Identifying and validating the strongest predictors of informed energy policy support across Europe

Identifying and validating the strongest predictors of informed energy policy support across Europe | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Identifying and validating the strongest predictors of informed energy policy support across Europe Morris Krainz, Valeria Sorgato, Izïa Vallaeys Mora, Evelina Trutnevyte, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7658135/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted You are reading this latest preprint version Abstract Widespread public policy support is crucial for efforts to decarbonize the energy system. Past research has proposed numerous variables as potential key predictors of support for mitigation policies but has failed to i) comprehensively evaluate and compare their predictive importance and ii) identify the strongest predictors within specific mitigation domains. We used machine learning models to identify the strongest predictors of energy policy support in informed citizens (Study 1), validate these results in a real-world referendum on renewable energy in Switzerland (Study 2) and test their generalizability to other climate mitigation policies across six European countries (Study 3). We identified affective responses, societal and environmental policy-impact beliefs, fairness perceptions and perceived trends in policy support over time as the strongest predictors of energy policy support. Using these predictors, we achieved high accuracy in predicting support for a real-world referendum as well as support for different climate mitigation measures across Europe. Social science/Psychology Scientific community and society/Social sciences/Communication Physical sciences/Energy science and technology/Renewable energy Scientific community and society/Energy and society/Energy policy Figures Figure 1 Introduction Various technically feasible and cost-comparable options are available to decarbonize the energy sector 1 . In democratic societies, the most appropriate approach is to implement the options preferred by the informed general public 2 . This is done for three reasons 3 : First, understanding which policies the public would support when aware of the benefits and drawbacks of the different options should help speed up the energy transition – which is supported by the strong observed association between public policy support and successful policy adoption 4 . Second, it is a normative expectation in democratic societies to implement policies that the public wants. Third, by combining insights about technically possible and publicly preferred options, we better understand which solutions are actually feasible 5 . To understand and explain policy preferences, previous research tested a large number of variables as predictors of policy support 6–9 . These variables can be organized into three categories: (1) Sociodemographic factors , including age 6 , gender 10 , education 6 and income 6 , (2) individual differences , including beliefs about climate change 6,9,11,12 , personal values 6 , trust in political institutions 6,13 and political affiliation 6 , and (3) policy-specific beliefs , including perceived policy effectiveness 6 , perceived environmental, societal and economic effects 14 , perceived fairness 6,15 , personal impacts 16,17 , perceived social consensus 18–20 , and subjective as well as objective knowledge of the policy 6 . While this has greatly contributed to our understanding of support for climate mitigation policies, several questions remain unanswered. First and foremost, so far, there is no comprehensive evaluation of the comparative predictive strength of this vast range of variables. Empirical evidence identifying the most important drivers of support would increase our understanding of policy preferences and help policymakers and communicators design targeted policies that better align with public concerns and preferences. This is especially important given that previous research often used broad and aggregated measures of support for “climate policies” in general by combining mitigation measures from different domains (e.g., measuring aggregated support for carbon taxes, renewable energy, infrastructure and public transport, consumer and dietary behaviour, etc.) 9,15,21,22 . While this approach measures overall support for all types of climate mitigation policies, it does not capture support for specific climate mitigation policies. Furthermore, the approach lacks ecological validity as the public is usually asked to vote on individual policies, rather than on a bundle of unconnected policy options. It is thus possible that the importance of some of the predictors identified with the aggregation approach is under- or overestimated 22 . Empirical evidence on the most important predictors of policy support in a specific domain can help align policy content with public interests and may accelerate their successful implementation. Previous acceptance research, moreover, tended to overly focus on cognitive factors 6 . Recently, researchers called for a stronger integration of the affective sciences into climate mitigation research 8,23,24 , highlighting the important role of factors such as general affect 9,25–28 and discrete emotions such as anger, worry, hope, and pride 21,29–35 in predicting policy support and climate-relevant actions and intentions. Yet, a comprehensive evaluation comparing the predictive strength of cognitive and affective factors is missing. The role of discrete emotions in policy support has so far mainly been investigated with a focus on emotions towards climate change, producing mixed results 21,32–35 . Almost no research exists on the role of emotions towards specific policies, which is however likely to result in more reliable associations 29–31,36,37 . Furthermore, several newly developed constructs have been associated with pro-environmental behaviour or policy support in other domains. General attitudes towards technology have been associated with pro-environmental behavioural intentions 38 . Perceived financial scarcity was associated with willingness to switch to more flexible energy consumption 39 . Dynamic social norms were positively associated with pro-environmental behaviour 40 and policy support for an equal pay policy 41 . These constructs may thus be additional predictors of climate mitigation policy support but remain untested in this context. Finally, while informed decision-making plays a crucial role in democracies 2 , most policy support research has not provided participants with sufficient information about the policies to help them make informed decisions on whether or not to support a policy (but see e.g. 42,43 ). Uninformed participants may provide unreliable insights into public preferences and policy support that are not aligned with their values and concerns. Hence, understanding which predictors drive support in informed participants 5,44,45 – those who are aware of the policy’s advantages and disadvantages – would provide robust and actionable insights into policy support by bridging the gap between how citizens think about a policy without being informed (i.e. the descriptive reality) and how they think about a policy when they are aware of the advantages and disadvantages (i.e. the normative ideal). Transparently communicating policy impacts may help to avoid decisions that may otherwise be taken based on misinformation 47 such as the British Brexit vote, where citizens seemed to either not be sufficiently informed or even misinformed about the consequences of the United Kingdom leaving the European Union 48–50 . Designing and implementing policies in accordance with informed public preferences is especially important in the energy domain, where a large number of potential options allow reaching a net-zero energy system 1,46 , and where taking public preferences into account could substantially facilitate policy implementation. To address these issues, here, we run three preregistered studies (total N = 5176) in seven different countries. Using machine-learning, we first test a comprehensive set of 50 predictors of support for different energy policies to identify the strongest predictors of energy policy support in informed citizens in Switzerland. We then validate these results by predicting the outcome of a real-world referendum on renewable energy in Switzerland and illustrate the generalizability of the identified predictors on support for other climate mitigation measures across six European countries. The work presented here highlights the value of using interpretable predictive machine-learning models to compare and identify the strongest predictors of climate mitigation support in different domains. Identifying the strongest predictors of energy policy support In Study 1, we randomly assigned Swiss participants (N = 1056) to receive transparent information (i.e. information aiming to inform rather than persuade participants; 5,44,45 ) about one of four policies that aim to increase the flexibility of the Swiss energy system and are being discussed in the context of the Swiss energy transition: a policy on the construction and use of Direct Air Capture and Carbon Storage (DACCS) facilities to allow the use of fossil fuel back up electricity plants which are not large enough for CCS (referred to as DACCS Policy from here on), a policy increasing electricity imports for flexibility (Electricity Trading Policy), a policy on national incentives for the construction of flexible EV charging infrastructure (Flexible EV Charging Policy), or a policy focusing on national incentives for the flexible operation of heat pumps (Flexible Heating Policy). Participants received transparent information about the environmental, energy system-related, economic and societal impacts of the respective policy (see supplementary materials S3 for the information provided to the participants). We measured sociodemographic factors and individual-level differences before information provision and policy support , policy-specific beliefs, affect and emotions towards policies after information provision (for a full list of variables see Table 6 in supplementary materials S2). We used regression random forests 51–54 to identify the strongest predictors of support across the four energy policies. This method is more suitable for exploratory data analysis than conventional statistical analyses such as multiple linear regression 52 (see methods section for details on hyperparameter settings and model specification). The model ranking the 50 predictors performed with a root mean squared error (RMSE) of 0.78 and explained 71.08 % of the variance in policy support (R² = 0.7108) in an independent test-set including 25 % of the participants. Mean RMSE loss after permutations was used as an indicator of variable importance. Twelve variables achieved a higher mean RMSE-loss based importance score than the categorical predictor which indicated the policy participants were assigned to (see Table 1), and were thus interpreted as the strongest domain-specific predictors for energy policy support. Table 1 Model-identified strongest domain-specific predictors of renewable energy policy support in Study 1 Predictor Example item Subgroup Rank (Random Forest / Ridge Regression) General affect “In general, how do you feel about this policy?” Affective responses 1 / 1 Perceived societal impacts “Overall, this policy will be beneficial to society and people.” Policy impact beliefs 2 / 2 Hope “This policy makes me hopeful.” Affective responses 3 / 3 Pride “This policy makes me proud.” Affective responses 4 / 5 Perceived environmental impacts “Overall, this policy will be beneficial for the environment.” Policy impact beliefs 5 / 4 Personal fairness “Overall, this policy is [completely unfair – completely fair] towards me.” Fairness perceptions 6 / 8 Fairness towards Swiss citizens “Overall, this policy is [completely unfair – completely fair] towards Swiss citizens.” Fairness perceptions 7 / 9 Perceived dynamic norms “Do you expect the number of people in Switzerland who would accept this policy rather to increase or decrease over the next 12 months?” Perceived social consensus 8 / 6 Perceived impact on personal wellbeing “Overall, this policy would have a [very negative – very positive] effect on my comfort.” Policy impact beliefs 9 / - Worry “This policy makes me worry.” Affective responses 10 / 7 Perceived impact on personal finances “Do you think that your household would win or lose financially from this policy?” Policy impact beliefs 11 / - Anger “This policy makes me angry.” Affective responses 12 / 11 Fairness towards people in other countries “Overall, this policy is [completely unfair – completely fair] towards people living outside of Switzerland.” Fairness perceptions - / 10 Fairness towards low-income earners “In your view, would the low-income earners win or lose if this policy was implemented in Switzerland?” Fairness perceptions - / 12 Note. All items in this table were rated on fully-labelled 7-point Likert scales. Affective responses, fairness perceptions, as well as perceived impacts on personal wellbeing and finances were measured with single items. Perceived societal impacts and perceived dynamic norms were measured with two items each and perceived environmental impacts were measured with three items. For more details on how predictor variables were measured, see the Methods section. Hyphens indicate that the variable in question was not included in the top 12 predictors of the random forest or ridge regression, respectively. Predictor weights (Permutation Variable Importance for the Random Forest and the mean absolute regression coefficient for ridge regression) are reported in the supplementary materials. An alternative analysis strategy running a ridge regression model was used to corroborate these results (see Methods for details). This model achieved slightly better performance metrics than the regression random forest (mean RMSE = 0.68; mean R² = 0.7812)[1]. When comparing the twelve strongest predictors of the ridge regression model (based on the mean absolute regression coefficient) with the twelve strongest predictors of the regression random forest, the models, by and large, converged on the same predictors of domain-specific policy support. The first three predictor variables were identical (and in the same order), and seven of the remaining nine model predictors overlapped (for more details on predictor ranks in both models, see Table 1). Combining the results of both analyses, we find that the strongest predictors of energy policy support include affective responses (general affect, hope, pride, worry, anger), policy-impact beliefs (regarding societal and environmental impacts, impact on personal wellbeing and finances), fairness perceptions (personal fairness, distributional fairness towards Swiss citizens, people living in other countries and low income earners) and perceived social norms (perceived dynamic norm). To provide a more intuitive interpretation of the model performance, we dichotomized the measure of policy support, separating participants into groups of supporters and non-supporters. We trained a random forest classifier to classify 75% of participants into these groups, and tested the accuracy of this model on the remaining 25% of participants. The random forest classifier showed excellent performance in the independent test-set (accuracy = 92.8%; AUC = 0.9746, true-positive rate = 0.9174; true-negative rate = 0.9355; positive predictive value = 0.9091; negative predictive value = 0.9416). Predictive model performance in a real-world referendum on renewable energy In Study 1, we identified the 14 strongest predictors of support for energy policies using a machine learning approach. We then tested the model’s predictive performance on real-world (rather than hypothetical) data related to the Swiss referendum on the Federal Act on Secure Electricity Supply from Renewable Energy Sources (from here on referred to as “Federal Act on Renewable Energy”), which proposed an infrastructure increase for solar, wind and hydropower and was voted on in a referendum on the 9 th of June 2024 . We collected data from the 30 th of May to the 7 th of June, measuring the 14 model-identified predictors as well as policy support for the Federal Act on Renewable Energy in a Swiss sample (N = 765). To mimic the information treatment in Study 1, we provided participants with a text summarising the content of the referendum as well as the key arguments put forward by proponents and opponents of the referendum. To define cut-offs for the interpretation of model accuracy, we used the training data from Study 1, applying a 20-fold cross validation approach to calculate the lower bound of the 95% Confidence Interval (CI) for the prediction accuracy of our random forest classifier (lower bound 95% CI = 0.882; mean accuracy over 20 folds = 0.895, standard deviation = 0.034). We pre-registered our interpretation of model-performance for Study 2 as follows: If prediction accuracy is at least as high as the lower bound of the 95% CI, the model generalizes to other types of renewable energy policies. If prediction accuracy is lower than the lower bound of the 95% CI but does not drop by more than 10%, the model performs reasonably well in predicting policy support in the domain of renewable energy. If prediction accuracy drops by more than 10 %, the model does not generalize to predict support for other renewable energy polices. The model was subsequently trained on all observations obtained in Study 1 (N = 1056) and tested on all observations obtained in Study 2 (N = 765) to predict dichotomized support for the Federal Act on Renewable Energy. The model achieved a prediction accuracy of 0.871 in predicting policy support (see Figure 1A). In line with our pre-registered model performance interpretation, we assess the models’ performance as reasonably well in generalizing to predict policy support for renewable energy policies. When exploratorily restricting the analysis to the participants who had either already cast their vote (N = 332; accuracy = 0.892) or who reported the intention to participate in the vote (N = 326; accuracy = 0.883), the model performed within the preregistered range, indicating generalizability. In contrast, model performance dropped below the preregistered range for participants who indicated no intention of voting (N = 107; accuracy = 0.766). We additionally used the predictions obtained from our model to provide a forecast for the results of the Swiss general referendum on the Federal Act on Renewable Energy. We predicted (and preregistered) a majority vote in favour of the Federal Act on Renewable Energy with 61.83%, matching national polls conducted about three to four weeks before the referendum 55,56 . The actual referendum resulted in a majority of votes in favour of the Federal Act on Renewable Energy (68.7%). Thus, our model predictions allowed a correct forecast of the decision the Swiss public would take, while also explaining which perceptions and affective responses were key in forming this decision. Model performance for climate mitigation policies across multiple countries In Study 3, we aimed to test to what extent the model generalizes to other types of climate mitigation measures and to what extent it generalizes to predict support in other countries. To this end, we presented participants in France, Germany, Italy, the Netherlands, Poland, and Spain with information about one out of four climate mitigation options frequently modelled in Integrated Assessment Models 1 : Sustainable diets , Bioenergy with Carbon Capture and Storage ( BECCS ), green hydrogen and, afforestation and reforestation . As in Study 2, participants first received transparent information on the respective mitigation measure, consisting of information on the measures’ general characteristics and its mitigation potential by 2050 (see supplementary materials). Subsequently, we collected data on the 14 model-relevant variables as well as on policy support. We then tested the predictive performance of our model in the whole sample (N = 3355) and a total of 34 subsamples: one subsample for each of the four climate mitigation solutions in each of the six countries (i.e. 24 subsamples), as well as overall performance by country (six subsamples) and by mitigation measure (four subsamples, see Figure 1B). Based on our preregistered cut-off criteria, the model performed reasonably well in predicting support for sustainable diets (overall accuracy = 0.88) and BECCS (overall accuracy = 0.87). Prediction accuracy decreased considerably regarding