; Koua and Kraak 2004), provides a way to identify distinct regions of climate beliefs that share many similarities in their general attitudes and policy preferences on climate change, yet may differ in their specific support for individual climate opinions.
2. Data and methods
a. Data
This study uses county-level climate opinion estimates provided by the Yale Program on Climate Change Communication (Howe et al. 2015) in order to identify regions that exhibit similarities and differences in their climate beliefs. There are a total of 56 opinion variables, which capture the estimated percentage of the population in support of a range of particular statements on climate change awareness, risk perceptions, and policy support (Table 1). Specifically, statements gauge different variations of the following questions: Is global warming happening and why? Whom will it harm? Who should do something about it? Who is regularly exposed to global warming issues? What do we do about it? The estimates are generated from a statistical model that uses multilevel regression to predict different climate opinions as a function of demographic and geographic variables, as well as poststratification, which weights the model coefficients by support for particular climate opinions in each area. The model was developed using data from a large national survey (>22 000 respondents) from 2008 to 2018 and was then cross validated and compared with other independent survey results. At the county level, it is accurate to approximately ±8 percentage points. Howe et al. (2015) provide more details on the survey and model implementation. In this paper, climate change opinion variables that exhibit high percentages in terms of their level of awareness, risk perception, and policy support are referred to as “proclimate opinions,” whereas climate change opinion variables that exhibit low percentages of awareness, risk perception, and policy support are referred to as “opposing-climate opinions.”
Climate opinion variable abbreviations and their descriptions. (Source: Yale Program on Climate Change Communication 2015.)
b. Implementation of the self-organizing map
The SOM (Kohonen 1995) is used as a data mining technique for pattern identification and knowledge discovery (Vesanto and Alhoniemi 2000; Agarwal and Skupin 2008) on the geography of climate change beliefs. It is capable of simultaneously handling both spatial and temporal characteristics of geographic data. It operates as an unsupervised artificial neural network and does not require advanced assumptions about the observations. This characteristic elevates it above conventional classification or clustering methods, which use manual or automated approaches, are more labor intensive, and require some knowledge of the data distribution in order to create clusters, groupings, or classes from the data (Andrienko et al. 2010).
The largest benefit of the SOM for climate opinion research, and geographic research in general (Agarwal and Skupin 2008), is its potential for data generalization and abstraction (Sester 2005), where it can be used to visualize high-dimensional and complex patterns in a simple two-dimensional component plane (Kohonen 1995). The plane maintains a topological grid of neurons, or nodes, that represent the entire data space. Counties with similar types of climate change beliefs are located closer together in the plane and likewise, counties with different types of beliefs are farther apart. The SOM eliminates the need to conduct the analysis across individual maps of all counties in the United States for each of the 56 climate opinion variables, thereby reducing the cognitive load during the task of pattern identification across the attribute space (Crampton 2002).
The first step in the implementation of the SOM is to identify the desired number of nodes, as well as the topology of the two-dimensional grid. In this study, an 8 × 22 hexagonal grid is used to represent the distribution of climate change opinions. This topological structure produces a vertically oriented rectangular shape that has some advantages. First, there are 176 individual nodes within the SOM, which forces, among the total counties, those with the most similar climate change beliefs to cluster together across the grid. Second, it is easier to identify extreme differences in climate change beliefs across the rectangular grid shape because of the differing county membership to nodes on opposite ends of the grid. A 13 × 13 square-shaped grid with 169 hexagonal nodes was visually compared to test this assumption (not shown), and there are many examples that discuss the influence of SOM size and shape on the resulting outcomes of the patterns (Gibson et al. 2016; Vesanto 1999; Lin et al. 1999).
The iterative training process of the SOM begins through random initialization, where a random observation is presented to the map and used to approximate the first guess, or arrangement, of nodes across the grid (Reusch et al. 2005). Each output node is assigned a reference coefficient vector and the dimension of the coefficients is equal to the dimension of the input nodes. The training process is also sequential, as each input coefficient vector from the observations is presented to the various reference coefficient vectors in each node. The winning node is chosen based on the minimum Euclidean distance between the reference vector and the input vector. The winning node, or best-matched unit, and its neighbors are then updated to reduce the difference between the reference and input vectors. During this training process, the map learns and self-organizes as the patterns in each node increasingly represent those in the data space. Training continues until there are no more changes in the county membership to each node. For a thorough description of the various training methods of the SOM, the reader is referred to Kohonen (1995).
In this study, the SOM is implemented using the Kohonen package in the R Language and Environment for Statistical Computing (R Core Team 2017; Wehrens and Buydens 2007; Wehrens and Kruisselbrink 2018) using the following parameters. The climate opinion estimates were standardized to equally weight all variables during the training process and then presented to the SOM grid 100 times to produce the final distribution of the nodes. The learning rate declined linearly from 0.05 to 0.01 over the course of the iterations. A 67% training radius was used across the grid to define the area of nodes that are updated after each iteration, which is equivalent to about two-thirds of the nodes in the array. The sum of squares is used as the distance decay function to shrink this area through each iteration until the winning node is reached.
c. SOM component planes, clustering, and mapping of climate change opinions
After training the SOM, component planes are used to iteratively determine the location of all climate change opinions that exhibit the most and the least support across the nodes. Because climate change opinions generally exhibit positive relationships with each other (Chryst et al. 2018), it is expected that counties with similar magnitudes of support will cluster together within the same or similarly located nodes. The component planes will thus readily show geographic outliers of opinion and whether any nonlinear relationships between variables exist at all (Delmelle et al. 2013). The nodes will also vary in their total county membership, so the component planes provide a way to assess how much of the country shares a similar opinion. A counts plot is used to summarize the county membership across the grid as well as the SOM training results (Fig. 1).
