Ecological Economics 187 (2021) 107092 Contents lists available at ScienceDirect Ecological Economics journal homepage: www.elsevier.com/locate/ecolecon Methodological and Ideological Options GEM: A short “Growth-vs-Environment” Module for survey research Ivan Savin a, b, *, Stefan Drews a, Jeroen van den Bergh a, c, d a Institute of Environmental Science and Technology, Universitat Autònoma de Barcelona, Spain b Graduate School of Economics and Management, Ural Federal University, Yekaterinburg, Russian Federation c ICREA, Barcelona, Spain d School of Business and Economics & Institute for Environmental Studies, VU University Amsterdam, the Netherlands A R T I C L E I N F O A B S T R A C T Keywords: Segmentation of survey respondents is a common tool in environmental communication as it helps to understand Agrowth opinions of people and to deliver targeted messages. Prior research has segmented people based on their opinions Degrowth about the relationship between economic growth and environmental sustainability. This involved an evaluation Green growth of 16 statements, which means considerable survey time and cost, particularly if administered by a third party, as Machine learning Public opinion well as cognitive burden on respondents, increasing the chance of incomplete responses. In this study, we apply a machine learning algorithm to results from past surveys among citizens and scientists to identify a robust, minimal set of questions that accurately segments respondents regarding their opinion on growth versus the environment. In particular, we distinguish three groups, called Green growth, Agrowth and Degrowth. To this end, we identify five perceptions, namely regarding ‘environmental protection’, ‘public services’, ‘life satisfac­ tion’, ‘stability’ and ‘development space’. Prediction accuracy ranges between 81% and 89% across surveys and opinion segments. We apply the proposed set of questions on growth-vs-environment to a new survey from 2020 to illustrate its use as an efficient instrument in future surveys. 1. Introduction uses a wider set of survey questions, arguably arriving at more robust and nuanced results (e.g., Tomaselli et al., 2019). Segmentation of survey respondents, also known as ‘audience seg­ We ourselves have contributed to this line of research in the form of mentation’, is a common tool in environmental communication (Metag surveys among citizens and scientists (Drews et al., 2019). These surveys and Schäfer, 2018). Various segmentation models have been developed built on prior theoretical analysis of the growth-environment debate, for different environmental issues, notably climate change (Hine et al., suggesting the existence of three main perspectives, sometimes labeled 2014). A somewhat neglected topic relevant to audience segmentation is as ‘Green growth’, ‘Agrowth’ and ‘Degrowth’ (van den Bergh and Kallis, the relationship between the environment and the economy. 2012; see also Jakob et al., 2020). One can generally say that supporters In the context of the long-standing debate on economic growth of Green growth argue that growth is needed to improve life satisfaction versus environmental sustainability (Meadows et al., 1972; Turner, and can be harmonized with protecting the environment, proponents of 2008; Jackson, 2009; Victor, 2012; van den Bergh and Kallis, 2012), Degrowth reject both propositions and therefore resist further growth or some research has begun to understand segments of public opinion even strive for economic contraction, and advocates of Agrowth are representing distinct attitudes and beliefs. The simplest, dichotomous indifferent about growth and hence refrain from any (de)growth aims. classification used in many large-scale surveys (e.g. World Values Sur­ Using 16 survey questions, we identified three similar segments in vey) is to group those who (do not) prefer environmental protection over samples of the general public and scientists (Drews et al., 2019), which economic growth, or who (do not) believe that economic growth is are fairly consistent with the theoretical perspectives in the debate compatible with environmental protection (Drews et al., 2018). These mentioned before. It may be noted that distinct labels have been used for classifications draw on single survey questions, which may fail to cap­ the same or approximately the same positions, such as sustainable ture important dimensions of public opinion related to the growth instead of green growth, or anti-growth, post-growth and zero-growth debate. Moreover, measurement of constructs works usually better with for degrowth. The appropriate label for degrowth has also been multiple survey questions (Churchill, 1979). In response, recent research debated (Drews and Antal, 2016; Drews and Reese, 2018; Raworth, * Corresponding author at: ICTA, Edifici Z, UAB Campus, 08193 Bellaterra, Spain. E-mail address: ivan.savin@uab.cat (I. Savin). https://doi.org/10.1016/j.ecolecon.2021.107092 Received 7 October 2020; Received in revised form 28 April 2021; Accepted 6 May 2021 Available online 17 May 2021 0921-8009/© 2021 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). I. Savin et al. Ecological Economics 187 (2021) 107092 2015). issues, but also broader public opinion surveys in which environmental Construction of shorter survey instruments has been identified as a awareness and concern are just one among other non-environmental research need in the field of environmental segmentation studies (Metag elements. and Schäfer, 2018). The quality of clustering public opinion depends on concise measures that reliably capture differences in people’s attitudes 2. Data and methods or beliefs. A parsimonious way to elicit opinions avoids cognitive overload and fatigue of respondents to a survey. Moreover, researchers 2.1. Data may hesitate to use relatively long sets of survey questions such as the 16-item instrument we developed, as they take up considerable survey We base our analysis here on two related online surveys on opinions time and increase survey costs. Here we use a machine learning algo­ regarding growth-versus-environment, with overlapping questions rithm to identify a concise and reliable instrument that will serve as a (Drews and van den Bergh, 2016; and Drews and van den Bergh, 2017). shortcut to the lengthy 16-item questionnaire being