Climate concern elasticity of carbon footprint

The income elasticity of carbon footprint is a summary variable often used to describe the relationship between income and carbon footprints. Previous studies primarily calculate this elasticity using emissions intensities per monetary unit. However, this study is based on a survey conducted in Nordic countries which allows us to directly calculate carbon footprints from responses about quantity and types of activities instead of from spending. As a result, we curtail an inbuilt relationship between income and carbon footprints. As a result, our method produces an income elasticity estimate that is approximately one-fourth of the highest estimates that exist, and 30% lower than the smallest current estimate. Furthermore, we introduce a new summary variable called the climate concern elasticity of carbon footprint. This variable provides a simple method to analyze the relationship between pro-climate attitudes, pro-climate behavior, and carbon footprints. This new parameter can serve as a framework that identifies key areas where the investigation of the relationship between people’s climate concern and their carbon footprint may be most useful. This framework and improved knowledge of income elasticities can guide policymakers and future research and provide new methods to estimate carbon footprint distributions.


Introduction
The scientific community has widely recognized anthropogenic climate change as a threat to the stability of the climate's regulating systems and, therefore, to current human living conditions (Masson-Delmotte et al 2018). The United Nations Framework Convention on Climate Change's Paris Agreement ratified in 2016 required that the increase in global average temperatures should be held within 2°C above pre-industrial levels, and efforts should be made to limit the temperature increase to 1.5°C (Wu et al 2020).
Recently, researchers have started examining what lifestyles are associated with small enough carbon footprints to reach this goal (van den Berg et al 2021, Akenji et al 2021). The study of consumption-based carbon footprints is essential to such endeavors; several studies suggest that consumption-based accounting captures carbon leakage, such as outsourcing greenhouse gas emissions outside the territorial boundaries, which other accounting measures, including production-based accounting, do not (Ottelin et al 2019, p. 2). The importance of the consumption-based carbon footprint methodology and its use for demand-side solutions was recently recognized by the IPCC for the first time (IPCC 2022).
Literature on consumption-based carbon footprint has shown that footprint is highly positively linked to wealth (e.g. Wiedenhofer et al 2017, Wiedmann et al 2020, Barros and Wilk 2021). Wealthier countries' percapita carbon footprint is considerably higher than in poorer countries (Hubacek et al 2017, Clarke et al 2017, and affluent individuals tend to have a higher carbon footprint than less affluent individuals (Heinonen et al 2013, Minx et al 2013. This pattern further tends to persist within cities (Fry et al 2018, Moran et al 2018. To reach the global average per-capita emissions needed by 2030 to stay within the IPCC 1.5°C limit, Oxfam estimates that the emissions of the wealthiest 10% should be reduced by approximately 90%, while the poorest 50% could still increase their carbon footprint by two or three-fold . This idea aligns with a set Baiocchi et al (2010) explicitly investigated this connection and found that income elasticity in the UK goes from 1.6 at the lowest income level to 0.6 at the mean income and up again to 1.2 at the highest income level. Expenditure elasticities (relating expenditure to carbon footprint in a given category) and coefficients of determination (R 2 ) are considerably higher than income elasticities and their R 2 (Pottier 2022, pp 4-5), which indicates that additional income is spent variably at different income levels. If expenditure is constantly related to carbon footprint, this divergence can be fully explained by changes in the savings rate. Since this is not necessarily the case, it is unclear what explanatory factors exactly cause this divergence. Almost every previous study on income elasticities uses a direct input-output (I/O) approach (Peters 2008, Miller andBlair 2009). This is based on a linearity assumption that higher spending directly leads to higher emissions at some constant rate that can only vary based on different sectoral intensities. However, more expensive options of a good or service often have a lower associated carbon footprint than their cheaper alternatives (Vringer andBlok 1995, Hertwich 2005, e.g.).
Our study supplies two primary novelties: It largely avoids the inbuilt relationship between income and carbon footprint that is common in income elasticity studies, and it provides a new method for looking at the link between pro-climate attitudes and carbon emissions. To curtail the issue of a direct link between spending and carbon footprint, we base carbon footprint calculations primarily on types of activities and quantities rather than expenditures and carbon intensities per monetary unit. For example, we consider fuel and vehicle types in calculating carbon footprint due to vehicle use. Therefore, our elasticities are less tightly bound to the inherent (and weak) linearity assumption characteristic of classic I/O assessments. Second, by studying climate concern elasticities of carbon footprint, we consider a possible important determinant of climate behavior and consumption. Attitudinal variables have seldom been available in carbon footprint studies, making our data ideal for examining the possible effects of pro-climate attitudes. In addition to these novelties, our study provides robust justification and discussion of different methodologies; it covers multiple countries, thereby increasing its generalizability; and it investigates whether, with this data and methodology, income elasticity remains constant or changes across income levels.
We discuss the survey design, the statistical methods, regression specifics, and other research design elements in section 2. Section 3 summarizes the main results, and a discussion and our conclusions follow in sections 4 and 5. This study has multiple countries, domains, and possible regressions. Therefore, we include a sizeable supplementary section with additional results and discussion.

