To assess accurate geographic differences in public CCOs, we used detailed information on households’ places of residence to the level of the 4,667 municipalities that represent a very fine-grained administrative level in Germany. Uncovering public opinions at a small spatial scale is methodologically challenging. One possibility is to disaggregate the data (i.e., simply calculating regional averages based on households’ locations). Disaggregation suffers from absent or imprecise estimates in low-population areas. Another approach is to use multilevel regression with poststratification (MRP)1,62. Besides the variable of interest, MRP requires the joint distribution of sociodemographic predictors at the same geographic scale63. In many research contexts, these information is not available at very small spatial scales. We therefore applied a spatial smoothing function with Actor-Based Clustering64 as an alternative approach.
In short, this approach uses the smallest scale of spatial information available in the data (here, which of the 4667 municipalities a household resides in) to depict geographic patterns without imposing any predetermined higher-level spatial boundaries. Specifically, the CCO scores for each municipality i were calculated based on CCO scores of households residing in municipality i as well as households in all municipalities j (j ≠ i) by considering spatial weights such that geographically proximate households receive higher weights than distant ones. For geographically weighting responses, we calculated the distance between all 4667 × 4667 German municipalities and transformed the geographic distance into spatial weights using a log-logistic distance-decay function following previous research64,65,66,67 (see Methods for details).
To quantify the accuracy of actor-based clustering and the applied spatial smoothing function, we followed previous work1,62 and cross-validated public CCOs as estimated with actor-based clustering by simulating small sample sizes based on subsets from large sample sizes in more populated regions. Cross-validation demonstrated that actor-based clustering produces highly accurate measures with mean absolute errors ranging between 0.38–1.35% (belief), 0.57–3.79% (concern), and 2.25–4.27% (importance) (see Supplementary Note 3 and Supplementary Figs. 1 and 2). The range of errors is comparable to methods applied in earlier research, such as MRP1. We excluded 96 municipalities for which the data was too sparse (<1% of total population in Germany) from the geographic mapping, reducing the number of municipalities to 4,571 (see Supplementary Note 4 and Supplementary Figs. 1 and 2).
We applied the spatial smoothing function to estimate public CCOs across the final sample of German municipalities (n = 4571) as depicted in Fig. 1. The figure shows the share of the local population (and the difference from the national average in percent) who believe that climate change has already begun (panels A and D), who are concerned about its consequences (panels B and E), and who perceive collective action against it as important (panels C and F). All maps show substantial and relatively stable geographic clustering across all CCO dimensions within Germany with regional shares varying between 71–89% (belief), 44–70% (concern) and 71–94% (importance). These observed magnitudes in public CCOs are comparable to the local variation reported in the US1. The combination of all three CCO dimensions (panel G), that is the average of panels D-F for every municipality, accounts for 83% of the observed geographic variance in public CCOs as indicated by principal component analysis, substantiating the robustness of the results across the dimensions.
The shares of local populations that (A) believe climate change has already begun, (B) are concerned, and (C) perceive collective responses to be important. The percentage difference from the national average is visualized in panel (D) for awareness, (E) concern, (F) importance. The average difference from the national average across the three dimensions reported in (D–F) is depicted in (G). The bold black line indicates the former division into East and West Germany. Solid black lines indicate the 16 federal states. Solid grey lines indicate the 96 planning regions used for the multilevel estimations. Municipalities with too little information are colored in grey (see Supplementary Note 5).
We validated our findings obtained from spatial smoothing by applying MRP at the level of 96 German planning regions (see Supplementary Note 5). Planning regions (“Raumordnungsregionen”) are functional units capturing socioeconomic interactions (e.g., core-periphery commuting) that cross administrative boundaries. The correlation r = 0.7, CI = [0.58; 0.79], and p < 0.001 of regional estimates obtained from MRP and spatial smoothing indicates that both methods produce acceptably similar results. We also validated our findings concerning the four geographic features using MRP. The revealed geographic patterns of public CCOs are therefore not an artefact of the spatial smoothing function but validated by using MRP as a fundamentally different methodology.
