Introduction

Only shortly interrupted during the COVID-19 pandemic (IEA, 2021a), the continuous rise in global emissions of greenhouse gases (GHG) calls for an effective response to fight anthropogenic climate change. Several measures have been proposed in international, EU and national strategies, such as improving energy efficiency or further increasing the share of renewable energies (IPCC, 2022). Moreover, the need to promote a substantial change in people’s behavior and lifestyles and to encourage them to display more pro-environmental behavior (PEB) in order to meet climate targets has been frequently emphasized (Bruderer Enzler et al., 2019; Creutzig et al., 2018; IEA, 2021b; UNEP, 2020). In the present study, we systematically aimed at identifying meaningful social-psychological predictors for specific PEB intentions.

Manifold forms of pro-environmental behavior

According to the so-called impact-oriented approach (Lange and Dewitte, 2019, Lange et al., 2021), PEBs are not a homogenous set of behaviors but vary in several ways. Among PEBs in the private sphere (Esfandiar et al., 2022), it is useful to distinguish forms of behavior that can have a higher, more significant impact (e.g., insulating a house) from forms of behavior that may have a lower, rather negligible impact (e.g., switching the light off when leaving a room) on GHG emissions. In this regard, recent studies (Ivanova et al., 2020; Lacroix, 2018; Wynes & Nicholas, 2017) have analyzed various common PEBs and categorized them by their GHG emission reduction potential in metric tons of carbon dioxide equivalents per capita (CO2 eq/cap). Based on these studies, we selected high-impact PEBs of 0.1 to 2 tons of CO2 eq/cap, which is equal to about 2 to 25 percent of the total annual GHG emissions produced by an average household in Austria or Germany (Climate Watch, 2020).

While PEBs can be distinguished by their climate impact, they can also be distinguished based on their generality (i.e., how many people in the population perform a certain type of behavior) and temporality (i.e., how regularly a person behaves in a specific way). For instance, renovating one’s house can be counted as a high-impact activity (reduction potential: mdn = 0.9 CO2 eq/cap, Ivanova et al., 2020), but it can only be performed by house owners and is likely to be done less frequently than once per decade. Recognizing such differences in PEBs while identifying the most important differences among individuals who display these PEBs is a challenge for scientists working in the fields of social sciences and psychology who are interested in PEB and climate change mitigation.

According to the different recognized types of PEB, these can also be measured in a rather broad context (e.g., such as PEBs in everyday life; Lades et al., 2021) or in a narrow one (e.g., PEBs in protected areas or parks; Esfandiar et al., 2022; Pearce et al., 2022). When understood in a broad and inclusive sense, this allows researchers to cover many PEB aspects at once (Fleiß et al., 2020; Kaiser & Wilson, 2000; Lades et al., 2021), while acknowledging their multidimensional facets (see also self-reported PEB, Kaiser, 1998). However, this approach neglects to consider that various underlying tendencies might influence individuals to display a PEB, and – again – that some forms of behavior have a (much) higher impact in terms of GHG emission reduction than others between the PEB domains (Stern, 2011; for a recent discussion, see Nielsen et al., 2021a and Lange et al., 2021). When measuring PEBs more narrowly, however, it is impossible to infer that the same people (e.g., who would renovate their house) would also display PEBs in other domains where high-impact PEBs are possible (e.g., avoiding airplane travel), even if their values and attitudes would suggest that they would do so (spillover effects, see, e.g., Nilsson et al., 2017). With regard to the measurement of PEBs, therefore, we attempted to find a middle ground in this study.

Energy-related PEBs

We wanted to focus on the domain of energy-related PEBs, which are highly relevant for lifestyle changes, as the current lifestyle of most people involves energy usage on a frequent basis. Energy-related PEBs also include different PEBs that significantly impact the GHG emission reduction potential (i.e., > 0.1 metric tons CO2 eq/cap) and can be performed regularly. We understood energy-related behavior as forms of behavior where people directly consume energy by interacting with devices, tools, or means of transportation in their daily lives. We argue that these PEBs are general enough, as many daily behaviors require various forms of direct energy use, whether this is the use of fossil fuel during a car ride or the use of electricity to charge a phone. At the same time, we avoided making our focus too broad, as many other potentially important PEBs, such as insulating a house, buying green and fair products and switching to a plant-based diet, were excluded. This exclusion also concerns otherwise crucial once-off PEBs (Lavelle et al., 2015), which cannot be performed habitually, but only on rare occasions (e.g., insulating a house is likely done once every few decades).

Nonetheless, we considered it necessary to further subdivide this domain into energy-related subdomains in order to take a more nuanced approach (see Lacroix, 2018; Ivanova et al., 2020), for example, by differentiating between transportation and energy consumption in the household. The latter enabled us to further categorize certain PEBs into the subdimensions of electricity consumption and heating. It was not possible to expect that high-impact PEBs in these subdomains would have similar GHG emission reduction potentials. In a systematic review, Ivanova et al. (2020) showed that, for instance, producing one’s own energy by using solar panels might have a larger impact (mdn ~ 1.3 tons CO2 eq/cap reduction) than using a smart meter tool (mdn ~ 0.3 tons CO2 eq/cap reduction). But the former also showed much more variation, ranging from close to zero to up to five tons CO2 eq/cap, depending on how the original study operationalized mitigation and how consistently the behavior was displayed by the participants.