Summary of the SOM training process, including (a) standardized sample input data that are used to train the SOM, (b) the location of these sample data points distributed across the SOM network of nodes after training, and (c) a SOM counts plot that provides the total county membership to each node.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Summary of the SOM training process, including (a) standardized sample input data that are used to train the SOM, (b) the location of these sample data points distributed across the SOM network of nodes after training, and (c) a SOM counts plot that provides the total county membership to each node.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Summary of the SOM training process, including (a) standardized sample input data that are used to train the SOM, (b) the location of these sample data points distributed across the SOM network of nodes after training, and (c) a SOM counts plot that provides the total county membership to each node.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
A statistical cluster analysis is used to group the most similar nodes together across the SOM and facilitate the interpretation of the patterns. Specifically, an agglomerative hierarchical clustering procedure using Ward’s minimum variance within classes method (Murtagh and Legendre 2014) compares the multidimensional weighting of climate change opinions in the output space from node to node. The hierarchical clustering procedure was chosen because it eliminates the need to specify the desired number of clusters in advance, as is the case with a k-means procedure (Kaufman and Rousseeuw 1990). The clustering process, instead, proceeds in bottom-up fashion with individual observations merging to form groups based on their similarity. The resulting dendrogram can be used for the exploration of the observations to find the most meaningful clusters.
An important final step in the visualization of geographic patterns in a SOM is to classify and map the original geographies according to their respective cluster membership in the node output space. A county-level choropleth map provides a visualization of the results, and an interactive web map is provided in the online supplemental material for a closer view of the regional patterns. Because some clusters will likely exhibit categorical differences in the character of their climate beliefs, a qualitative color palette is used to do the shading (Harrower and Brewer 2003). A value-by-alpha technique is applied to the color scheme to equalize the influence of population in each county (Roth et al. 2010; Gao et al. 2019). Regardless of their area size, the counties with low populations are not as visually important and fade into the background. The counties with high populations, on the other hand, are highly visible since they may provide a substantial contribution to general climate belief in a particular cluster. Specifically, definitions for core-based statistical areas from the U.S. Census Bureau (Ratcliffe et al. 2016) were used as a population classification to adjust the alpha channel of each county, representing rural (<10 000), micropolitan (10 000–50 000), and metropolitan (>50 000) variations. Differences in the support for specific climate opinions across the clusters are explored graphically in the context of the choropleth maps.
3. Results
a. Comparison of climate change beliefs
Component planes are frequently used to visualize the results from a SOM because they provide a cross-sectional view of the spread of values that are associated with each variable (Delmelle et al. 2013; Vesanto 1999). In Figs. 2 and 3, the component planes show the relative support for the 56 specific climate opinions across the SOM output space. The greatest support for proclimate opinions is located near the upper part of the plane while the least support for proclimate opinions is located in the lower part. This indicates that in general, the vertical dimension of the SOM in this study can be interpreted as the ordinal ranking, from low to high, of support for nearly all climate opinions. For example, the estimated percentage who think that global warming is happening (happening), that it is caused by human activities (human), and that there is a scientific consensus (consensus) about it increases toward the top of the plane. The support for these opinions also decreases toward the bottom of the plane, to varying degrees. Only two climate opinions exhibited the opposite pattern, including the percentage who support drilling for oil in the Arctic National Wildlife Refuge (drillANWR) and who support expanding offshore drilling for oil and natural gas off the U.S. coast (drilloffshore). This result is to be expected since support for these two variables generally aligns with lower awareness, risk perception, and policy support for climate change overall. The component planes show that the vast majority of proclimate opinions exhibit positive relationships. This relationship is also present among the opposing-climate opinions in Fig. 3, though they generally exhibit an inverse of the spatial pattern in Fig. 2, with high support for opposing opinions in the lower part of the plane and low support for them across the top.
Component planes showing the distribution and weighting of values for all 28 proclimate opinion variables across the SOM network of nodes.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Component planes showing the distribution and weighting of values for all 28 proclimate opinion variables across the SOM network of nodes.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Component planes showing the distribution and weighting of values for all 28 proclimate opinion variables across the SOM network of nodes.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Component planes showing the distribution and weighting of values for all 28 opposing-climate opinion variables across the SOM network of nodes.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Component planes showing the distribution and weighting of values for all 28 opposing-climate opinion variables across the SOM network of nodes.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Component planes showing the distribution and weighting of values for all 28 opposing-climate opinion variables across the SOM network of nodes.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
The spatial patterns in the vertical dimension of the SOM are useful for understanding the common attributes and ranking of broad climate beliefs in U.S. counties. However, some partial correlations are also present in the horizontal dimension across the SOM. For some climate opinions, increased or decreased support occurs further left or right in the plane, indicating that the attributes, or character of belief, is different in each part of the plane. For example, there is a high percentage who think that global warming will harm them personally (personal), and that the governor and local officials (governor, local officials) should be doing more to address it in the upper-right part of the plane, yet lower support for these opinions in the upper-left part. In contrast, there is instead a high percentage who think that global warming will harm future generations (futuregen), rather than themselves personally, in the upper-left part of the plane. In this region, a higher percentage also hear about it in the media and discuss (media, discuss) it more frequently.
The lower part of the SOM also exhibits variation in the magnitude of support for different climate opinions although the broadscale support for them is lower. For example, the lower-right part of the plane exhibits the lowest percentages of support for nearly all proclimate opinions, with the exception of drillANWR and drilloffshore, which garner widespread support. The lower-right part of the plane also contains the highest percentages of support for opposing-climate opinions. Of the counties in this region, there are majorities of adults who do not think that global warming is happening and attribute it to natural changes in the environment. These majorities are also not very worried about it (worried). The lower-left part of the plane also exhibits low broadscale support for proclimate opinions, but on the other hand, there is less opposition to some of them. For example, there are relatively higher percentages who think that global warming is affecting the weather in the United States (affectweather), who support funding research into renewable energy sources (fundrenewables), and who support setting stricter limits on existing coal-fire power plants (CO2limits). There are also higher percentages who hear about global warming at least weekly in the media (mediaweekly) and who support requiring fossil fuel companies to pay a carbon tax (reducetax) when compared to the lower-right part of the plane. Overall, a heterogeneity of climate change awareness, risk perceptions, and policy preferences is present across the SOM.
Because there are 176 nodes across the SOM with varying degrees of county membership, it becomes an onerous task to decode and evaluate individual differences in climate opinion from node to node. The cluster analysis used in this study, instead takes advantage of the SOM spatial–topological organization and groups together the nodes with the most similar multidimensional attributes of climate opinion. It thereby increases the interpretability of the patterns of climate opinion at the aggregate level across the country for all nodes. Six groups of climate belief are identified in this study.