prohibitive for The first was conducted in 2014 and covered 1008 respondents using many research studies. In doing this, we follow the method in Chryst quota sampling. It is representative of the general population of Spain in et al. (2018), who proposed a four-question shortcut for segmenting terms of age, gender, income, education and geographical regions (see climate change views, which was originally comprised of 15 survey Table A1 in the Appendix for descriptive statistics). The second survey items. was executed in 2015 and involved 814 researchers from many coun­ Our principal aim is to identify a subset of questions capable of tries and a wide range of self-identified academic backgrounds, identifying the three segments of growth-vs-environment debate with including general economics, environmental economics, economic sufficient accuracy, namely at the minimum 80% accurately categorized growth studies, ecological economics, environmental social sciences, respondents in each of the three segments on out-of-sample validation and environmental (natural and engineering) sciences (see Table A2 in data (Fawcett, 2006; Kleinbaum and Klein, 2010). As a result, we the Appendix for descriptive statistics). identify here a five-question instrument that satisfies this requirement. Both surveys included the same set of 16 statements to elicit views on This can serve as a useful module to be part of a wide range of future the debate on economic growth versus environmental sustainability. surveys and experiments, not only those focused on environmental Table 1 lists these along with summarizing labels as used in the previous Table 1 Main statements of both surveys serving as input to the segmentation analysis. Note: The five statements providing an instrument to elicit opinions on the debate are marked with shading. 2 I. Savin et al. Ecological Economics 187 (2021) 107092 studies. For all the statements, respondents could provide answers on a regression trees are fast and intuitive in exploring complex datasets, they 7-point Likert scale, from 1 (strongly disagree) to 7 (strongly agree). typically suffer from low robustness. With the evolution of computer One minor difference between the designs of the two surveys might power and machine learning algorithms, new methods based on be noted. Whereas in the scientific opinion survey respondents could regression trees such as ‘random forests’ and GBM have been proposed. select “no opinion” for each of the 16 statements, in the public opinion Whereas random forests build an ensemble of independent trees, GBM survey there was no such option. This difference explains why there are builds an ensemble of simple successive trees with each tree learning 143 missing observations in the scientists’ survey. In Drews et al. (2019) and improving on the previous and ensuring robustness of results. We we tested if dropping these observations biases the sample by means of decided to use GBM because it has proven its high performance in Kolmogorov-Smirnov test and a Wilcox rank sum test and both tests general (Brownlee, 2019; Friedman, 2001) and particularly in devel­ indicated there was no statistically significant bias. oping short instruments replacing a 15-item instrument to just four Subsequently, Drews et al. (2019) used a Latent Class Analysis al­ items (more than 70% reduction). gorithm to segment the responses into distinct clusters of opinions. This In line with Chryst et al. (2018), we implement a three-step established that in the scientists’ survey 31%, 44% and 25% of re­ approach. First, we randomly withdraw from our data 10% of obser­ spondents belonged to the Green growth, Agrowth and Degrowth clus­ vations (held-out sample) to test the predictive power of our model in a ters, while in the public survey these shares were 29%, 43% and 18%, third stage. Second, the remaining 90% of observations are analysed respectively. In the public survey, Drews et al. (2019) also identified a with the GBM algorithm. To this end, we use the associated package in R fourth cluster of “indifferent” respondents (10% of the sample or 97 (Greenwell et al., 2019). Specifically, we use 1000 trees2 and multino­ respondents) who exhibited little variation in their responses to the 16 mial data distribution to build a prediction model. To validate the statements. In particular, over 99% of their responses were “4” on the model, we use a 10-fold cross-validation. This means that GBM splits the Likert scale (i.e. neither disagree nor agree) for 14 out of 16 questions.1 90% sample again into training and testing sets ten times randomly, This result was interpreted as those people being indifferent, undecided, trains the model on the first set and then evaluates its performance on neutral or lacking awareness about the issue. It is possible that these the testing set, reporting average results. As an outcome, GBM ranks the people would have preferred to not respond since the public survey did 16 statements in their predictive power, which we later use for defining not include a non-response option, as discussed above. For this reason, in a sufficient set of statements with required prediction accuracy. the following we will use 671 complete responses of the scientists’ and In the third step, we select the top ranked statements and evaluate 907 responses of the public survey. To test if omitting the indifferent their prediction quality on the 10% held-out sample by means of a cluster biases our results, we conducted a robustness test by applying the multinomial logistic regression. In particular, using the three segments method to all 1008 respondents of the public survey and found that our of opinions as the dependent variable, we successively fit to the 90% results are robust (see Fig. A1 in the Appendix). data sample a model consisting of the single best predicting variable Comparing the three clusters produced based on the responses from ranked by GBM, and subsequently add further ones. Based on the the general public and scientists’ community, Drews et al. (2019) found regression results, we construct predictions of being classified to one of stronger polarization between clusters of opinions among scientists than the three opinion segments and compare the predictions with the true among the general public on statement like environmental