Despite these differences in GHG emission potential, such categorization into subdomains can be useful, because the individuals’ personal capabilities (e.g., financial resources or social status) or existing contextual factors (such as laws and regulations, policies, technologies) might encourage or prevent them from actually displaying a specific PEB (Smiley et al., 2022; Stern, 2000). For example, installing solar panels may depend on the individuals’ housing type, financial resources, or subsidies provided by the government, whereas an EU-wide smart meter roll-out could be an important prerequisite for effective energy management at the grid level in theory, despite the small individual financial benefits (e.g., Department for Business, Energy & Industrial Strategy, 2019). Eventually, the tendency to install solar panels or to use smart meters efficiently (i.e., the specific behavior) may also depend on whether the people see a personal need to do so, value the benefits, or hold positive attitudes towards these measures. In other words, their engagement in terms of energy-related PEB may depend on their person-specific, psychological factors.

Potential factors

Just like person-specific factors, personal capabilities and contextual factors might influence whether individuals display specific PEBs. In this study, we focused on several factors that have been identified by social scientists and psychologists as important in relation to broader categories of PEBs and narrower categories of energy-related PEBs (e.g., Carrus et al., 2021). Moreover, we were especially interested in meaningful predictors that show medium and larger standardized effects (Lovakov & Agadullina, 2021), as these would indicate that large shares of the observed variance in the behavioral intentions could be explained. Here, we summarize the reasoning for our selection of factors in bullet-point format. More detailed theoretical descriptions can be found in the Supplemental Document.

  • Socio-economic status (SES): Despite stronger environmental values, people with a high socio-economic often have a higher than average carbon footprint (see Umweltbundesamt, 2016; Moser & Kleinhückelkotten, 2018; Nielsen et al., 2021b).

  • Biospheric Values (BIOS, De Groot & Steg, 2008): Individuals strive for and uphold biospheric values as guiding principles in order to reach desirable end states: caring for and saving nature. These values appear to be robust predictors of pro-environmental attitude, intentions and, most of all, self-reported behavior (Bouman et al., 2018; De Groot & Steg, 2009; Ojea & Loureiro, 2007; Tamar et al., 2021; Van der Werff & Steg, 2016).

  • Consideration of Future Consequences (CFC, Strathman et al., 1994): some individuals may evaluate future outcomes as more worthwhile, which is crucial in the decision-making process, as they would choose display future-oriented PEB, rather than opting for a personally comfortable, but more environmentally harmful alternative (Carmi & Arnon, 2014).

  • Consideration of Immediate Consequences (CIC, Strathman et al., 1994): if people prefer the outcomes of their behavior to have an impact on their lives in the short term, this might have a more negative impact on their individual GHG emission in several domains (Arnocky et al., 2014; Joireman et al., 2004; Khachatryan et al., 2013; Sun et al., 2021a, 2021b).

  • Perceived Behavioral Control (PBC; Ajzen, 2002): even if people might be willing to adapt their lifestyles in theory, their current personal situation may not allow them to exhibit more PEBs (e.g., if someone lives on the countryside with limited access to public transport). Hence, whether individuals perceive that they have personal control over their behavior has frequently been shown to make a difference for PEB intentions in the literature (De Leeuw, et al., 2015; Jugert et al., 2016; Morren & Grinstein, 2016).Footnote 1

  • Willingness to Sacrifice (WTS;Davis et al., 2011; Thaller et al., 2020): People who show a high WTS are more likely to accept discomfort and costs associated with fulfilling an immediate personal interest, investing an effort, or covering an expense (see also Lavelle et al., 2015; Han & Hyun, 2017; Lesic et al., 2018). Moreover, environmental problems are often described as social dilemmas in which the short-term interests of the individual are at odds with the long-term collective interests of humans and the environment (Milfont & Gouveia, 2006; Van Lange et al., 2013; Pinker, 2021; see also Evans, 2011). As it stands, achieving environmental goals will require an increasing number of individuals to adopt collective interests and sacrifice individual gains to ensure the security of natural resources.Footnote 2 Although WTS has sometimes been used as an outcome variable (Bilandzic et al., 2017; Bord et al., 1997; Davis et al., 2011), it might be crucial in bridging the gap between actionless values and actual PEBs. Hence, WTS is likely to take on a mediating role.

The present study

Our goal with study was to test whether we could identify psychological predictors that could explain relevant energy-related behaviors with a non-negligible GHG emission potential. For this large Austrian survey, we focused on person-specific predictors that we have summarized above. Moreover, we employ a confirmatory approach with specific hypotheses based on the aforementioned theoretical considerations and an exploratory approach via machine learning. With this latter data-driven method we aim to get valuable additional information, based on variables that have been collected in addition to the variables mentioned above.