The dendrogram in Fig. 4 provides a way to assess the clustering performance and evaluate the number of groups. The height of the fusion of each observation in the dendrogram reflects the degree of dissimilarity between them as distinct groups. A very high fusion point near 150 would result in two groups with a marked difference of opinion (i.e., for climate support and opposing climate support). A fusion point of 0 would result in the creation of 176 groups with each node belonging to its own cluster. While each group exhibits its own character, the dendrogram shows that groups 1, 2, and 3 are markedly different from groups 4, 5, and 6. Furthermore, six groups were chosen because they revealed subtle differences in the magnitude of support for different opinions across all parts of the SOM, including variability between regions with low and high climate support as well as the formation of a group with moderate climate support.
Dendrogram view of the 127 leaves representing individual SOM nodes and their agglomerative hierarchical clustering membership using Ward’s minimum variance technique. Branches show the joining points of nodes or groups of nodes based on their similarity.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Dendrogram view of the 127 leaves representing individual SOM nodes and their agglomerative hierarchical clustering membership using Ward’s minimum variance technique. Branches show the joining points of nodes or groups of nodes based on their similarity.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Dendrogram view of the 127 leaves representing individual SOM nodes and their agglomerative hierarchical clustering membership using Ward’s minimum variance technique. Branches show the joining points of nodes or groups of nodes based on their similarity.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
The six groups are mapped onto the SOM output space in Fig. 5. All clusters maintained a relatively compact shape. Node cluster memberships were spatially contiguous, which facilitates the interpretation of each group’s climate belief based on the patterns that are present in each corresponding region of the component planes. The results are summarized in Table 2, which provides a qualitative description of the character of climate change belief between the groups.
Clustering of the SOM nodes into six groups of climate change belief.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Clustering of the SOM nodes into six groups of climate change belief.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Clustering of the SOM nodes into six groups of climate change belief.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Qualitative descriptions of the character of belief for six groups identified through cluster analysis. Variable definitions are provided in Table 1.
Group 1, which is located in the lower-right part of the SOM, exhibits the lowest awareness, risk perception, and support for climate policies when compared to any other group. In fact, it contains the greatest percentages who hold opposing viewpoints. Group 2 is located in the lower-left part of the SOM, with low levels of awareness, risk perception, and climate policy support overall. However, it is located slightly higher in the vertical dimension of the SOM than group 1 because there is less opposition to some of the climate opinions. Despite the lowered resistance to some of these proclimate opinions in group 2, many of the other proclimate variables still exhibit low support in common with group 1.
Whereas groups 1 and 2 exhibit the lowest support and the greatest opposition in terms of awareness, risk perception, and policy preference, group 3 exhibits the most moderate levels of support in its climate belief. Its nodes occupy the entire middle portion of the SOM and do not show much variation in the horizontal dimension. Group 3 is also the largest group, containing the majority of nodes across the SOM. Support for proclimate opinions within this group transitions from lower to higher percentages relative to the mean when moving from the lower toward the upper part of the SOM.
The remaining groups 4, 5, and 6 are smaller and include nodes that occupy the entire upper part of the SOM, with the greatest awareness, risk perception, and policy support. Together, they comprise close to the same number of nodes in groups 1 and 2 combined, or in group 3 alone. This distribution indicates that there is more heterogeneity in the character of climate change beliefs for groups that exhibit the highest percentages of awareness, risk perception, and policy support overall. Group 4, for instance, is located in the upper-left part of the SOM, whereas group 5 is located in the upper-right part, so there are some relative differences in climate change belief between them. Group 6 nodes are clustered in the uppermost center part of the SOM. As a result, it exhibits high percentages of support for nearly all proclimate opinion variables and low percentages of support for nearly all opposing-climate opinion variables.
b. Geographic variability of climate change beliefs
The six groups of climate change beliefs and their respective population totals are mapped across the continental United States in Fig. 6. A comma-separated-values (CSV) file of the individual county-level results, including population classes and cluster membership, is included in the online supplemental material. Although the cluster analysis produced spatially contiguous and compact regions across the SOM, the distribution of the clusters by their county membership is more dispersed. However, there are still some notable geographic concentrations in contiguous counties and some differences between the groups throughout regions of the country, overall. Counties in groups 1 and 2, where climate change awareness, risk perceptions, and policy support are the lowest, are distributed throughout interior portions and noncoastal areas of the continental United States. Group 1 includes the second lowest population (Fig. 7) and is located in parts of the American South and south-central regions, with large concentrations in Alabama, Kentucky, Oklahoma, Tennessee, and Texas. There is also a concentration across the Intermountain West in Nevada, Utah, and Wyoming. Specific county examples with membership to this group include Johnson County, Tennessee and Mohave County, Arizona. Group 2 has the smallest population and its counties are more dispersed, with some of them clustered farther north of group 1 in the Rust Belt states, extending from Pennsylvania to Illinois. Another portion of counties in group 2 extend west into Kansas and Nebraska, but also include large sections of Idaho, Montana, Oregon, and Wyoming. Although these groups occur in many rural areas across the country, the counties are primarily composed of micropolitan-sized populations (Fig. 7). For example, this group includes Mesa County, Colorado, Flathead County, Montana, and Scioto County, Ohio.
Value-by-alpha mapping of the distribution of climate change beliefs on the basis of county cluster membership and population.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Value-by-alpha mapping of the distribution of climate change beliefs on the basis of county cluster membership and population.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Value-by-alpha mapping of the distribution of climate change beliefs on the basis of county cluster membership and population.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Population class distribution in each cluster. The color shading is as in Fig. 6.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Population class distribution in each cluster. The color shading is as in Fig. 6.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Population class distribution in each cluster. The color shading is as in Fig. 6.
Citation: Weather, Climate, and Society 13, 1; 10.1175/WCAS-D-20-0071.1
Group 3 contains the third highest population, and it is geographically the most diverse group. For example, metropolitan-sized counties in this group are located in central Florida, western New York and Pennsylvania, southern Arizona, and Michigan and Wisconsin. There are concentrations of micropolitan counties in this group across the Carolinas and upper-midwestern regions. Many of the rural counties in this group extend across the southeastern United States into western Texas and southern New Mexico as well as the Dakotas region into Montana. There are very few counties belonging to group 3 along the Pacific coast states. Overall, notable counties in group 3 include Marion County, Florida, Kent County, Michigan, and Carlton County, Minnesota.