Research design 2.1. Data and survey design
We provide an overview of the survey design and data, but a more detailed description can be found in Heinonen et al (2022a). The data consists of more than 5,000 responses to a carbon footprint calculator survey conducted in five Nordic countries: Denmark, Finland, Iceland, Sweden, and Norway. The survey was administered on the University of Iceland's web server at carbonfootprint.hi.is and was tailored to each country in the footprint assessment, income level questions, and language of the questions, with all primary languages of these countries being included (see table 1). The Nordic countries' affluence brings a large part of the population above a subsistence level, underpinning the role of lifestyle choices. In these circumstances, one could expect climate attitudes to have a considerable effect. Our method is different from the standard bottom-up survey in a few ways. First, carbon footprints are calculated from direct answers to questions rather than being built from microdata. Second, the large number of questions provides an opportunity to link climate concern to carbon footprints and link climate concern to the relationship between income and carbon footprints. The survey was intended to collect consumption-based carbon footprints, climate attitudes, engagement in pro-climate behavior, and self-reported quality of life of the respondents. Each country had a slightly different survey since footprint assessments, language options, and income level questions were tailored to the country in question. Previous studies have based carbon footprint calculations on spending and income data from household budget surveys and average carbon intensities within goods categories. Carbon intensities are, in turn, obtained from input-output techniques that transform emissions intensities of production into emission intensities of final products. Our assessment has important improvements over these traditional carbon footprint assessments. For example, expenditure data do not capture air travel emissions well. Instead, our assessment is based on trip numbers and distances and includes the significant effect of short-term radiative forcers (e.g. Lee et al 2021). Second, our car travel emissions estimate is based on reported kilometers traveled and allows for various fuel types instead of using an average carbon intensity per monetary unit. Third, rather than spending on public and long-distance ground transport, we utilize travel modes and distances. Lastly, energy spending is often inseparable from rent and other housing management fees (Heinonen et al 2020). Our method captures home sizes and building ages, which are more reliable.

Distribution of the survey
An online marketing company distributed information about the survey and invitations to participate on Facebook and other social media platforms, through advertisements and posts in Facebook groups. The survey was distributed between the fall of 2021 and the spring of 2022, for approximately 2 months in each country. Survey participation was limited to adults residing in one of the surveyed countries that participate in the household's financials in some form. Respondents were encouraged to share their carbon footprint results, prompting others to take it. A few news outlets shared the survey, leading to further responses. The aim was to get as many complete responses as possible rather than to achieve socio-demographic representativeness for the respective countries (representativeness can be gauged partially from table B1 in Supplementary section B).
Possible biases due to this are discussed in section 4.3.

Survey respondents
The number of unique respondents between autumn 2021 and spring 2022 was 13,924, of which 7,682 answered the whole questionnaire. Each participant gave their consent to use the responses in this study. Furthermore, no information is presented that would enable the identification of individual respondents. We avoid multiple responses from the same people in the data by asking at the start if the respondent had taken the survey before. If they had, we kept the first response and erased the duplicate response from the data. We furthermore removed a few obviously erroneous responses. Incomplete responses were removed, so missing data was not included in the regression. Responses were required for carbon footprint-related responses for the carbon footprint calculator, so there were no missing responses to these questions. However, responses were not required for the climate concern-related questions necessary to construct our climate concern variable (see section 2.2.2). The final number that responded to each of the questions used to create the climate concern variable was 7,431. Finally, to decrease the likelihood of including extreme cases of over-and under-reporting, we removed carbon footprint outliers outside a z-score of -3 and 3 for the total footprint and each domain footprint, which would leave around 99.7% of the responses if they were normally distributed. However, since the carbon footprint calculations result in discrete estimates, many points can be in the range's maximum and minimum. Therefore, we used a threshold of 2%, where if the number of data points identified with our z-score exceeded that threshold, these points would not be removed. The final sample consisted of 5,161 responses, with 334 responses from Denmark, 1384 from Finland, 1023 from Iceland, 939 from Norway, and 1481 from Sweden. Respondents provided extensive background information, but the primary focus of this study is responses relating to income levels, carbon footprints, and climate concern.
2.1.3. Capturing income Income was captured both by asking about personal income and total household income. In this study primarily use the response to the latter, as described in section 2.2.1. The respondents had to choose from the ranges for each income decile for that country. Deciles were determined based on survey data aggregated by Eurostat and exchange rates for 2018. The 10th (highest) decile was highly heterogenous, so it was split in two to reach a smaller range of possible incomes. We term the most affluent bracket, which we split from the 10th decile, the 11th decile. We used the midpoint of income in the inner deciles to code brackets of this categorical income variable into single values. The resulting income variable is, therefore, discrete.