Methods

We followed three steps to pursue our goals. First, we described the Austrian population’s propensity to engage in a range of medium and high-impact PEBs. Second, we tested two confirmatory hypotheses that deal with psychological and socio-economic predictors of three different kinds of PEB intention and were identified as relevant in the research cited above. The most relevant predictors were then synthesized. In the third step, we systematically explored further variables via machine learning. This approach is beneficial to identify predictors that can meaningfully explain PEB intentions through several cross-validation steps. It will provide a ranked set of important predictors along with an R2 model fit.

Hypotheses

We tested the following hypotheses in the confirmatory part of this study. 1) We hypothesized that biospheric values (BIOS) would predict the willingness to sacrifice (mediator), which in turn predicts energy-related PEB intentions across domains. We also assumed that these effects would not be similar for all socio-economic status groups, assuming a moderating role of SES. 2) We hypothesized that considerations of both future and immediate consequences predict energy-related PEB intentions due to the willingness to sacrifice (mediator), whereby the mediator is positively related to future considerations (CFC) and negatively related to immediate considerations (CIC). Moreover, we assume that the effect of biospheric values (BIOS) and perceived behavior control (PBC) would be present in both models. Models for both hypotheses are presented in Fig. 1.

Fig. 1
figure 1

Hypothesized conceptual models. Note: BIOS = biospheric values, WTS = willingness to sacrifice, SES = socio-economic status, PBC = perceived behavioral control, CFC = consideration of future consequences, CIC = consideration of immediate consequences, DV = dependent variable, which are traveling, electricity consumption, heating

Full size image

Data exploration using machine learning

The systematic exploration performed in the third part of this study was carried out to identify how well energy-related PEBs could be predicted by the hypothesized variables and an additional set of social-psychological variables. This exploration was performed by taking a supervised machine-learning approach (Kuhn, 2019, 2022). The additional variables were gathered as part of this large survey and cover, among others, skepticism about climate change (van Boven et al., 2018), power orientation, or benevolence values (as part of the basic human values, Schwartz, 2003). More detailed information about this approach can be found in the preprocessing and analysis section below.

Survey and variables

We set up a survey in LimeSurvey (version 3, LimeSurvey Development Team, 2012), which is a widely used open-source online tool for surveys and experiments (see Klieve et al., 2010). The survey consisted of four parts: The first part contained a consent form in line with the GDPR and questions that enabled us to assess general socio-demographic variables, such as participants’ gender, age and education. The second part featured different questions addressing policies for low-carbon passenger transport, which are part of this article.Footnote 3

The third part asked the participants to express general opinions on the broad topics of energy saving and reducing GHG emissions. Specifically, on respective pages of the survey participants had to indicate whether they would consider performing specific PEBs (from 1 “Would not consider at all” to 5 “Would definitely consider”), referring to the energy-related behaviors traveling (b1), electricity consumption (b2) and heating (b3). Below each of the PEB questions on the same page, we asked participants to report any external constraints (with regard to money or time) to actually displaying the specific behavior, which constituted PBC. The complete list of items and variables for the confirmatory analyses can be found in Table 1.

Table 1 Variable list for N = 1083
Full size table

In the fourth part of the survey, we gathered data on more psychological constructs. These variables were measured at the end of the survey to prevent these items from influencing the participants’ ratings of the DVs. These constructs included biospheric values (De Groot & Steg, 2008) along with self-enhancement and self-transcendence values (i.e., universalism, benevolence, achievement orientation and power; same scales as biospheric values, Schwartz, 2003), as well as CFC and CIC (Joireman et al., 2012; Strathman et al., 1994) and WTS (International Social Survey Programme, 2010; Thaller et al., 2020). For the exploratory part, we asked questions that assessed the participants’ skepticism towards climate change, their political orientation, level of trust in the government, household income (as proxy for SES), household size, type of housing, housing conditions and residential area. The survey setup is depicted in Figure S1 and detailed information about all variables is found in the supplementary document on the OSF (https://osf.io/5gb8m/?view_only=92ec94be08cd4a85a25afc194696ed27).

Research design, sample and power

To test the two hypotheses of 1) a moderated mediation and 2) a parallel mediation, which were described in detail above, we ran structural equation models (SEM; more specifically: a structural regression model, see Kline, 2011) and emphasized the paths leading to the specific energy-related PEB intentions (i.e., the three blocks traveling, electricity and heating). This was crucial for the power analysis, as one could argue that the model fit is most important for SEM. However, a model can have good fit even when the individual path coefficients do not show the anticipated effects; hence, we decided that path coefficients should be of central interest. These path models corresponding to our hypotheses are depicted in Fig. 1.