Groups 4, 5, and 6, which overall have the greatest climate change awareness, risk perceptions, and policy support are more coastal and less rural than the other groups. For example, group 4, which contains the largest population, is primarily concentrated along the Pacific coast states and it includes the second largest proportion of metropolitan-sized counties, like Sacramento County, California, and Clark County, Nevada. There are also some counties in group 4 that are concentrated in the northeastern states ranging from New York to Maine, including Worcester County, Massachusetts, and Penobscot County, Maine. In contrast, group 5, which has the third lowest population, is distributed from coastal North Carolina southwestward through Georgia, Alabama, and Mississippi in a region known as the southern Black Belt. It includes Jefferson County, Alabama, and Shelby County, Tennessee, in addition to many other counties with predominate Black or African American racial and ethnic groups. Areas of southwestern Texas along the U.S.-Mexico border, which include a large proportion of Hispanic residents, are part of group 5, and it also includes Dallas and Harris Counties in Texas. Finally, group 6 contains the counties with the largest populated cities in the country, including New York, Los Angeles, and Chicago, as well as densely populated counties in California and other urban centers around the country. Group 6 has the second largest population.
4. Discussion
a. Discussion of research findings
In an effort to understand why Americans disagree about climate change, research has focused on many causal explanations, including political polarization, disinformation campaigns, partisan media coverage, communication of scientific findings, and environmental attitudes. This body of research has done much to show that climate change opinions are largely framed by prevailing cultural, political, or social identities (Unsworth and Fielding 2014; Fielding and Hornsey 2016) and that tailored communication efforts can reinforce opposing viewpoints when the goal is to shift public opinion (Hart and Nisbet 2012). Climate change invokes different meanings, values, and risk perceptions for people across different social and cultural contexts (Hulme 2009), so a continued exploration of these complex opinions and where they exist is vital to understanding the nature of disagreement. This study provides some confirmation of the known differences of climate change opinions and assesses their broadscale level of disagreement for different groups across the country. It also highlights many of the large commonalities of opinion between the different groups of belief.
A majority of the proclimate opinions in this study exhibited positive relationships with each other in terms of their magnitude of support in each county. Likewise, many of the opposing-climate opinions exhibited negative relationships in terms of their support. These relationships are evident in the SOM vertical distribution of variables across the component planes. The idea that a few important opinion variables are predictive of an entity’s larger stance on climate change in terms of awareness, risk perception, and policy support, is well supported in different areas of the research literature. Chryst et al. (2018) showed that four screener questions can be used to segment individuals in an audience according to climate change belief. Evidence in sociological research suggests that ideological preferences about climate change can be predicted through county-level voting records (Hamilton et al. 2016). Similarly, counties with membership to nodes in opposing regions of the SOM are also indicative of these broadscale disagreements.
A major finding of this study is that despite the prevailing binary divide over climate change that has been framed in numerous ways (e.g., for or against, alarmist or denier, Democrat or Republican), there is a heterogeneity of climate belief across the country. The cluster analysis partitioned groups of counties across the SOM where the strength of the relationships between climate change opinion variables was weaker. These partial correlations occurred between different groups that were otherwise located in the same regions of the SOM, described above. In regions that exhibit broadscale opposition in their awareness, risk perception, and policy preference, there were subtle differences in the magnitude of support for specific opinions. For example, counties in group 2 are characterized by higher percentages who are less opposed to regulating emissions or funding renewable energy research. Much research has traditionally focused on understanding the political, psychological, and social context of climate denial that is found in regions that exhibit beliefs similar to groups 1 and 2 in order to generate tailored messages about the threats and risks of climate change (McCright and Dunlap 2011; Joslyn and Demnitz 2019; Rolfe-Redding et al. 2011). Although many of these rural and micropolitan-sized counties may vote along party lines in a way that reinforces this traditional framing of climate change, some of them hold opinions that are more in line with moderate climate beliefs. Mildenberger et al. (2017) likewise explain that Democrats consistently exhibit higher average support for climate policy reforms, but there are many areas of the country where Republicans support them as well. Similar to other findings in the literature (Maibach et al. 2011), here, a large proportion of the total population demonstrate more moderate beliefs in their climate change awareness, risk perception, and policy preferences.
In regions that are typically associated with high awareness, risk perceptions, and policy support for climate change, there was also a heterogeneity in the types of climate belief. Whereas majorities in group 4 are discussing global warming, hearing about it in the media frequently, and expressing concern for its effects on future generations, they do not think as strongly that it will harm them personally and do not believe as strongly that governors or local officials should be doing anything about it relative to the other groups. Group 5 exhibited some of the opposite supporting beliefs relative to group 4. Notably, majorities in group 5 do think that global warming will harm them personally and that governors and local officials should do more to address it. There are lower percentages who think that global warming is human caused and that there is scientific consensus about it, as well as lower percentages who think it will harm future generations and developing countries. Climate change threats appear to be interpreted with a greater sense of awareness and lower risk perception in group 4 yet a lower sense of awareness and greater risk perception in group 5. In terms of their policy preferences, majorities in group 4 counties may place a larger emphasis on federal regulation by congress or private corporations, whereas in group 5, they may prefer regulation to be handled at state or local levels.
There are several potential explanations that may be tied to the geographic distribution and difference of beliefs for counties in groups 4, 5, and 6. Climate change is frequently viewed as a psychologically distant threat by inland residents (Milfont et al. 2014; Retchless 2018), which suggests that those in coastal areas would exhibit a higher risk perception to its threats. However, group 4 counties are primarily distributed throughout coastal states, and there were lower percentages who thought that global warming would harm them personally. These counties may exhibit a psychological distance that is tied to other dimensions beyond geographic distance, including temporal or social distance (Brügger et al. 2015; Spence et al. 2012). The impacts of climate change are also likely to be felt in areas of socioeconomic inequality or with disparities between racial and ethnic groups (Leiserowitz and Akerlof 2010). Group 5, which is located throughout many rural parts of the southeastern United States, may exhibit some of these differences that are tied to its character of climate change belief compared to group 4. In 2011, Maibach et al. found that a large majority (70%) of Americans were to varying degrees concerned about global warming and supportive of policy responses. Since then, the number of Americans that are alarmed or concerned about global warming has increased to make up the largest segment of the population (Goldberg et al. 2020), so it is important to consider the heterogeneity that exists across these groups. Group 6, which exhibited strong support for all climate opinions and very little variation among them, has likely increased in size as a result of these changes.
b. Limitations
There are several limitations that should be taken into account when interpreting the results. First, this study identified the character of belief based on the combined magnitude of support for different opinions in each county of the United States. However, the reader is cautioned from making direct interpretations at any smaller scale of analysis. While much of the research literature has focused on psychological differences of belief at the individual level, the climate beliefs estimated for each county should not be attributed to any individual resident within that county; rather, they represent the aggregate beliefs of the adult population of that county. Furthermore, the county-level scale of analysis yielded the aforementioned patterns, but the same methodology applied to the congressional district or state level of analysis would likely show some different and interesting results.