protection, values (the classes assigned earlier on the full set of 16 statements). This development space and life satisfaction. We will further address this in is done until a satisfactory accuracy of 80% was achieved. The reason to Section 3. select the 80% accuracy threshold is that one typically has to strike a In addition, we use a third survey conducted between June–July balance between the accuracy and the length of the survey module: the 2020 in Spain to demonstrate that using our survey module allows shorter it is, the lower is its accuracy. For example, Maibach et al. (2011) obtaining the same three clusters as derived from the original study. reduced 36 items to 15 (i.e. a 60% reduction) reaching 80% overall Sampling was again done by using quotas on age, gender and accuracy threshold (though the accuracy of distinct opinion clusters was geographical distribution, making the survey sample representative of lower). Chryst et al. (2018) reduced 15 items to just 4 (73% reduction), the general population on these characteristics. The sample included and in doing this employed a 70% accuracy threshold. We aim to reduce 2200 respondents who took on average 19 min to finish. The response 16 items to 3–5 items (70–80% reduction) and adopt the 80% accuracy rate was 68%. More details on the survey can be found in Savin et al. threshold both for aggregate accuracy as well as for accuracy for each (2020a, 2020b). See Table A3 in the Appendix for descriptive statistics opinion cluster. of the sample. Furthermore, recognizing the limited amount of data we have for training and validating our results, we repeat steps 1 and 3 described 2.2. Method above for 100 different random seeds in separating the held-out sample, and measure prediction accuracy of the preselected subset of statements The literature offers different approaches to develop short in­ by fitting them via multinomial logistic regression and comparing struments (Siblini et al., 2019). One example is linear discriminant associated predictions with the results. This way, given our relatively analysis (Welling, 2005). It was used, e.g., in a study by Maibach et al. small dataset, one keeps the training data sample larger (90% vs 80% in (2011) on climate change. The authors found that using a limited set of Chryst et al. (2018)), while estimating the prediction accuracy not on a questions one can separate distinct opinion segments on climate change. single held-out sample but on 100 random ones, i.e. varying the allo­ The resulting combination replaced 36 items by only 15, thus consid­ cation of observations between the in- and out-of-samples each time. erably reducing the length of the questionnaire. Another popular tech­ This approach is called cross-validation and is common in machine nique is Principal Component Analysis (Jolliffe and Cadima, 2016), learning (Arlot and Celisse, 2010). Results reported in Section 3 pertain which was used by Vogel et al. (2020) to reduce a 21-item questionnaire to averages and standard errors over the 100 restarts. This approach on public leadership to just 11 items. More recently, Chryst et al. (2018) contributes to robustness of the findings. applied the Gradient Boosting Machines (GBM) algorithm (Natekin and Knoll, 2013) to further shorten the list of questions by Maibach et al. (2011) to identify opinion segments on climate change. GBM is a ma­ chine learning algorithm based on the principle of classification and regression trees (Breiman et al., 1984; Strobl et al., 2009). While basic 2 Typically, GBM converges much faster, with 50–250 trees, but we keep the number larger as it involves relatively little computational time (GBM 1 The two exceptions are life satisfaction and income inequality. convergence took only a few minutes). 3 I. Savin et al. Ecological Economics 187 (2021) 107092 Fig. 1. Results from applying GBM on scientists’ survey. The left panel shows the rank of 16 statements, while the right panel provides the average accuracy in predicting the true segmentation. Note: Results for accuracy are averaged over 100 restarts with different starting seed. 3. Results Table 2 Accuracy, precision and recall of predictions (in %) based on survey among Given the differences in the audiences of our two surveys addressed scientists using the top three statements for each of the three opinion segments. in Section 2.1, in the following section we will apply the procedure Green growth Agrowth Degrowth described in Section 2.2 first on each of the two primary surveys sepa­ Accuracy 88.9 (0.7) 82.1 (0.6) 88.1 (0.6) rately, and then pool the two data samples together to find a parsimo­ Precision 81.7 (1.1) 80.6 (0.9) 83.7 (1.1) nious instrument that can correctly classify the survey respondents for Recall 84.2 (1.2) 79.5 (0.9) 83.4 (1.1) different audiences. To illustrate how the survey module works, we Note: Means with standard errors in parentheses. subsequently apply it to a recent survey conducted in Spain in 2020. Accordingly, this section consists of four subsections for each of the accuracy stated earlier. The standard error in prediction accuracy over exercises, with the third providing the overall five-statement instrument 100 restarts was 0.46%, which implies that the confidence band of mean and the fourth validating the instrument on a recent dataset. ± two standard errors also lies well above the 80% requirement. Furthermore, to make sure that prediction accuracy is high across all the 3.1. Scientists’ survey three opinion segments, in Table 2 below we report means and standard errors of accuracy for Green growth, Agrowth and Degrowth separately. Results for the survey of scientists are presented in Fig. 1. The left As one can see, our model has a slightly better accuracy for Degrowth panel shows the most influential variables in explaining the clustering and Green growth. Table 3 reports estimates of the resulting multino­ membership on the y-axis, while the x-axis captures their relative in­ mial regression model. Interpretation of the coefficients is as follows. If a fluence, i.e. the improvement that variables produce in the regression respondent chose the answer “strongly agree” instead of “strongly trees (equivalent for mean squared error for a standard regression) disagree” for the statement on life satisfaction (“Continued economic averaged over all the regression trees used. The relative influence is growth is essential for improving people’s life satisfaction.”), then the normalized to sum up to 1, i.e. expressed in %. The variables producing odds of being classified in the Green growth (vs Degrowth) cluster in­ the largest improvement may be considered as the most important ones. crease by factor 3.4 to 1. We can see that the ‘life satisfaction’ statement leads with a big margin followed by ‘environmental protection’ and ‘development space’. 