the other two investigated measures, green hydrogen (overall accuracy = 0.73) and afforestation and reforestation (overall accuracy = 0.65). This pattern is explained by a substantial negative correlation between mean reported support and model accuracy ( r = - 0.80, 95% CI [-0.89, 0.63], t (33) = -7.63, p < .001; see also supplementary materials), indicating that the model generally performs better for mitigation measures with lower levels of policy support. Discussion We developed and tested a data-driven domain-specific model of the strongest predictors of energy policy support. We identified affective responses , policy-impact beliefs , fairness perceptions, and perceived trends in policy support over time as the strongest predictors of energy policy support. We then trained a random forest classifier to predict support for a real-world political decision and validated the classifier in a broader European context by testing model performance for four different climate mitigation solutions in six European countries. Our model achieved high accuracy in predicting support for the referendum, allowed correctly forecasting the majority vote, and achieved reasonable accuracy in predicting support for other climate mitigation measures in samples across Europe. These results, gained from informed citizens, can be used to inform policy design and communication strategies as well as to generate and test hypotheses about causal relationships of variables within the model. The findings expand research in decision-science 5 (Wong-Parodi et al., 2016) and transparent evidence communication 44,45 and have direct policy implications. The importance ranking identified in Study 1 indicates that transparent information on environmental and societal impacts may have a stronger influence on policy support than transparent information regarding energy system or economic impacts (see Table 1). The high importance of perceived societal impacts and fairness is in line with research highlighting the important role of climate justice beliefs for policy support 15 . Energy policies should thus be designed taking not only effects on the energy system and the economy into account, but especially effects on the environment and on society. Additionally, our work extends the literature on predictors of climate mitigation policy support by showing that the discrete emotions hope, pride, worry and anger as well as dynamic social norms play an important role in predicting support for energy policies, even when directly comparing them to predictors of policy support highlighted in past research 6 . Finally, this work contributes to methodological developments at the intersection of energy and social sciences, demonstrating both the value and the limitations of predictive machine-learning models in this field. Our approach offers insights into the overall level of public support as well as the concrete reasons for why a policy receives this support (or not), while allowing to forecast future referendum results. However, the model performed considerably worse for mitigation measures with very high levels of overall support ( green hydrogen and afforestation and reforestation ) in the six European countries in Study 3. While it is arguably more important to accurately predict support for measures that are not already highly accepted by the public, it does highlight the limitations of our approach. The decrease in model performance for subsamples in which overall policy support was substantially higher than in our training data underscores the importance of using training data that is maximally appropriate for the prediction context, as this can substantially impact model performance. Additionally, while our approach allowed identifying the most important predictors of renewable energy policy support, it did not allow insights into the causal connections between variables. Some variables that received lower importance scores in our models may still be important antecedents of variables ranked higher in importance. As such, it is not possible to conclude that these variables can be ignored when investigating policy support. Rather, their role as antecedents of more proximally related predictors should be investigated. Overall, we believe that the results presented here illustrate that predictive machine-learning models can be a valuable extension to the methodological toolbox of social scientists. We thus add to the growing literature on interpretable machine-learning 22,52,54,57,58 by showing that predictive models can: i) be extremely valuable tools for exploratory research, ii) achieve high accuracy in forecasting outcomes of interest and iii) be used for data-driven hypothesis generation to inform subsequent research. Methods Study 1 Participants and sample size justification A total of N = 1096 Swiss participants completed the survey. The sample size was based on an a-priori power analysis for a different research question not relevant to the purposes of this paper. Given that i) machine-learning models such as random forests were developed to deal with high dimensional data 51,53 and ii) the predictive performance of the models employed in this study was high (see Results), we deem this sample size as sufficient for our purposes. We excluded 40 participants who were either flagged as duplicates, had a straightlining percentage of more than 80%, indicated that they answered randomly in the open comment section (one participant), failed a honeypot question testing for potential bots or had a survey completion time of less than 10 minutes. To ensure our results do not depend on the exclusion of these participants, analyses were also run with N = 1096 and are reported in supplementary materials S4. Participants were randomly assigned to one of the four experimental conditions. Sample sizes per condition are as follows: Direct Air Capture and Carbon Storage (DACCS) Policy: N = 272 (without quality checks) / 264 (with quality checks); Electricity Trading Policy: N = 273 / 257; Flexible EV Charging Policy: N = 275 / 267; Flexible Heating Policy: N = 276 / 268. Descriptive sample characteristics are reported in supplementary materials S1. Study design and procedure The survey was designed in four blocks. In block one, we checked participants’ commitment and attention and collected sociodemographic data. Participants who either did not commit to paying attention in the survey, failed the attention check, were younger than 18 years old or did not hold Swiss citizenship were screened out here. In block two, we measured variables relating to individual differences in climate change beliefs, trust in political institutions, general attitudes towards technology, political affiliation, perceived financial scarcity, and personal values in randomized order (see below for details). In block three, participants first read a short explanation on the topic of “Flexibility of the electricity system”. Subsequently, they were randomly assigned to either receive information on the DACCS Policy, the Electricity Trading Policy, the Flexible EV Charging Policy, or the Flexible Heating Policy and received a short description of the respective policy. Following this, we measured baseline levels of subjective and objective knowledge on this policy, as well as policy support. Next, participants received detailed information on impacts of the policy on the environment (including CO 2 emissions, health and infrastructure impacts), the energy sector (including renewable generation integration, electricity supply reliability, and energy independence), the economy (including costs, economic inequalities and jobs) and the society (comfort, privacy and distributional impacts). All informational materials were pre-rated for understandability and trustworthiness of the information by a separate sample of German and French speaking Swiss citizens (N total = 65; N German = 32; N French = 33). While all materials were rated as understandable and trustworthy, we used the comments provided by some participants to adapt small parts of the materials to increase understandability. Informational materials are provided through the OSF pre-registration (https://doi.org/10.17605/OSF.IO/ZN72G)After providing this detailed information we measured informed levels of policy support as well as subjective and objective knowledge. In the last block, we then measured policy-specific beliefs and affective variables. Ethics statement The study was approved by the ethics committee of the University of Geneva. Measures Note: all items mentioned in this section are listed in detail on the OSF (https://doi.org/10.17605/OSF.IO/ZN72G). All items were translated to both German and French by the first author and subsequently back-translated to English by a German and a French native speaker, respectively. In case back-translations did not capture the meaning of the original item, the translated version was adapted in accordance with the translator. Climate change beliefs Climate change beliefs were measured using five unipolar subscales of the Climate Change Perceptions scale 59 : Reality, Human causes, Negative consequences, Spatial proximity, and Temporal distance . All five subscales showed good to excellent internal consistency in our data ( Reality : Cronbach’s α = 0.90 (N = 1096; i.e. without data exclusions)/0.91 (N = 1056; i.e. with data exclusions); Human causes : Cronbach’s α = 0.94/0.95; Negative consequences : Cronbach’s α = 0.93/0.93; Spatial proximity : Cronbach’s α = 0.91/0.91; Temporal distance : Cronbach’s α = 0.83/0.83). Additionally, we measured climate change worry and personal responsibility to mitigate climate change with three items each (all on a 7-point Likert scale). Both scales were self-constructed based on previous literature 60,61 . Both scales showed excellent internal consistency in our data ( Climate Change Worry : Cronbach’s α = 0.95/0.95; Personal responsibility : Cronbach’s α = 0.92/0.92). Trust in political institutions Trust was measured using three subscales with two items each, all on 7-point Likert scales: Value similarity (Cronbach’s α = 0.79/0.79), Credibility (Cronbach’s α = 0.82/0.82) and Competence (Cronbach’s α = 0.73/0.73). The overall six-item measure of trust showed excellent internal consistency (Cronbach’s α = 0.91/0.91). Items were adapted from previous research 62 . General attitudes towards technology (GATT) Attitudes towards technology were measured using an adapted version of the GATT scale 38 . For each subscale (techno-pessimism, techno-optimism, techno-fix), we chose to use the three items with the highest factor loadings. Additionally, each item was measured on a 7-point Likert scale, instead of the original 6-point scale. The subscale on techno-optimism showed good internal consistency (Cronbach’s α = 0.79/0.80), whereas subscales on techno-pessimism (Cronbach’s α = 0.71/0.71) and techno-fix attitudes (Cronbach’s α = 0.72/0.71) showed acceptable internal consistency. Personal values Personal values were measured using an adapted version of the Environmental Portrait Value Questionnaire (E-PVQ) 63 . We used the three items with the highest factor loadings per subscale (biospheric, altruistic, hedonic, and egoistic values). All four subscales showed good internal consistency (Biospheric values: Cronbach’s α = 0.87/0.87; Altruistic values: Cronbach’s α = 0.80.0.79; Hedonic values: Cronbach’s α = 0.80/0.80; Egoistic values: Cronbach’s α = 0.82/0.82). Perceived financial scarcity We used a three-item scale created by Berthold et al. 39 to assess perceptions of financial scarcity. Responses were measured on a 7-point Likert scale. The scale showed excellent internal consistency (Cronbach’s α = 0.90/0.90). Political affiliation Political affiliation was measured using a single item on a 7-point Likert scale from “completely left” (coded as -3) to “completely right (coded as +3) with a neutral anchor “neither left nor right” coded as 0. The item was adapted from the left-to-right placement scale used in the European Social Survey round 8 64 . Additionally, participants were able to choose “prefer not to answer” (coded as 99). Based on these responses, we created an additional dichotomous variable which was coded as 0 = answered the question; 1 = preferred not to answer. Subsequently, 99 was recoded to 0 in the numeric political affiliation variable. Both the numeric and the dichotomous variable are included in the statistical models to control for participants who did not answer the question. Knowledge measures Subjective knowledge was measured using a single-item on a 7-point Likert scale adapted from previous research 65 , asking participants how much they felt they knew about the respective policy from “nothing” (1) to “a lot” (7). Objective knowledge was measured using five self-constructed “True-False” questions per policy. For each statement, participants could either choose “True” (1), “False” (-1) or “I don’t know” (0). Reverse-coded statements (i.e. statements that were false) were recoded so that participants who chose “True” would receive a score of -1, while participants who chose “False” would receive a score of 1. After recoding, we calculated the composite score for objective knowledge per participant by building the sum over all five items and dividing by five. Policy support Based on recommendations by Kyselá and colleagues 66 , we measured policy support with two sub-scales: policy acceptability (Cronbach’s α pre-test = 0.95/0.96; Cronbach’s α post-test = 0.97/0.97) and potential to support (Cronbach’s α pre-test = 0.88/0.89; Cronbach’s α post-test = 0.90/0.90). While we had planned to report analyses separately for these two subscales, we opted to combine both subscales into one variable. We took this decision as a) we found a strong correlation ( r pre-test = 0.80/0.81; r post-test = 0.86/0.87;) between both subscales and b) the internal consistency of a combined scale was excellent (Cronbach’s α pre-test = 0.95/0.95; Cronbach’s α post-test = 0.96/0.96). Thus, we interpreted the distinction of the two subscales as not practically meaningful in our data. Policy-specific beliefs We assessed the following policy-specific beliefs: perceived social norms (perceived descriptive and dynamic norms; two items each with one item relating to perceived acceptability and perceived potential to support in the Swiss public as well as perceived trends in acceptability and potential to support), perceived effectiveness in mitigating climate change and protecting the environment (for simplicity referred to as perceived environmental impacts), perceived effects on the Swiss electricity system, perceived economic impacts, perceived societal impacts, perceived personal impacts, and perceived distributional impacts (i.e. perceived fairness). Perceived social norm items were self-constructed based on recommendations by Constantino et al. 67 . Descriptive norms were measured on a slider scale from 0-100 (indicating the perceived percentage of the Swiss population who would accept/support the respective policy), while dynamic norms were measured on a 7-point Likert scale. Perceived environmental impacts were assessed with three items on a 7-point Likert scale. We initially intended to construct a separate variable for net environmental impacts based on two of these three items (perceived environmental harms and perceived environmental benefits; 68 ). However, given that net environmental impacts correlated strongly with perceived effectiveness in mitigating climate change ( r = .71/.72), we decided to build a composite score over all three items. Perceived effects on the Swiss electricity system included perceived effects on electricity supply reliability, perceived effects on flexibility of the electricity system, perceived effects on the share of renewables in the electricity system and perceived effects on energy independence, all assessed with a single item on a 7-point Likert scale. Perceived economic and societal impacts were measured with two 7-point Likert scale items each (based on Geiger et al. 68 ). We found strong negative correlations between societal benefits and societal harms ( r benefits-harms = -.69/-.71) as well as between economic benefits and economic harms ( r benefits-harms = -.67/-.70). We subsequently built a score for net societal and net economic impacts, respectively, by subtracting scores for perceived harms from scores for perceived benefits (i.e. Benefits minus Harms), creating a score where a positive value indicates perceived net positive impacts, and a negative score indicates perceived net negative impacts. Perceived personal impacts included one item each for perceived coerciveness, perceived (personal) financial impact, perceived personal fairness and perceived impacts on quality of life (one item each for health, comfort and privacy). All items were measured on a 7-point Likert scale. Finally, perceived distributional impacts included three items, each measured on a 7-point Likert scale for impacts on low-income earners in Switzerland, Swiss citizens in general and people living outside of Switzerland. Items on perceived personal and perceived distributional impacts were self-constructed. Affect and emotions Affect was measured using a single-item on a 7-point Likert scale (adapted from Spampatti et al. 28 ). Additionally, we measured the discrete emotions of worry, anger, hope and pride. Each emotion was assessed on a 7-point Likert scale. Items on discrete emotions were self-constructed. Statistical analyses To identify the strongest predictors of policy support, we used random forests 51,53 . We chose this method as we believe it to be more suitable for exploratory data analysis than conventional statistical analyses such as multiple linear regression 52 . More specifically, we ran a regression random forest to predict overall post-test (i.e. informed) policy support. We set the following random forest hyperparameters based recommendations of Pargent et al. 54 : number of trees = 500, number of features (i.e. predictors) to consider at each split = 18 (based on = p /3, with p = number of predictors in the model), and minimum node size to continue splitting a tree = 5. We split our data into a separate training (75% of data) and test set (25% of data) to evaluate model performance. Reported performance measures are the root mean squared error ( RMSE ; that is, the mean prediction error on the scale of the outcome measure) as well as the percentage of explained variance ( R² ). The random forest was subsequently trained on the training set and model performance was evaluated on the test-set. The model used 53 predictors, including three randomly created variables with a gaussian, uniform and binomial distribution, respectively. To identify the strongest predictors of policy support, we subsequently calculated the mean permutation variable importance (PVI) 52,54,69 based on ten permutations per predictor in the model. A PVI score is then obtained by comparing the difference of the RMSE in the actual model compared to a model with permuted values for variable j . This process is then repeated for every variable in the model. The random predictors were included to allow a hard cut-off rule for generally relevant and irrelevant predictors. That is, predictors that have a lower mean PVI than any of the three random predictors were deemed irrelevant for predicting policy support. However, given that random forests are “greedy” algorithms 51 the PVI for some predictors may be inflated when using the hyperparameter settings specified above. This is because at every split, our random forest could choose between 18 predictors. Here, it will always prioritize the predictor that leads to the highest reduction in the RMSE, even though other predictors in this subset may only be marginally worse at predicting support. It is therefore possible that the PVI for some predictors is inflated, while it is deflated for other predictors. To mitigate this, we additionally ran a regression random forest where we set the number of features to consider at each split to 1, effectively forcing the algorithm to always use one randomly selected predictor to predict policy support at each split. To obtain stable predictions, we additionally increased the number of trees to 100.000. PVI is thus reported based on a random forest with these adapted hyperparameter settings, as we believe these PVI scores more accurately reflect the actual variable importance. We preregistered to corroborate these results with standard linear regression. However, given that the mean variance inflation factor in this model was at 1.6, we were concerned about potential multicollinearity biasing regression coefficients 70 . We instead corroborated