Second, the groups of climate belief that emerged from the clustering procedure and the distribution of the counties making up the groups across the SOM similarly represent the combination of percentages in each unit who hold certain opinions, as reflected in the original model estimates. Together, they provide a characterization of the combined magnitude of support for all climate opinions across the population within the dataset based on their location and distribution. The SOM is used to organize and visualize the multivariate distribution of the climate opinions across a two-dimensional surface and the clustering procedure is implemented to produce distinct cluster boundaries (Hagenauer and Helbich 2013). These factors must be weighed in the interpretation since both the SOM and the cluster algorithms ultimately force an observation to join a particular node and cluster, and any changes to parameters will likely result in different versions of a similar pattern (Budayan et al. 2009).
c. Significance and implications for future research
During the decade of study (2008–18), the rhetoric around climate change arguably shifted to the extremes (Hulme 2019). As Hulme (2020) notes, there were numerous interventions that shaped the new public discourse, including the U.S. official withdrawal from the Paris Agreement, the Green New Deal, and the release of the IPCC Special Report on 1.5°. The time period also marked the approaching 10-yr anniversary of the 2009 Climategate scandal, which at the time damaged public climate opinion and scientific trust (Leiserowitz et al. 2013). Each of these events occurred as scientific consensus on climate change was increasingly quantified to promote policy making (Pearce et al. 2017) and the partisan presentation of these events in various media outlets only served to exacerbate the political paralysis over the issue (Boykoff 2013). In the research literature, the lack of climate policy progress in the polarized environment has largely been attributed to organized denial campaigns, which promoted climate skepticism and misinformation to the public (Kahan 2012; Dunlap and McCright 2011; Dunlap 2013). While this research focus has produced some understandings of the psychology of climate denial as well targeted communication strategies for different groups, there are still many unaddressed questions that could examine how populations that disagree in the polarized environment might actually lend support for different policies or technology approaches to the climate issue (Nisbet 2009; McCright et al. 2016).
Understanding the geographic distribution and character of American climate change opinions is important because it sets a baseline of climate belief in the context of the recent period. A snapshot of the current combined climate opinion in aggregate for the whole country is useful as research moves to examine any longitudinal changes in the character of belief over time. Recent events will likely influence opinion, as well. There are already efforts to identify relationships between COVID-19 and climate (Bashir et al. 2020; Şahin 2020; Tosepu et al. 2020; Hepburn et al. 2020) and emergency decision-making in the context of the pandemic is increasingly being invoked as a model for climate change response (Hulme 2019). Furthermore, policy makers, planners, educators, and scientists are also working at different spatial scales, so this study takes advantage of the continued research efforts by (Howe et al. 2015) to address the challenges of understanding climate change beliefs on a local basis.
The SOM is frequently used as an exploratory data visualization method to uncover patterns in complex multidimensional data and for hypotheses generation about relationships that may exist. Instead of iteratively examining the individual opinions from the climate opinion estimates across the counties, the SOM environment provides a way to combine them into groups. This geovisualization method is useful because it quickly shows the amount of variability in opinions for a large group and identifies heterogeneous beliefs in areas that are commonly framed in simpler terms. The groups of climate belief identified in this study may be used to examine relationships between climate opinion and political decision-making, extreme weather events, and the cultural, demographic, psychological, or social characteristics of the regions where different types of belief exist.
5. Conclusions
This study used the exploratory geovisualization environment of a self-organizing map (SOM) to examine the character and distribution of American climate change beliefs. County-level climate opinion estimates, which represent the percentages in each county who support a range of statements on climate change, were used to train the SOM neural network. Agglomerative hierarchical cluster analysis was used to further classify the regions of similar climate opinions across the SOM network of nodes into groups of climate belief across the country. This study finds that among the six groups of climate belief, there is a substantial heterogeneity of the support for different climate opinions occurring in different geographic regions of the country. Two of the groups exhibited absolute support or opposition to the range of climate opinions. Yet, despite the overall framing or characterization of climate change as a binary issue that falls along partisan lines, this study supports recent research that examines the commonalities and differences of opinion among a range of groups. Research efforts on climate change, public opinion, and perception should thus address how climate change may engage the geographically diverse public across the range of these shared values and how they connect in new and different ways with policy making.
Acknowledgments
The author thanks W. B. Capell, who assisted with figure formatting, as well as the reviewers of this paper, who considerably improved its content and mapping components. The author declares no conflict of interest.
Data availability statement
Data used in this study were provided by the Yale Program on Climate Change Communication. The Yale Program on Climate Change Communication bears no responsibility for the analyses or interpretations of the data presented here.
REFERENCES
Agarwal, P., and A. Skupin, 2008: Self-Organizing Maps: Applications in Geographic Information Science. John Wiley and Sons, 214 pp.
Andrienko, G., N. Andrienko, S. Bremm, T. Schreck, T. Von Landesberger, P. Bak, and D. Keim, 2010: Space-in-time and time-in-space self-organizing maps for exploring spatiotemporal patterns. Comput. Graph. Forum, 29, 913–922, https://doi.org/10.1111/j.1467-8659.2009.01664.x.
Barnes, A. P., and L. Toma, 2012: A typology of dairy farmer perceptions towards climate change. Climatic Change, 112, 507–522, https://doi.org/10.1007/s10584-011-0226-2.
Bashir, M. F., B. M. Bilal, B. Komal, M. A. Bashir, D. Tan, and M. Bashir, 2020: Correlation between climate indicators and COVID-19 pandemic in New York, USA. Sci. Total Environ., 728, 138835, https://doi.org/10.1016/j.scitotenv.2020.138835.