3.2. Public survey Together these three variables produce an accuracy3 of 85.7% on average over 100 restarts, which satisfies our requirement on prediction When we repeat the procedure on the public survey, we find that the rank of most predictive statements, shown in Fig. 2, is different from 3 what we find for scientists. While the ‘environmental protection’ state­ Accuracy measures the fraction of instances correctly predicted. The ment is still among the top three, the other two most predictive items maximum value it can take is 100%, meaning that observations with predicted with a considerable margin are ‘public services’ and ‘stability’. These class A should actually be assigned to that class, while no observation of class A three indicators together achieve an average accuracy of 83.0% (with a is assigned to a different class. Average accuracy is then obtained as an average per class accuracy of the survey instrument. Precision measures the effective­ standard error of 0.3%) satisfying our requirement on prediction qual­ ness of the classifier to avoid false positives (i.e. classify observations from other ity. The fact that statements like ‘development space’ score lower here classes to the focal class) while recall measures the effectiveness in limiting may be that among the general public economic growth is primarily false negatives (i.e. classify observations from the focal class to other classes). associated with low unemployment and a good standard of living. On See Sokolova and Lapalme (2009) for more details. the other hand, scientists name among their first associations with 4 I. Savin et al. Ecological Economics 187 (2021) 107092 Table 3 Table 4 Results of the multinomial logistic model based on the survey among scientists. Accuracy, precision and recall of predictions (in %) based on a survey among the Agrowth Green growth general public, using the top three statements for each of the three opinion segments. Life satisfaction Disagree 1.589*** (0.547) − 1.357 (1.380) Somewhat disagree 3.342*** (0.645) 1.673 (1.137) Green growth Agrowth Degrowth Neutral 3.940*** (0.941) 3.488*** (1.263) Accuracy 82.6 (0.5) 82.1 (0.4) 84.6 (0.4) Somewhat agree 3.512*** (0.892) 4.577*** (1.163) Precision 79.6 (1.1) 79.5 (0.6) 79.1 (0.7) Agree 23.553*** (0.441) 25.513*** (0.441) Recall 72.1 (1.0) 83.8 (0.6) 78.9 (0.7) Strongly agree − 0.128 (1.794) 3.421** (1.731) Environmental Disagree 0.611 (0.510) − 0.529 (0.988) Note: Means with standard errors in parentheses. protection Somewhat disagree 1.879*** (0.607) 0.387 (1.023) Neutral 2.449*** (0.909) − 0.280 (1.296) Somewhat agree 4.179*** (0.895) 3.717*** (1.135) Table 5 Agree 26.136*** (0.418) 27.328*** (0.418) Strongly agree 17.272*** (0.469) 19.114*** (0.469) Results of the multinomial logistic model based on the survey among general Development Disagree 2.947* (1.515) 1.585 (1.535) public. space Somewhat disagree 3.131** (1.484) 0.487 (1.523) Agrowth Green growth Neutral 2.298 (1.429) − 1.943 (1.516) Somewhat agree 1.383 (1.381) − 2.562* (1.436) Public services Disagree 10.321*** (0.925) − 0.612 (1.283) Agree 1.319 (1.370) − 2.306 (1.441) Somewhat disagree 14.930*** (0.385) − 16.416*** (0.000) Strongly agree − 2.178 (1.434) − 4.474*** (1.527) Neutral 14.666*** (0.373) 0.728*** (1.194) Somewhat agree 15.550*** (0.368) 1.526 (1.119) Constant − 3.985*** (1.440) − 2.129 (1.515) Agree 15.558*** (0.368) 3.250*** (1.122) Akaike Inf. Crit. 570.662 570.662 Strongly agree 17.065*** (0.870) 5.839** (1.432) Stability Disagree 15.438*** (0.562) 2.009* (1.211) Note: Results are reported as odds-ratios. Degrowth is a reference group. Somewhat disagree 18.166*** (0.380) 2.049 (1.275) Reference response is 1 (“strongly disagree”). Asterisks ***, **, and * denote 1%, Neutral 18.362*** (0.325) 0.643 (1.230) 5%, and 10% significance, respectively. Somewhat agree 19.434*** (0.337) 2.794** (1.164) Agree 19.910*** (0.459) 5.141*** (1.202) economic growth also the side effect of increasing pollution and envi­ Strongly agree 16.145*** (1.063) 6.167*** (1.454) Environmental Disagree 1.665* (0.851) 0.659 (0.789) ronmental degradation, pointing out that this growth is unsustainable protection Somewhat disagree 3.346** (0.848) 0.710 (0.814) (Savin et al., 2020a; Savin et al., 2020b). Examining prediction accuracy Neutral 4.575*** (0.860) 2.284*** (0.798) for each of the three opinion segments (Table 4), we find slightly better Somewhat agree 5.255*** (0.878) 2.159*** (0.810) results for the Degrowth cluster and worse for the Green growth cluster. Agree 4.936*** (0.952) 3.285*** (0.861) The resulting multinomial model is reported in Table 5. Strongly agree 2.856*** (1.085) 2.758*** (0.982) Constant − 36.300*** (0.638) − 6.573*** (1.613) Akaike Inf. Crit. 911.286 911.286 3.3. Pooled dataset Note: Results are reported as odds-ratios. Degrowth is a reference group. Next, we pool the two datasets together to see what would be the Reference response is 1 (“strongly disagree”). Asterisks ***, **, and * denote 1%, rank of statements and their accuracy on the whole set of observations 5%, and 10% significance, respectively. we have. Fig. 3 demonstrates the results. The top five statements reach Fig. 2. Results from applying GBM on public survey. The left panel shows the rank of 16 statements, while the right panel provides the average accuracy in pre­ dicting the true segmentation. Note: Results for accuracy are averaged over 100 restarts with different starting seed. 5 I. Savin et al. Ecological Economics 187 (2021) 107092 Fig. 3. Results from applying GBM pooled dataset. The left panel shows the rank of 16 statements, while the right panel provides the average accuracy in predicting the true segmentation. Note: Results for accuracy are averaged over 100 restarts with different starting seed. 82.8% accuracy (with standard error 0.2%) and comprise the same