Carbon footprint calculations
Carbon footprints were calculated using a personal consumption-based approach (Baynes and Wiedmann 2012) where the emissions are allocated to the end consumer of the good or service, regardless of the location of the purchase or where the emissions take place (Heinonen et al 2022b) combined with a hybrid I/O assessment model in which key components are calculated using information about the process through all steps of creating the good or service, but smaller impact categories are calculated using a direct I/O approach. Survey respondents estimated their private consumption of goods and services over a period of one year, meaning that the survey's scope includes all personal consumption except for purchases of durable goods, homes, and vehicles. In line with our previous work and suggestions that the scope of the system and classification of consumption-based carbon footprints needs to be clear for the replication and comparison of studies (Heinonen et al 2022b), we calculated footprints in nine domains: diet, housing energy, vehicle use, public transport, leisure travel, goods and services, second homes, and pets. We succinctly describe these domains here, but more specific information on calculation methods and formulae for each domain can be found in Heinonen et al (2022a) as well as online at carbonfootprint.hi.is. The food domain is based on responses about diet from vegan to omnivore; housing energy is based on respondents' answers about housing type, the decade of construction, heating mode, electricity, and size of the home; private vehicle possession and use is based on the number of vehicles in possession of their household and respondent's answers about the fuel efficiency, fuel type, and distance driven in the past 12 months for each vehicle; public transport was based on respondent's estimates about personal average weekly use of public transport in kilometers; leisure travel is based on the number and length of trips taken for leisure, and mode of travel; for goods and services respondents were asked to provide information on their purchases in multiple domains in line with the Classification of Individual Consumption According to Purpose (United Nations 2018); the pets domain is based on answers about the number of cats and dogs owned; second home footprint is based on Ottelin et al (2015).
Emissions were calculated using a lifecycle approach (LCA) based on current literature. We applied the I/O approach to goods and services, composing approximately 10%-20% of the carbon footprint in the study for each country. Emission intensities of purchased goods and services were calculated using the Exiobase I/O model (Stadler et al 2018), which is a multiregional I/O model including almost all European countries as well as other parts of the world at a lower resolution. To match the Exiobase sectors with purchasing data, a concordance matrix was created following Ottelin et al (2020). Potential uncertainties and limitations from our assumptions are discussed in section 4.3 (Limitations).

Constructed variables
In addition to calculating domain-specific and total carbon footprints, we construct a variable named climate concern and household income per capita.

Household income per capita
The affluence of an individual is dependent both on the affluence of the household as well as the household's size. Therefore, our study constructed the household income per capita variable by dividing household income by the number of people in the household and used this variable in carbon intensity calculations. This variable of interest uses household income since it indicates the purchasing power of the household, within which consumption is shared. This is also the most typical income variable used in carbon footprint studies (Heinonen et al 2020). Because the original household income variable was discrete, this household income per capita is discrete but with more possible values. This variable is also qualitatively different from household income. For household income, the highest earners are older households with or without children. However, for household income per capita, the highest earners are often younger adults that do not have children. The exact exchange rates used for currency conversion can be found online at carbonfootprint.hi.is/methodology.

Climate concern
Respondents were asked ten questions relating to their climate attitudes. Answer choices to each question were 'not at all,' 'slightly,' 'moderately,' 'very,' and 'extremely,' which were respectively given numeric values from 1-5. Subsequently, factor analysis based on these ten attitude variables was performed which returned five critical variables. The questions for each of these five variables are shown in table 2. The arithmetic mean of these five chosen variables constitutes the climate concern value for any given respondent. The final variable is, therefore, discrete and in the range between 1 and 5, but not necessarily an integer. The construction of the climate concern variable is described in greater detail by Abdirova (2022).

Multivariate regression analysis
Regression analysis is a highly popular method for analyzing the relationship between variables. One variable is the so-called dependent variable-which is the total or domain-specific carbon footprint in our study-and the other variables are so-called regressor variables-which are primarily income and climate concerns in our study. The regressor variables (regressors) are assumed to relate with the dependent variable linearly, and the regression estimates that linear relationship. Our estimates of interest are elasticities, which can be obtained with a log-log regression where we take the natural logarithm of the dependent variable and regressors. Elasticities are useful because they provide us with the percentage change in the dependent variable associated with a 1% change in an independent variable rather than a numerical change given a numerical change of 1 in the other. We look at a linear regression for the relationship between climate concern, income, and carbon footprint. The supplementary section will include results with all combinations of regressing total or domain carbon footprint regressing on one, some, or all of income, climate concern, and a constant. The main body of the paper will include all of these, i.e. the primary regression equation is: )is the carbon footprint in domain d (including total carbon footprint), income elasticity is β 1 and climate concern elasticity is β 2 . Coefficients β 1 , and β 2 are estimated by minimizing the mean squared error. The survey is slightly different across countries in language and response options where needed due to differences in availability, such as methods to heat one's home. Therefore, the independence of observations (i.i.d.) occurs only within a stratum of countries. Accordingly, we repeat this regression for each country. To clarify that this is the standard case with the whole sample, particularly people with all incomes, we call the estimates from this model the 'across-decile' elasticities. For across-decile elasticities, we remove the 1st and '11th' deciles since we cannot take a midpoint of each range to estimate the income. By removing them, we intend to minimize the risk of skewing the results due to choosing an unrealistic 'midpoint.' To further inspect how income and climate concern elasticities behave as a function of income, we use 'indecile' elasticity estimates. These estimates restrict the regression equation to an income group, and we can plot the progression of these estimates over income groups. Papers have shown various behavior of income elasticity as a function of income (Pottier 2022, p. 9). For example, a related elasticity-the expenditure elasticity of carbon footprint-is U-shaped in the UK but bell-shaped over income in Brazil (Baiocchi et al 2010, Cohen et al 2005. Similarly, we seek to identify this shape in the Nordic countries when we control for climate concern and to investigate the behavior of climate concern elasticity over income similarly. Therefore, the results include the standard' across-decile and in-decile estimates of income and climate concern elasticities. To confirm the major directions and results, we also execute what we call the single regression method. This regression is similar to our standard regression ( * ), but instead of limiting the across-decile regression to a country, we include the whole set of responses and regress on country. Thus, the regression equation is: Where ∑ j=1 β j × Country j is a sum of indicator variables for each country except for one country (to avoid multicollinearity) which acts as the base. Other symbols are the same as in equation ( * ). To be able to combine countries into one regression, we must first convert all currencies to the same monetary unit. To this end, since exchange rates do not capture differences in purchasing power, we use Eurostat's purchasing power parity (PPP) database (Eurostat 2022a). This merely incorporates the total PPP between countries, thereby leaving sectoral price differences unnoticed. For the elasticity of total carbon footprint, this sort of comparison is still logical, but this precludes domain-based regressions using PPP.