We calculated a priori that a sample of n = 1,000 is sufficiently large, even when standardized estimates for the paths are very small (e.g., β ~ 0.1), based on the G*Power settings of linear multiple regression, fixed model, R2 deviation from zero and sensitivity analysis: α = 0.05, power = 0.80, n = 1,000, tested predictors = 1, resulting in f2 = 0.008 (≘ f = 0.09, r = 0.09, d = 0.18) (Faul et al., 2007). Even if we analyzed sub-samples with half the sample size (n = 500), there would be sufficient power to find small estimates of β ~ 0.13 (f2 = 0.016). Moreover, we paid careful attention to effect sizes that are conventionally considered as medium-sized and larger. In this study, “medium-sized” was considered to begin at standardized coefficient of β ~ 0.24 (Lovakov & Agadullina, 2021, see also Cohen, 1992), as these effects explain variance in PEBs over five percent.

Preprocessing and analysis

Data pre-processing and analyses were conducted in R (R Core Team, 2021). The reported analyses were performed using the following packages: lavaan (Rosseel, 2022), semTools (Jorgesen, 2022), jmv (Selker & Love, 2022), tidyverse (Wickham, 2021), GGally (Schloerke, 2021) and likert (Bryer, 2016).

With regard to the DVs, all items represented behavioral intentions for PEBs in the future, ranging from "I would not do this at all" (1) to "I would definitely do this" (5). However, for the items in block 2 (electricity) and block 3 (heating), we gave participants the option to indicate "I do this already". To align this answer with behavioral intentions, we planned to merge it with answer option 5 for the respective items. We then planned to create mean composite scores for the DVs (i.e., traveling, electricity and heating) based on the items. Moreover, we planned to do the same steps for all relevant independent variables that consisted of multiple items in our data set. All these variables are shown in Table 1 (for more information, see Table S2 in the supplemental materials).

We aimed at performing several structural equation models (SEMs), placing a focus on the relevant path coefficients. Model fit indices were improved only by adding free correlations between variables in two steps: First, we checked the originally hypothesized SEM without including additional correlations. When these models revealed that residual correlations between certain variables were of medium size (i.e., r > 0.3), we added these correlations to the model, but kept smaller correlations at zero, thereby not depleting the degrees of freedom of the model. Finally, we synthesized the information from the two confirmatory models for the optimal model per DV.

Beside the confirmatory part, we aimed at exploring the data systematically with several machine learning algorithms using the caret R package (Kuhn, 2019; Kuhn, 2022; for a tutorial see Szabelska et al., 2021). We run these algorithms simultaneously with all variables from the confirmatory analysis and additional variables (described in Table 2). The full data set was split repeatedly into training and test data subsets before it was cross-validated in several iterations. The algorithms vary in their approach how the data split is performed, leading to slightly different results. Before variables were fed into these different machine learning models, we determined which variable should be explained. Applying these so-called supervised machine learning approaches enabled us to distill a ranked list of variables (or features) that show the best fit (in R2 metric) for our DVs. Additional to these feature importance rankings, these models provided a root mean squared error (RMSE), enabling us to quantify the amount of noise. Typically, the model with a good fit and the lowest possible noise would be selected.

Table 2 Importance scores from conditional random forestes
Full size table

We also decided to split three blocks of energy-related PEB intentions into the individual indicators, as this would enable us to see whether the machine learning algorithms could be used to identify similar predictors for outcome variables from the same blocks. In other words, this would.

constitute a consistency check. The algorithms that appropriate for large between-subjects data sets are the “K-Nearest Neighbor”, “Regression Tree”, “Conditional Inference Tree”, “Random Forest”, “Conditional Random Forest”, “Neural Network”, and “Neural Network with feature extraction” algorithms, all provided within caret (Kuhn, 2022). More descriptions are presented in the supplemental document.

Results

Participants were recruited by the market research institute Talk Online (https://talkonlinepanel.com/de). The sample is quota-representative for Austria’s population for the sociodemographic variables of age (five age groups between 16 and 69 years old) and gender, but not quota-representative for education (with/without high school diploma) and size of the residential municipality (below/above 10,000 residents), as based on estimates from Statistics Austria (https://www.statistik.at/web_en/statistics/index.html). Detailed information about these quotas can be found on the OSFFootnote 4 and in Table S1.

We recruited 1,465 participants in total, but several did not reach the second block of the study due to two failed attention checks (see OSF) or because they had to be excluded due to a lack of response on key variables, leaving us with a final sample of N = 1,083, of which 49.9% (n = 541) were men, 49.8% (n = 540) were women, and three identified as non-binary or did not wish to indicate their gender. On average, respondents were 44 years old (SD = 14.7, min = 16, max = 70). In total, 341 participants (31.4%) indicated that their highest education degree was an A-level equivalent (Matura) or higher, 38% reported that they lived in a rural area, and 62.2% of the sample reported a net household income between 1,501 and 5,000 Euro. Another n = 195 did not want to report their income—our proxy for SES (see Figure S2 for more information). Hence, we ran part of our analysis with a sample N = 888, which was still sufficiently powered (i.e., power = 0.80) to keep the β = 0.1 as a threshold for small, but notable effects.

Descriptive statistics for our final sample of N = 1,083 (see sample description above) for all variables of the confirmatory analysis, along with reliability measures for the composite scores, can be found in Table S2, and zero-order correlations can be found in Figure S3. These show that correlations and reliability scores are not optimally large, especially for block 1 (traveling, r = 0.19) and block 2 (electricity, α = 0.65), indicating some heterogeneity across these items.