Boykoff, M. T., 2013: Public Enemy No. 1?: Understanding media representations of outlier views on climate change. Amer. Behav. Sci., 57, 796–817, https://doi.org/10.1177/0002764213476846.
Brügger, A., S. Dessai, P. Devine-Wright, T. A. Morton, and N. F. Pidgeon, 2015: Psychological responses to the proximity of climate change. Nat. Climate Change, 5, 1031–1037, https://doi.org/10.1038/nclimate2760.
Budayan, C., I. Dikmen, and M. T. Birgonul, 2009: Comparing the performance of traditional cluster analysis, self-organizing maps and fuzzy C-means method for strategic grouping. Expert Syst. Appl., 36, 11 772–11 781, https://doi.org/10.1016/j.eswa.2009.04.022.
Capstick, S., L. Whitmarsh, W. Poortinga, N. Pidgeon, and P. Upham, 2015: International trends in public perceptions of climate change over the past quarter century. Wiley Interdiscip. Rev.: Climate Change, 6, 35–61, https://doi.org/10.1002/wcc.321.
Carmichael, J. T., R. J. Brulle, and J. K. Huxster, 2017: The great divide: Understanding the role of media and other drivers of the partisan divide in public concern over climate change in the USA, 2001–2014. Climatic Change, 141, 599–612, https://doi.org/10.1007/s10584-017-1908-1.
Chryst, B., J. Marlon, S. van der Linden, A. Leiserowitz, E. Maibach, and C. Roser-Renouf, 2018: Global warming’s “Six Americas Short Survey”: Audience segmentation of climate change views using a four question instrument. Environ. Commun., 12, 1109–1122, https://doi.org/10.1080/17524032.2018.1508047.
Crampton, J. W., 2002: Interactivity types in geographic visualization. Cartogr. Geogr. Inf. Sci., 29, 85–98, https://doi.org/10.1559/152304002782053314.
Delmelle, E. C., J. C. Thill, O. Furuseth, and T. Ludden, 2013: Trajectories of multidimensional neighbourhood quality of life change. Urban Stud., 50, 923–941, https://doi.org/10.1177/0042098012458003.
Detenber, B. H., S. Rosenthal, Y. Liao, and S. S. Ho, 2016: Audience segmentation for campaign design: Addressing climate change in Singapore. Int. J. Commun., 10, 4736–4758.
Dunlap, R. E., 2013: Climate change skepticism and denial: An introduction. Amer. Behav. Sci., 57, 691–698, https://doi.org/10.1177/0002764213477097.
Dunlap, R. E., and A. M. McCright, 2011: Organized climate change denial. The Oxford Handbook of Climate Change and Society, J. S. Dryzek, R. B. Norgaard, and D. Schlosberg, Eds., Oxford University Press, 144–160.
Feldman, L., E. W. Maibach, C. Roser-Renouf, and A. Leiserowitz, 2012: Climate on cable: The nature and impact of global warming coverage on Fox News, CNN, and MSNBC. Int. J. Press., 17, 3–31, https://doi.org/10.1177/1940161211425410.
Feldman, L., P. S. Hart, and T. Milosevic, 2017: Polarizing news? Representations of threat and efficacy in leading US newspapers’ coverage of climate change. Public Understanding Sci., 26, 481–497, https://doi.org/10.1177/0963662515595348.
Fielding, K. S., and M. J. Hornsey, 2016: A social identity analysis of climate change and environmental attitudes and behaviors: Insights and opportunities. Front. Psychol., 7, 1–12, https://doi.org/10.3389/fpsyg.2016.00121.
Gao, P., Z. Li, and Z. Qin, 2019: Usability of value-by-alpha maps compared to area cartograms and proportional symbol maps. J. Spat. Sci., 64, 239–255, https://doi.org/10.1080/14498596.2018.1440649.
Gibson, P. B., P. Uotila, S. E. Perkins-Kirkpatrick, L. V. Alexander, and A. J. Pitman, 2016: Evaluating synoptic systems in the CMIP5 climate models over the Australian region. Climate Dyn., 47, 2235–2251, https://doi.org/10.1007/s00382-015-2961-y.
Goldberg, M., A. Gustafson, S. Rosenthal, J. Kotcher, E. Maibach, and A. Leiserowitz, 2020: For the first time, the alarmed are now the largest of global warming’s six Americas. Yale Program on Climate Change Communication, https://climatecommunication.yale.edu/publications/for-the-first-time-the-alarmed-are-now-the-largest-of-global-warmings-six-americas/.
Hagenauer, J., and M. Helbich, 2013: Hierarchical self-organizing maps for clustering spatiotemporal data. Int. J. Geogr. Inf. Sci., 27, 2026–2042, https://doi.org/10.1080/13658816.2013.788249.
Hamilton, L. C., and B. Keim, 2009: Regional variation in perceptions about climate change. Int. J. Climatol., 29, 2348–2352, https://doi.org/10.1002/joc.1930.
Hamilton, L. C., C. P. Wake, J. Hartter, T. G. Safford, and A. J. Puchlopek, 2016: Flood realities, perceptions and the depth of divisions on climate. Sociology, 50, 913–933, https://doi.org/10.1177/0038038516648547.
Harrower, M., and C. A. Brewer, 2003: ColorBrewer.org: An online tool for selecting colour schemes for maps. Cartogr. J., 40, 27–37, https://doi.org/10.1179/000870403235002042.
Hart, P. S., and E. C. Nisbet, 2012: Boomerang effects in science communication: How motivated reasoning and identity cues amplify opinion polarization about climate mitigation policies. Commun. Res., 39, 701–723, https://doi.org/10.1177/0093650211416646.
Hepburn, C., B. O’Callaghan, N. Stern, J. Stiglitz, and D. Zenghelis, 2020: Will COVID-19 fiscal recovery packages accelerate or retard progress on climate change? Smith School Working Paper 20-02, 48 pp., https://www.smithschool.ox.ac.uk/publications/wpapers/workingpaper20-02.pdf.
Hine, D. W., J. P. Reser, M. Morrison, W. J. Phillips, P. Nunn, and R. Cooksey, 2014: Audience segmentation and climate change communication: Conceptual and methodological considerations. Wiley Interdiscip. Rev.: Climate Change, 5, 441–459, https://doi.org/10.1002/wcc.279.