Differences between urban and rural areas
To first investigate whether differences between urban and rural areas exist, we used a predefined categorization of municipalities into five settlement types as provided by the Federal Institute for Research on Building, Urban Affairs, and Spatial Development (BBSR):68 (1) large cities (n = 79, 1.8% of total population), (2) medium-sized cities (n = 621, 14.1%), (3) towns (n = 868, 19.7%), (4) small towns (n = 1211, 27.5%), and (5) rural municipalities (n = 1624, 36.9%). This categorization is based on municipalities’ population sizes, their functions, and settlement structures. 168 municipalities (0.01% of total population) were not classified. We calculated the percentage difference from the national average and corresponding 95% confidence intervals for each settlement type across the individual CCO dimensions. Results are depicted in Fig. 2 and indicate significant differences between more urban areas and more rural ones. The overall difference in public CCOs (panel D) between large cities (type 1) and rural municipalities (type 5) is ~2%. The results of one-way analysis of variance (ANOVA) indicate that group differences across the five settlement types are significant for all CCO dimensions (belief: F = 10.4, p < 0.001; concern: F = 33.4, p < 0.001; importance: F = 26.6, p < 0.001; overall: F = 28.8, p < 0.001).
Local shares of people who believe climate change has already begun, are concerned, and perceive collective responses to be important were calculated at the municipality level as percentage difference from the national mean. Error bars represent 95% confidence intervals of SEM.
Differences between prospering vs. declining areas
Our second hypothesis concerns geographic dispersion of public CCOs between prospering vs. declining areas. We again relied on a classification provided by the BBSR68 that groups municipalities into five categories: (1) strongly prospering (n = 867, 19.8% of total population), (2) prospering (n = 1484, 33.9%), (3) stable (n = 605, 13.8%), (4) declining (n = 876, 20%), and (5) strongly declining (n = 542, 12.4%). This categorization is derived from six indicators (population development, net migration, development of workforce aged 20–64, workplace development, unemployment rate, development of commercial tax income) and therefore provides a comprehensive distinction between prospering and declining areas. 197 municipalities (0.01% of the total population) were not classified. The results, depicted in Fig. 2, indicate a significant difference between prospering and declining municipalities in public CCOs. While the prospering municipalities all show positive differences from the national average, the percentage difference is negative for declining areas. The overall differences between the most prospering and most declining regions is 6.4% and hence exceeds the observed urban-rural divide by a factor of 3. As indicated by a one-way ANOVA test, the group differences are significant for every CCO dimension (belief: F = 405.8, p < 0.001; concern: F = 371.7, p < 0.001; importance: F = 326.1, p < 0.001; overall: F = 477.1, p < 0.001).
Local green cultures
We examined the relationship between pre-existing differences in local green political cultures and current CCOs to address hypothesis three. To approximate green political cultures, we relied on local variation in votes for the Green Party during the general election in 1994. The 1994 federal election was the first election after the Greens from the West merged with the civil rights party Alliance 90 from the East and ran as an all-German party (Alliance 90/The Greens) with a clear focus on environmental topics53. Although climate change became an established topic for the German public during the late 1980s and early 1990s43, the Green Party’s focus on global warming and the fight against climate change did not play a meaningful role in national elections before 200269,70. Hence, local variation in green votes in 1994 indicates to what extent local populations favored green policies as an expression of general environmental concern roughly twenty years before the household survey was conducted. Figure 3 displays the local share of green votes in 1994 (panel A), with a national average of 5.8% and local shares ranging between 0.8% and 22%. To test group differences, we divided the sample into “green” (above-average vote shares) and “non-green” municipalities (below-average vote shares). As depicted in Fig. 2, green municipalities have distinct CCOs than nongreen municipalities. The results of two sample t-tests indicate that group differences are significant for all CCO dimensions (belief: t = 31.8, p < 0.001; concern: t = 37.1, p < 0.001; importance: t = 27.2, p < 0.001; overall: t = 37.3, p < 0.001).