For this sample, we report the energy-related PEB intentions for all items in Fig. 2. With regard to traveling (see Fig. 2A), only 50% of the people surveyed expressed a high or very high intention to avoid flying in the future and only 37% indicated that they would avoid using a car. Perhaps not surprisingly, the percentages of people with these intentions are even lower in the subgroups of people who had flown before (only 27% would avoid flying in the future) and of people who owned a car or an e-car (29% to 41% would avoid using a car or an e-car in the future). With regard to electricity (see Fig. 2B), about two-thirds of the participants expressed that they would use solar panels, smart meters, or look for certified energy providers (61%) in the future to reduce their GHG emissions, but less than 45% indicated that they would consider using energy management systems or stop using extra energy-intense appliances in the household. Finally, with regard to heating (see Fig. 2C), 57% stated that they would reduce their hot water usage and 55% would regulate their room temperature in the cold seasons, respectively.

Fig. 2
figure 2

Raw rating answers of the DV items. Panel A shows item score for DV traveling and split by subgroups of people who did not fly and flew before and car users, e-car users and non-users; Panel B shows items scores for DV electricity; panel C shows items scores for DV heating

Full size image

Confirmatory analysis

For the first model (N = 888), we hypothesized that biospheric values would predict three energy-related PEB intentions via the mediator willingness to sacrifice. Moreover, this mediation should have been moderated by the socio-economic status of the household and have controlled for perceived behavioral control. As depicted in Fig. 3A, this hypothesis was only partially supported by the data: All standardized direct path coefficients for BIOS on WTS and BIOS, PBC and WTS on the three DVs passed our threshold (all β ≥ 0.10, p < 0.001, a priori power > 0.08). Mediation effects via WTS were small for traveling and electricity (β = 0.09 and β = 0.06, ps < 0.001) and negligible for heating (β = 0.04, 95%CI [0.02, 0.06], z = 3.77, p < 0.001). Those effects were driven by the large path coefficient from BIOS to WTS (β = 0.36, 95%CI [0.30, 0.42], z = 12.53, p < 0.001). One important consideration with regard to Hypothesis 1 was that SES did not show meaningfully large moderation effects (all β ≤ 0.05) and that varying degrees of the main effects of SES on the DVs were observed. The largest of these was observed on traveling (β = -0.20, 95%CI [-0.26, -0.14], z = -6.93, p < 0.001). Although both WTS and BIOS showed effects in the range of β = 0.11 to 0.26, PBC’s effects fell most consistently within the medium to large spectrum (all β ≥ 0.23, p < 0.001). Model fit indices mostly indicated acceptable fit (i.e., RMSEs ≤ 0.08, p > 0.05), with an exception of the Tucker-Lewis Index for DV electricity (TLI = 0.77). All variables combined explained only 20–23% of the total variances in the DVs.

Fig. 3
figure 3

Confirmatory analysis and synthesized model. Note: (A) confirmatory SEM for H1 for all three DVs; (B) confirmatory SEM for H2 for all three DVs; (C) synthesized model for DV electricity tested with two random sets of the original sample; TLI = Tucker-Lewis Index, SRMR = standardized root mean squared residual, RMSEA = root mean square error of approximation; error variances are omitted for simplicity; free correlations were added post-hoc for r ≥ 0.3 (Cohen, 1992) to avoid arbitrarily inflated parameters (Hallquist, 2017; Landis et al., 2008); zero-order correlations can be found in Figure S3 (https://osf.io/5gb8m/?view_only=92ec94be08cd4a85a25afc194696ed27)

Full size image

For the second model (see Fig. 3B), we hypothesized that – in addition to BIOS – the consideration of immediate and future consequences would explain the variance in both WTS and the three DVs. Hence, again, we predicted a mediation via WTS. As this multiple mediation was specified with three predictors (i.e., BIOS, CFC, and CIC) and one mediator (WTS), we describe the path coefficients step by step below.

The initial effect of BIOS on WTS from the previous model shrank drastically (β = 0.11, 95%CI [-0.05, 0.17], z = 3.53, p < 0.001) through the inclusion of CFC, which in turn showed a large effect on WTS (β = 0.45, 95%CI [0.39, 0.51], z = 14.87, p < 0.001), whereas CIC did not predict WTS (β = 0.00, 95%CI [-0.05, 0.05], z = 0.05, p = 0.957). However, the coefficients in the second part of the model varied in the area of small effects: WTS showed small effects on traveling and electricity (β ≥ 0.17, p < 0.001; mediations: β ≥ 0.07, p < 0.001), but not on heating (β = 0.03, 95%CI [-0.02, 0.09], z = 1.23, p = 0.210, mediation: β = 0.01, 95%CI [-0.01, 0.04], z = 1.22, p = 0.220). The effects of BIOS and CIC also varied within a similar range of small and tiny effects. Effects by CFC (β ≥ 0.13, p < 0.001) and PBC (β ≥ 0.20, p < 0.001) were more consistent, with the latter being again the superior predictor out of all three independent variables. One important aspect with regard to Hypothesis 2 was that a mediating effect from CFC to WTS to PEB intentions was observed, but that no such effect from CIC was found. The model fit indices indicated a mediocre to weak fit overall (RMSEAs ≥ 0.08, p > 0.001). All variables combined explained 19–25% of the variance in the DVs. Next, we ran a synthesized model for the DVs. We included variables from the previous models that had been identified as relevant and consistently predicted the DVs (i.e., CFC and PBC), but excluded those that were only relevant for some variables (e.g., SES was a predictor for traveling and heating, but not for electricity) and added large free correlations when indicated.