Howe, P. D., and A. Leiserowitz, 2013: Who remembers a hot summer or a cold winter? The asymmetric effect of beliefs about global warming on perceptions of local climate conditions in the U.S. Global Environ. Change, 23, 1488–1500, https://doi.org/10.1016/j.gloenvcha.2013.09.014.
Howe, P. D., M. Mildenberger, J. R. Marlon, and A. Leiserowitz, 2015: Geographic variation in opinions on climate change at state and local scales in the USA. Nat. Climate Change, 5, 596–603, https://doi.org/10.1038/nclimate2583.
Hulme, M., 2009: Why We Disagree about Climate Change: Understanding Controversy, Inaction and Opportunity. Cambridge University Press, 432 pp.
Hulme, M., 2019: Climate emergency politics is dangerous. Issues Sci. Technol., 36, 23–25, https://doi.org/10.17863/CAM.46233.
Hulme, M., 2020: Is it too late (to stop dangerous climate change)? An editorial. Wiley Interdiscip. Rev.: Climate Change, 11, e619, https://doi.org/10.1002/wcc.619.
Janetzko, H., D. Jäckle, and T. Schreck, 2013: Comparative visual analysis of large customer feedback based on self-organizing sentiment maps. Proc. Third Int. Conf. on Advances in Information Mining and Management, Lisbon, Portugal, IARIA, 12–17.
Jones, M. D., and G. Song, 2014: Making sense of climate change: How story frames shape cognition. Polit. Psychol., 35, 447–476, https://doi.org/10.1111/pops.12057.
Joslyn, S., and R. Demnitz, 2019: Communicating climate change: Probabilistic expressions and concrete events. Wea. Climate Soc., 11, 651–664, https://doi.org/10.1175/WCAS-D-18-0126.1.
Kahan, D., 2012: Why we are poles apart on climate change. Nature, 488, 255, https://doi.org/10.1038/488255a.
Kahan, D., 2015: Climate-science communication and the measurement problem. Polit. Psychol., 36, 1–43, https://doi.org/10.1111/pops.12244.
Kaufman, L., and P. J. Rousseeuw, 1990: Finding Groups in Data: An Introduction to Cluster Analysis. John Wiley and Sons, 342 pp.
Kohonen, T., 1995: Self-Organizing Maps. Springer, 392 pp.
Koua, E. L., 2003: Using self-organizing maps for information visualization and knowledge discovery in complex geospatial datasets. Proc. 21st Int. Cartographic Conf., Durban, South Africa, International Cartographic Association, 1694–1702.
Koua, E. L., and M.-J. Kraak, 2004: Alternative visualization of large geospatial datasets. Cartogr. J., 41, 217–228, https://doi.org/10.1179/000870404X13283.
Koua, E. L., A. Maceachren, and M.-J. Kraak, 2006: Evaluating the usability of visualization methods in an exploratory geovisualization environment. Int. J. Geogr. Inf. Sci., 20, 425–448, https://doi.org/10.1080/13658810600607550.
Lee, T. M., E. M. Markowitz, P. D. Howe, C. Y. Ko, and A. A. Leiserowitz, 2015: Predictors of public climate change awareness and risk perception around the world. Nat. Climate Change, 5, 1014–1020, https://doi.org/10.1038/nclimate2728.
Leiserowitz, A., and K. Akerlof, 2010: Race, ethnicity and public responses to climate change. Yale Project on Climate Change and the George Mason University Center for Climate Change Communication Rep., 70 pp., https://climatecommunication.yale.edu/publications/race-ethnicity-and-public-responses-to-climate-change/.
Leiserowitz, A., E. Maibach, and C. Roser-Renouf, 2009: Global warming’s six Americas: An audience segmentation analysis. Yale Project on Climate Change and the George Mason University Center for Climate Change Communication, https://climatecommunication.yale.edu/about/projects/global-warmings-six-americas/.
Leiserowitz, A., E. Maibach, C. Roser-Renouf, N. Smith, and E. Dawson, 2013: Climategate, public opinion, and the loss of trust. Amer. Behav. Sci., 57, 818–837, https://doi.org/10.1177/0002764212458272.
Leiserowitz, A., E. Maibach, C. Roser-Renouf, M. Cutle, and S. Rosenthal, 2017: Trump voters & global warming. Yale Project on Climate Change and the George Mason University Center for Climate Change Communication Rep., 16 pp., http://climatecommunication.yale.edu/publications/trump-voters-global-warming/.
Lin, C., H. Chen, and J. F. Nunamaker, 1999: Verifying the proximity and size hypothesis for self-organizing maps. J. Manage. Inf. Syst., 16, 57–70, https://doi.org/10.1080/07421222.1999.11518256.
Lyons, B. A., A. Hasell, and N. J. Stroud, 2018: Enduring extremes? Polar vortex, drought, and climate change beliefs. Environ. Commun., 12, 876–894, https://doi.org/10.1080/17524032.2018.1520735.
Maibach, E. W., A. Leiserowitz, C. Roser-Renouf, and C. K. Mertz, 2011: Identifying like-minded audiences for global warming public engagement campaigns: An audience segmentation analysis and tool development. PLOS ONE, 6, e17571, https://doi.org/10.1371/journal.pone.0017571.
Marquart-Pyatt, S. T., R. L. Shwom, T. Dietez, R. E. Dunlap, S. A. Kaplowitz, A. M. McCright, and S. Zahran, 2011: Understanding public opinion on climate change: A call for research. Environ. Sci. Policy Sustainable Dev., 53, 38–42, https://doi.org/10.1080/00139157.2011.588555.
Marquart-Pyatt, S. T., A. M. McCright, T. Dietz, and R. E. Dunlap, 2014: Politics eclipses climate extremes for climate change perceptions. Global Environ. Change, 29, 246–257, https://doi.org/10.1016/j.gloenvcha.2014.10.004.
McCright, A. M., and R. E. Dunlap, 2011: Cool dudes: The denial of climate change among conservative white males in the United States. Global Environ. Change, 21, 1163–1172, https://doi.org/10.1016/j.gloenvcha.2011.06.003.
McCright, A. M., M. Charters, K. Dentzman, and T. Dietz, 2016: Examining the effectiveness of climate change frames in the face of a climate change denial counter-frame. Top. Cognit. Sci., 8, 76–97, https://doi.org/10.1111/tops.12171.