the results of the random forest using ridge regression 71 . We ran 10-fold cross-validated ridge regression models with 1000 repeats. To prevent potential overfitting, we extracted the ridge regression coefficients with the highest lambda-value within one standard error of the cross-validation fold with the lowest lambda-value (i.e. “lambda.1se” of the “cv.glmnet” function) for each repeat, thus yielding 1000 regression coefficients per predictor. We subsequently ranked the predictors based on absolute size of the average ridge regression coefficients and compared this ranking to the PVI ranking resulting from the random forest. Statistical software Data were analyzed using RStudio (R version 4.3.2; R Core Team (2023)) and the following packages (in alphabetical order):broom 72 (version 1.0.5), car 73 (v. 3.1-2), DALEX 74 (v. 2.4.3, Biecek, 2018), DALEXtra 75 (v. 2.3.0, Maksymiuk et al., 2020), easystats 76 (v. 0.7.4, Lüdecke et al., 2022), ggbeeswarm 77 (v. 0.7.2, Clarke et al., 2023), ggdist 78 (v. 3.3.1; Kay, 2024), ggpubr 79 (v.0.6.0, Kassambara, 2023), glmnet 80, 81, 82 (v. 4.1-8, Friedman et al. (2010); Simon et al. (2011); Tay et al. (2023)), MASS 83 (v.7.3-60.0.1, Venables & Ripley, 2002), mlr3verse 84 (v. 0.2.8, Lang & Schratz, 2023), psych 85 (v. 2.4.3, Revelle, W., 2024), tidyverse 86 (v. 2.0.0, Wickham et al., 2019), and yhat 87 (v. 2.0-4). Study 2 Participants and sample size justification N = 771 participants completed the survey. Sample size was based on a different research question, investigating the role of emotion norms in renewable energy policy support (see OSF pre-registration https://osf.io/qjr4s/?view_only=37a2db2601314d558c0acdaeea80287a). Based on pre-registered exclusion criteria, we excluded six participants from further analyses, given that they were flagged as duplicates. We thus ended up with a final sample of N = 765 participants, representative for the German and French speaking Swiss voting population based on a crossed quota in age, gender and language region. We deem this sample size as sufficient for our purposes as it allows testing predictive performance on a dataset about 72% the size of the original training data with comparable population characteristics as in Study 1. For detailed information on sample characteristics, see Table 2 in supplementary materials S1. Study design and procedure We used a correlational study design with no experimental manipulation. In the first part of the survey, we collected sociodemographic data. Here, participants who were below the age of 18 or did not possess Swiss citizenship were screened out. We additionally screened out participants who failed an attention check on the subsequent survey page (see OSF pre-registration) or had a reCAPTCHA score below 0.5. Next, we presented a short text summarising the content of the referendum on the Federal Act on Renewable Energy (see supplementary materials S3). The text was intended to be as neutral as possible, while summarising key arguments of both the proponents (i.e. the Swiss government) and opponents (i.e. the referendum committee) of the Federal Act on Renewable Energy. After reading this text, participants completed items on model-relevant predictors, which were measured in the same way as in Study 1. The survey also included several additional items which were not the focus of this paper (see OSF). Finally, we asked participants if they had heard of the referendum before taking part in this survey (yes/no), whether they intend to vote (yes/no/already voted by mail) and depending on their answer, how they intend to vote/have voted (in favour/against/prefer not to answer). On the last page of the survey, participants rated their political affiliation. Ethics statement The study was approved by the ethics committee of the University of Geneva. Measures All items relevant to the analyses reported here were measured in the same way as in Study 1. Statistical analyses To test the external validity of the model identified in Study 1 we used classification random forests 54 . To this end we dichotomized our measure of policy support so that values higher than zero were coded as “support” and values equal to or lower than zero were coded as “no support”. We subsequently used a cross-validated classification forest with 20 folds to calculate the lower bound of the 95% CI of the accuracy of our model in classifying participants as supporters or non-supporters. We pre-registered our interpretation of model-performance as follows: If the prediction accuracy for data from Study 2 is at least as high as the lower bound of the 95% CI, the model generalizes to other types of renewable energy policies. If the prediction accuracy for data from Study 2 is lower than the lower bound of the 95% CI but does not drop by more than 10%, the model still performs reasonably well in predicting renewable energy policy support in general. If the prediction accuracy for data from Study 2 drops by more than 10 percent, the model does not generalize well to predict support for other renewable energy polices. The model was subsequently trained on all observations obtained in Study 1 (N = 1056; clean data) and tested on all observations obtained in Study 2 (N = 765). Statistical software Data were analysed using RStudio (R version 4.3.2; R Core Team (2023)) and the following packages (in alphabetical order): ggpubr 77 (v.0.6.0, Kassambara, 2023), mlr3verse 84 (v. 0.2.8, Lang & Schratz, 2023), psych 85 (v. 2.4.3, Revelle, W., 2024), and tidyverse 86 (v. 2.0.0, Wickham et al., 2019). Data, code and materials used are available on the associated OSF project: https://osf.io/bpgx2/?view_only=0e9d765dead2418bbb6e723e3cea6eb3 Study 3 Participants and sample size justification N = 3812 participants completed Study 3. Our required sample size was determined for a different research question (see OSF pre-registration: https://doi.org/10.17605/OSF.IO/TNQFR). We sampled participants from France, Germany, Italy, the Netherlands, Poland and Spain with a market research institute, employing a crossed-quota for age and gender and a separate quota for education level on the country-level. We excluded participants with a straightlining percentage higher than 80% (n = 91), participants flagged as duplicates (n = 196) as well as participants with a survey duration of 5.93 minutes or less (i.e. the fastest 5% of participants; n = 177). We thus ended up with an analysis sample of N = 3355 participants. For detailed information on sample characteristics, see Tables 3 and 4 in supplementary materials S1. Study design and procedure The survey was structured in two parts. Part one of the survey employed a between-subjects design. Participants first answered demographic items, followed by an instructional attention check. Participants who failed this check were screened out of the survey. Participants who passed the attention check were randomly assigned to receive information about one of the four climate mitigation measures (bioenergy with carbon capture and storage, green hydrogen, afforestation and reforestation or switching to sustainable diets). Participants were then asked how much they think they know about the respective mitigation measure. Following this, we presented basic information on this mitigation measure (including information on the Intergovernmental Panel on Climate Change (IPCC) and the global goal of reaching net-zero greenhouse gas emissions by 2050, how the respective mitigation measures works and the estimated emissions reductions until 2050 if employed, based on the IPCC 1 ). Here, we used an attention check asking participants one true-false question regarding the text they just read. Participants who provided the correct response to this question then went on answer items relevant to the analyses in this paper (policy support, perceived social norms, policy-specific beliefs, affect and emotions). The survey then continued with part two, in which additional information was presented and items on policy support and emotions were presented again. All data used for the analyses was obtained in part one of the survey. We will therefore not describe part two in detail. Ethics statement The study was approved by the ethics committee of the University of Geneva. Measures Items relevant to the analyses reported here were measured in the same way as in Studies 1 and 2. However, there were two exceptions to this. First, the item asking participants whether they would vote in favour or against the implementation of the mitigation measure “in a general referendum” was changed to “in a direct vote”, given that referendums are not as typical in representative democracies as in direct democracies. Second, all references to “Switzerland” in the items from Studies 1 and 2 were changed so that the name of the country the participant was from was used instead. Statistical analyses We used the same analysis approach and cut-off criteria as in Study 2. Model accuracy was tested on the entire data and 34subsamples. These were created based on the country participants were from and the mitigation measure they were assigned to (resulting in 24 subsamples) as well as overall subsamples for each country and each mitigation measure. Statistical software Data were analyzed using RStudio (R version 4.3.2; R Core Team (2023)) and the following packages (in alphabetical order): easystats 74 (v. 0.7.4, Lüdecke et al., 2022), ggpubr 77 (v.0.6.0, Kassambara, 2023), mlr3verse 84 (v. 0.2.8, Lang & Schratz, 2023), psych 85 (v. 2.4.3, Revelle, W., 2024), and tidyverse 86 (v. 2.0.0, Wickham et al., 2019). Declarations all participants had to provide explicit consent to taking part in the study as well as to the use and publication of their anonymised data and aggregated results for scientific purposes. Acknowledgements We would like to thank Ben Meuleman for his feedback and advice on our modelling approach. Additionally, we would also like to thank the members of the SWEET PATHFNDR consortium for their valuable feedback on the content of the policy information in Study 1, especially Christian Moretti, Siobhan Powell, Christian Winzer and Florian Baader as well as the members of the Consumer Decision and Sustainable Behavior Lab, Mario Herberz and Manon Gouiran, for their valuable feedback on our manuscript. Finally, we would like to thank PRISMA project experts, colleagues from the University of Geneva’s Renewable Energy Systems group, Joanna Sobocińska, Marta Brucka, Mellano Ocèane, and Francesco Sasso for the translation and revision of the technical information used in Study 3. Author Contributions: CRediT Conceptualization: MK, ET & TB (Studies 1-3), IV (Study 2), VS (Study 3) Data curation: MK (Studies 1-2), IV (Study 2), VS (Study 3) Formal analysis: MK (Studies 1-3) Funding acquisition: ET & TB Methodology: MK, ET & TB Visualization: MK (Studies 1-3) Writing – first draft: MK Writing – review & editing: TB, ET,VS & IV Funding statement This work received funding from the Swiss Federal Office of Energy (SFOE) as part of the SWEET project PATHFNDR and from the Swiss State Secretariat for Education, Research, and Innovation SEFRI for the project PRISMA “Net zero pathway research through integrated assessment model advancements” (project no. 101081604). The authors bear sole responsibility for the conclusions and the results. References Climate Change 2022: Mitigation of Climate Change. (IPCC, Geneva, 2022). Dahl, R. A. On Democracy . (Yale University Press, New Haven London, 1998). Fiorino, D. J. Environmental Risk and Democratic Process: A Critical Review. Columbia J. Environ. Law 14 , 501–548 (1989). Yeganeh, A. J., McCoy, A. P. & Schenk, T. Determinants of climate change policy adoption: A meta-analysis. Urban Clim. 31 , 100547 (2020). Wong-Parodi, G., Krishnamurti, T., Davis, A., Schwartz, D. & Fischhoff, B. A decision science approach for integrating social science in climate and energy solutions. Nat. Clim. Change 6 , 563–569 (2016). Bergquist, M., Nilsson, A., Harring, N. & Jagers, S. C. Meta-analyses of fifteen determinants of public opinion about climate change taxes and laws. Nat. Clim. Change 12 , 235–240 (2022). Drews, S. & Van Den Bergh, J. C. J. M. What explains public support for climate policies? A review of empirical and experimental studies. Clim. Policy 16 , 855–876 (2016). Ejelöv, E. & Nilsson, A. Individual Factors Influencing Acceptability for Environmental Policies: A Review and Research Agenda. Sustainability 12 , 2404 (2020). Goldberg, M. H., Gustafson, A., Ballew, M. T., Rosenthal, S. A. & Leiserowitz, A. Identifying the most important predictors of support for climate policy in the United States. Behav. Public Policy 5 , 480–502 (2021). Clulow, Z., Ferguson, M., Ashworth, P. & Reiner, D. Comparing public attitudes towards energy technologies in Australia and the UK: The role of political ideology. Glob. Environ. Change 70 , 102327 (2021). Ding, D., Maibach, E. W., Zhao, X., Roser-Renouf, C. & Leiserowitz, A. Support for climate policy and societal action are linked to perceptions about scientific agreement. Nat. Clim. Change 1 , 462–466 (2011). Syropoulos, S. & Markowitz, E. M. Perceived responsibility to address climate change consistently relates to increased pro-environmental attitudes, behaviors and policy support: Evidence across 23 countries. J. Environ. Psychol. 83 , 101868 (2022). Fairbrother, M., Johansson Sevä, I. & Kulin, J. Political trust and the relationship between climate change beliefs and support for fossil fuel taxes: Evidence from a survey of 23 European countries. Glob. Environ. Change 59 , 102003 (2019). Swim, J. K. & Geiger, N. Policy attributes, perceived impacts, and climate change policy preferences. J. Environ. Psychol. 77 , 101673 (2021). Ogunbode, C. A. et al. Climate justice beliefs related to climate action and policy support around the world. Nat. Clim. Change 14 , 1144–1150 (2024). Huber, R. A., Wicki, M. L. & Bernauer, T. Public support for environmental policy depends on beliefs concerning effectiveness, intrusiveness, and fairness. Environ. Polit. 29 , 649–673 (2020). Thaller, A. et al. When perceived fairness and acceptance go hand in hand–Drivers of regulatory and economic policies for low-carbon mobility. PLOS Clim. 2 , e0000157 (2023). Boon-Falleur, M., Grandin, A., Baumard, N. & Chevallier, C. Leveraging social cognition to promote effective climate change mitigation. Nat. Clim. Change 12 , 332–338 (2022). Goldberg, M. H., Van Der Linden, S., Leiserowitz, A. & Maibach, E. W. Perceived Social Consensus Can Reduce Ideological Biases on Climate Change. Environ. Behav. 52 , (2019). Sokoloski, R., Markowitz, E. M. & Bidwell, D. Public estimates of support for offshore wind energy: False consensus, pluralistic ignorance, and partisan effects. Energy Policy 112 , 45–55 (2018). Smith, N. & Leiserowitz, A. The Role of Emotion in Global Warming Policy Support and Opposition: The Role of Emotion in Global Warming Policy Support and Opposition. Risk Anal. 34 , 937–948 (2014). Todorova, B. et al. Machine learning identifies key individual and nation-level factors predicting climate-relevant beliefs and behaviors. Preprint at https://doi.org/10.31219/osf.io/um69t (2025). Brosch, T. & Steg, L. Leveraging emotion for sustainable action. One Earth 4 , 1693–1703 (2021). Brosch, T. & Sauter, D. Emotions and the Climate Crisis: A Research Agenda for an Affective Sustainability Science. Emot. Rev. 15 , 253–257 (2023). Jobin, M. & Siegrist, M. We choose what we like – Affect as a driver of electricity portfolio choice. Energy Policy 122 , 736–747 (2018). Jobin, M., Visschers, V. H. M., Van Vliet, O. P. R., Árvai, J. & Siegrist, M. Affect or information? Examining drivers of public preferences of future energy portfolios in Switzerland. Energy Res. Soc. Sci. 52 , 20–29 (2019). Rinscheid, A. & Wüstenhagen, R. Divesting, Fast and Slow: Affective and Cognitive Drivers of Fading Voter Support for a Nuclear Phase-Out. Ecol. Econ. 152 , 51–61 (2018). Spampatti, T., Hahnel, U. J. J., Trutnevyte, E. & Brosch, T. Short and long-term dominance of negative information in shaping public energy perceptions: The case of shallow geothermal systems. Energy Policy 167 , 113070 (2022). Geiger, N., Dwyer, T. & Swim, J. K. Hopium or empowering hope? A meta-analysis of hope and climate engagement. Front. Psychol. 14 , 1139427 (2023). Gregersen, T., Andersen, G. & Tvinnereim, E. The strength and content of climate anger. Glob. Environ. Change 82 , 102738 (2023). Marlon, J. R. et al. How Hope and Doubt Affect Climate Change Mobilization. Front. Commun. 4 , 20 (2019). Myers, T. A., Roser-Renouf, C., Leiserowitz, A. & Maibach, E. Emotional signatures of climate policy support. PLOS Clim. 3 , e0000381 (2024). Schneider, C. R., Zaval, L. & Markowitz, E. M. Positive emotions and climate change. Curr. Opin. Behav. Sci. 42 , 114–120 (2021). Wang, S., Leviston, Z., Hurlstone, M., Lawrence, C. & Walker, I. Emotions predict policy support: Why it matters how people feel about climate change. Glob. Environ. Change 50 , 25–40 (2018). Weber, E. U. & Constantino, S. M. All Hearts and Minds on Deck: Hope Motivates Climate Action by Linking the Present and the Future. Emot. Rev. 15 , 293–297 (2023). Swim, J. K., Guerriero, J. G., Gasper, K., DeCoster, J. & Lengieza, M. L. Emotions and policy information predicting water-quality policy support. J. Environ. Psychol. 98 , 102385 (2024). Zeelenberg, M., Nelissen, R. M. A., Breugelmans, S. M. & Pieters, R. On emotion specificity in decision making: Why feeling is for doing. Judgm. Decis. Mak. 3 , 18–27 (2008). Cologna, V., Berthold, A., Kreissel, A. L. & Siegrist, M. Attitudes towards technology and their relationship with pro-environmental behaviour: Development and validation of the GATT scale. J. Environ. Psychol. 95 , 102258 (2024). Berthold, A., Cologna, V., Hardmeier, M. & Siegrist, M. Drop some money! The influence of income and subjective financial scarcity on pro-environmental behaviour. J. Environ. Psychol. 91 , 102149 (2023). Sparkman, G. & Walton, G. M. Dynamic Norms Promote Sustainable Behavior, Even if It Is Counternormative. Psychol. Sci. 28 , 1663–1674 (2017). Sparkman, G. & Walton, G. M. Witnessing change: Dynamic norms help resolve diverse barriers to personal change. J. Exp. Soc. Psychol. 82 , 238–252 (2019). Kunnas, S. & Trutnevyte, E. Informed minds, opinions aligned? Informed perceptions towards solar PV locations in Switzerland. Environ. Res. Commun. 7 , 075030 (2025). Volken, S. P., Xexakis, G. & Trutnevyte, E. Perspectives of Informed Citizen Panel on Low-Carbon Electricity Portfolios in Switzerland and Longer-Term Evaluation of Informational Materials. Environ. Sci. Technol. acs.est.8b01265 (2018) doi:10.1021/acs.est.8b01265. Blastland, M., Freeman, A. L. J., Van Der Linden, S., Marteau, T. M. & Spiegelhalter, D. Five rules for evidence communication. Nature 587 , 362–364 (2020). Kerr, J. R., Schneider, C. R., Freeman, A. L. J., Marteau, T. & van der Linden, S. Transparent communication of evidence does not undermine public trust in evidence. PNAS Nexus 1 , pgac280 (2022). Pickering, B., Lombardi, F. & Pfenninger, S. Diversity of options to eliminate fossil fuels and reach carbon neutrality across the entire European energy system. Joule 6 , 1253–1276 (2022). Ecker, U., Roozenbeek, J. & Lewandowsky, S. Misinformation remains a threat to democracy. Nature (2024) doi:https://doi.org/10.1038/d41586-024-01587-3. Brexit was wrong, say 57% of British voters. The Economist . Why most people regret Brexit. The Economist . Mitchell, M. C., David Luhnow and Josh. Essay | Brexit Backlash: Brits Now Regret Their Populist Revolt. WSJ https://www.wsj.com/world/uk/brexit-british-regret-uk-election-837cbf4c (2024). Breiman, L. Random Forests. Mach. Learn. 45 , 5–32 (2001). Fife, D. A. & D’Onofrio, J. Common, uncommon, and novel applications of random forest in psychological research. Behav. Res. Methods 55 , 2447–2466 (2022). James, G., Witten, D., Hastie, T. & Tibshirani, R. An introduction to statistical learning with applications in R. Stat. Theory Relat. Fields 6 , 87–87 (2022). Pargent, F., Schoedel, R. & Stachl, C. Best Practices in Supervised Machine Learning: A Tutorial for Psychologists. Adv. Methods Pract. Psychol. Sci. 6 , 1–35 (2023). Freiburghaus, R., Leemann, L., Wasserfallen, F., Willi, T. & Yin, J. 20 Minuten-/Tamedia- Abstimmungsumfrage - Eidgenössische Volksabstimmungen vom 9. Juni . https://www.tamedia.ch/de/unternehmen/umfragen (2024). Mousson, M. 2. SRG-Trendumfrage zur eidg. Abstimmung vom 9. Juni 2024. https://www.gfsbern.ch/de/news/2-srg-trendumfrage-zur-eidg-abstimmung-vom-9-juni-2024/ (2024). Lee, T. M., Markowitz, E. M., Howe, P. D., Ko, C.