Zero values and the log-of-0 fix
One issue for log-log regressions is that when a variableʼs value approaches 0 from above, it approaches negative infinity. However, a 0 carbon footprint in a domain is important because it indicates those who have chosen not to participate-e.g., someone with 0 carbon footprint in the domain of vehicle possession is someone who has no car, potentially due to their concern for the climate. To ameliorate the issue that the logarithm of 0 does not exist, we use what Bellégo et al (2022) term the 'popular fix,' which involves adding a small term to all variables before taking their logarithm. We add 0.0001 to all values in that country's carbon footprint domain if it has a 0 value. We implemented this for the following domains: vehicle use, public transport, leisure travel, pets, and second homes. We also removed 12 erroneous observations which contained a 0 carbon footprint in the housing energy and goods and services domains. The popular log-of-0 fix will give us smaller-than-true standard errors and might result in a bias (Bellégo et al 2022). However, its simplicity and widespread use is valuable for reproducing these results and for further use. For the domains where this 'popular fix' was implemented, we found non-normal residuals, violating the assumption of normality. This does not necessarily bias our estimated elasticities, but it may bias standard errors (as well as confidence intervals and p-values) unless the sample size is large enough (Schmidt and Finan 2018). We discuss our choice of methodology over other methods that are often implemented when residuals are non-normal in section 4.2, and the possible limitations of our choice in section 4.3.

Controlling for More socio-demographic variables
Our regression model does not include a vast array of controls. Adding more controls, such as age, gender, and urbanity, has no significant effect on the elasticity estimates. Therefore, given the large number of regressions and results, we opted for clarity by not reporting these controls. The limitations for doing this and suggestions for improvements in future studies are discussed in the limitations section (section 4.3).
2.4. Income and climate concern elasticity of carbon footprint The income elasticity allows us to look at the relationship between income and carbon footprints. The log-log transformation in our multivariate regression allows us to identify elasticities (the percentage change in our outcome variable given a percentage change in an independent variable). We look at this elasticity in multiple domains of footprints to identify areas where further research could be conducted to establish effectual links. Furthermore, we control for climate concern to see if doing so affects the income elasticity estimate. In section 2.3.2 we described our reason for not including more control variables, and in section 4, we discuss the limitations of doing so.
Our dataset has environmental attitudes allowing us to investigate the relationship between pro-climate attitudes and carbon footprints through a variable similarly constructed to income elasticity. That is, we investigate a new benchmark elasticity-the climate concern elasticity of carbon footprint. We use the same methods for climate concern elasticity of carbon footprints as for the income elasticity.

Results
Using the methodology and data described above, we find a low but statistically significant income elasticity, and similarly sized and statistically significant negative climate concern elasticity of total carbon footprint. We find that the relationships between income or climate concern and carbon footprint vary greatly between domains, giving elastic, inelastic, positive, and negative elasticities in different domains. The climate concern elasticity of total carbon footprint is slightly U-shaped over income (i.e., is more negative for the middle income deciles), but seems to be the same across income deciles in each carbon footprint domain. We are unable to determine the relationship of income elasticity over income. Finally, controlling for climate concern has no significant effect on the income elasticity estimate. The following sections provide more detailed results which contribute to these key findings. Section 3.1 provides a brief and simple overview of the relationship between carbon footprint and income or climate concern. The following section (3.2) introduces the results from the most straightforward regression-the single regression method. Subsequently, we look at the across-decile income and climate concern elasticity estimates in section 3.3, focusing on important domains and interesting cases. In this section (3.3), we also briefly discuss R 2 values, the key results accrued by calculating in-decile elasticity estimates, and briefly describe the differences in income elasticity when controlled for climate concern. Greater detail for some results can be found in the supplementary section below. Note that the responses in our final data for Denmark are only 334. This results in very large confidence intervals for almost all estimates. Therefore, we excluded Danish estimates from most results and discussions.
3.1. Descriptive analysis of mean carbon footprints, incomes, and climate concern Figure 1 shows us the size and split by domain of carbon footprint per capita in each country, based on the survey. From this figure, we see that the largest domains for emissions are diets, vehicle use, leisure travel, and goods and services. Housing Energy and public transport emissions vary a lot by country due to very low carbon intensities in Iceland and Norway, and higher but still relatively low carbon intensities in Sweden, compared to Denmark and Finland. In a global comparison, these Danish and Finnish carbon intensities are low. Second homes and pets are minor components of the average carbon footprint. Finally, cleaning the data leads to the footprints being lower than in the original pre-cleaned data.
There is no evident second-order trend in how carbon footprint relates to climate concern, as seen in figure 2. The figure plots all observations with carbon footprint on the y-axis and climate concern on the x-axis, with different coloring and trendlines for each country. The polynomial fit has a slightly concave curve for each country, except Denmark and Sweden. Figure 3 shows the relationships between climate concern, carbon footprint, and income deciles for all the studied countries. We see that climate concern is fairly constant across income deciles; carbon footprint is higher with higher income deciles; and carbon footprint is roughly, but slowly, decreasing with climate concern. Figure A.1 in supplementary section A includes a more detailed version of this graph, where it is split into countries. Table 3 shows the across-decile results when we use regression ( ** ), including significance levels and confidence intervals. We have excluded an indicator variable for Sweden to avoid multicollinearity. Therefore, the countryindicator variables compare against Sweden, which has the lowest carbon footprint per capita. We get similar results to those we get from the regression ( * ), which performs the regression only within a certain country. For  3. Boxplots graphing carbon footprints, income, and climate concern over each other.  Note. The two numbers in parenthesis below income and climate concern coefficients are 95% confidence intervals.