Local green cultures and their spatial relationship with public CCOs as depicted by local green vote shares in 1994 (A), the local pairing of above/below average green vote shares and above/below average CCO shares (B). The color scheme of panel B is illustrated in panel C.
To provide a visual illustration of the relationship between local green vote shares and overall CCOs (panel G of Fig. 1), we created a map (panel B of Fig. 3) highlighting the local co-occurrence of either positive or negative differences from the national average in both variables (70% of all municipalities). In this map, we highlighted the cities Freiburg and Tübingen in Germany’s Federal State Baden-Wuerttemberg (red border in Fig. 3), which show above-average values in local green vote shares and local CCOs. In these cities, local policymakers have long-established green development strategies71. Baden-Wuerttemberg is the first federal state in German history to have a Minister-President from the Green Party. The local variation in public CCOs today seems to be associated with the manifestation of local green political cultures in the past. More generally, the correlation between local shares of green votes and CCOs are constant across all CCO dimensions (belief: r = 0.5, p < 0.001, concern: r = 0.47, p < 0.001, importance: r = 0.4, p < 0.001). Between 16% and 25% of the observed local variation in public CCOs at the municipality level is explained by the presence of local green political cultures.
Differences between East and West Germany
In our fourth hypothesis, we expect distinct CCOs between East and West Germany. Over 20 years after German reunification, the geographic variation in public CCOs (see maps in Fig. 1) shows a striking East-West divide. The polarization between East and West is prevalent in each CCO dimension, although to varying degrees. Municipalities in West Germany show decisively higher levels of belief (West: 82.9%, CI = [82.9–83.0%] vs. East: 78.0%, CI = [77.8–78.3%]), concern (West: 58.3%, CI = [58.2–58.4%] vs. East: 51.2%, CI = [50.9–51.5%]), and perceived importance of collective responses (West: 85.3%, CI = [85.3–85.4%] vs. East: 79.8%, CI = [79.5–80.0%]). Group differences are significant as indicated by the results of two sample t-tests (belief: t = 41.6, p < 0.001; concern: t = −41.5, p < 0.001; importance: t = −43.0, p < 0.001; overall: t = −47.7, p < 0.001).
The systematic polarization of public CCOs between municipalities in East (N = 927, 20.3% of all municipalities) and West (N = 3,644, 79.7%) becomes even more evident by investigating the distribution of local CCOs in more detail. In Fig. 4, municipalities are ranked according to their deviation from the national average across the three CCO dimensions of belief (panel A), concern (panel B), and importance (panel C). We added a random ordering of municipalities (panel D) to highlight the systematic difference in CCOs between East and West. Of the 2,950 municipalities with an above-average belief in climate change (positive difference from the national average as shown in panel A of Fig. 4), 116 are located in the East representing 12.5% of all Eastern municipalities. Regarding concern (panel B) and importance (panel C), the share of Eastern regions with above-average shares is 17% and 5%, respectively. Specifically, in 95% of the municipalities located in East Germany, local populations place less importance on collective responses to climate change than the national average. Dividing the country once again into East and West explains 53% of the observed variance in public CCOs. Hence, public polarization in climate opinions manifests to a large degree between East and West Germany.
Ranked deviations from the national average in public CCOs across municipalities in East and West Germany for (A) belief, (B) concern, and (C) importance. To ease interpretation, we added a random ordering of municipalities (D).
As outlined above, we investigated how contextual factors (urban vs. rural, prospering vs. declining, green vs. non-green areas, East vs. West) relate to aggregate opinion at the local scale. Individual-level research, however, has shown that individuals’ sociodemographic features, such as their age or education, predict their CCOs20. To finally test our four hypotheses, we thus estimated multilevel regression models with households nested in their locations to consider individual-level alongside context-level factors. As the level of municipalities is too fine-grained to estimate multilevel regressions, we nested households into German planning regions (N = 96). This regional scale guarantees sufficiently large sample sizes per region (5 < n < 431) (see Methods for details). To model households’ opinions across the three CCO dimensions, we created a variable equaling 1 if a household believes in climate change, its members are concerned, and they perceive collective responses as important, and 0 otherwise. Regression results are robust for each CCO dimension separately and (also for) using the original ordinal response scales (see Supplementary Note 6 for additional regression results). Based on previous research4,20,21, we included a rich set of household-level variables (gender, age, education, income, political affiliation, environmental concern) representing the most important predictors for individual opinions on climate change. At the regional level, we created four variables each representing one of the four contextual factors (urban vs. rural, prospering vs. declining, East vs. West, green vs. non-green) outlined above (see Methods for a detailed theoretical reasoning of all variables, their definition, and data sources).