-Y. & Leiserowitz, A. A. Predictors of public climate change awareness and risk perception around the world. Nat. Clim. Change 5 , 1014–1020 (2015). Yarkoni, T. & Westfall, J. Choosing Prediction Over Explanation in Psychology: Lessons From Machine Learning. Perspect. Psychol. Sci. 12 , 1100–1122 (2017). Van Valkengoed, A. M., Steg, L. & Perlaviciute, G. Development and validation of a climate change perceptions scale. J. Environ. Psychol. 76 , 101652 (2021). Bouman, T. et al. When worry about climate change leads to climate action: How values, worry and personal responsibility relate to various climate actions. Glob. Environ. Change 62 , 102061 (2020). Wilson, R. S., Zwickle, A. & Walpole, H. Developing a Broadly Applicable Measure of Risk Perception. Risk Anal. 39 , 777–791 (2019). Spampatti, T., Brosch, T., Trutnevyte, E. & Hahnel, U. J. J. A Preregistered Field Study of the Trust Inoculation Against a Negative Event Involving Geothermal Energy Systems. Collabra Psychol. 9 , 89755 (2023). Bouman, T., Steg, L. & Kiers, H. A. L. Measuring Values in Environmental Research: A Test of an Environmental Portrait Value Questionnaire. Front. Psychol. 9 , 564 (2018). European Social Survey. ESS Round 8 Source Questionnaire . (2016). Wong-Parodi, G. & Berlin Rubin, N. Exploring how climate change subjective attribution, personal experience with extremes, concern, and subjective knowledge relate to pro-environmental attitudes and behavioral intentions in the United States. J. Environ. Psychol. 79 , 101728 (2022). Kyselá, E., Ščasný, M. & Zvěřinová, I. Attitudes toward climate change mitigation policies: a review of measures and a construct of policy attitudes. Clim. Policy 19 , 878–892 (2019). Constantino, S. M. et al. Scaling Up Change: A Critical Review and Practical Guide to Harnessing Social Norms for Climate Action. Psychol. Sci. Public Interest 23 , 50–97 (2022). Geiger, N., Swim, J. K. & Benson, L. Using the three-pillar model of sustainability to understand lay reactions to climate policy: A multilevel approach. Environ. Sci. Policy 126 , 132–141 (2021). Henninger, M., Debelak, R., Rothacher, Y. & Strobl, C. Interpretable machine learning for psychological research: Opportunities and pitfalls. Psychol. Methods Advance online publication (2023) doi:10.1037/met0000560. Bowerman, B. L. & O’Connell, R. T. Linear Statistical Models : An Applied Approach . (Boston : PWS-Kent Pub. Co., 1990). Hoerl, A. E. & Kennard, R. W. Ridge Regression: Biased Estimation for Nonorthogonal Problems. Technometrics 12 , 55–67 (1970). R Packages used (Studies 1-3) Robinson, D., Hayes, A., & Couch, S. (2023). broom: Convert Statistical Objects into Tidy Tibbles. R package version 1.0.5, https://CRAN.R-project.org/package=broom. Fox, J., & Weisberg, S. (2019). An R Companion to Applied Regression, Third edition. Sage, Thousand Oaks CA. https://socialsciences.mcmaster.ca/jfox/Books/Companion/. Biecek, P. (2018). DALEX: Explainers for Complex Predictive Models in R. Journal of Machine Learning Research, 19(84), 1-5. https://jmlr.org/papers/v19/18-416.html. Maksymiuk, S., Gosiewska, A., & Biecek, P. (2020). Landscape of R packages for eXplainable Artificial Intelligence. arXiv. Pages 6, 7, 11, 15, https://arxiv.org/abs/2009.13248. Lüdecke, D., Patil, I., Ben-Shachar, M. S., Wiernik, B. M., Bacher, E., Thériault, R., & Makowski, D. (2022). easystats: Framework for Easy Statistical Modeling, Visualization, and Reporting. CRAN. doi:10.32614/CRAN.package.easystats Clarke, E., Sherrill-Mix, S., & Dawson, C. (2023). ggbeeswarm: Categorical Scatter (Violin Point) Plots. R package version 0.7.2, https://CRAN.R-project.org/package=ggbeeswarm. Kay, M. (2024). “ggdist: Visualizations of Distributions and Uncertainty in the Grammar of Graphics.” IEEE Transactions on Visualization and Computer Graphics, 1-11. doi:10.1109/TVCG.2023.3327195. Kassambara, A. (2023). ggpubr: 'ggplot2' Based Publication Ready Plots. R package version 0.6.0, https://CRAN.R-project.org/package=ggpubr. Friedman, J., Tibshirani, R., & Hastie, T. (2010). Regularization Paths for Generalized Linear Models via Coordinate Descent. Journal of Statistical Software, 33(1), 1-22. doi:10.18637/jss.v033.i01. Simon, N., Friedman, J., Tibshirani, R., & Hastie ,T. (2011). Regularization Paths for Cox's Proportional Hazards Model via Coordinate Descent. Journal of Statistical Software, 39(5), 1-13. doi:10.18637/jss.v039.i05 Tay, J.K., Narasimhan, B., & Hastie, T. (2023). Elastic Net Regularization Paths for All Generalized Linear Models. Journal of Statistical Software, 106(1), 1-31. doi:10.18637/jss.v106.i01 Venables, W. N. & Ripley, B. D. (2002) Modern Applied Statistics with S. Fourth Edition. Springer, New York. ISBN 0-387-95457-0 Lang, M., & Schratz, P. (2023). mlr3verse: Easily Install and Load the 'mlr3' Package Family. R package version 0.2.8, https://CRAN.R-project.org/package=mlr3verse Revelle, W. (2024). psych: Procedures for Psychological, Psychometric, and Personality Research. Northwestern University, Evanston, Illinois. R package version 2.4.3, https://CRAN.R-project.org/package=psych. Wickham, H., Averick, M., Bryan, J., Chang, W., McGowan, L.D., François, R., Grolemund, G., Hayes, A., Henry, L., Hester, J., Kuhn, M., Pedersen, T.L., Miller, E., Bache, S.M., Müller, K., Ooms, J., Robinson, D., Seidel, D.P., Spinu, V., Takahashi, K., Vaughan, D., Wilke, C., Woo, K., & Yutani, H. (2019). Welcome to the tidyverse. Journal of Open Source Software, 4(43), 1686. doi:10.21105/joss.01686. Additional Declarations There is NO Competing Interest. Supplementary Files supplementarymaterials.docx Supplementary information supplementarymaterials.docx Supplementary Materials_24.9.2025 Cite Share Download PDF Status: Under Review Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7658135","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":520474609,"identity":"d4469889-2b2c-42e2-b96c-8d6d421d548a","order_by":0,"name":"Morris Krainz","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+UlEQVRIiWNgGAWjYBACgwNAIoGBIbGBgbGBgaGCgYENJAxi4wKWUC3JEGVniNBifwBCp4FJxjaoMD4tZsd7n3148Ichx7z9cNvDn/MO5/HxLz7A8HMHHi1njhvPSGxjqJA5k9huILntcDGbxLMExt4zeLTcSGMG+b1CgiGxTcJw22EgecaAGe5CLMAApCUB6DAJ/odtEolzQFrOfyBCCxtDmoQEUPHBBqAW/h4G/FrOHGMGOSlZQuJhm2TDsXQgm83gYC8+LcfbmBl//LFJnMGf/kzyR4114vz+ww8f/MSjBQokkNkJDAcIakAF/KRqGAWjYBSMguEOAGiXVXNj5nSoAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0009-0003-8146-763X","institution":"University of Geneva","correspondingAuthor":true,"prefix":"","firstName":"Morris","middleName":"","lastName":"Krainz","suffix":""},{"id":520474610,"identity":"3f252082-67ef-495f-8499-369d1f1c2adb","order_by":1,"name":"Valeria Sorgato","email":"","orcid":"","institution":"Renewable Energy Systems, University of Geneva","correspondingAuthor":false,"prefix":"","firstName":"Valeria","middleName":"","lastName":"Sorgato","suffix":""},{"id":520474611,"identity":"a34456de-4755-4b5a-8df3-b71c8788cabd","order_by":2,"name":"Izïa Vallaeys Mora","email":"","orcid":"","institution":"Department of Psychology and Swiss Center for Affective Sciences, University of Geneva","correspondingAuthor":false,"prefix":"","firstName":"Izïa","middleName":"Vallaeys","lastName":"Mora","suffix":""},{"id":520474612,"identity":"7e4d9257-551e-42ad-a32c-86518746aa40","order_by":3,"name":"Evelina Trutnevyte","email":"","orcid":"https://orcid.org/0000-0002-1716-6192","institution":"University of Geneva","correspondingAuthor":false,"prefix":"","firstName":"Evelina","middleName":"","lastName":"Trutnevyte","suffix":""},{"id":520474613,"identity":"715660f7-1a8d-49f1-8a79-e603af3a107b","order_by":4,"name":"Tobias Brosch","email":"","orcid":"https://orcid.org/0000-0001-6883-0383","institution":"University of Geneva","correspondingAuthor":false,"prefix":"","firstName":"Tobias","middleName":"","lastName":"Brosch","suffix":""}],"badges":[],"createdAt":"2025-09-19 11:35:08","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7658135/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7658135/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":92536680,"identity":"cf737093-02ed-426a-8c15-7b5ceef54683","added_by":"auto","created_at":"2025-09-30 17:31:38","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":32040,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eModel prediction accuracy in Studies 2 and 3 \u003c/strong\u003eThe dotted lines represent the pre-registered cut-off at the lower bound of the 95% CI for model accuracy in the training data. The black lines represent the cut-off for reasonable model performance based on our pre-registered evaluation criteria. Points falling to the left of this line show subsamples for which the model did not generalize. Points falling between both lines show subsamples for which the model performed \u003cem\u003ereasonably well\u003c/em\u003e. Points falling to the right of the dotted line show subsamples for which the model generalized. Panel A shows model performance in Study 2 for the full sample as well as subsamples of participants who had already voted, intended to vote and did not intend to vote. Panel B shows model performance for each mitigation measure and country investigated in Study 3.\u003c/p\u003e","description":"","filename":"Picture19.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7658135/v1/38d595d40b4a8233d687d60d.jpg"},{"id":92537261,"identity":"b7b37ec4-88e7-4a42-9caf-e181d3dc5e74","added_by":"auto","created_at":"2025-09-30 17:39:39","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1128682,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7658135/v1/2b2c66ce-3c79-4149-98db-6831f42524a8.pdf"},{"id":92536682,"identity":"8cff0f8c-6d05-4ea8-bd0a-b54f504c3a9b","added_by":"auto","created_at":"2025-09-30 17:31:39","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":764437,"visible":true,"origin":"","legend":"Supplementary information","description":"","filename":"supplementarymaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-7658135/v1/b65ee8bccdc34d93c1422a9d.docx"},{"id":92536681,"identity":"423fbc08-2a56-45c0-8506-bdb524d3d8ab","added_by":"auto","created_at":"2025-09-30 17:31:39","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":765452,"visible":true,"origin":"","legend":"Supplementary Materials_24.9.2025","description":"","filename":"supplementarymaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-7658135/v1/8ad56199cd6c54d3f9751a42.docx"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Identifying and validating the strongest predictors of informed energy policy support across Europe","fulltext":[{"header":"Introduction","content":"\u003cp\u003eVarious technically feasible and cost-comparable options are available to decarbonize the energy sector\u003csup\u003e1\u003c/sup\u003e. In democratic societies, the most appropriate approach is to implement the options preferred by the informed general public\u003csup\u003e2\u003c/sup\u003e. This is done for three reasons\u003csup\u003e3\u003c/sup\u003e: First, understanding which policies the public would support when aware of the benefits and drawbacks of the different options should help speed up the energy transition \u0026ndash; which is supported by the strong observed association between public policy support and successful policy adoption\u003csup\u003e4\u003c/sup\u003e. Second, it is a normative expectation in democratic societies to implement policies that the public wants. Third, by combining insights about technically possible and publicly preferred options, we better understand which solutions are actually feasible\u003csup\u003e5\u003c/sup\u003e.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo understand and explain policy preferences, previous research tested a large number of variables as predictors of policy support\u003csup\u003e6\u0026ndash;9\u003c/sup\u003e. These variables can be organized into three categories: (1) \u003cem\u003eSociodemographic factors\u003c/em\u003e, including age\u003csup\u003e6\u003c/sup\u003e, gender\u003csup\u003e10\u003c/sup\u003e, education\u003csup\u003e6\u003c/sup\u003e and income\u003csup\u003e6\u003c/sup\u003e, (2) \u003cem\u003eindividual differences\u003c/em\u003e, including beliefs about climate change\u003csup\u003e6,9,11,12\u003c/sup\u003e, personal values\u003csup\u003e6\u003c/sup\u003e, trust in political institutions\u003csup\u003e6,13\u003c/sup\u003e and political affiliation \u003csup\u003e6\u003c/sup\u003e, and (3) \u003cem\u003epolicy-specific beliefs\u003c/em\u003e, including perceived policy effectiveness\u003csup\u003e6\u003c/sup\u003e, perceived environmental, societal and economic effects\u003csup\u003e14\u003c/sup\u003e, perceived fairness\u003csup\u003e6,15\u003c/sup\u003e, personal impacts\u003csup\u003e16,17\u003c/sup\u003e, perceived social consensus \u003csup\u003e18\u0026ndash;20\u003c/sup\u003e, and subjective as well as objective knowledge of the policy\u003csup\u003e6\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eWhile this has greatly contributed to our understanding of support for climate mitigation policies, several questions remain unanswered. First and foremost, so far, there is no comprehensive evaluation of the comparative predictive strength of this vast range of variables. Empirical evidence identifying the most important drivers of support would increase our understanding of policy preferences and help policymakers and communicators design targeted policies that better align with public concerns and preferences. This is especially important given that previous research often used broad and aggregated measures of support for \u0026ldquo;climate policies\u0026rdquo; in general by combining mitigation measures from different domains (e.g., measuring aggregated support for carbon taxes, renewable energy, infrastructure and public transport, consumer and dietary behaviour, etc.) \u003csup\u003e9,15,21,22\u003c/sup\u003e. While this approach measures overall support for all types of climate mitigation policies, it does not capture support for specific climate mitigation policies. Furthermore, the approach lacks ecological validity as the public is usually asked to vote on individual policies, rather than on a bundle of unconnected policy options. It is thus possible that the importance of some of the predictors identified with the aggregation approach is under- or overestimated \u003csup\u003e22\u003c/sup\u003e. Empirical evidence on the most important predictors of policy support in a specific domain can help align policy content with public interests and may accelerate their successful implementation.\u003c/p\u003e\n\u003cp\u003ePrevious acceptance research, moreover, tended to overly focus on cognitive factors\u003csup\u003e6\u003c/sup\u003e. Recently, researchers called for a stronger integration of the affective sciences into climate mitigation research\u003csup\u003e8,23,24\u003c/sup\u003e, highlighting the important role of factors such as \u003cem\u003egeneral affect\u003c/em\u003e\u003csup\u003e9,25\u0026ndash;28\u003c/sup\u003e and discrete emotions such as \u003cem\u003eanger, worry, hope,\u003c/em\u003e and \u003cem\u003epride\u0026nbsp;\u003c/em\u003e\u003csup\u003e21,29\u0026ndash;35\u003c/sup\u003e in predicting policy support and climate-relevant actions and intentions. Yet, a comprehensive evaluation comparing the predictive strength of cognitive and affective factors is missing. The role of discrete emotions in policy support has so far mainly been investigated with a focus on emotions towards climate change, producing mixed results\u003csup\u003e21,32\u0026ndash;35\u003c/sup\u003e. Almost no research exists on the role of emotions towards specific policies, which is however likely to result in more reliable associations\u003csup\u003e29\u0026ndash;31,36,37\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eFurthermore, several newly developed constructs have been associated with pro-environmental behaviour or policy support in other domains. \u003cem\u003eGeneral attitudes towards technology\u003c/em\u003e have been associated with pro-environmental behavioural intentions\u003csup\u003e38\u003c/sup\u003e. \u003cem\u003ePerceived financial scarcity\u003c/em\u003e was associated with willingness to switch to more flexible energy consumption\u003csup\u003e39\u003c/sup\u003e. \u003cem\u003eDynamic social norms\u003c/em\u003e were positively associated with pro-environmental behaviour\u003csup\u003e40\u003c/sup\u003e and policy support for an equal pay policy\u003csup\u003e41\u003c/sup\u003e. These constructs may thus be additional predictors of climate mitigation policy support but remain untested in this context.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFinally, while informed decision-making plays a crucial role in democracies\u003csup\u003e2\u003c/sup\u003e, most policy support research has not provided participants with sufficient information about the policies to help them make \u003cem\u003einformed\u003c/em\u003e decisions on whether or not to support a policy (but see e.g.\u003csup\u003e42,43\u003c/sup\u003e). Uninformed participants may provide unreliable insights into public preferences and policy support that are not aligned with their values and concerns. Hence, understanding which predictors drive support in informed participants\u003csup\u003e5,44,45\u003c/sup\u003e \u0026ndash; those who are aware of the policy\u0026rsquo;s advantages and disadvantages \u0026nbsp;\u0026ndash; would provide robust and actionable insights into policy support by bridging the gap between how citizens think about a policy without being informed (i.e. the descriptive reality) and how they think about a policy when they are aware of the advantages and disadvantages (i.e. the normative ideal). Transparently communicating policy impacts may help to avoid decisions that may otherwise be taken based on misinformation\u003csup\u003e47\u003c/sup\u003e such as the British Brexit vote, where citizens seemed to either not be sufficiently informed or even misinformed about the consequences of the United Kingdom leaving the European Union \u003csup\u003e48\u0026ndash;50\u003c/sup\u003e. Designing and implementing policies in accordance with informed public preferences is especially important in the energy domain, where a large number of potential options allow reaching a net-zero energy system \u003csup\u003e1,46\u003c/sup\u003e, and where taking public preferences into account could substantially facilitate policy implementation.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo address these issues, here, we run three preregistered studies (total N = 5176) in seven different countries. Using machine-learning, we first test a comprehensive set of 50 predictors of support for different energy policies to identify the strongest predictors of energy policy support in informed citizens in Switzerland. We then validate these results by predicting the outcome of a real-world referendum on renewable energy in Switzerland and illustrate the generalizability of the identified predictors on support for other climate mitigation measures across six European countries. The work presented here highlights the value of using interpretable predictive machine-learning models to compare and identify the strongest predictors of climate mitigation support in different domains.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eIdentifying the strongest predictors of energy policy support\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eIn Study 1, we randomly assigned Swiss participants (N = 1056) to receive transparent information (i.e. information aiming to inform rather than persuade participants;\u003csup\u003e5,44,45\u003c/sup\u003e) about one of four policies that aim to increase the flexibility of the Swiss energy system and are being discussed in the context of the Swiss energy transition: a policy on the construction and use of Direct Air Capture and Carbon Storage (DACCS) facilities to allow the use of fossil fuel back up electricity plants which are not large enough for CCS (referred to as DACCS Policy from here on), a policy increasing electricity imports for flexibility (Electricity Trading Policy), a policy on national incentives for the construction of flexible EV charging infrastructure (Flexible EV Charging Policy), or a policy focusing on national incentives for the flexible operation of heat pumps (Flexible Heating Policy). Participants received transparent information about the environmental, energy system-related, economic and societal impacts of the respective policy (see supplementary materials S3 for the information provided to the participants). We measured \u003cem\u003esociodemographic factors\u003c/em\u003e and \u003cem\u003eindividual-level differences\u003c/em\u003e before information provision and \u003cem\u003epolicy support\u003c/em\u003e, \u003cem\u003epolicy-specific beliefs,\u003c/em\u003e \u003cem\u003eaffect\u003c/em\u003e and \u003cem\u003eemotions\u003c/em\u003e towards policies after information provision (for a full list of variables see Table 6 in supplementary materials S2).