Single regression method
example, for total carbon footprint as a dependent variable, we get 0.22 for income elasticity and -0.21 for climate concern elasticity. That is, carbon footprint increases by 2.2% for a 10% increase in income, and carbon footprint decreases by 2.1% when the climate concern variable increases by 10%. Income elasticities ranged from 0.17 to 0.25, and climate concern elasticities ranged from −0.23 to −0.18 with regression ( ** ). The same goes for each domain of carbon footprint; the estimate in this regression falls in the range of estimates from regression ( ** ). Table C1 in supplementary section C is the same as table 3 but includes interaction terms between countries and income and countries and climate concern. In section C, we find that there is little variability in the income and climate concern elasticity effects between countries. Table 4 shows the income and climate concern elasticities of carbon footprint from our primary regression equation ( * ) for all countries and domains, including significance levels and confidence intervals. Note that income and climate concern elasticities are statistically significant in most cases. We also have mostly similar patterns of direction and size comparison in elasticities. Elasticities of total carbon footprints are very similar across each country: income elasticity estimates span from 0.17 in Iceland to 0.25 in Norway, and climate concern elasticity estimates span from -0.18 in Norway down to −0.23 in Denmark. Income elasticity is the largest positive for leisure travel in all countries but Denmark, which has a very large confidence interval. Income elasticity is the most negative for pets in all countries but Denmark, but Denmark and Finland have statistically insignificant estimates. Although climate concern elasticity estimates are quite different in size between countries, the size and sign comparison between domains is even more similar in all countries than income elasticities. Climate concern elasticity is the most negative for vehicle use in all countries, the highest positive for public transport, and the second-highest positive for leisure travel. We highlight some findings for the more major domains:

Across-decile elasticities
Diet: • We fail to reject the hypothesis that income elasticity of diet footprint is non-zero at alpha = 0.05, except for Denmark, where the income elasticity is slightly negative.
• However, climate concern elasticity is negative and statistically significant.
Vehicle use: • Only in Finland can we reject at the alpha=0.05 level that income elasticity is 0.
• As we mentioned, vehicle use is strongly negatively related to climate concern but is differentially related to income.
Leisure travel: • Income elasticity of leisure travel footprint is positive and is well above 1 when statistically significant.
• Climate concern elasticity is large and positive in all countries, except in Iceland, where the estimate is not statistically significantly different from 0.
Goods and services: • Income elasticity of goods and services footprint is positive but well below parity and statistically significant at the alpha = 0.01 level.
• The climate concern elasticity is close to 0 in all countries but Denmark, where it is negative.
Housing Energy: • Income elasticity of housing energy footprint is positive and statistically significant for all countries. The elasticity is lower in the countries where the domain is a significant part of the total carbon footprint-in Finland (0.13), Denmark (0.32), and Sweden (0.33)-than where it is a minor proportion of the total footprint, such as in Iceland (0.49) and Norway (0.48).
• Climate concern elasticity of housing energy footprint is negative and similar in size in all countries but Denmark, where it is positive.
Supplementary section F provides tables F2-F9 with results for other variations of this regression-including one, two, or all three of climate concern, income, and a constant as regressors. Coefficient of determination (R 2 ) Note that the adjusted R 2 is generally fairly low and sometimes even close to or equal to 0. Nonetheless, provides tables F1 through F9 in the appendix shows that R 2 behaves as it should since it is higher when we add climate concern as a control. We also see, however, in tables X-Z that R 2 is much higher (often close to or equal to 1) when we do not include an intercept. We discuss this behavior in the discussion section.
In-Decile elasticities Supplementary sections D and E provide figures D1 and D2 of in-decile climate concern elasticities and figure E1 for in-decile income elasticities, with a brief discussion. Section E also includes table E1, which shows income variances in subgroups in our sample, which is relevant to the discussion of indecile income elasticities. In-decile climate concern elasticities of total carbon footprint are slightly U-shaped. For the carbon footprint domains, however, climate concern elasticity remains fairly constant over income. Indecile estimates of income elasticities have considerable uncertainty due to a small variation in incomes when the data is reduced to deciles. Therefore, we cannot adequately decipher any trend or non-trend from our data.
Income elasticity and controlling for concern As mentioned before, supplementary section F displays results for other regression variations, which shows us what happens to income elasticities when we control for climate concern. The effect is minuscule and without any clear trend; controlling for climate concern increases or decreases income elasticity by a small amount. Since this amount is insignificant, we conclude that the effect is close to none.