top t is worth noting that the results for the pooled dataset are slightly statements discussed earlier. The three statements with the highest worse than on any of the two surveys taken separately, which is due to it relative influence, above 10%, are ‘environmental protection’, ‘public being less homogeneous than each individual dataset. As we know from services’ and ‘life satisfaction’. The other two statements on ‘stability’ Section 3.1 and 3.2, same statements serve differently in their prediction and ‘development space’ have a lower relative influence of around accuracy in the two surveys. For example, the ‘stability’ statement dis­ 8–10%. tinguishes well respondents among the general public but functions less Table 6 summarizes the prediction accuracy of this instrument both well for scientists. Logically, their estimates in the multinomial regres­ on the pooled dataset but also on the two surveys separately. It shows sion differ considerably. Hence, when fitting the multinomial regression that the proposed five-item instrument scores well on all three datasets model to both samples at the same time results in certain averaging of (individual surveys and the pooled sample), for each reaching 80% the estimates, which moderately reduces the quality of out-of-sample prediction accuracy or higher. This demonstrates that the 5-item in­ prediction (Table 7). strument obtained for the pooled dataset also works well when applied to each of the survey samples separately. Note that by using a 5- rather 3.4. Applying the survey module on a recent survey from 2020 than 3-item instrument (Section 3.1–3.2), we can reach an accuracy that is either the same (for the scientific survey) or higher (for the general In the following section we present results of applying the identified public survey) on both surveys separately. survey module to recent survey data for Spain from 2020 which has not been used for constructing it. The data contains only the selected five statements, and in line with Drews et al. (2019) we use Latent Class Table 6 Analysis to cluster the observations. The number of clusters is selected Accuracy, precision and recall of predictions (in %) made with five-item in­ on the basis of three commonly used information criteria: the consistent strument for each of the three opinion segments. Akaike criterion (cAIC), the Bayesion criterion (BIC) and the adjusted Prediction based Green Agrowth Degrowth BIC (aBIC). All three criteria suggest three clusters (see Fig. A2 in the on growth Appendix). The three derived clusters have the following proportions: Pooled dataset Average 82.8 (0.2) Green growth (22.3%), Agrowth (50.2%) and Degrowth (27.6%). This accuracy demonstrates that since 2014 the share of people supporting Green Accuracy 82.1 (0.3) 81.0 (0.2) 86.1 (0.3) growth has fallen, while the share of Agrowth has risen. As one can see Precision 77.9 (0.5) 77.5 (0.4) 82.5 (0.5) from Fig. 4, the clusters are different across all five survey items except Recall 71.8 (0.7) 82.9 (0.4) 80.0 (0.5) Scientists’ survey Average 86.5 (0.4) for ‘development space’. The latter perhaps can be explained by two accuracy arguments. First, development space has been ranked high in our survey Accuracy 87.1 (0.6) 84.1 (0.5) 89.4 (0.5) tool particularly on the sample of scientists and the pooled sample of Precision 83.9 (1.1) 81.1 (0.9) 85.7 (0.9) survey respondents, while for the general public it was considered less Recall 79.0 (1.2) 83.3 (0.8) 85.7 (1.1) Public survey Average 86.0 (0.3) important (see Fig. 2). Bearing this in mind we try to segment the gen­ accuracy eral public with a three-item module below showing that it can reach Accuracy 87.3 (0.5) 84.8 (0.4) 87.0 (0.4) almost the same result. Second, the timing of the questionnaire (COVID- Precision 83.3 (0.9) 82.9 (0.6) 84.1 (0.7) 19 in 2020) may have stimulated stronger solidarity between people so Recall 78.4 (1.0) 86.7 (0.5) 81.3 (0.8) that they tend to respond consistently high on this dimension irre­ Note: Means with standard errors in parentheses. spective of the cluster membership. As the sole purpose here is to show 6 I. Savin et al. Table 7 Results of the multinomial logistic model with five-item instrument. Pooled dataset Scientists’ survey Public survey Agrowth Green growth Agrowth Green growth Agrowth Green growth Environmental protection Disagree 0.885** (0.400) 0.712 (0.642) 0.668 (0.611) − 1.241 (1.526) 1.700* (0.973) 1.632 (1.072) Somewhat disagree 1.986*** (0.421) 1.138* (0.660) 2.024*** (0.714) 0.006 (1.515) 3.518*** (0.972) 1.761 (1.086) Neutral 2.533*** (0.448) 1.396** (0.649) 1.882* (0.974) − 1.783 (1.820) 5.118*** (1.007) 3.577*** (1.078) Somewhat agree 3.303*** (0.458) 2.328*** (0.631) 3.749*** (0.985) 2.259 (1.562) 6.351*** (1.068) 4.032*** (1.117) Agree 3.362*** (0.627) 3.802*** (0.748) 27.381*** (0.700) 27.648*** (0.700) 5.726*** (1.195) 4.837*** (1.204) Strongly agree 2.258*** (0.809) 3.771*** (0.883) 22.697*** (0.792) 24.176*** (0.792) 2.876** (1.388) 3.615*** (1.371) Public services Disagree 12.396*** (0.302) − 2.761** (1.114) 15.633*** (0.505) − 2.156 (2.216) 7.039*** (1.186) − 1.820 (2.885) Somewhat disagree 13.540*** (0.254) − 2.569** (1.042) 16.019*** (0.513) − 3.934* (2.069) 14.393*** (0.472) − 28.109*** (0.000) Neutral 13.791*** (0.260) − 1.020 (0.933) 16.498*** (0.495) − 2.450 (2.033) 14.166*** (0.451) 2.091 (2.777) Somewhat agree 13.988*** (0.254) 0.274 (0.868) 17.203*** (0.563) 0.832 (1.912) 14.685*** (0.391) 2.744 (2.747) Agree 13.891*** (0.299) 1.133 (0.872) 17.765*** (0.782) 2.479 (2.065) 14.496*** (0.436) 4.621* (2.750) Strongly agree 14.312*** (0.560) 2.474** (0.962) 16.512*** (1.230) 1.690 (2.271) 15.961*** (1.014) 7.116** (2.884) Life Satisfaction Disagree 1.491*** (0.516) 0.316 (0.993) 1.947*** (0.654) − 0.518 (1.568) 28.669*** (0.788) 15.971*** (1.107) Somewhat disagree 2.211*** (0.529) 0.041 (0.973) 3.739*** (0.764) 0.945 (1.386) 31.560*** (0.449) 15.764*** (1.360) Neutral 1.974*** (0.548) 1.852** (0.901) 3.910*** (0.992) 3.423** (1.504) 32.103*** (0.443) 17.968*** (0.885) 7 Somewhat agree 1.610*** (0.524) 2.329*** (0.815) 3.841*** (1.078) 4.610*** (1.431) 31.608*** (0.355) 18.323*** (0.648) Agree 3.145*** (0.592) 4.288*** (0.848) 39.569*** (0.545) 41.154*** (0.545) 34.317*** (0.484) 21.696*** (0.709) Strongly agree 1.508** (0.666) 3.635*** (0.889) − 0.651 (2.471) 2.691 (2.176) 33.248*** (0.638) 21.944*** (0.778) Development space Disagree 3.302*** (0.927) 0.408 (0.697) 5.041*** (1.851) 2.512 (1.913) 16.510*** (0.596) − 0.322 (1.361) Somewhat disagree 3.456*** (0.903) − 1.132* (0.684) 4.256** (1.746) 0.345 (1.862) 16.836*** (0.430) − 3.013** (1.361) Neutral 3.880*** (0.900) − 2.102*** (0.686) 3.461** (1.697) − 2.537 (1.901) 17.567*** (0.413) − 2.020 (1.322) Somewhat agree 3.017*** (0.881) − 2.666*** (0.663) 2.519 (1.633) − 3.645** (1.808) 15.953*** (0.354) − 3.061** (1.285) Agree 2.636*** (0.884) − 2.647*** (0.667) 2.615 (1.624) − 2.709 (1.816) 15.072*** (0.419) − 3.754*** (1.334) Strongly agree 0.562 (0.906) − 3.247*** (0.686) − 1.276 (1.645) − 5.999*** (1.822) 13.171*** (0.559) − 3.486*** (1.328) Stability Disagree 16.916*** (0.281) 1.057 (0.949) 14.110*** (0.466) − 1.477 (2.440) 15.542*** (0.739) 0.832 (1.697) Somewhat disagree 17.310*** (0.275) 0.589 (0.960) 14.532*** (0.521) 0.417 (2.472) 19.277*** (0.485) 0.486 (1.730) Neutral 17.159*** (0.259) − 0.169 (0.920) 15.053*** (0.536) 0.236 (2.432) 18.717*** (0.395) − 1.510 (1.667) Somewhat agree 18.341*** (0.252) 1.669* (0.882) 15.641*** (0.471) 1.801 (2.340) 20.320*** (0.416) 1.911 (1.544) Agree 17.876*** (0.309) 2.621*** (0.887) 14.131*** (0.595) 1.391 (2.385) 20.457*** (0.549) 3.918** (1.573) Strongly agree 16.002*** (0.622) 3.509*** (1.046) 11.965*** (1.258) 2.008 (2.770) 17.364*** (1.211) 6.132*** (2.030) Constant − 36.755*** (0.776) − 3.810*** (1.220) − 36.356*** (1.379) − 1.767 (2.431) − 84.624*** (0.597) − 25.012*** (2.622) Ecological Economics 187 (2021) 107092 Akaike Inf. Crit. 1464.604 1464.604 494.932 494.932 699.566 699.566 Note: Results are reported as odds-ratios. Degrowth is a reference group. Reference response is 1 (“strongly disagree”). Asterisks ***, **, and * denote 1%, 5%, and 10% significance, respectively. I. Savin et al. Ecological Economics 187 (2021) 107092 Fig. 4. Response distribution among the three clusters for the five items making up the survey module. Fig. 5. Distribution of responses on the three items from our survey module in the public opinion survey. Table 8 Table 9 Confusion matrix for clustered constructed with two versions of the survey Multinomial logit regression analysis of cluster membership and additional module based on the survey from 2020. survey variables. Clusters based on three-items Clusters based on five-items Clusters based on three-items instrument instrument instrument Green Agrowth Degrowth Degrowth Green Degrowth Green growth growth growth Clusters based on five- Green 441 42 6 Age − 0.024*** 0.011*** − 0.021*** 0.014*** items instrument growth (0.004) (0.004) (0.004) (0.004) Agrowth 24 1060 20 Gender − 0.515*** − 0.016 − 0.514*** 0.008 Degrowth 1 38 568 (0.123) (0.127) (0.123) (0.129) Growth-vs- 0.819*** − 0.607*** 0.811*** − 0.605*** environment (0.078) (0.123) (0.077) (0.125) that clusters can be developed reliably by using the five items, we refrain strategy Political − 0.041 0.157*** − 0.029 0.172*** from any interpretation of differences in the sizes of the clusters orientation (0.027) (0.025) (0.026) (0.025) compared to earlier results. Constant − 0.682* − 0.749* − 0.873** − 1.080** Since here we are working with survey data from the general public, (0.375) (0.438) (0.374) (0.447) one may then ask if using only three out of five items (public services, Akaike Inf. Crit. 3622.98 3622.98 3574.37 3574.37 stability, environmental protection) would produce distinct results. To Note: Agrowth is the reference group. Standard deviations in parentheses. As­ test this, we apply an instrument limited to these three items. The cluster terisks ***, **, and * denote 1%, 5%, and 10% significance, respectively. proportions in this case look very similar: Green growth (23.4%), Agrowth (47.6%) and Degrowth (29%). In this case, clusters differ well To demonstrate that the clusters identified using 5-item instrument on all three dimensions used for classifying the responses (Fig. 5). have similar characteristics to clusters identified in Drews et al. (2019), To see how the three clusters using the 5- and 3-item instrument we regress the cluster membership using multinomial logit model on coincide, we build a confusion matrix (Table 8). As only 131 out of 2200 age, gender, political orientation and preferred growth-vs-environment responses (i.e. 6%) are classified differently by the two versions of the strategy (single-item question present in both surveys). The results are instrument, we conclude that the results for the 5-item instrument are summarized in Table 9. In line with Drews et al. (2019), Table 3), we robust. find that the Green growth cluster prefers a public policy that combines 8 I. Savin et al. Ecological Economics 187 (2021) 107092 economic growth and environmental sustainability, assuming that both A limitation of the present study is that we used only two surveys, objectives are compatible (see the negative coefficient of the preferred one for Spanish citizens and one for scientists worldwide. Hence, it “growth-vs-environment strategy”), while the Degrowth cluster prefers would be good to test the proposed instrument for public surveys in a public policy that has the goal of stopping economic growth (largest other countries. Another limitation is that most items in our survey value of the associated variable). This result can be seen as a validation module (in contrast to the full set of 16 items) involve ‘positive’ wording of the clustering made by means of our survey module, as these pref­ about growth. This might introduce so-called acquiescence bias of re­ erences reflect the main beliefs of each cluster. Also, similar to Drews sponses (Krosnick, 1999). Further research might test whether using et al. (2019), we find that members of the Green growth cluster are older alternative expressions of the same underlying ideas provide similar and those of the Degrowth younger than members of the Agrowth results. cluster. Moreover, compared to Agrowth, supporters of Degrowth Our survey module GEM can serve as a basis for a variety of future include more men and slightly more left-wing political views. Overall, empirical studies that aim to address views on academic and societal these results can be viewed as further support for the validity of our debate surrounding growth and environment. In addition, it can