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWe used regression random forests\u003csup\u003e51\u0026ndash;54\u003c/sup\u003e to identify the strongest predictors of support across the four energy policies. This method is more suitable for exploratory data analysis than conventional statistical analyses such as multiple linear regression\u003csup\u003e52\u003c/sup\u003e (see methods section for details on hyperparameter settings and model specification). The model ranking the 50 predictors performed with a root mean squared error (RMSE) of 0.78 and explained 71.08 % of the variance in policy support (R\u0026sup2; = 0.7108) in an independent test-set including 25 % of the participants. Mean RMSE loss after permutations was used as an indicator of variable importance. Twelve variables achieved a higher mean RMSE-loss based importance score than the categorical predictor which indicated the policy participants were assigned to (see Table 1), and were thus interpreted as the strongest domain-specific predictors for energy policy support.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel-identified strongest domain-specific predictors of renewable energy policy support in Study 1\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 23.7542%;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePredictor\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eExample item\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7575%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSubgroup\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRank\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e(Random Forest / Ridge Regression)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 23.7542%;\"\u003e\n \u003cp\u003eGeneral affect\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;In general, how do you feel about this policy?\u0026rdquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7575%;\"\u003e\n \u003cp\u003eAffective responses\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e1 / 1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 23.7542%;\"\u003e\n \u003cp\u003ePerceived societal impacts\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;Overall, this policy will be beneficial to society and people.\u0026rdquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 22.7575%;\"\u003e\n \u003cp\u003ePolicy impact beliefs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e2 / 2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 23.7542%;\"\u003e\n \u003cp\u003eHope\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;This policy makes me hopeful.\u0026rdquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7575%;\"\u003e\n \u003cp\u003eAffective responses\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e3 / 3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 23.7542%;\"\u003e\n \u003cp\u003ePride\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;This policy makes me proud.\u0026rdquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7575%;\"\u003e\n \u003cp\u003eAffective responses\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e4 / 5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 23.7542%;\"\u003e\n \u003cp\u003ePerceived environmental impacts\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;Overall, this policy will be beneficial for the environment.\u0026rdquo;\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 22.7575%;\"\u003e\n \u003cp\u003ePolicy impact beliefs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e5 / 4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 23.7542%;\"\u003e\n \u003cp\u003ePersonal fairness\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;Overall, this policy is [completely unfair \u0026ndash; completely fair] towards me.\u0026rdquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 22.7575%;\"\u003e\n \u003cp\u003eFairness perceptions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e6 / 8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 23.7542%;\"\u003e\n \u003cp\u003eFairness towards Swiss citizens\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;Overall, this policy is [completely unfair \u0026ndash; completely fair] towards Swiss citizens.\u0026rdquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7575%;\"\u003e\n \u003cp\u003eFairness perceptions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e7 / 9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 23.7542%;\"\u003e\n \u003cp\u003ePerceived dynamic norms\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;Do you expect the number of people in Switzerland who would accept this policy rather to increase or decrease over the next 12 months?\u0026rdquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 22.7575%;\"\u003e\n \u003cp\u003ePerceived social consensus\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e8 / 6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 23.7542%;\"\u003e\n \u003cp\u003ePerceived impact on personal wellbeing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;Overall, this policy would have a [very negative \u0026ndash; very positive] effect on my comfort.\u0026rdquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7575%;\"\u003e\n \u003cp\u003ePolicy impact beliefs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e9 / -\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 23.7542%;\"\u003e\n \u003cp\u003eWorry\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;This policy makes me worry.\u0026rdquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7575%;\"\u003e\n \u003cp\u003eAffective responses\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e10 / 7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 23.7542%;\"\u003e\n \u003cp\u003ePerceived impact on personal finances\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;Do you think that your household would win or lose financially from this policy?\u0026rdquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7575%;\"\u003e\n \u003cp\u003ePolicy impact beliefs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e11 / -\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 23.7542%;\"\u003e\n \u003cp\u003eAnger\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;This policy makes me angry.\u0026rdquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7575%;\"\u003e\n \u003cp\u003eAffective responses\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e12 / 11\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 23.7542%;\"\u003e\n \u003cp\u003eFairness towards people in other countries\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;Overall, this policy is [completely unfair \u0026ndash; completely fair] towards people living outside of Switzerland.\u0026rdquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7575%;\"\u003e\n \u003cp\u003eFairness perceptions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e- / 10\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 23.7542%;\"\u003e\n \u003cp\u003eFairness towards low-income earners\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 38.0399%;\"\u003e\n \u003cp\u003e\u0026ldquo;In your view, would the low-income earners win or lose if this policy was implemented in Switzerland?\u0026rdquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7575%;\"\u003e\n \u003cp\u003eFairness perceptions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 15.4485%;\"\u003e\n \u003cp\u003e- / 12\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eNote.\u0026nbsp;\u003c/em\u003eAll items in this table were rated on fully-labelled 7-point Likert scales. Affective responses, fairness perceptions, as well as perceived impacts on personal wellbeing and finances were measured with single items. Perceived societal impacts and perceived dynamic norms were measured with two items each and perceived environmental impacts were measured with three items. For more details on how predictor variables were measured, see the Methods section. Hyphens indicate that the variable in question was not included in the top 12 predictors of the random forest or ridge regression, respectively. Predictor weights (Permutation Variable Importance for the Random Forest and the mean absolute regression coefficient for ridge regression) are reported in the supplementary materials.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAn alternative analysis strategy running a ridge regression model was used to corroborate these results (see Methods for details). This model achieved slightly better performance metrics than the regression random forest (mean RMSE = 0.68; mean R\u0026sup2; = 0.7812)[1]. When comparing the twelve strongest predictors of the ridge regression model (based on the mean absolute regression coefficient) with the twelve strongest predictors of the regression random forest, the models, by and large, converged on the same predictors of domain-specific policy support. The first three predictor variables were identical (and in the same order), and seven of the remaining nine model predictors overlapped (for more details on predictor ranks in both models, see Table 1). Combining the results of both analyses, we find that the strongest predictors of energy policy support include \u003cem\u003eaffective responses\u003c/em\u003e (general affect, hope, pride, worry, anger), \u003cem\u003epolicy-impact beliefs\u003c/em\u003e (regarding societal and environmental impacts, impact on personal wellbeing and finances), \u003cem\u003efairness perceptions\u003c/em\u003e (personal fairness, distributional fairness towards Swiss citizens, people living in other countries and low income earners) and \u003cem\u003eperceived social norms\u003c/em\u003e (perceived dynamic norm).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo provide a more intuitive interpretation of the model performance, we dichotomized the measure of policy support, separating participants into groups of supporters and non-supporters. We trained a random forest classifier to classify 75% of participants into these groups, and tested the accuracy of this model on the remaining 25% of participants. The random forest classifier showed excellent performance in the independent test-set (accuracy = 92.8%; AUC = 0.9746, true-positive rate = 0.9174; true-negative rate = 0.9355; positive predictive value = 0.9091; negative predictive value = 0.9416).\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003ePredictive model performance in a real-world referendum on renewable energy\u0026nbsp;\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eIn Study 1, we identified the 14 strongest predictors of support for energy policies using a machine learning approach. We then tested the model\u0026rsquo;s predictive performance on real-world (rather than hypothetical) data related to the Swiss referendum on the Federal Act on Secure Electricity Supply from Renewable Energy Sources (from here on referred to as \u0026ldquo;Federal Act on Renewable Energy\u0026rdquo;), which proposed an infrastructure increase for solar, wind and hydropower and was voted on in a referendum on the \u003cem\u003e9\u003csup\u003eth\u003c/sup\u003e of June 2024\u003c/em\u003e. We collected data from the 30\u003csup\u003eth\u003c/sup\u003e of May to the 7\u003csup\u003eth\u003c/sup\u003e of June, measuring the 14 model-identified predictors as well as policy support for the Federal Act on Renewable Energy in a Swiss sample (N = 765). To mimic the information treatment in Study 1, we provided participants with a text summarising the content of the referendum as well as the key arguments put forward by proponents and opponents of the referendum.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo define cut-offs for the interpretation of model accuracy, we used the training data from Study 1, applying a 20-fold cross validation approach to calculate the lower bound of the 95% Confidence Interval (CI) for the prediction accuracy of our random forest classifier (lower bound 95% CI = 0.882; mean accuracy over 20 folds = 0.895, standard deviation = 0.034). We pre-registered our interpretation of model-performance for Study 2 as follows:\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003eIf prediction accuracy is at least as high as the lower bound of the 95% CI, the model \u003cem\u003egeneralizes\u003c/em\u003e to other types of renewable energy policies.\u003c/li\u003e\n \u003cli\u003eIf prediction accuracy is lower than the lower bound of the 95% CI but does not drop by more than 10%, the model performs \u003cem\u003ereasonably well\u003c/em\u003e in predicting policy support in the domain of renewable energy.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eIf prediction accuracy drops by more than 10 %, the model \u003cem\u003edoes not generalize\u003c/em\u003e to predict support for other renewable energy polices.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eThe model was subsequently trained on all observations obtained in Study 1 (N = 1056) and tested on all observations obtained in Study 2 (N = 765) to predict dichotomized support for the Federal Act on Renewable Energy. The model achieved a prediction accuracy of 0.871 in predicting policy support (see Figure 1A). In line with our pre-registered model performance interpretation, we assess the models\u0026rsquo; performance as reasonably well in generalizing to predict policy support for renewable energy policies. When exploratorily restricting the analysis to the participants who had either already cast their vote (N = 332; accuracy = 0.892) or who reported the intention to participate in the vote (N = 326; accuracy = 0.883), the model performed within the preregistered range, indicating generalizability. In contrast, model performance dropped below the preregistered range for participants who indicated no intention of voting (N = 107; accuracy = 0.766). We additionally used the predictions obtained from our model to provide a forecast for the results of the Swiss general referendum on the Federal Act on Renewable Energy. We predicted (and preregistered) a majority vote in favour of the Federal Act on Renewable Energy with 61.83%, matching national polls conducted about three to four weeks before the referendum\u003csup\u003e55,56\u003c/sup\u003e. The actual referendum resulted in a majority of votes in favour of the Federal Act on Renewable Energy (68.7%). Thus, our model predictions allowed a correct forecast of the decision the Swiss public would take, while also explaining which perceptions and affective responses were key in forming this decision.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eModel performance for climate mitigation policies across multiple countries\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eIn Study 3, we aimed to test to what extent the model generalizes to other types of climate mitigation measures and to what extent it generalizes to predict support in other countries. To this end, we presented participants in France, Germany, Italy, the Netherlands, Poland, and Spain with information about one out of four climate mitigation options frequently modelled in Integrated Assessment Models\u003csup\u003e1\u003c/sup\u003e: \u003cem\u003eSustainable diets\u003c/em\u003e, \u003cem\u003eBioenergy with Carbon Capture and Storage\u003c/em\u003e (\u003cem\u003eBECCS\u003c/em\u003e), \u003cem\u003egreen hydrogen\u003c/em\u003e and, \u003cem\u003eafforestation and reforestation\u003c/em\u003e. As in Study 2, participants first received transparent information on the respective mitigation measure, consisting of information on the measures\u0026rsquo; general characteristics and its mitigation potential by 2050 (see supplementary materials). Subsequently, we collected data on the 14 model-relevant variables as well as on policy support. We then tested the predictive performance of our model in the whole sample (N = 3355) and a total of 34 subsamples: one subsample for each of the four climate mitigation solutions in each of the six countries (i.e. 24 subsamples), as well as overall performance by country (six subsamples) and by mitigation measure (four subsamples, see Figure 1B).\u003c/p\u003e\n\u003cp\u003eBased on our preregistered cut-off criteria, the model performed reasonably well in predicting support for \u003cem\u003esustainable diets\u003c/em\u003e (overall accuracy = 0.88) and \u003cem\u003eBECCS\u003c/em\u003e (overall accuracy = 0.87). Prediction accuracy decreased considerably regarding the other two investigated measures, \u003cem\u003egreen hydrogen\u003c/em\u003e (overall accuracy = 0.73) and \u003cem\u003eafforestation and reforestation\u003c/em\u003e (overall accuracy = 0.65). This pattern is explained by a substantial negative correlation between mean reported support and model accuracy (\u003cem\u003er\u003c/em\u003e = - 0.80, 95% CI [-0.89, 0.63], \u003cem\u003et\u003c/em\u003e(33) = -7.63, \u003cem\u003ep\u003c/em\u003e \u0026lt; .001; see also supplementary materials), indicating that the model generally performs better for mitigation measures with lower levels of policy support.\u0026nbsp;\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eWe developed and tested a data-driven domain-specific model of the strongest predictors of energy policy support. We identified \u003cem\u003eaffective responses\u003c/em\u003e, \u003cem\u003epolicy-impact beliefs\u003c/em\u003e, \u003cem\u003efairness perceptions,\u003c/em\u003e and\u003cem\u003e\u0026nbsp;perceived trends in policy support over time\u003c/em\u003e as the strongest predictors of energy policy support. We then trained a random forest classifier to predict support for a real-world political decision and validated the classifier in a broader European context by testing model performance for four different climate mitigation solutions in six European countries. Our model achieved high accuracy in predicting support for the referendum, allowed correctly forecasting the majority vote, and achieved reasonable accuracy in predicting support for other climate mitigation measures in samples across Europe. These results, gained from informed citizens, can be used to inform policy design and communication strategies as well as to generate and test hypotheses about causal relationships of variables within the model.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe findings expand research in decision-science\u003csup\u003e5\u003c/sup\u003e (Wong-Parodi et al., 2016) and transparent evidence communication\u003csup\u003e44,45\u003c/sup\u003e and have direct policy implications. The importance ranking identified in Study 1 indicates that transparent information on environmental and societal impacts may have a stronger influence on policy support than transparent information regarding energy system or economic impacts (see Table 1). The high importance of perceived societal impacts and fairness is in line with research highlighting the important role of climate justice beliefs for policy support\u003csup\u003e15\u003c/sup\u003e. Energy policies should thus be designed taking not only effects on the energy system and the economy into account, but especially effects on the environment and on society. Additionally, our work extends the literature on predictors of climate mitigation policy support by showing that the discrete emotions hope, pride, worry and anger as well as dynamic social norms play an important role in predicting support for energy policies, even when directly comparing them to predictors of policy support highlighted in past research\u003csup\u003e6\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eFinally, this work contributes to methodological developments at the intersection of energy and social sciences, demonstrating both the value and the limitations of predictive machine-learning models in this field. Our approach offers insights into the overall level of public support as well as the concrete reasons for \u003cem\u003ewhy\u003c/em\u003e a policy receives this support (or not), while allowing to forecast future referendum results. However, the model performed considerably worse for mitigation measures with very high levels of overall support (\u003cem\u003egreen hydrogen\u003c/em\u003e and \u003cem\u003eafforestation and reforestation\u003c/em\u003e) in the six European countries in Study 3. While it is arguably more important to accurately predict support for measures that are not already highly accepted by the public, it does highlight the limitations of our approach. The decrease in model performance for subsamples in which overall policy support was substantially higher than in our training data underscores the importance of using training data that is maximally appropriate for the prediction context, as this can substantially impact model performance. Additionally, while our approach allowed identifying the most important predictors of renewable energy policy support, it did not allow insights into the causal connections between variables. Some variables that received lower importance scores in our models may still be important antecedents of variables ranked higher in importance. As such, it is not possible to conclude that these variables can be ignored when investigating policy support. Rather, their role as antecedents of more proximally related predictors should be investigated.