Because these synthesized models were not preregistered, we randomly split our file into two data sets for cross validation. We used data set 1 with N = 558 to identify potential path coefficients.

We then compared these to and verified these with data set 2 with N = 525. In Fig. 3E, we present the analysis for electricity only, as this was the only DV consisting of more than two items. However, the global findings were similar across models: Model fit parameters slightly improved, as only the relevant path coefficients remained in these models. However, these standardized coefficients fell within the small to medium spectrum (i.e., mostly between 0.10 and 0.25); PBC was an exception, as it showed somewhat larger effects.

In sum, in all of the confirmatory and synthesized models, we found that most of the explained variance was relatively low (20–30%) and was distributed among several variables with equally small to medium effects and with PBC usually contributing the most to the models.

Exploratory analysis using machine learning

The previously described finding implies that most energy-related PEB intentions to reduce GHG emissions have not yet been accounted for (i.e., 70 to 80%). Therefore, we considered that it would be beneficial to identify variables that might have individual explanatory power beyond the effects of PBC.

Hence, we conducted exploratory analyses to identify better (sets of) predictors for the DVs via machine learning as described above. A comparison of the error rates across all algorithms showed that the lowest values were reached when the two random forest approaches were taken (0.92 ≥ RMSE ≤ 1.63), whereas the neural network solutions showed the highest average error rates across models (1.39 ≥ RMSE ≤ 3.43). Therefore, we selected the conditional random forest approach for the default ten-fold cross-validation procedure that would enable us to rank the potential variables (i.e., features). An important consideration was that most variable extractions suggested that fewer rather than more variables were important for a robust model fit. In Table 2 we provide an overview of the extracted features from the superior conditional random forests along with the importance scores.Footnote 5

In all but one model, PBC outperformed all other variables that were included in the algorithms. Moreover, a comparison of these models suggests that the confirmatory analyses already included the most important variables, namely the consideration of future consequences and willingness to sacrifice (see last column in Table 2). Other variables, such as age, climate change skepticism, or values other than biospheric values did not occur consistently and were ranked mostly in the lower end of the top five importance scores.

We then split the data for PBC and re-ran the CRF models to see if different variables would become prevalent for people who ranked high or low in perceived behavior control. However, the sample sizes for low-PBC people (Table S3) were very small, and diverse results were accordingly obtained; therefore, an interpretation of these results would be premature. Results for high-PBC people – with a considerably higher sample size – showed that CFC and CIC were the best predictors, but the explained variance was below the estimate of the full data set (0.08 ≥ R2 ≤ 0.23). Hence, no additional relevant variables beyond the ones from the confirmatory models could be identified.

Discussion

Identifying social and psychological indicators that can consistently and reliably predict different forms of relevant energy-related PEBs to reduce GHG emissions is an important goal for researchers in the social sciences and in psychology. Importantly, effects of these variables should also be meaningfully large, account for large shares in the variation in PEB. This is especially important, as virtually all countries around the globe are striving to reach net zero carbon emission within the next 30 to 40 years (IPCC, 2022), which will require large-scale behavioral changes by their people in less than two generations.

In a quota-representative sample of the Austrian population, we evaluated different energy-related PEB intentions, arguing that a focus on this domain is useful, as it encompasses several forms of behavior linked to significant reductions in GHG emissions. Moreover, the energy domain might not be too broad of a measure, especially in comparison to previous studies, which often inappropriately combined PEBs that may not be related.

We included a set of social-psychological variables that are often used in both large-scale surveys and smaller studies. Research argues that these variables (i.e., biospheric values, the consideration of future and immediate consequences, willingness to sacrifice, and perceived behavioral control) are important for explaining various kinds of PEBs (e.g., Joireman et al., 2004; Khachatryan et al., 2013; Jugert et al., 2016; Steg, 2016; Van der Werff and Steg et al., 2016, van Valkengoed & Steg, 2019; Thaller et al., 2020). Hence, we expected that these variables would explain many of the energy-related PEB intentions to reduce GHG emissions.