Discussion
Our study had two aims: first, to provide an improved estimate of income elasticity using novel data, and second, to use this novel data to create a variable to summarize the relationship between climate attitudes and carbon footprints succinctly that mirrors income elasticity literature. To do this, we conducted log-log regressions and found that income may be less important in individuals' consumption decisions than previously thought. Furthermore, we found that individuals' concerns about the climate may be important in their consumption decisions. The small income elasticity is a sign that in the common (typically I/O-driven) methods used to estimate income elasticity, there is inherently a very strong income-emissions relationship that may misconstrue the importance of households' individual income to carbon emissions. This severely destabilizes common methodology and indicates that any efforts to affect income to reduce carbon footprints may be ineffective. The significant climate concern elasticities indicate that consumption preferences can greatly influence carbon emissions, and investigating this variable can highlight the areas where preferences are a large driver of emissions.

Meaning of numerical results
The range of definitive estimates of the income elasticity of carbon footprint in previous studies is from 0.34 (lowest with controls is 0.31) to around 0.86 (Pottier 2022, pp. 4-5). Therefore, our estimate of the income elasticity of total carbon footprint is approximately 30% lower than the lowest previous estimate and is only about one-fourth of the highest current estimate (0.86). A few factors may be at play. First, as described in section 2.1, previous studies have a considerably inbuilt connection between income and carbon footprints, which our survey method ameliorates. Second, few studies have covered the topic of income elasticity in terms of consumption-based carbon footprint, as we do in our study. As mentioned in the introduction, this is beneficial since consumption-based accounting has been argued to capture carbon leakage, which other accounting measures, such as production-based accounting, do not (Ottelin et al 2019, p. 2). Lastly, the Nordic countries are highly affluent, so income levels in these countries are less restrictive to consumption than in most other countries. This result has been found before for leisure travel; A 2019 study found that income did not restrict international travel among Icelanders (Czepkiewicz et al 2019). We note, once again, that carbon footprints in the Nordic countries remain high compared to the global average and are far from those suggested as compatible with the 1.5-degree warming target by Akenji et al (2021) (which are also in line with the IPCC (2021) SSP1-1.9 pathway). Our scope is, therefore, on an area where some of the most significant reductions in carbon footprint must occur, making this research area vital for reducing emissions.
The climate concern elasticity of total carbon footprint is approximately similarly sized to the income elasticity, in the negative direction. Based on our single regression results, a 10% increase in income is related to a 2.2% increase in carbon footprint and a 10% increase in climate concern is related to a 2.1% decrease in carbon footprint. Some in-domain elasticities are, however, highly negative, some highly positive, and some close to 0. These differences between in-domain elasticities might align with previous evidence on compensatory beliefs. Studies have found compensatory beliefs, which means that when some people reduce high-emission behavior in one domain, they believe they have done 'enough' and can therefore do less in other domains (Nayum and Thøgersen 2022). An even lower climate concern elasticity might be expected in less affluent countries, as income restricts the available consumption bundles. Investigating whether this holds could provide insight into the restrictive effects of income. We conclude that domain-based climate concern elasticities may provide valuable insights and that the high statistical significance justifies further investigating climate concern elasticities. Domain-restricted elasticities provide a more detailed insight into the relationship between income or climate concern and emissions. We discuss the major results for both income and climate concern elasticities domain-by-domain for those domains that constitute a large portion of the average carbon footprint: Diet: • Income elasticity: Diet income elasticities are close to 0, which is reconcilable with the idea that income is not largely restrictive to a choice of diet. This is in line with a 2014 study that found that income had little explanatory power on sustainable eating in the Nordic countries (Niva et al 2014, p. 474).
• Climate concern elasticity: The negative and statistically significant climate concern elasticity is in line with a study that showed that sustainable food consumption in the Nordic countries was positively correlated to support for environmental policies (a proxy for pro-climate attitudes) (Niva et al 2014).
Vehicle use: • Income elasticity: Havranek and Kokes (2015) estimated the mean long-run income elasticity of gasoline demand as 0.66, establishing a positive relationship. Only our Finnish income elasticity is statistically significantly positive in our standard regression. The confidence intervals are large for the other estimates, perhaps due to a high inherent variability or uncertainty from our estimation method. However, the single regression estimate is statistically significant (0.91), supporting the result of Havranek and Kokes (2015).
Although there do not seem to be any mobility poverty studies covering Nordic countries, it is likely that mobility poverty is low, which would also partially explain our finding. Furthermore, property prices are much higher in locations with good public transport access (Weisbrod and Reno 2009), leading to a trade-off between housing costs and gasoline purchases.
• Climate concern elasticity: Vehicle use footprint is highly sensitive to climate concern. The negative relation between climate concern and vehicle use might indicate that climate-concerned people tend to avoid private vehicles in all Nordic countries. In line with our finding, Árnadóttir et al (2019) found that pro-environmental attitudes partially drive the use of public transport, and environmental benefits seem to be key motivators for the mass adoption of EVs (Singh et al 2020).