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eOverall, we believe that the results presented here illustrate that predictive machine-learning models can be a valuable extension to the methodological toolbox of social scientists. We thus add to the growing literature on interpretable machine-learning\u003csup\u003e22,52,54,57,58\u003c/sup\u003e by showing that predictive models can: i) be extremely valuable tools for exploratory research, ii) achieve high accuracy in forecasting outcomes of interest and iii) be used for data-driven hypothesis generation to inform subsequent research.\u0026nbsp;\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eStudy 1\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eParticipants and sample size justification\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of N = 1096 Swiss participants completed the survey. The sample size was based on an a-priori power analysis for a different research question not relevant to the purposes of this paper. Given that i) machine-learning models such as random forests were developed to deal with high dimensional data\u003csup\u003e51,53\u003c/sup\u003e and ii) the predictive performance of the models employed in this study was high (see Results), we deem this sample size as sufficient for our purposes. We excluded 40 participants who were either flagged as duplicates, had a straightlining percentage of more than 80%, indicated that they answered randomly in the open comment section (one participant), failed a honeypot question testing for potential bots or had a survey completion time of less than 10 minutes. To ensure our results do not depend on the exclusion of these participants, analyses were also run with N = 1096 and are reported in supplementary materials S4.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eParticipants were randomly assigned to one of the four experimental conditions. Sample sizes per condition are as follows: Direct Air Capture and Carbon Storage (DACCS) Policy: N = 272 (without quality checks) / 264 (with quality checks); Electricity Trading Policy: N = 273 / 257; Flexible EV Charging Policy: N = 275 / 267; Flexible Heating Policy: N = 276 / 268. Descriptive sample characteristics are reported in supplementary materials S1.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStudy design and procedure\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe survey was designed in four blocks. In block one, we checked participants\u0026rsquo; commitment and attention and collected sociodemographic data. Participants who either did not commit to paying attention in the survey, failed the attention check, were younger than 18 years old or did not hold Swiss citizenship were screened out here. In block two, we measured variables relating to individual differences in climate change beliefs, trust in political institutions, general attitudes towards technology, political affiliation, perceived financial scarcity, and personal values in randomized order (see below for details). In block three, participants first read a short explanation on the topic of \u0026ldquo;Flexibility of the electricity system\u0026rdquo;. Subsequently, they were randomly assigned to either receive information on the DACCS Policy, the Electricity Trading Policy, the Flexible EV Charging Policy, or the Flexible Heating Policy and received a short description of the respective policy. Following this, we measured baseline levels of subjective and objective knowledge on this policy, as well as policy support. Next, participants received detailed information on impacts of the policy on the environment (including CO\u003csub\u003e2\u003c/sub\u003e emissions, health and infrastructure impacts), the energy sector (including renewable generation integration, electricity supply reliability, and energy independence), the economy (including costs, economic inequalities and jobs) and the society (comfort, privacy and distributional impacts). All informational materials were pre-rated for understandability and trustworthiness of the information by a separate sample of German and French speaking Swiss citizens (N\u003csub\u003etotal\u003c/sub\u003e = 65; N\u003csub\u003eGerman\u003c/sub\u003e = 32; N\u003csub\u003eFrench\u0026nbsp;\u003c/sub\u003e= 33). While all materials were rated as understandable and trustworthy, we used the comments provided by some participants to adapt small parts of the materials to increase understandability. Informational materials are provided through the OSF pre-registration (https://doi.org/10.17605/OSF.IO/ZN72G)After providing this detailed information we measured informed levels of policy support as well as subjective and objective knowledge. In the last block, we then measured policy-specific beliefs and affective variables.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was approved by the ethics committee of the University of Geneva.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMeasures\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNote: all items mentioned in this section are listed in detail on the OSF (https://doi.org/10.17605/OSF.IO/ZN72G). All items were translated to both German and French by the first author and subsequently back-translated to English by a German and a French native speaker, respectively. In case back-translations did not capture the meaning of the original item, the translated version was adapted in accordance with the translator.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eClimate change beliefs\u003c/p\u003e\n\u003cp\u003eClimate change beliefs were measured using five unipolar subscales of the \u003cem\u003eClimate Change Perceptions scale\u003c/em\u003e\u003csup\u003e59\u003c/sup\u003e:\u0026nbsp;\u003cem\u003eReality, Human causes, Negative consequences, Spatial proximity, and Temporal distance\u003c/em\u003e. All five subscales showed good to excellent internal consistency in our data (\u003cem\u003eReality\u003c/em\u003e: Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.90 (N = 1096; i.e. without data exclusions)/0.91 (N = 1056; i.e. with data exclusions); \u003cem\u003eHuman causes\u003c/em\u003e: Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.94/0.95; \u003cem\u003eNegative consequences\u003c/em\u003e: Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.93/0.93; \u003cem\u003eSpatial proximity\u003c/em\u003e: Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.91/0.91; \u003cem\u003eTemporal distance\u003c/em\u003e: Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.83/0.83). Additionally, we measured climate change worry and personal responsibility to mitigate climate change with three items each (all on a 7-point Likert scale). Both scales were self-constructed based on previous literature\u003csup\u003e60,61\u003c/sup\u003e. Both scales showed excellent internal consistency in our data (\u003cem\u003eClimate Change Worry\u003c/em\u003e: Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.95/0.95; \u003cem\u003ePersonal responsibility\u003c/em\u003e: Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.92/0.92).\u003c/p\u003e\n\u003cp\u003eTrust in political institutions\u003c/p\u003e\n\u003cp\u003eTrust was measured using three subscales with two items each, all on 7-point Likert scales: \u003cem\u003eValue similarity\u003c/em\u003e (Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.79/0.79), \u003cem\u003eCredibility\u003c/em\u003e (Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.82/0.82) and \u003cem\u003eCompetence\u003c/em\u003e (Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.73/0.73). The overall six-item measure of trust showed excellent internal consistency (Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.91/0.91). Items were adapted from previous research\u003csup\u003e62\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eGeneral attitudes towards technology (GATT)\u003c/p\u003e\n\u003cp\u003eAttitudes towards technology were measured using an adapted version of the GATT scale\u003csup\u003e38\u003c/sup\u003e. For each subscale (techno-pessimism, techno-optimism, techno-fix), we chose to use the three items with the highest factor loadings. Additionally, each item was measured on a 7-point Likert scale, instead of the original 6-point scale. The subscale on techno-optimism showed good internal consistency (Cronbach\u0026rsquo;s \u0026alpha; = 0.79/0.80), whereas subscales on techno-pessimism (Cronbach\u0026rsquo;s \u0026alpha; = 0.71/0.71) and techno-fix attitudes (Cronbach\u0026rsquo;s \u0026alpha; = 0.72/0.71) showed acceptable internal consistency.\u003c/p\u003e\n\u003cp\u003ePersonal values\u003c/p\u003e\n\u003cp\u003ePersonal values were measured using an adapted version of the Environmental Portrait Value Questionnaire (E-PVQ)\u003csup\u003e63\u003c/sup\u003e. We used the three items with the highest factor loadings per subscale (biospheric, altruistic, hedonic, and egoistic values). All four subscales showed good internal consistency (Biospheric values: Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.87/0.87; Altruistic values: Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.80.0.79; Hedonic values: Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.80/0.80; Egoistic values: Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.82/0.82).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePerceived financial scarcity\u003c/p\u003e\n\u003cp\u003eWe used a three-item scale created by Berthold et al.\u003csup\u003e39\u003c/sup\u003e to assess perceptions of financial scarcity. Responses were measured on a 7-point Likert scale. The scale showed excellent internal consistency (Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003c/em\u003e = 0.90/0.90).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePolitical affiliation\u003c/p\u003e\n\u003cp\u003ePolitical affiliation was measured using a single item on a 7-point Likert scale from \u0026ldquo;completely left\u0026rdquo; (coded as -3) to \u0026ldquo;completely right (coded as +3) with a neutral anchor \u0026ldquo;neither left nor right\u0026rdquo; coded as 0. The item was adapted from the left-to-right placement scale used in the European Social Survey round 8\u003csup\u003e64\u003c/sup\u003e. Additionally, participants were able to choose \u0026ldquo;prefer not to answer\u0026rdquo; (coded as 99). Based on these responses, we created an additional dichotomous variable which was coded as 0 = answered the question; 1 = preferred not to answer. Subsequently, 99 was recoded to 0 in the \u003cem\u003enumeric\u003c/em\u003e political affiliation variable. Both the numeric and the dichotomous variable are included in the statistical models to control for participants who did not answer the question.\u003c/p\u003e\n\u003cp\u003eKnowledge measures\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eSubjective knowledge\u003c/em\u003e was measured using a single-item on a 7-point Likert scale adapted from previous research\u0026nbsp;\u003csup\u003e65\u003c/sup\u003e, asking participants how much they felt they knew about the respective policy from \u0026ldquo;nothing\u0026rdquo; (1) to \u0026ldquo;a lot\u0026rdquo; (7). \u003cem\u003eObjective knowledge\u003c/em\u003e was measured using five self-constructed \u0026ldquo;True-False\u0026rdquo; questions per policy. For each statement, participants could either choose \u0026ldquo;True\u0026rdquo; (1), \u0026ldquo;False\u0026rdquo; (-1) or \u0026ldquo;I don\u0026rsquo;t know\u0026rdquo; (0). Reverse-coded statements (i.e. statements that were false) were recoded so that participants who chose \u0026ldquo;True\u0026rdquo; would receive a score of -1, while participants who chose \u0026ldquo;False\u0026rdquo; would receive a score of 1. After recoding, we calculated the composite score for objective knowledge per participant by building the sum over all five items and dividing by five.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePolicy support\u003c/p\u003e\n\u003cp\u003eBased on recommendations by Kysel\u0026aacute; and colleagues\u0026nbsp;\u003csup\u003e66\u003c/sup\u003e, we measured policy support with two sub-scales: policy acceptability (Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003csub\u003epre-test\u003c/sub\u003e\u0026nbsp;\u003c/em\u003e= 0.95/0.96; Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003csub\u003epost-test\u003c/sub\u003e\u003c/em\u003e = 0.97/0.97) and potential to support (Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003csub\u003epre-test\u003c/sub\u003e\u003c/em\u003e= 0.88/0.89; Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003csub\u003epost-test\u003c/sub\u003e\u003c/em\u003e= 0.90/0.90). While we had planned to report analyses separately for these two subscales, we opted to combine both subscales into one variable. We took this decision as a) we found a strong correlation (\u003cem\u003er\u003csub\u003epre-test\u003c/sub\u003e\u003c/em\u003e = 0.80/0.81; \u003cem\u003er\u003csub\u003epost-test\u003c/sub\u003e\u003c/em\u003e = 0.86/0.87;) between both subscales and b) the internal consistency of a combined scale was excellent (Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003csub\u003epre-test\u003c/sub\u003e\u003c/em\u003e= 0.95/0.95; Cronbach\u0026rsquo;s \u003cem\u003e\u0026alpha;\u003csub\u003epost-test\u003c/sub\u003e\u003c/em\u003e= 0.96/0.96). Thus, we interpreted the distinction of the two subscales as not practically meaningful in our data.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePolicy-specific beliefs\u003c/p\u003e\n\u003cp\u003eWe assessed the following policy-specific beliefs: perceived social norms (perceived descriptive and dynamic norms; two items each with one item relating to perceived acceptability and perceived potential to support in the Swiss public as well as perceived trends in acceptability and potential to support), perceived effectiveness in mitigating climate change and protecting the environment (for simplicity referred to as perceived environmental impacts), perceived effects on the Swiss electricity system, perceived economic impacts, perceived societal impacts, perceived personal impacts, and perceived distributional impacts (i.e. perceived fairness).\u003c/p\u003e\n\u003cp\u003ePerceived social norm items were self-constructed based on recommendations by Constantino et al.\u003csup\u003e67\u003c/sup\u003e. Descriptive norms were measured on a slider scale from 0-100 (indicating the perceived percentage of the Swiss population who would accept/support the respective policy), while dynamic norms were measured on a 7-point Likert scale. Perceived environmental impacts were assessed with three items on a 7-point Likert scale. We initially intended to construct a separate variable for net environmental impacts based on two of these three items (perceived environmental harms and perceived environmental benefits;\u003csup\u003e68\u003c/sup\u003e). However, given that net environmental impacts correlated strongly with perceived effectiveness in mitigating climate change (\u003cem\u003er\u003c/em\u003e = .71/.72), we decided to build a composite score over all three items. Perceived effects on the Swiss electricity system included perceived effects on electricity supply reliability, perceived effects on flexibility of the electricity system, perceived effects on the share of renewables in the electricity system and perceived effects on energy independence, all assessed with a single item on a 7-point Likert scale. Perceived economic and societal impacts were measured with two 7-point Likert scale items each (based on Geiger et al.\u003csup\u003e68\u003c/sup\u003e). We found strong negative correlations between societal benefits and societal harms (\u003cem\u003er\u003csub\u003ebenefits-harms\u003c/sub\u003e\u003c/em\u003e = -.69/-.71) as well as between economic benefits and economic harms (\u003cem\u003er\u003csub\u003ebenefits-harms\u003c/sub\u003e\u003c/em\u003e = -.67/-.70). We subsequently built a score for net societal and net economic impacts, respectively, by subtracting scores for perceived harms from scores for perceived benefits (i.e. Benefits minus Harms), creating a score where a positive value indicates perceived net positive impacts, and a negative score indicates perceived net negative impacts. Perceived personal impacts included one item each for perceived coerciveness, perceived (personal) financial impact, perceived personal fairness and perceived impacts on quality of life (one item each for health, comfort and privacy). All items were measured on a 7-point Likert scale. Finally, perceived distributional impacts included three items, each measured on a 7-point Likert scale for impacts on low-income earners in Switzerland, Swiss citizens in general and people living outside of Switzerland. Items on perceived personal and perceived distributional impacts were self-constructed.\u003c/p\u003e\n\u003cp\u003eAffect and emotions\u003c/p\u003e\n\u003cp\u003eAffect was measured using a single-item on a 7-point Likert scale (adapted from Spampatti et al.\u003csup\u003e28\u003c/sup\u003e). Additionally, we measured the discrete emotions of worry, anger, hope and pride. Each emotion was assessed on a 7-point Likert scale. Items on discrete emotions were self-constructed.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical analyses\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo identify the strongest predictors of policy support, we used random forests\u003csup\u003e51,53\u003c/sup\u003e. We chose this method as we believe it to be more suitable for exploratory data analysis than conventional statistical analyses such as multiple linear regression\u0026nbsp;\u003csup\u003e52\u003c/sup\u003e. More specifically, we ran a regression random forest to predict overall post-test (i.e. informed) policy support. We set the following random forest hyperparameters based recommendations of Pargent et al.\u0026nbsp;\u003csup\u003e54\u003c/sup\u003e: number of trees = 500, number of features (i.e. predictors) to consider at each split = 18 (based on = \u003cem\u003ep\u003c/em\u003e/3, with \u003cem\u003ep\u003c/em\u003e = number of predictors in the model), and minimum node size to continue splitting a tree = 5. We split our data into a separate \u003cem\u003etraining\u003c/em\u003e (75% of data) and \u003cem\u003etest set\u003c/em\u003e (25% of data) to evaluate model performance. Reported performance measures are the \u003cem\u003eroot mean squared error\u003c/em\u003e (\u003cem\u003eRMSE\u003c/em\u003e; that is, the mean prediction error on the scale of the outcome measure) as well as the percentage of explained variance (\u003cem\u003eR\u0026sup2;\u003c/em\u003e). The random forest was subsequently trained on the training set and model performance was evaluated on the test-set. The model used 53 predictors, including three randomly created variables with a gaussian, uniform and binomial distribution, respectively. To identify the strongest predictors of policy support, we subsequently calculated the mean permutation variable importance (PVI)\u003csup\u003e52,54,69\u003c/sup\u003e based on ten permutations per predictor in the model. A PVI score is then obtained by comparing the difference of the RMSE in the actual model compared to a model with permuted values for variable \u003cem\u003ej\u003c/em\u003e. This process is then repeated for every variable in the model. The random predictors were included to allow a hard cut-off rule for generally relevant and irrelevant predictors. That is, predictors that have a lower mean PVI than any of the three random predictors were deemed irrelevant for predicting policy support. However, given that random forests are \u0026ldquo;greedy\u0026rdquo; algorithms\u003csup\u003e51\u003c/sup\u003e the PVI for some predictors may be inflated when using the hyperparameter settings specified above. This is because at every split, our random forest could choose between 18 predictors. Here, it will always prioritize the predictor that leads to the highest reduction in the RMSE, even though other predictors in this subset may only be marginally worse at predicting support. It is therefore possible that the PVI for some predictors is inflated, while it is deflated for other predictors. To mitigate this, we additionally ran a regression random forest where we set the number of features to consider at each split to 1, effectively forcing the algorithm to always use one randomly selected predictor to predict policy support at each split. To obtain stable predictions, we additionally increased the number of trees to 100.000. PVI is thus reported based on a random forest with these adapted hyperparameter settings, as we believe these PVI scores more accurately reflect the actual variable importance.