become instrument. a part of larger questionnaires that have a broader aim, such as assessing environmental attitudes, support for climate policy, or opinions about 4. Conclusion the state of the economy. Despite the importance of the debate on growth versus the environment, empirical studies of public opinion on This study has identified a reduced set of survey questions to be used environmental issues often lack a thorough tool to identify opinions in it. for segmenting an audience into three clusters of the growth-vs- A short survey module like GEM could easily be incorporated in a larger environment debate, namely Green growth, Agrowth and Degrowth. survey questionnaire. This would have several advantages; it would To segment an audience consisting of the general public, a three-item avoid ad hoc formulations that hamper comparability of findings be­ solution is sufficient to achieve good accuracy. These three items tween studies; and it would save other researchers time and efforts to address aspects of public services, economic stability and environmental examine the relevant literature in order to develop an own set of ques­ protection. To segment an audience made up of scientists, or more tions to elicit growth-versus-environment opinions. Such incorporation generally people familiar with the debate, another three-item instru­ of standard instruments is already common in psychological research, ment seems appropriate. It covers environmental protection, but in and could benefit environmental studies. contrast to the three-item instrument for the general public, addresses the link to people’s life satisfaction and the question of development Funding space for rich and poor countries. We also derived a five-item instrument based on pooling data on opinions by the general public and scientists. This work was funded by an ERC Advanced Grant from the European This instrument, which represents a combination of solutions for the Research Council (ERC) under the European Union’s Horizon 2020 separate samples, achieves sufficient prediction accuracy for the pooled Research and Innovation Programme [grant agreement n◦ 741087]. I.S. dataset, and the same or a higher accuracy for the two samples. Indeed, acknowledges financial support from the Russian Science Foundation it is especially useful when one is uncertain about the type of audience [RSF grant number 19-18-00262]. one has in mind for segmentation and communication. Moreover, while earlier studies tested their instrument only among the general public, our instrument passes a different, arguably harder, test for validity: Declaration of Competing Interest namely, we show that this instrument can be used even for structurally different populations and still delivers a good performance. None. Appendix A. Appendix Table A1 Key socio-demographic characteristics of the public opinion survey in 2014 (N = 907). Variables Description Mean (SD) or % Gender Female 48.1% Age 18 to 64 years 40.82 (12.46) Household income 1 (≤1000 €) to 5 (≥3000 €) 2.78 (1.25) Educational attainment 1 (primary education) to 4 (postgraduate degree) 2.62 (0.76) Political orientation 1 (left-wing) to 9 (right-wing) 3.95 (2.01) Table A2 Key socio-demographic characteristics of the scientific opinion survey (N = 671). Variable n Variable n Variable n Age Research field Political ideology <30 years 19 GrowEc 34 Very left 46 30–39 218 GrowEnv 31 Left 206 40–49 200 OthEc 75 Slightly left 190 50–59 113 EnvEc 228 Center 108 ≥60 87 EcoEc 131 Slightly right 50 Gender EnvSoc 156 Right 28 Female 162 EnvSci 16 Very right 3 Male 498 # publications grow/env Don’t know 31 9 I. Savin et al. Ecological Economics 187 (2021) 107092 Table A2 (continued ) Variable n Variable n Variable n Education 0 213 Citizenship PhD 586 1–3 187 North America 160 Other 85 4–10 185 EU 337 Professional affiliation 11–29 54 Asia 65 Academia 553 ≥30 32 Africa 16 Government 36 # publications growth Australia and Oceania 27 Private 33 0 355 Central and Southern America 15 Other 47 1–3 77 Other 51 Income of country of origina 4–9 40 High 589 10–19 23 Middle/low 82 ≥20 19 a We use the classification of The World Bank for high and middle/low income countries: http://data.worldbank.org/about/country-and-lending-gro ups. The research fields are described in Section 2 in Drews and van den Bergh (2017). Not all numbers add up to N = 671 due to missing data. Table A3 Key socio-demographic characteristics of the public opinion survey in 2020 (N = 2200). Variables Description Mean (SD) or % Gender Female 51% Age 18 to 88 years 45.50 (15.09) Household income 1 (No income) to 5 (≥4000€) 2.58 (1.14) Educational attainment 1 (Less than 5 years of school) to 8 (University) 5.02 (1.32) Political orientation 1 (left-wing) to 10 (right-wing) 4.44 (2.41) Fig. A1. Results from applying GBM on public survey keeping 101 respondents from the Indifferent cluster. The left panel shows the rank of 16 statements, while the right panel provides the average accuracy in predicting the true segmentation. Note: Results for accuracy are averaged over 100 restarts with different starting seed. As one can see from comparing Fig. A1 and Fig. 2, the top three statements with highest relative influence remain the same, while average accuracy of the survey module including those statements reaches 82.3%, which is above our threshold. Table A4 below further illustrates that accuracy of predicting each of the four clusters with the survey instrument remains high. This evidence supports the robustness of our results obtained earlier, namely that our proposed instrument has a good accuracy in classifying opinions on the growth-vs-environment debate. Table A4 Accuracy, precision and recall of predictions (in %) based on survey among general pubic survey, keeping 101 respondents from the Indifferent cluster and using top three statements for each of the four opinion segments. Green growth Agrowth Degrowth Indifferent Accuracy 81.9 (0.6) 81.1 (0.3) 82.5 (0.4) 81.6 (0.7) Precision 75.5 (1.2) 75.8 (0.6) 78.0 (0.6) 73.5 (1.5) Recall 72.4 (1.1) 80.4 (0.6) 77.8 (0.8) 73.1 (1.5) Note: Means with standard errors in parentheses. 