The results of our survey are two-fold. On the one hand, we found statistically significant path coefficients in the SEMs which partially corroborated our hypotheses and which exceeded our preregistered minimum threshold of β = 0.10. On the other hand, these effects were small to medium in size (Cohen, 1992) for most of the psychological variables across PEBs, and usually between β = 0.10 and 0.25, which constituted explained degrees of variance between approximately 1 to 6% per variable. One exception was the perceived behavioral control, which we measured in close relation to the PEB intentions at hand. Potentially due to the similarities to the PEBs, standardized coefficients mostly explained between 0.2 and 0.37 or 4 to 14% of explained variance, which constitute medium-sized effects. This result indicates that personally perceived barriers (or their absence) are the most crucial factor for individuals to perform energy-efficient behavior. If such a behavior is too time-consuming, too financial costly, or inconveniently interferes in their daily life, it is unlikely that a person is changing it for the common good (see also Pinker, 2021, Chapter 3). These effects are similar to those reported by some other studies that investigated intentions for PEBs or self-reported PEBs (e.g., Aral & López-Sintas, 2022; Gkargkavouzi et al., 2019; Korfiatis et al, 2004; Miller et al., 2022; Wang et al., 2021), but probably not sufficiently large to justify a future intervention for meaningful behavior change.

Another notably small effect results from the mediation via willingness to sacrifice. WTS has sometimes been treated as a proxy PEB intention (i.e., a DV), but our results show that future researchers should refrain from doing so: Although WTS was highly correlated with BIOS (Model 1) and CFC (Model 2), its own effect on specific intentions was more than halved when models included these two predictors. This highlights the fact that relatively vague intentions to constrain oneself in response to the environment might not be reflected in salient, concrete behavior, thereby demonstrating a large intention vs. behavior gap (see Kormos & Gifford, 2014; Lange et al., 2018) or, in our case, a vague vs. concrete intention gap.

Overall, although our confirmatory models account for an accumulated 20 to 30% of the degrees of variance in the dependent variables, this, in turn, implies that 70 to 80% of the variance in PEB intentions could not be explained consistently by using four and more explanatory variables. Therefore, we hesitated to consider these as statistically significant and sufficiently powered effects as meaningful and subsequently explored other variables with effects that equaled or exceeded the one from the most relevant predictor, PBC.

Taking an exploratory machine learning approach, however, did not contribute further to the identification of variables. We could neither identify a variable that systematically outcompeted PBC in the top-ranked variables nor any combination of selected top-ranked variables that could explain much more overall variance. Moreover, these variables largely resembled the ones that had already been selected for the confirmatory analysis. In other words, the theoretically important social-psychological predictors were indeed the best possible predictors, but we believe that their practical importance is limited.Footnote 6 For future research on specific intentions and energy behavior, our results imply to focus on perceived behavior control, consideration of future and immediate consequences, and willingness to sacrifice and central psychological predictors – but less on biospheric values, as this predictor was constantly outperformed in the synthesized and machine learning models.

Limitations

Although the results of the present study are partially representative for Austria, replication studies are required to evaluate whether these findings hold in larger samples. As our effects were small and the machine learning approach was exploratory in nature, we would suggest to back up our results in another Austrian sample with at least twice as many participants. A sample size of N ~ 2000 would even allow to follow up machine learning by testing relevant extracted features with new data with adequate power (Pargent et al., 2023; Szabelska et al., 2021). Moreover, a larger sample would also allow to account for other relevant predictors that we might have missed (e.g., social norms, see below).

Afterwards, one should test, whether these results can be transferred to other industrialized countries where societal and political dynamics are different. For instance, a large consensus about climate change occurs in many EU countries (European Commission, 2021), but more polarization between left and right is observed in the US on this topic (Pew Research Center, 2021). This could possibly affect the estimates for both political orientation and biospheric values, yielding larger effects. Therefore, country-specific replication studies with similar variables are required. Likewise, analyses of developing countries (see Morren & Grinstein, 2016) with different cultural contexts and where the national authorities do not yet provide far-reaching subsidies for green and sustainable solutions to achieve the energy transition (e.g., Coady et al., 2019; Papadimitriou et al., 2020), might also yield different results.

Due to study-specific constraints, we might have failed to integrate theoretically important variables. As we argued in the introduction, we selected variables that we assumed were important, central and relatively independent person-specific predictors. As one could see in the case of CFC and BIOS, however, the effects of BIOS were reduced to a large degree when CFC was included in the model. This could occur again when other variables are accounted for. Certainly, many other constructs could be examined along with the variables we used, such as the connectedness-to-nature scale (Martin & Czellar, 2017; Mayer & Frantz, 2004), pragmatic prospection (Baumeister et al., 2016), or social identity and collective norms (e.g., Fritsche et al., 2018; Römpke et al., 2019). Especially the latter are crucial to understand societies long-term changes in values and behaviors, although their incorporation in cross-sectional surveys may be challenging. Moreover, the size and direction of the effects of additional variables, along with potential confounding factors, should be carefully considered beforehand (see Rohrer, 2018).

Moreover, our PEB measures addressed intentions to perform energy-related behaviors. One could argue that, although the behavioral intentions were formulated clearly, they would not represent real behavior. Although this is true, we argue that the relatively low levels of association between these specific intentions and the vaguer framing of willingness to reduce GHG emissions demonstrate that participants really did envision specific types of behavior. Nonetheless, future studies should test if actual behavior in this domain is indeed strongly related to specific intentions (as compared to vague ones), but this is likely to be achieved in smaller study designs.