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWe preregistered to corroborate these results with standard linear regression. However, given that the mean variance inflation factor in this model was at 1.6, we were concerned about potential multicollinearity biasing regression coefficients\u003csup\u003e70\u003c/sup\u003e. We instead corroborated the results of the random forest using ridge regression\u003csup\u003e71\u003c/sup\u003e. We ran 10-fold cross-validated ridge regression models with 1000 repeats. To prevent potential overfitting, we extracted the ridge regression coefficients with the highest lambda-value within one standard error of the cross-validation fold with the lowest lambda-value (i.e. \u0026ldquo;lambda.1se\u0026rdquo; of the \u0026ldquo;cv.glmnet\u0026rdquo; function) for each repeat, thus yielding 1000 regression coefficients per predictor. We subsequently ranked the predictors based on absolute size of the average ridge regression coefficients and compared this ranking to the PVI ranking resulting from the random forest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical software\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData were analyzed using RStudio (R version 4.3.2; R Core Team (2023)) and the following packages (in alphabetical order):broom\u003csup\u003e72\u003c/sup\u003e (version 1.0.5), car\u003csup\u003e73\u003c/sup\u003e (v. 3.1-2), DALEX\u003csup\u003e74\u003c/sup\u003e (v. 2.4.3, Biecek, 2018), DALEXtra\u003csup\u003e75\u003c/sup\u003e (v. 2.3.0, Maksymiuk et al., 2020), easystats\u003csup\u003e76\u003c/sup\u003e (v. 0.7.4, L\u0026uuml;decke et al., 2022), ggbeeswarm\u003csup\u003e77\u003c/sup\u003e (v. 0.7.2, Clarke et al., 2023), ggdist\u003csup\u003e78\u003c/sup\u003e (v. 3.3.1; Kay, 2024), ggpubr\u003csup\u003e79\u003c/sup\u003e (v.0.6.0, Kassambara, 2023), glmnet\u003csup\u003e80, 81, 82\u003c/sup\u003e (v. 4.1-8, Friedman et al. (2010); Simon et al. (2011); Tay et al. (2023)), MASS\u003csup\u003e83\u003c/sup\u003e (v.7.3-60.0.1, Venables \u0026amp; Ripley, 2002), mlr3verse\u003csup\u003e84\u003c/sup\u003e (v. 0.2.8, Lang \u0026amp; Schratz, 2023), psych\u003csup\u003e85\u003c/sup\u003e (v. 2.4.3, Revelle, W., 2024), tidyverse\u003csup\u003e86\u003c/sup\u003e (v. 2.0.0, Wickham et al., 2019), and yhat\u003csup\u003e87\u003c/sup\u003e (v. 2.0-4).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStudy 2\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eParticipants and sample size justification\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eN = 771 participants completed the survey. Sample size was based on a different research question, investigating the role of emotion norms in renewable energy policy support (see OSF pre-registration https://osf.io/qjr4s/?view_only=37a2db2601314d558c0acdaeea80287a). Based on pre-registered exclusion criteria, we excluded six participants from further analyses, given that they were flagged as duplicates. We thus ended up with a final sample of N = 765 participants, representative for the German and French speaking Swiss voting population based on a crossed quota in age, gender and language region. We deem this sample size as sufficient for our purposes as it allows testing predictive performance on a dataset about 72% the size of the original training data with comparable population characteristics as in Study 1. For detailed information on sample characteristics, see Table 2 in supplementary materials S1.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStudy design and procedure\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe used a correlational study design with no experimental manipulation. In the first part of the survey, we collected sociodemographic data. Here, participants who were below the age of 18 or did not possess Swiss citizenship were screened out. We additionally screened out participants who failed an attention check on the subsequent survey page (see OSF pre-registration) or had a reCAPTCHA score below 0.5. Next, we presented a short text summarising the content of the referendum on the Federal Act on Renewable Energy (see supplementary materials S3). The text was intended to be as neutral as possible, while summarising key arguments of both the proponents (i.e. the Swiss government) and opponents (i.e. the referendum committee) of the Federal Act on Renewable Energy. After reading this text, participants completed items on model-relevant predictors, which were measured in the same way as in Study 1. The survey also included several additional items which were not the focus of this paper (see OSF). Finally, we asked participants if they had heard of the referendum before taking part in this survey (yes/no), whether they intend to vote (yes/no/already voted by mail) and depending on their answer, how they intend to vote/have voted (in favour/against/prefer not to answer). On the last page of the survey, participants rated their political affiliation.\u0026nbsp;\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eEthics statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was approved by the ethics committee of the University of Geneva.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMeasures\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll items relevant to the analyses reported here were measured in the same way as in Study 1.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical analyses\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo test the external validity of the model identified in Study 1 we used classification random forests\u003csup\u003e54\u003c/sup\u003e. To this end we dichotomized our measure of policy support so that values higher than zero were coded as \u0026ldquo;support\u0026rdquo; and values equal to or lower than zero were coded as \u0026ldquo;no support\u0026rdquo;. We subsequently used a cross-validated classification forest with 20 folds to calculate the lower bound of the 95% CI of the accuracy of our model in classifying participants as supporters or non-supporters. We pre-registered our interpretation of model-performance as follows:\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003eIf the prediction accuracy for data from Study 2 is at least as high as the lower bound of the 95% CI, the model generalizes to other types of renewable energy policies.\u003c/li\u003e\n \u003cli\u003eIf the prediction accuracy for data from Study 2 is lower than the lower bound of the 95% CI but does not drop by more than 10%, the model still performs reasonably well in predicting renewable energy policy support in general.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eIf the prediction accuracy for data from Study 2 drops by more than 10 percent, the model does not generalize well to predict support for other renewable energy polices.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eThe model was subsequently trained on all observations obtained in Study 1 (N = 1056; clean data) and tested on all observations obtained in Study 2 (N = 765).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical software\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData were analysed using RStudio (R version 4.3.2; R Core Team (2023)) and the following packages (in alphabetical order): ggpubr\u003csup\u003e77\u003c/sup\u003e (v.0.6.0, Kassambara, 2023), mlr3verse\u003csup\u003e84\u003c/sup\u003e (v. 0.2.8, Lang \u0026amp; Schratz, 2023), psych\u003csup\u003e85\u003c/sup\u003e (v. 2.4.3, Revelle, W., 2024), and tidyverse\u003csup\u003e86\u003c/sup\u003e (v. 2.0.0, Wickham et al., 2019).\u003c/p\u003e\n\u003cp\u003eData, code and materials used are available on the associated OSF project: https://osf.io/bpgx2/?view_only=0e9d765dead2418bbb6e723e3cea6eb3\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStudy 3\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eParticipants and sample size justification\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eN = 3812 participants completed Study 3. Our required sample size was determined for a different research question (see OSF pre-registration: https://doi.org/10.17605/OSF.IO/TNQFR). We sampled participants from France, Germany, Italy, the Netherlands, Poland and Spain with a market research institute, employing a crossed-quota for age and gender and a separate quota for education level on the country-level. We excluded participants with a straightlining percentage higher than 80% (n = 91), participants flagged as duplicates (n = 196) as well as participants with a survey duration of 5.93 minutes or less (i.e. the fastest 5% of participants; n = 177). We thus ended up with an analysis sample of N = 3355 participants. For detailed information on sample characteristics, see Tables 3 and 4 in supplementary materials S1.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStudy design and procedure\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe survey was structured in two parts. Part one of the survey employed a between-subjects design. Participants first answered demographic items, followed by an instructional attention check. Participants who failed this check were screened out of the survey. Participants who passed the attention check were randomly assigned to receive information about one of the four climate mitigation measures (bioenergy with carbon capture and storage, green hydrogen, afforestation and reforestation or switching to sustainable diets). Participants were then asked how much they think they know about the respective mitigation measure. Following this, we presented basic information on this mitigation measure (including information on the Intergovernmental Panel on Climate Change (IPCC) and the global goal of reaching net-zero greenhouse gas emissions by 2050, how the respective mitigation measures works and the estimated emissions reductions until 2050 if employed, based on the IPCC\u003csup\u003e1\u003c/sup\u003e). Here, we used an attention check asking participants one true-false question regarding the text they just read. Participants who provided the correct response to this question then went on answer items relevant to the analyses in this paper (policy support, perceived social norms, policy-specific beliefs, affect and emotions). The survey then continued with part two, in which additional information was presented and items on policy support and emotions were presented again. All data used for the analyses was obtained in part one of the survey. We will therefore not describe part two in detail.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was approved by the ethics committee of the University of Geneva.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMeasures\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eItems relevant to the analyses reported here were measured in the same way as in Studies 1 and 2. However, there were two exceptions to this. First, the item asking participants whether they would vote in favour or against the implementation of the mitigation measure \u0026ldquo;in a general referendum\u0026rdquo; was changed to \u0026ldquo;in a direct vote\u0026rdquo;, given that referendums are not as typical in representative democracies as in direct democracies. Second, all references to \u0026ldquo;Switzerland\u0026rdquo; in the items from Studies 1 and 2 were changed so that the name of the country the participant was from was used instead.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical analyses\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe used the same analysis approach and cut-off criteria as in Study 2. Model accuracy was tested on the entire data and 34subsamples. These were created based on the country participants were from and the mitigation measure they were assigned to (resulting in 24 subsamples) as well as overall subsamples for each country and each mitigation measure.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical software\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData were analyzed using RStudio (R version 4.3.2; R Core Team (2023)) and the following packages (in alphabetical order): easystats\u003csup\u003e74\u003c/sup\u003e (v. 0.7.4, L\u0026uuml;decke et al., 2022), ggpubr\u003csup\u003e77\u003c/sup\u003e (v.0.6.0, Kassambara, 2023), mlr3verse\u003csup\u003e84\u003c/sup\u003e (v. 0.2.8, Lang \u0026amp; Schratz, 2023), psych\u003csup\u003e85\u003c/sup\u003e (v. 2.4.3, Revelle, W., 2024), and tidyverse\u003csup\u003e86\u003c/sup\u003e (v. 2.0.0, Wickham et al., 2019).\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eall participants had to provide explicit consent to taking part in the study as well as to the use and publication of their anonymised data and aggregated results for scientific purposes.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe would like to thank Ben Meuleman for his feedback and advice on our modelling approach. Additionally, we would also like to thank the members of the SWEET PATHFNDR consortium for their valuable feedback on the content of the policy information in Study 1, especially Christian Moretti, Siobhan Powell, Christian Winzer and Florian Baader as well as the members of the Consumer Decision and Sustainable Behavior Lab, Mario Herberz and Manon Gouiran, for their valuable feedback on our manuscript. Finally, we would like to thank PRISMA project experts, colleagues from the University of Geneva’s Renewable Energy Systems group, Joanna Sobocińska, Marta Brucka, Mellano Ocèane, and Francesco Sasso for the translation and revision of the technical information used in Study 3.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eAuthor Contributions: CRediT\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Conceptualization: MK, ET \u0026amp; TB\u0026nbsp;\u003c/em\u003e(Studies 1-3), IV (Study 2), VS (Study 3)\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Data curation: MK\u0026nbsp;\u003c/em\u003e(Studies 1-2), IV (Study 2), VS (Study 3)\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Formal analysis:\u0026nbsp;\u003c/em\u003eMK (Studies 1-3)\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Funding acquisition: ET \u0026amp; TB\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Methodology:\u0026nbsp;\u003c/em\u003eMK, ET \u0026amp; TB\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Visualization:\u0026nbsp;\u003c/em\u003eMK (Studies 1-3)\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Writing – first draft:\u0026nbsp;\u003c/em\u003eMK\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Writing – review \u0026amp; editing:\u003c/em\u003e TB, ET,VS \u0026amp; IV\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work received funding from the Swiss Federal Office of Energy (SFOE) as part of the SWEET project PATHFNDR and from the Swiss State Secretariat for Education, Research, and Innovation SEFRI for the project PRISMA “Net zero pathway research through integrated assessment model advancements” (project no. 101081604). The authors bear sole responsibility for the conclusions and the results.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eClimate Change 2022: Mitigation of Climate Change. (IPCC, Geneva, 2022).\u003c/li\u003e\n\u003cli\u003eDahl, R. A. \u003cem\u003eOn Democracy\u003c/em\u003e. (Yale University Press, New Haven London, 1998).\u003c/li\u003e\n\u003cli\u003eFiorino, D. J. Environmental Risk and Democratic Process: A Critical Review. \u003cem\u003eColumbia J. Environ. Law\u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 501\u0026ndash;548 (1989).\u003c/li\u003e\n\u003cli\u003eYeganeh, A. J., McCoy, A. P. \u0026amp; Schenk, T. Determinants of climate change policy adoption: A meta-analysis. \u003cem\u003eUrban Clim.\u003c/em\u003e\u003cstrong\u003e31\u003c/strong\u003e, 100547 (2020).\u003c/li\u003e\n\u003cli\u003eWong-Parodi, G., Krishnamurti, T., Davis, A., Schwartz, D. \u0026amp; Fischhoff, B. A decision science approach for integrating social science in climate and energy solutions. \u003cem\u003eNat. Clim. Change\u003c/em\u003e\u003cstrong\u003e6\u003c/strong\u003e, 563\u0026ndash;569 (2016).\u003c/li\u003e\n\u003cli\u003eBergquist, M., Nilsson, A., Harring, N. \u0026amp; Jagers, S. C. Meta-analyses of fifteen determinants of public opinion about climate change taxes and laws. \u003cem\u003eNat. Clim. Change\u003c/em\u003e\u003cstrong\u003e12\u003c/strong\u003e, 235\u0026ndash;240 (2022).\u003c/li\u003e\n\u003cli\u003eDrews, S. \u0026amp; Van Den Bergh, J. C. J. M. What explains public support for climate policies? A review of empirical and experimental studies. \u003cem\u003eClim. Policy\u003c/em\u003e\u003cstrong\u003e16\u003c/strong\u003e, 855\u0026ndash;876 (2016).\u003c/li\u003e\n\u003cli\u003eEjel\u0026ouml;v, E. \u0026amp; Nilsson, A. Individual Factors Influencing Acceptability for Environmental Policies: A Review and Research Agenda. \u003cem\u003eSustainability\u003c/em\u003e\u003cstrong\u003e12\u003c/strong\u003e, 2404 (2020).\u003c/li\u003e\n\u003cli\u003eGoldberg, M. H., Gustafson, A., Ballew, M. T., Rosenthal, S. A. \u0026amp; Leiserowitz, A. Identifying the most important predictors of support for climate policy in the United States. \u003cem\u003eBehav. Public Policy\u003c/em\u003e\u003cstrong\u003e5\u003c/strong\u003e, 480\u0026ndash;502 (2021).\u003c/li\u003e\n\u003cli\u003eClulow, Z., Ferguson, M., Ashworth, P. \u0026amp; Reiner, D. Comparing public attitudes towards energy technologies in Australia and the UK: The role of political ideology. \u003cem\u003eGlob. Environ. Change\u003c/em\u003e\u003cstrong\u003e70\u003c/strong\u003e, 102327 (2021).\u003c/li\u003e\n\u003cli\u003eDing, D., Maibach, E. W., Zhao, X., Roser-Renouf, C. \u0026amp; Leiserowitz, A. Support for climate policy and societal action are linked to perceptions about scientific agreement. \u003cem\u003eNat. Clim. Change\u003c/em\u003e\u003cstrong\u003e1\u003c/strong\u003e, 462\u0026ndash;466 (2011).\u003c/li\u003e\n\u003cli\u003eSyropoulos, S. \u0026amp; Markowitz, E. M. Perceived responsibility to address climate change consistently relates to increased pro-environmental attitudes, behaviors and policy support: Evidence across 23 countries. \u003cem\u003eJ. Environ. Psychol.\u003c/em\u003e\u003cstrong\u003e83\u003c/strong\u003e, 101868 (2022).\u003c/li\u003e\n\u003cli\u003eFairbrother, M., Johansson Sev\u0026auml;, I. \u0026amp; Kulin, J. Political trust and the relationship between climate change beliefs and support for fossil fuel taxes: Evidence from a survey of 23 European countries. \u003cem\u003eGlob. Environ. Change\u003c/em\u003e\u003cstrong\u003e59\u003c/strong\u003e, 102003 (2019).\u003c/li\u003e\n\u003cli\u003eSwim, J. K. \u0026amp; Geiger, N. Policy attributes, perceived impacts, and climate change policy preferences. \u003cem\u003eJ. Environ. Psychol.\u003c/em\u003e\u003cstrong\u003e77\u003c/strong\u003e, 101673 (2021).\u003c/li\u003e\n\u003cli\u003eOgunbode, C. A. \u003cem\u003eet al.\u003c/em\u003e Climate justice beliefs related to climate action and policy support around the world. \u003cem\u003eNat. Clim. Change\u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 1144\u0026ndash;1150 (2024).\u003c/li\u003e\n\u003cli\u003eHuber, R. A., Wicki, M. L. \u0026amp; Bernauer, T. Public support for environmental policy depends on beliefs concerning effectiveness, intrusiveness, and fairness. \u003cem\u003eEnviron. Polit.\u003c/em\u003e\u003cstrong\u003e29\u003c/strong\u003e, 649\u0026ndash;673 (2020).\u003c/li\u003e\n\u003cli\u003eThaller, A. \u003cem\u003eet al.\u003c/em\u003e When perceived fairness and acceptance go hand in hand\u0026ndash;Drivers of regulatory and economic policies for low-carbon mobility. \u003cem\u003ePLOS Clim.\u003c/em\u003e\u003cstrong\u003e2\u003c/strong\u003e, e0000157 (2023).\u003c/li\u003e\n\u003cli\u003eBoon-Falleur, M., Grandin, A., Baumard, N. \u0026amp; Chevallier, C. Leveraging social cognition to promote effective climate change mitigation. \u003cem\u003eNat. Clim. Change\u003c/em\u003e\u003cstrong\u003e12\u003c/strong\u003e, 332\u0026ndash;338 (2022).