10 I. Savin et al. Ecological Economics 187 (2021) 107092 Fig. A2. Plot of information criteria for 1- to 9-cluster solutions for the public opinion survey in 2020. References Jolliffe, I.T., Cadima, J., 2016. Principal component analysis: a review and recent developments. Philos. Trans. R. Soc. 374, 20150202. Kleinbaum, D.G., Klein, M., 2010. Introduction to logistic regression. In: Logistic Arlot, S., Celisse, 2010. A survey of cross-validation procedures for model selection. Stat. Regression. Springer, pp. 1–39. Surv. 4, 40–79. https://doi.org/10.1214/09-SS054. Krosnick, J.A., 1999. Survey research. Annu. Rev. Psychol. 50 (1), 537–567. https://doi. Breiman, L., Friedman, J.H., Olshen, R.A., Stone, C.J., 1984. Classification and org/10.1146/annurev.psych.50.1.537. Regression Trees. Wadsworth & Brooks/Cole Advanced Books & Software, Maibach, E.W., Leiserowitz, A., Roser-Renouf, C., Mertz, C.K., 2011. Identifying like- Monterey, CA. minded audiences for global warming public engagement campaigns: an audience Brownlee, J., 2019. Comparing 13 Algorithms on 165 Datasets. Machine Learning segmentation analysis and tool development. PLoS ONE 6, e17571. Mastery. https://machinelearningmastery.com/start-with-gradient-boosting/. Meadows, D.H., Meadows, D.L., Randers, J., Behrens, W.W., 1972. The Limits to Growth: Chryst, B., Marlon, J., van der Linden, S., Leiserowitz, A., Maibach, E., Roser-Renouf, C., A Report for the Club of Rome’s Proyect on the Predicament of Mankind, Vol. 102. 2018. Global warming’s “six Americas short survey”: audience segmentation of Universe Books. climate change views using a four-question instrument. Environ. Commun. 12 (8), Metag, J., Schäfer, M.S., 2018. Audience segments in environmental and science 1109–1122. https://doi.org/10.1080/17524032.2018.1508047. communication: recent findings and future perspectives. Environ. Commun. 12 (8), Churchill, G.A., 1979. A paradigm for developing better measures of marketing 995–1004. https://doi.org/10.1080/17524032.2018.1521542. constructs. J. Market. Res. 16 (1), 64–73. JSTOR. https://doi.org/10.2307/3150876. Natekin, A., Knoll, A., 2013. Gradient boosting machines, a tutorial. Front. Neurorobot. Drews, S., Antal, M., 2016. Degrowth: a “missile word” that backfires? Ecol. Econ. 126, 7, 21. 182–187. https://doi.org/10.1016/j.ecolecon.2016.04.001. Raworth, K. (2015). Why degrowth has outgrown its own name. https://www.kateraw Drews, S., Reese, G., 2018. “Degrowth” vs. other types of growth: labeling affects orth.com/2015/12/01/degrowth/. emotions but not attitudes. Environ. Commun. 12 (6), 763–772. https://doi.org/ Savin, I., Drews, S., van den Bergh, J., 2020a. Free associations of citizens and scientists 10.1080/17524032.2018.1472127. with economic and green growth: a computational linguistics analysis. Ecol. Econ. Drews, S., van den Bergh, J.C.J.M., 2016. Public views on economic growth, the 180, 106878. https://doi.org/10.1016/j.ecolecon.2020.106878. environment and prosperity: results of a questionnaire survey. Glob. Environ. Chang. Savin, I., Drews, S., van den Bergh, J., Villamayor-Tomas, S., 2020b. People’s 39, 1–14. https://doi.org/10.1016/j.gloenvcha.2016.04.001. Expectations about the Impact of COVID-19 on Climate Action by Citizens and Drews, S., van den Bergh, J.C.J.M., 2017. Scientists’ views on economic growth versus Government. Working paper. ICTA UAB. the environment: a questionnaire survey among economists and non-economists. Siblini, W., Kuntz, P., Meyer, F., 2019. A review on dimensionality reduction for multi- Glob. Environ. Chang. 46, 88–103. https://doi.org/10.1016/j. label classification. IEEE Trans. Knowl. Data Eng. 1–20. https://doi.org/10.1109/ gloenvcha.2017.08.007. TKDE.2019.2940014. Drews, S., Antal, M., van den Bergh, J.C.J.M., 2018. Challenges in assessing public Sokolova, M., Lapalme, G., 2009. A systematic analysis of performance measures for opinion on economic growth versus environment: considering European and US classification tasks. Inf. Process. Manag. 45 (4), 427–437. https://doi.org/10.1016/j. data. Ecol. Econ. 146, 265–272. https://doi.org/10.1016/j.ecolecon.2017.11.006. ipm.2009.03.002. Drews, S., Savin, I., van den Bergh, J.C.J.M., 2019. Opinion clusters in academic and Strobl, C., Malley, J., Tutz, G., 2009. An introduction to recursive partitioning: rationale, public debates on growth-vs-environment. Ecol. Econ. 157, 141–155. https://doi. application, and characteristics of classification and regression trees, bagging, and org/10.1016/j.ecolecon.2018.11.012. random forests. Psychol. Methods 14 (4), 323. Fawcett, T., 2006. An introduction to ROC analysis. Pattern Recogn. Lett. 27 (8), Tomaselli, M.F., Sheppard, S.R.J., Kozak, R., Gifford, R., 2019. What do Canadians think 861–874. about economic growth, prosperity and the environment? Ecol. Econ. 161, 41–49. Friedman, J.H., 2001. Greedy function approximation: a gradient boosting machine. https://doi.org/10.1016/j.ecolecon.2019.03.007. Ann. Stat. 29 (5), 1189–1232. Turner, G., 2008. A comparison of the limits to growth with 30 years of reality. Glob. Greenwell, B., B. Boehmke, J. Cunningham, and GBM Developers. 2019. Package ‘gbm’. Environ. Chang. 18 (3), 397–411. https://doi.org/10.1016/j. https://cran.r-project.org/web/packages/gbm/gbm.pdf. gloenvcha.2008.05.001. Hine, D.W., Reser, J.P., Morrison, M., Phillips, W.J., Nunn, P., Cooksey, R., 2014. van den Bergh, J.C.J.M., Kallis, G., 2012. Growth, A-growth or degrowth to stay within Audience segmentation and climate change communication: conceptual and planetary boundaries? J. Econ. Issues 46 (4), 909–920. https://doi.org/10.2753/ methodological considerations: audience segmentation and climate change JEI0021-3624460404. communication. Wiley Interdiscip. Rev. Clim. Chang. 5 (4), 441–459. https://doi. Victor, P.A., 2012. Growth, degrowth and climate change: a scenario analysis. Ecol. org/10.1002/wcc.279. Econ. 84, 206–212. https://doi.org/10.1016/j.ecolecon.2011.04.013. Jackson, T., 2009. Prosperity without Growth: Economics for a Finite Planet. Routledge. Vogel, D., Reuber, A., Vogel, R., 2020. Developing a short scale to assess public Jakob, M., Lamb, W. F., Steckel, J. C., Flachsland, C., & Edenhofer, O. (2020). leadership. Public Adm. 98 (4), 958–973. https://doi.org/10.1111/padm.12665. Understanding different perspectives on economic growth and climate policy. WIREs Welling, M., 2005. Fisher Linear Discriminant Analysis, vol. 3. Department of Computer Clim. Change, n/a(n/a), e677. doi:https://doi.org/10.1002/wcc.677. Science, University of Toronto. 11