Public transport: • Income elasticity: The connection between public transport and income has not been studied extensively. The statistically significantly positive elasticity in Norway and Sweden indicates that higher income is related to higher carbon footprints from public transport. In Denmark, Finland, and Iceland, the uncertainty is high, so we cannot reject that the elasticity is 0. This result is in line with qualitative results from Heinonen et al (2021), which indicated that affluent Icelanders do not use public transport.
• Climate concern elasticity: Climate concern elasticity of public transport footprints is positive and relatively large. Intuitively, this indicates that climate-concerned individuals choose public transport to a higher degree. This result is consistent with the strong negative climate concern elasticity of vehicle use; climate-concerned individuals may choose public transport over private vehicles. However, if this story holds, since the positive elasticity for public transport is larger in size than the negative vehicle use elasticity, the overall emissions are increased. However, the two may be completely unrelated. Future studies should try to determine if this hypothesis holds.
Leisure travel: • Income elasticity: The positive income elasticity estimates indicate that higher incomes are considerably related to more leisure travel. However, the elasticity is lower and not statistically significant from 0 in Iceland. Since Iceland is an island, it could be that people have a higher urge to fly out of the country. Árnadóttir et al (2021) found that Icelanders save in other areas in order not to have to compromise on their flights.
• Climate concern elasticity: The high climate concern elasticity for leisure travel indicates that more climateconcerned individuals go on more leisure trips. This could be due to compensatory behavior. Icelanders indicated in interviews that they allow themselves to fly more since they are reducing their carbon footprint in other areas (Árnadóttir et al 2021, p. 285).
Goods and services: • Income elasticity: Incomes allow for higher spending. Elasticities below one could be due to an increased savings rate. Our data shows that savings rates do not increase significantly across deciles. There is uncertainty due to the vagueness of our question, but it is nonetheless in the same range as a different data source in the Nordic countries (Eurostat 2022b, OECD 2022, European Federation of Building Societies 2020, p. 48). Therefore, the difference could instead be explained by different types of goods bought with higher income.
• Climate concern elasticity: Since climate concern elasticity is close to 0, it might be that climate-concerned people do not compromise on their consumption of goods and services despite their concern. Confirming the reasons behind these relationships could be of high value to policies. For example, insufficient regulation of producers might restrict the effect of climate concern due to greenwashing or lack of more sustainable options.
Housing energy: • Income elasticity: The income elasticities here align with a paper based on a household consumption survey which found an income elasticity of housing energy greenhouse gas (GHG) emissions of 0.577 in Finland when controlling for urbanity and household size (Ala-Mantila et al 2014).
• Climate concern elasticity: Small negative climate concern elasticities of housing energy footprints indicate a relationship between concern and housing energy carbon footprint, but it is minuscule. This could indicate that pro-environmental housing energy choices, such as installing heat pumps or improving insulation, do not have a large effect. Furthermore, many climate-concerned people may not compromise on the size of their homes. On the other hand, choices might be restricted and different when renting or owning homes, and there may not be many types of homes available.
Overall, we see that the domain-specific elasticities provide information that can highlight interesting future research areas. We note that the single regression method ( ** ) aligns with our standard regression's ( * ) domain and total carbon footprint-related results. It, therefore, serves as a confirmation of our results. As mentioned in the results section, supplementary sections D and E provide information and a brief discussion of in-decile elasticities. We only note a few things here. First, income elasticity estimates were too uncertain to be used for our analysis, leaving little to discuss. Second, in-decile climate concern elasticities also have fairly large confidence intervals, but they are small enough to provide us with some information. We do not see any evident trends in climate concern elasticities over income. It is, therefore, possible that this variable does not significantly change based on income. Only for the total carbon footprint is there some indication of a slight U-shape, but this needs to be investigated by others. Climate concern elasticity does not change with income, and controlling for climate concern has a minuscule, if any, effect on income elasticity, which indicates the possibility that income and climate concern are relatively independent actors on carbon footprint. Figure 3 intuitively supports the potential independence of these two variables since it shows that climate concern tends to be similar across all incomes, indicating a low covariance. Pottier (2022) proposes that adding a variability unrelated to income can be used to recover the full distribution (rather than only the concentration) of carbon footprint in a population. Our results suggest that climate concern might fulfill that role.
Our low R 2 is a testament to the high variance in carbon footprints unrelated to income and climate concern. Nonetheless, the often high statistical significance indicates that those relations exist. Thus, we see that income and climate concern are often definitively related to carbon footprint. However, they are also relatively small determinants of carbon footprint since their relation to carbon footprint is small. Supplementary section G gives a more detailed discussion of our R 2 values and possible explanations.