Finally, and on a technical level, we had initially decided to place our specific PEBs and social-psychological variables at the end of this survey, whereas the former part consisted of different Austrian policies regarding mobility. Although it is unlikely, this could have influenced the answer behavior of participants on the latter part of the survey, as they could have felt less motivated and potentially answered our questions on GHG emissions less thoroughly.

Suggestions

If social scientists and psychologists have the goal to identify which variables may have a meaningful effect on energy-related PEB intentions in the future, it will be useful and potentially necessary to 1) define specific intentions and their measures, 2) pre-specify predictors along with thresholds for meaningful effects beyond statistical significance, and 3) do this in systematic studies (e.g., large and representative surveys, multi-lab studies, as well as experimental studies and longitudinal intervention).

The first point is crucial, because PEBs have many different impacts on GHG emissions, and not everyone would or could perform the same behavior. This became obvious in the present study within the domain of energy-related behavior (see, e.g., Fig. 2A). Similarly, we would advise researchers to select PEBs that have the potential to make a significant impact (in terms of GHG emission reduction). Although our benchmark of more than one-tenth of a ton CO2 eq/cap was based on previous studies (Ivanova et al., 2020; Lacroix, 2018), one could certainly opt for other, even higher benchmarks that could be in accordance with net zero-emission goals (Koide et al., 2021). In this study, our goal was to set a preliminary threshold and hope that other studies will follow.

The second point is equally important, as we want to identify avenues to adapt people's energy-related PEBs, and social-psychological predictors can be central for future interventions. The success of future interventions will rely heavily on how meaningful these predictors really are. Here, we have set a statistical threshold that was not just based on the common alpha-error rate of 0.05, but also on statistical power of at least 80%. This threshold was exceeded by most variables in our model. However, we also paid attention to the size of the effects obtained, which did not account for more than 30% of the variance with mostly small effect sizes. As most of the variation in energy-related PEB will not be caught by any combination of our psychological variables, we suggest that future researchers should confidently set even higher benchmarks by, for example, building on effect sizes either in standardized or unstandardized metrics (Lovakov & Agadullina, 2021; Pek & Flora, 2018), when other socially relevant variables are tested (e.g., from social practice theory, see Shove & Walker, 2014). In this regard, one important take-away message from our cross-sectional survey is that no individual person-specific variable is sufficient to explain energy-related behavior – instead, it is often a set of variables that differs by the behavior of interest.

The third reason for performing systematic studies is to receive unbiased information about how people within a country truly differ and to identify potentially relevant subsamples. This will not be possible by solely relying on small Western student samples tested in single studies, which is still common in psychological research (Newson et al., 2018). Instead, studies need to scale up either in terms of their participants or measures. The former can be achieved either by carrying out large representative surveys (as we did) or multi-lab studies, where a single paradigm is tested across several labs and countries simultaneously (see, e.g., Moshontz et al., 2018; Miller et al., 2022). In this way, contextual factors, such as the country-specific policies and regulations, can be taken into account and compared (e.g., one contextual factor could be that not all national authorities provide green subsidies; in these countries, participants might feel less control over their behavior—lower PBC -, which in turn might have a larger negative effect on their pro-environmental behavior). The latter can be achieved by conducting experimental and longitudinal studies with interventions directed at actual energy-related PEBs instead of intentions (e.g., Lange et al., 2018). Those studies are often resource-intensive, but could take contextual factors more effectively into account. They would go beyond self-reported PEBs and intentions and, in combination with registered report protocols, research design and diagnostic strength would potentially improve (Lange et al., 2020; Soderberg et al., 2021).

On a methodological note, newly developed methods allow us to identify predictors for specific PEBs more efficiently. We have outlined one of these methods in the exploratory part of this paper: One could apply machine learning (e.g., supervised machine learning defining DVs) to large datasets to explore relevant predictor variables, especially if many variables are included in the survey. These data-driven exploratory approaches may also help to refine the impact-oriented approach, where variations in relevant PEBs are defined first based on real-world observations (i.e., behaviors that reduce GHG emissions) before variation, correlations and causations of these PEBs are investigated (Nielsen et al., 2021a). This would help to complement current research in the field of PEB, which is mostly theory-driven (Lange and Dewitte, 2019), but has not supported an efficient search for both significant PEBs and meaningful predictors that practitioners can build on.

Conclusion

This survey provided a snapshot of the energy-related behavioral intentions of the Austrian population. We could show that no single person-specific social-psychological predictor, but a set of predictors – foremost perceived behavior control, consideration of future and immediate consequences, and willingness to sacrifice – account for less than a 30 percent of the variance for most intentions. This indicates that these behaviors should not be seen as one pro-environmental behavior, but need to be studied individually and systematically. For future survey like ours, we suggest using multi-lab approaches and machine learning to draw broader conclusions.