Merkley, E., and D. A. Stecula, 2018: Party elites or manufactured doubt? The informational context of climate change polarization. Sci. Commun., 40, 258–274, https://doi.org/10.1177/1075547018760334.
Mildenberger, M., J. R. Marlon, P. D. Howe, and A. Leiserowitz, 2017: The spatial distribution of Republican and Democratic climate opinions at state and local scales. Climatic Change, 145, 539–548, https://doi.org/10.1007/s10584-017-2103-0.
Milfont, T. L., L. Evans, C. G. Sibley, J. Ries, and A. Cunningham, 2014: Proximity to coast is linked to climate change belief. PLOS ONE, 9, e103180, https://doi.org/10.1371/journal.pone.0103180.
Murtagh, F., and P. Legendre, 2014: Ward’s hierarchical agglomerative clustering method: Which algorithms implement Ward’s criterion? J. Classif., 31, 274–295, https://doi.org/10.1007/s00357-014-9161-z.
Neme, A., S. Hernández, and O. Neme, 2011: An electoral preferences model based on self-organizing maps. J. Comput. Sci., 2, 345–352, https://doi.org/10.1016/j.jocs.2011.08.003.
Nisbet, M., 2009: Communicating climate change: Why frames matter for public engagement. Environment, 51, 12–23, https://doi.org/10.3200/ENVT.51.2.12-23.
Pearce, W., R. Grundmann, M. Hulme, S. Raman, E. Hadley Kershaw, and J. Tsouvalis, 2017: Beyond counting climate consensus. Environ. Commun., 11, 723–730, https://doi.org/10.1080/17524032.2017.1333965.
Pew Research Center, 2014: Political polarization in the American public. https://www.pewresearch.org/politics/2014/06/12/political-polarization-in-the-american-public/.
Ratcliffe, M., C. Burd, K. Holder, and A. Fields, 2016: Defining rural at the U.S. Census Bureau: American Community Survey and Geography Brief. U.S. Census Bureau Doc., 8 pp., https://www2.census.gov/geo/pdfs/reference/ua/Defining_Rural.pdf.
R Core Team, 2017: R: A language and environment for statistical computing. R Foundation for Statistical Computing, https://www.R-project.org/.
Rentfrow, P. J., S. D. Gosling, and J. Potter, 2008: A theory of the emergence, persistence, and expression of geographic variation in psychological characteristics. Perspect. Psychol. Sci., 3, 339–369, https://doi.org/10.1111/j.1745-6924.2008.00084.x.
Retchless, D. P., 2018: Understanding local sea level rise risk perceptions and the power of maps to change them: The effects of distance and doubt. Environ. Behav., 50, 483–511, https://doi.org/10.1177/0013916517709043.
Reusch, D. B., R. B. Alley, and B. C. Hewitson, 2005: Relative performance of self-organizing maps and principal component analysis in pattern extraction from synthetic climatological data. Polar Geogr., 29, 188–212, https://doi.org/10.1080/789610199.
Rolfe-Redding, J. C., E. W. Maibach, L. Feldman, and A. A. Leiserowitz, 2011: Republicans and climate change: An audience analysis of predictors for belief and policy preferences. SSRN, 48 pp., https://doi.org/10.2139/ssrn.2026002.
Roth, R. E., A. W. Woodruff, and Z. F. Johnson, 2010: Value-by-alpha maps: An alternative technique to the cartogram. Cartogr. J., 47, 130–140, https://doi.org/10.1179/000870409X12488753453372.
Şahin, M., 2020: Impact of weather on COVID-19 pandemic in Turkey. Sci. Total Environ., 728, 138810, https://doi.org/10.1016/j.scitotenv.2020.138810.
Sester, M., 2005: Optimization approaches for generalization and data abstraction. Int. J. Geogr. Inf. Sci., 19, 871–897, https://doi.org/10.1080/13658810500161179.
Sharma, A., and S. Dey, 2013: Using self-organizing maps for sentiment analysis. arXiv 1309.3946, 13 pp., https://arxiv.org/abs/1309.3946.
Smith, T. W., J. Kim, and J. Son, 2017: Public attitudes toward climate change and other environmental issues across countries. Int. J. Sociol., 47, 62–80, https://doi.org/10.1080/00207659.2017.1264837.
Spence, A., W. Poortinga, and N. Pidgeon, 2012: The psychological distance of climate change. Risk Anal., 32, 957–972, https://doi.org/10.1111/j.1539-6924.2011.01695.x.
Stover, D., 2017: Bob Inglis: A conservative for climate action. Bull. At. Sci., 73, 364–367, https://doi.org/10.1080/00963402.2017.1388661.
Tosepu, R., J. Gunawan, D. S. Effendy, L. O. A. I. Ahmad, H. Lestari, H. Bahar, and P. Asfian, 2020: Correlation between weather and Covid-19 pandemic in Jakarta, Indonesia. Sci. Total Environ., 725, 138436, https://doi.org/10.1016/j.scitotenv.2020.138436.
Unsworth, K. L., and K. S. Fielding, 2014: It’s political: How the salience of one’s political identity changes climate change beliefs and policy support. Global Environ. Change, 27, 131–137, https://doi.org/10.1016/j.gloenvcha.2014.05.002.
Van Boven, L., P. J. Ehret, and D. K. Sherman, 2018: Psychological barriers to bipartisan public support for climate policy. Perspect. Psychol. Sci., 13, 492–507, https://doi.org/10.1177/1745691617748966.
Vesanto, J., 1999: SOM-based data visualization methods. Intell. Data Anal., 3, 111–126, https://doi.org/10.3233/IDA-1999-3203.
Vesanto, J., and E. Alhoniemi, 2000: Clustering of the self-organizing map. IEEE Trans. Neural Networks, 11, 586–600, https://doi.org/10.1109/72.846731.
Wehrens, R., and L. M. C. Buydens, 2007: Self- and super-organizing maps in R: The Kohonen package. J. Stat. Software, 21 (5), https://doi.org/10.18637/jss.v021.i05.
Wehrens, R., and J. Kruisselbrink, 2018: Flexible self-organizing maps in Kohonen 3.0. J. Stat. Software, 87 (7), https://doi.org/10.18637/jss.v087.i07.