\u003c/li\u003e\n\u003cli\u003eGoldberg, M. H., Van Der Linden, S., Leiserowitz, A. \u0026amp; Maibach, E. W. Perceived Social Consensus Can Reduce Ideological Biases on Climate Change. \u003cem\u003eEnviron. Behav.\u003c/em\u003e\u003cstrong\u003e52\u003c/strong\u003e, (2019).\u003c/li\u003e\n\u003cli\u003eSokoloski, R., Markowitz, E. M. \u0026amp; Bidwell, D. Public estimates of support for offshore wind energy: False consensus, pluralistic ignorance, and partisan effects. \u003cem\u003eEnergy Policy\u003c/em\u003e\u003cstrong\u003e112\u003c/strong\u003e, 45\u0026ndash;55 (2018).\u003c/li\u003e\n\u003cli\u003eSmith, N. \u0026amp; Leiserowitz, A. The Role of Emotion in Global Warming Policy Support and Opposition: The Role of Emotion in Global Warming Policy Support and Opposition. \u003cem\u003eRisk Anal.\u003c/em\u003e\u003cstrong\u003e34\u003c/strong\u003e, 937\u0026ndash;948 (2014).\u003c/li\u003e\n\u003cli\u003eTodorova, B. \u003cem\u003eet al.\u003c/em\u003e Machine learning identifies key individual and nation-level factors predicting climate-relevant beliefs and behaviors. Preprint at https://doi.org/10.31219/osf.io/um69t (2025).\u003c/li\u003e\n\u003cli\u003eBrosch, T. \u0026amp; Steg, L. Leveraging emotion for sustainable action. \u003cem\u003eOne Earth\u003c/em\u003e\u003cstrong\u003e4\u003c/strong\u003e, 1693\u0026ndash;1703 (2021).\u003c/li\u003e\n\u003cli\u003eBrosch, T. \u0026amp; Sauter, D. Emotions and the Climate Crisis: A Research Agenda for an Affective Sustainability Science. \u003cem\u003eEmot. Rev.\u003c/em\u003e\u003cstrong\u003e15\u003c/strong\u003e, 253\u0026ndash;257 (2023).\u003c/li\u003e\n\u003cli\u003eJobin, M. \u0026amp; Siegrist, M. We choose what we like \u0026ndash; Affect as a driver of electricity portfolio choice. \u003cem\u003eEnergy Policy\u003c/em\u003e\u003cstrong\u003e122\u003c/strong\u003e, 736\u0026ndash;747 (2018).\u003c/li\u003e\n\u003cli\u003eJobin, M., Visschers, V. H. M., Van Vliet, O. P. R., \u0026Aacute;rvai, J. \u0026amp; Siegrist, M. Affect or information? Examining drivers of public preferences of future energy portfolios in Switzerland. \u003cem\u003eEnergy Res. Soc. Sci.\u003c/em\u003e\u003cstrong\u003e52\u003c/strong\u003e, 20\u0026ndash;29 (2019).\u003c/li\u003e\n\u003cli\u003eRinscheid, A. \u0026amp; W\u0026uuml;stenhagen, R. Divesting, Fast and Slow: Affective and Cognitive Drivers of Fading Voter Support for a Nuclear Phase-Out. \u003cem\u003eEcol. Econ.\u003c/em\u003e\u003cstrong\u003e152\u003c/strong\u003e, 51\u0026ndash;61 (2018).\u003c/li\u003e\n\u003cli\u003eSpampatti, T., Hahnel, U. J. J., Trutnevyte, E. \u0026amp; Brosch, T. Short and long-term dominance of negative information in shaping public energy perceptions: The case of shallow geothermal systems. \u003cem\u003eEnergy Policy\u003c/em\u003e\u003cstrong\u003e167\u003c/strong\u003e, 113070 (2022).\u003c/li\u003e\n\u003cli\u003eGeiger, N., Dwyer, T. \u0026amp; Swim, J. K. Hopium or empowering hope? A meta-analysis of hope and climate engagement. \u003cem\u003eFront. Psychol.\u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 1139427 (2023).\u003c/li\u003e\n\u003cli\u003eGregersen, T., Andersen, G. \u0026amp; Tvinnereim, E. The strength and content of climate anger. \u003cem\u003eGlob. Environ. Change\u003c/em\u003e\u003cstrong\u003e82\u003c/strong\u003e, 102738 (2023).\u003c/li\u003e\n\u003cli\u003eMarlon, J. R. \u003cem\u003eet al.\u003c/em\u003e How Hope and Doubt Affect Climate Change Mobilization. \u003cem\u003eFront. Commun.\u003c/em\u003e\u003cstrong\u003e4\u003c/strong\u003e, 20 (2019).\u003c/li\u003e\n\u003cli\u003eMyers, T. A., Roser-Renouf, C., Leiserowitz, A. \u0026amp; Maibach, E. Emotional signatures of climate policy support. \u003cem\u003ePLOS Clim.\u003c/em\u003e\u003cstrong\u003e3\u003c/strong\u003e, e0000381 (2024).\u003c/li\u003e\n\u003cli\u003eSchneider, C. R., Zaval, L. \u0026amp; Markowitz, E. M. Positive emotions and climate change. \u003cem\u003eCurr. Opin. Behav. Sci.\u003c/em\u003e\u003cstrong\u003e42\u003c/strong\u003e, 114\u0026ndash;120 (2021).\u003c/li\u003e\n\u003cli\u003eWang, S., Leviston, Z., Hurlstone, M., Lawrence, C. \u0026amp; Walker, I. Emotions predict policy support: Why it matters how people feel about climate change. \u003cem\u003eGlob. Environ. Change\u003c/em\u003e\u003cstrong\u003e50\u003c/strong\u003e, 25\u0026ndash;40 (2018).\u003c/li\u003e\n\u003cli\u003eWeber, E. U. \u0026amp; Constantino, S. M. All Hearts and Minds on Deck: Hope Motivates Climate Action by Linking the Present and the Future. \u003cem\u003eEmot. Rev.\u003c/em\u003e\u003cstrong\u003e15\u003c/strong\u003e, 293\u0026ndash;297 (2023).\u003c/li\u003e\n\u003cli\u003eSwim, J. K., Guerriero, J. G., Gasper, K., DeCoster, J. \u0026amp; Lengieza, M. L. Emotions and policy information predicting water-quality policy support. \u003cem\u003eJ. Environ. Psychol.\u003c/em\u003e\u003cstrong\u003e98\u003c/strong\u003e, 102385 (2024).\u003c/li\u003e\n\u003cli\u003eZeelenberg, M., Nelissen, R. M. A., Breugelmans, S. M. \u0026amp; Pieters, R. On emotion specificity in decision making: Why feeling is for doing. \u003cem\u003eJudgm. Decis. Mak.\u003c/em\u003e\u003cstrong\u003e3\u003c/strong\u003e, 18\u0026ndash;27 (2008).\u003c/li\u003e\n\u003cli\u003eCologna, V., Berthold, A., Kreissel, A. L. \u0026amp; Siegrist, M. Attitudes towards technology and their relationship with pro-environmental behaviour: Development and validation of the GATT scale. \u003cem\u003eJ. Environ. Psychol.\u003c/em\u003e\u003cstrong\u003e95\u003c/strong\u003e, 102258 (2024).\u003c/li\u003e\n\u003cli\u003eBerthold, A., Cologna, V., Hardmeier, M. \u0026amp; Siegrist, M. Drop some money! The influence of income and subjective financial scarcity on pro-environmental behaviour. \u003cem\u003eJ. Environ. Psychol.\u003c/em\u003e\u003cstrong\u003e91\u003c/strong\u003e, 102149 (2023).\u003c/li\u003e\n\u003cli\u003eSparkman, G. \u0026amp; Walton, G. M. Dynamic Norms Promote Sustainable Behavior, Even if It Is Counternormative. \u003cem\u003ePsychol. Sci.\u003c/em\u003e\u003cstrong\u003e28\u003c/strong\u003e, 1663\u0026ndash;1674 (2017).\u003c/li\u003e\n\u003cli\u003eSparkman, G. \u0026amp; Walton, G. M. Witnessing change: Dynamic norms help resolve diverse barriers to personal change. \u003cem\u003eJ. Exp. Soc. Psychol.\u003c/em\u003e\u003cstrong\u003e82\u003c/strong\u003e, 238\u0026ndash;252 (2019).\u003c/li\u003e\n\u003cli\u003eKunnas, S. \u0026amp; Trutnevyte, E. Informed minds, opinions aligned? Informed perceptions towards solar PV locations in Switzerland. \u003cem\u003eEnviron. Res. Commun.\u003c/em\u003e\u003cstrong\u003e7\u003c/strong\u003e, 075030 (2025).\u003c/li\u003e\n\u003cli\u003eVolken, S. P., Xexakis, G. \u0026amp; Trutnevyte, E. Perspectives of Informed Citizen Panel on Low-Carbon Electricity Portfolios in Switzerland and Longer-Term Evaluation of Informational Materials. \u003cem\u003eEnviron. Sci. Technol.\u003c/em\u003e acs.est.8b01265 (2018) doi:10.1021/acs.est.8b01265.\u003c/li\u003e\n\u003cli\u003eBlastland, M., Freeman, A. L. J., Van Der Linden, S., Marteau, T. M. \u0026amp; Spiegelhalter, D. Five rules for evidence communication. \u003cem\u003eNature\u003c/em\u003e\u003cstrong\u003e587\u003c/strong\u003e, 362\u0026ndash;364 (2020).\u003c/li\u003e\n\u003cli\u003eKerr, J. R., Schneider, C. R., Freeman, A. L. J., Marteau, T. \u0026amp; van der Linden, S. Transparent communication of evidence does not undermine public trust in evidence. \u003cem\u003ePNAS Nexus\u003c/em\u003e\u003cstrong\u003e1\u003c/strong\u003e, pgac280 (2022).\u003c/li\u003e\n\u003cli\u003ePickering, B., Lombardi, F. \u0026amp; Pfenninger, S. Diversity of options to eliminate fossil fuels and reach carbon neutrality across the entire European energy system. \u003cem\u003eJoule\u003c/em\u003e\u003cstrong\u003e6\u003c/strong\u003e, 1253\u0026ndash;1276 (2022).\u003c/li\u003e\n\u003cli\u003eEcker, U., Roozenbeek, J. \u0026amp; Lewandowsky, S. Misinformation remains a threat to democracy. \u003cem\u003eNature\u003c/em\u003e (2024) doi:https://doi.org/10.1038/d41586-024-01587-3.\u003c/li\u003e\n\u003cli\u003eBrexit was wrong, say 57% of British voters. \u003cem\u003eThe Economist\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eWhy most people regret Brexit. \u003cem\u003eThe Economist\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eMitchell, M. C., David Luhnow and Josh. Essay | Brexit Backlash: Brits Now Regret Their Populist Revolt. \u003cem\u003eWSJ\u003c/em\u003e https://www.wsj.com/world/uk/brexit-british-regret-uk-election-837cbf4c (2024).\u003c/li\u003e\n\u003cli\u003eBreiman, L. Random Forests. \u003cem\u003eMach. Learn.\u003c/em\u003e\u003cstrong\u003e45\u003c/strong\u003e, 5\u0026ndash;32 (2001).\u003c/li\u003e\n\u003cli\u003eFife, D. A. \u0026amp; D\u0026rsquo;Onofrio, J. Common, uncommon, and novel applications of random forest in psychological research. \u003cem\u003eBehav. Res. Methods\u003c/em\u003e\u003cstrong\u003e55\u003c/strong\u003e, 2447\u0026ndash;2466 (2022).\u003c/li\u003e\n\u003cli\u003eJames, G., Witten, D., Hastie, T. \u0026amp; Tibshirani, R. An introduction to statistical learning with applications in R. \u003cem\u003eStat. Theory Relat. Fields\u003c/em\u003e\u003cstrong\u003e6\u003c/strong\u003e, 87\u0026ndash;87 (2022).\u003c/li\u003e\n\u003cli\u003ePargent, F., Schoedel, R. \u0026amp; Stachl, C. Best Practices in Supervised Machine Learning: A Tutorial for Psychologists. \u003cem\u003eAdv. Methods Pract. Psychol. Sci.\u003c/em\u003e\u003cstrong\u003e6\u003c/strong\u003e, 1\u0026ndash;35 (2023).\u003c/li\u003e\n\u003cli\u003eFreiburghaus, R., Leemann, L., Wasserfallen, F., Willi, T. \u0026amp; Yin, J. \u003cem\u003e20 Minuten-/Tamedia- Abstimmungsumfrage - Eidgen\u0026ouml;ssische Volksabstimmungen vom 9. \u003c/em\u003e\u003cem\u003eJuni\u003c/em\u003e. https://www.tamedia.ch/de/unternehmen/umfragen (2024).\u003c/li\u003e\n\u003cli\u003eMousson, M. 2. SRG-Trendumfrage zur eidg. Abstimmung vom 9. Juni 2024. https://www.gfsbern.ch/de/news/2-srg-trendumfrage-zur-eidg-abstimmung-vom-9-juni-2024/ (2024).\u003c/li\u003e\n\u003cli\u003eLee, T. M., Markowitz, E. M., Howe, P. D., Ko, C.-Y. \u0026amp; Leiserowitz, A. A. Predictors of public climate change awareness and risk perception around the world. \u003cem\u003eNat. Clim. Change\u003c/em\u003e\u003cstrong\u003e5\u003c/strong\u003e, 1014\u0026ndash;1020 (2015).\u003c/li\u003e\n\u003cli\u003eYarkoni, T. \u0026amp; Westfall, J. Choosing Prediction Over Explanation in Psychology: Lessons From Machine Learning. \u003cem\u003ePerspect. Psychol. Sci.\u003c/em\u003e\u003cstrong\u003e12\u003c/strong\u003e, 1100\u0026ndash;1122 (2017).\u003c/li\u003e\n\u003cli\u003eVan Valkengoed, A. M., Steg, L. \u0026amp; Perlaviciute, G. Development and validation of a climate change perceptions scale. \u003cem\u003eJ. Environ. Psychol.\u003c/em\u003e\u003cstrong\u003e76\u003c/strong\u003e, 101652 (2021).\u003c/li\u003e\n\u003cli\u003eBouman, T. \u003cem\u003eet al.\u003c/em\u003e When worry about climate change leads to climate action: How values, worry and personal responsibility relate to various climate actions. \u003cem\u003eGlob. Environ. Change\u003c/em\u003e\u003cstrong\u003e62\u003c/strong\u003e, 102061 (2020).\u003c/li\u003e\n\u003cli\u003eWilson, R. S., Zwickle, A. \u0026amp; Walpole, H. Developing a Broadly Applicable Measure of Risk Perception. \u003cem\u003eRisk Anal.\u003c/em\u003e\u003cstrong\u003e39\u003c/strong\u003e, 777\u0026ndash;791 (2019).\u003c/li\u003e\n\u003cli\u003eSpampatti, T., Brosch, T., Trutnevyte, E. \u0026amp; Hahnel, U. J. J. A Preregistered Field Study of the Trust Inoculation Against a Negative Event Involving Geothermal Energy Systems. \u003cem\u003eCollabra Psychol.\u003c/em\u003e\u003cstrong\u003e9\u003c/strong\u003e, 89755 (2023).\u003c/li\u003e\n\u003cli\u003eBouman, T., Steg, L. \u0026amp; Kiers, H. A. L. Measuring Values in Environmental Research: A Test of an Environmental Portrait Value Questionnaire. \u003cem\u003eFront. Psychol.\u003c/em\u003e\u003cstrong\u003e9\u003c/strong\u003e, 564 (2018).\u003c/li\u003e\n\u003cli\u003eEuropean Social Survey. \u003cem\u003eESS Round 8 Source Questionnaire\u003c/em\u003e. (2016).\u003c/li\u003e\n\u003cli\u003eWong-Parodi, G. \u0026amp; Berlin Rubin, N. Exploring how climate change subjective attribution, personal experience with extremes, concern, and subjective knowledge relate to pro-environmental attitudes and behavioral intentions in the United States. \u003cem\u003eJ. Environ. Psychol.\u003c/em\u003e\u003cstrong\u003e79\u003c/strong\u003e, 101728 (2022).\u003c/li\u003e\n\u003cli\u003eKysel\u0026aacute;, E., \u0026Scaron;časn\u0026yacute;, M. \u0026amp; Zvěřinov\u0026aacute;, I. Attitudes toward climate change mitigation policies: a review of measures and a construct of policy attitudes. \u003cem\u003eClim. Policy\u003c/em\u003e\u003cstrong\u003e19\u003c/strong\u003e, 878\u0026ndash;892 (2019).\u003c/li\u003e\n\u003cli\u003eConstantino, S. M. \u003cem\u003eet al.\u003c/em\u003e Scaling Up Change: A Critical Review and Practical Guide to Harnessing Social Norms for Climate Action. \u003cem\u003ePsychol. Sci. Public Interest\u003c/em\u003e\u003cstrong\u003e23\u003c/strong\u003e, 50\u0026ndash;97 (2022).\u003c/li\u003e\n\u003cli\u003eGeiger, N., Swim, J. K. \u0026amp; Benson, L. Using the three-pillar model of sustainability to understand lay reactions to climate policy: A multilevel approach. \u003cem\u003eEnviron. Sci. Policy\u003c/em\u003e\u003cstrong\u003e126\u003c/strong\u003e, 132\u0026ndash;141 (2021).\u003c/li\u003e\n\u003cli\u003eHenninger, M., Debelak, R., Rothacher, Y. \u0026amp; Strobl, C. Interpretable machine learning for psychological research: Opportunities and pitfalls. \u003cem\u003ePsychol. Methods\u003c/em\u003e Advance online publication (2023) doi:10.1037/met0000560.\u003c/li\u003e\n\u003cli\u003eBowerman, B. L. \u0026amp; O\u0026rsquo;Connell, R. T. \u003cem\u003eLinear Statistical Models\u003c/em\u003e\u003cem\u003e \u003c/em\u003e\u003cem\u003e: An Applied Approach\u003c/em\u003e. (Boston : PWS-Kent Pub. Co., 1990).\u003c/li\u003e\n\u003cli\u003eHoerl, A. E. \u0026amp; Kennard, R. W. Ridge Regression: Biased Estimation for Nonorthogonal Problems. \u003cem\u003eTechnometrics\u003c/em\u003e\u003cstrong\u003e12\u003c/strong\u003e, 55\u0026ndash;67 (1970). \u003cstrong\u003eR Packages used (Studies 1-3)\u003c/strong\u003e\u003c/li\u003e\n\u003c/ol\u003e\n\u003col start=\"72\"\u003e\n\u003cli\u003eRobinson, D., Hayes, A., \u0026amp; Couch, S. (2023). broom: Convert Statistical Objects into Tidy Tibbles. R package version 1.0.5, https://CRAN.R-project.org/package=broom.\u003c/li\u003e\n\u003cli\u003eFox, J., \u0026amp; Weisberg, S. (2019). An R Companion to Applied Regression, Third edition. Sage, Thousand Oaks CA. https://socialsciences.mcmaster.ca/jfox/Books/Companion/.\u003c/li\u003e\n\u003cli\u003eBiecek, P. (2018). DALEX: Explainers for Complex Predictive Models in R. Journal of Machine Learning Research, 19(84), 1-5. https://jmlr.org/papers/v19/18-416.html.\u003c/li\u003e\n\u003cli\u003eMaksymiuk, S., Gosiewska, A., \u0026amp; Biecek, P. (2020). Landscape of R packages for eXplainable Artificial Intelligence. arXiv. Pages 6, 7, 11, 15, https://arxiv.org/abs/2009.13248.\u003c/li\u003e\n\u003cli\u003eL\u0026uuml;decke, D., Patil, I., Ben-Shachar, M. S., Wiernik, B. M., Bacher, E., Th\u0026eacute;riault, R., \u0026amp; Makowski, D. (2022). easystats: Framework for Easy Statistical Modeling, Visualization, and Reporting. CRAN. doi:10.32614/CRAN.package.easystats\u003c/li\u003e\n\u003cli\u003eClarke, E., Sherrill-Mix, S., \u0026amp; Dawson, C. (2023). ggbeeswarm: Categorical Scatter (Violin Point) Plots. R package version 0.7.2, https://CRAN.R-project.org/package=ggbeeswarm.\u003c/li\u003e\n\u003cli\u003eKay, M. (2024). \u0026ldquo;ggdist: Visualizations of Distributions and Uncertainty in the Grammar of Graphics.\u0026rdquo; IEEE Transactions on Visualization and Computer Graphics, 1-11. doi:10.1109/TVCG.2023.3327195.\u003c/li\u003e\n\u003cli\u003eKassambara, A. (2023). ggpubr: 'ggplot2' Based Publication Ready Plots. R package version 0.6.0, https://CRAN.R-project.org/package=ggpubr.\u003c/li\u003e\n\u003cli\u003eFriedman, J., Tibshirani, R., \u0026amp; Hastie, T. (2010). Regularization Paths for Generalized Linear Models via Coordinate Descent. Journal of Statistical Software, 33(1), 1-22. doi:10.18637/jss.v033.i01.\u003c/li\u003e\n\u003cli\u003eSimon, N., Friedman, J., Tibshirani, R., \u0026amp; Hastie ,T. (2011). Regularization Paths for Cox's Proportional Hazards Model via Coordinate Descent. Journal of Statistical Software, 39(5), 1-13. doi:10.18637/jss.v039.i05\u003c/li\u003e\n\u003cli\u003eTay, J.K., Narasimhan, B., \u0026amp; Hastie, T. (2023). Elastic Net Regularization Paths for All Generalized Linear Models. Journal of Statistical Software, 106(1), 1-31. doi:10.18637/jss.v106.i01\u003c/li\u003e\n\u003cli\u003eVenables, W. N. \u0026amp; Ripley, B. D. (2002) Modern Applied Statistics with S. Fourth Edition. Springer, New York. ISBN 0-387-95457-0\u003c/li\u003e\n\u003cli\u003eLang, M., \u0026amp; Schratz, P. (2023). mlr3verse: Easily Install and Load the 'mlr3' Package Family. R package version 0.2.8, https://CRAN.R-project.org/package=mlr3verse\u003c/li\u003e\n\u003cli\u003eRevelle, W. (2024). psych: Procedures for Psychological, Psychometric, and Personality Research. Northwestern University, Evanston, Illinois. R package version 2.4.3, https://CRAN.R-project.org/package=psych.\u003c/li\u003e\n\u003cli\u003eWickham, H., Averick, M., Bryan, J., Chang, W., McGowan, L.D., Fran\u0026ccedil;ois, R., Grolemund, G., Hayes, A., Henry, L., Hester, J., Kuhn, M., Pedersen, T.L., Miller, E., Bache, S.M., M\u0026uuml;ller, K., Ooms, J., Robinson, D., Seidel, D.P., Spinu, V., Takahashi, K., Vaughan, D., Wilke, C., Woo, K., \u0026amp; Yutani, H. (2019). Welcome to the tidyverse. Journal of Open Source Software, 4(43), 1686. doi:10.21105/joss.01686.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"nature-portfolio","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Nature Portfolio","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"ejp","reportingPortfolio":"","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7658135/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7658135/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Widespread public policy support is crucial for efforts to decarbonize the energy system. Past research has proposed numerous variables as potential key predictors of support for mitigation policies but has failed to i) comprehensively evaluate and compare their predictive importance and ii) identify the strongest predictors within specific mitigation domains. We used machine learning models to identify the strongest predictors of energy policy support in informed citizens (Study 1), validate these results in a real-world referendum on renewable energy in Switzerland (Study 2) and test their generalizability to other climate mitigation policies across six European countries (Study 3). We identified affective responses, societal and environmental policy-impact beliefs, fairness perceptions and perceived trends in policy support over time as the strongest predictors of energy policy support. Using these predictors, we achieved high accuracy in predicting support for a real-world referendum as well as support for different climate mitigation measures across Europe.","manuscriptTitle":"Identifying and validating the strongest predictors of informed energy policy support across Europe","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-30 17:31:34","doi":"10.21203/rs.3.rs-7658135/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"nature-energy","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"nenergy","sideBox":"Learn more about [Nature Energy](http://www.nature.com/nenergy/)","snPcode":"","submissionUrl":"","title":"Nature Energy","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature Research","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"666cf482-20dc-47ac-82b2-121665cd11a8","owner":[],"postedDate":"September 30th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":55322490,"name":"Social science/Psychology"},{"id":55322491,"name":"Scientific community and society/Social sciences/Communication"},{"id":55322492,"name":"Physical sciences/Energy science and technology/Renewable energy"},{"id":55322493,"name":"Scientific community and society/Energy and society/Energy policy"}],"tags":[],"updatedAt":"2026-03-31T14:15:42+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-30 17:31:34","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7658135","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7658135","identity":"rs-7658135","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}