Table 1 reports the results. The goodness of fit (R2) ranges between .058 (marginal) and .066 (conditional). Consistent with previous individual-level research4,20,21, the individual-level predictors are associated with individual CCOs as expected. Females (exp(β) = 1.22, p < 0.001), people with higher education (exp(β) = 1.13, p = 0.03), those affiliated to the Green Party (exp(β) = 3.58, p < 0.001) and environmentally concerned (exp(β) = 1.89, p < 0.001) show higher odds of being aware of climate change. Individual income (exp(β) = 1.98, p < 0.001) did not play a meaningful role in our estimations. Political affiliation (~258% higher odds of being aware of climate change) and environmental concern (~89% higher odds) show the strongest relationship with climate change awareness validating previous research20.
Compared with the strongest individual-level predictors, the coefficients of the regional-level predictors range at lower levels but are nevertheless meaningful. The coefficient of urban areas is positive (exp(β) = 1.05) but not statistically significant at the 5% level (p = 0.068). Given the 95% CI of 1.00 to 1.11, it would be hasty to conclude that an urban-rural divide does not exist72. However, the other context factors seem more important as indicated by the regression results. The divide between prospering and declining areas is statistically significant (exp(β) = 1.13, p < 0.001) and meaningful. If respondents live in more prosperous regional contexts, their odds of being aware of climate change increases by ~13%.
To test hypothesis 3, we included the regional green vote share in 1994 to approximate the existence of local green cultures (see Supplementary Note 7 for similar results for six subsequent general elections). The corresponding coefficient (exp(β) = 1.12, p <0 .001) indicates that we expect to see a ~12% increase in respondents’ odds of being aware of climate change, for a one-unit increase in the regional green vote share. In theories on local cultures, it is further hypothesized that established values and norms (e.g. environmental concern) are associated with people’s beliefs and behaviors35,38. These works raise the important question whether people are more likely aware of climate change if they live in local contexts emphasizing “green” values although they do not hold a green attitude personally. This question has not been systematically investigated regarding public CCOs at a local scale. We investigated the hypothesis by including cross-level interaction terms between the individual-level feature that households tend to vote green and the regional-level predictor of green vote shares. The corresponding results show that local green cultures are particularly associated with people’s opinions about climate change if they personally do not hold a green attitude. That is, the CCOs of a large part of the population (those that do not hold a green attitude a priori) depend on their local context.
Lastly, the divide between East and West remains striking in the multilevel setting. Respondents in East Germany (exp(β) = 0.71, p < 0.001) have ~30% lower odds of being aware of climate change than Western respondents. Using the individual-level estimate of respondents’ sex as a benchmark (~22 higher odds of being aware of climate change for females) indicates that especially the East-West divide is a meaningful regional-level predictor for individual CCOs. The results for the prospering vs. declining and green vs. non-green areas remained robust even when controlling for the prevalent East-West divide that confounds many socioeconomic variables, including economic development for example (see Supplementary Note 8 for additional model specifications)73.
We estimated binary logistic multilevel models with households nested in German planning regions. The dependent variable (CCO) equals 1 if household believes in climate change, is concerned, and perceives collective responses as important; and 0 otherwise. 53% were coded as 1 and 47% as 0. Coefficients represent odd ratios. 95% confidence intervals in square brackets. See Methods for a detailed description of all variables included in the model. All continuous variables were z-standardized beforehand. See Supplement for additional regression results for each CCO dimension.