Methodological considerations
As is standard, the assumption of normality (and homoscedasticity) of the errors (residuals) was empirically evaluated by examining the distribution of residuals, which is done by analyzing their density or conducting residual analysis (Pek et al 2018). For regressions where the dependent variable was a carbon footprint domain with non-erroneous 0-values, we found that the regression residuals were non-normal when using the 'popular fix.' When the assumption of normality of residuals is violated, the following approaches are used: data transformations (applying non-linear functions to the data), invoking the central limit theorem (CLT), rankbased non-parametric approaches (e.g., the sign test), the bootstrap, trimming to remove outliers, Winsorizing (i.e. changing the values of outliers to be less extreme), heteroscedastic consistent covariance matrices (HCCMs; which allows for heteroscedasticity), and non-linear models (e.g., logistic regression) (Pek et al 2018). Lastly, we consider the option of removing 0-values instead of implementing the 'popular fix.' Only one of these methods can be used appropriately for this study.
Data transformations are often applied to force the residuals to follow a normal distribution. However, doing so can adversely affect the interpretability of the resulting estimated effect, its statistical power and confidence intervals, and can bias slope coefficients (Pek et al 2017). Furthermore, our study aims at improving the income elasticity of carbon footprint and creating a new variable that mirrors that methodology, which necessitates, in particular, a log transformation of dependent and independent variables. Transforming the data would, therefore, go against this goal.
Similarly, we opted to use a linear model rather than a non-linear model. It would be good for future research to look at whether the relationship between climate concern and carbon footprint is non-linear in specific domains. However, since our results show that every domain that does not allow non-participation in the activity works well with linear models, there is no reason to believe that the relationship is non-linear rather than that the non-normal residuals result from the issue of taking logarithms of 0.
Another option, instead of the 'popular fix' of adding 0.0001 to 0-values before taking the logarithm (as described in section 2.3.1), would be to remove 0-values from the domains where they appear. In fact, removing 0-values in the domains where non-participation is viable makes the residuals normally distributed. However, as discussed in section 4, removing 0-values also obfuscates how to interpret the results since non-participation is an important consumption choice. Removing 0-values led to similar income elasticities in most domainssupporting our income elasticity results-but lower climate concern elasticity, which is uninterpretable because it takes out those who choose not to participate in an activity due to their emissions (e.g., those who do not own a car because of their concern for the climate). Results when 0-values are removed only describe a relationship among those who participate in the activity, reducing the usefulness of its interpretation and biasing the regression coefficient toward zero. Therefore, although removing 0-values rather than using the 'popular fix' may improve the fit between the variables and the model, it reduces the interpretability so heavily that it becomes unproductive. However, the fact that a linear regression model fits well when non-participants are removed indicates that a linear relationship exists among those who participate. Therefore, the relationship may be piecewisely linear or linear amongst participants but non-linear between non-participants and participants.
Rank-based non-parametric approaches circumvent the non-normality of residuals because they abandon the linear model to analyze the data ranks. This could be an interesting avenue for future research, but greatly affects the interpretability.
For both the bootstrap and HCCMs, the data is analyzed with a linear model, and the data is not altered, but the non-normality of residuals is treated as non-informative and a nuisance to be addressed. Instead, the bootstrap utilizes the less restrictive assumption that the sample is representative of the population. However, by definition, a small sample cannot be representative, thereby making the bootstrap a large sample method akin to the CLT. Furthermore, the study did not aim for representativeness (see sections 2.1.1 and 4.3). Both of these reasons make the bootstrap a subpar choice in our situation. Nonetheless, this machine-learning method may be interesting for future analyses of the issues discussed in this study but does not mirror literature on the income elasticity of carbon footprints which traditionally use log-log linear regressions. The HCCM, however, is a smallsample method, which we exclude as an option due to our considerably large data size (Pek et al 2018).
Trimming and Winsorizing both assume outliers contaminate the observed data (Wilcox 2011, Pek et al 2018). To avoid heavily biased estimates by trimming or Winsorizing more data if extreme cases are true parts of the population rather than outliers, only a small percentage can be altered or removed. As described in section 2.1.2, outliers were trimmed according to certain specifications meant not to remove extreme cases common enough for them to possibly be true measurements. Doing this did not resolve the issue of non-normal residuals.
Research has found that, in line with the CLT, violations of the normality assumption of residuals in linear regression analyses do not necessarily bias point estimates. However, the normality assumption is needed to estimate standard errors unbiasedly, thereby affecting confidence intervals and P-values. However, when the sample is large, violations of the normality assumption often do not noticeably impact results (Schmidt and Finan 2018). The CLT justifies our choice of methodology due to our large sample size. According to the CLT, the normality assumption is uninformative when the sample size is large. That is, violating the assumption of normal errors is not necessarily fatal when the sample size, N, is large enough for the CLT to be at work. The CLT assures both the robustness of model results and assures that the sampling distribution of estimates will converge toward a normal distribution as N increases to infinity, when errors are independent and identically distributed, and when the standard deviation is finite. Exactly what N is large enough is uncertain. Rules of thumb are inaccurate because the size needed for the CLT to apply depends on the number of independent variables in the regression as well as the extent of non-normality of the errors. Schmidt and Finan (2018) showed that for different data-generating processes, the coverage (i.e., the number of times the 95% confidence interval included the true slope coefficient) increases quickly with larger N. Pek et al (2017) simulates populations to identify the size of N needed when the skew from normality is higher, finding that a ratio of 10 or lower observations N per parameter may suffice. Still, the ratio may need to be larger (e.g., 50) when variables are correlated or when variable distributions are in localized sparse data settings. This ratio of the number of subjects over the number of variables is 2580.5 5161 2 = for our primary regression equation ( * ) with all countries, and even if we would only take Denmark, which had the fewest observations, it would be 167. Lastly, if we add more controls to the regression (e.g., age, gender, urbanity), this would be well above 50. Therefore, we conclude that our sample size is high enough for accepting the distribution assumption violation. We discuss the limitations of using the 'popular fix' and having non-normal residuals in the domains where this fix was implemented in further detail in section 4.3 (Limitations).
We also note that we found a slightly 3-peaked rather than fully normal distribution of residuals for the diet domain. Although this is not due to the log-of-0 fix, we believe the same applies: it is still useful despite violating the normality assumption. Knief and Forstmeier (2021) shows, using Monte Carlo simulations, that despite even the most dramatic violations of the normality assumption, there is no increased risk of obtaining false-positive results (Type I Error). According to the simulations conducted by Knief and Forstmeier (2021), regression coefficients for larger sample sizes (above 100) should be unbiased, which follows from the Lindeberg-Feller Central Limit theorem (Lumley et al 2002). In fact, P-values are highly robust to even extreme violations of the normality assumption and can be trusted, except when there are extreme outliers in the variable distributions. However, as described in section 2.1.2, we removed any potential outliers. Therefore, we take our results in the diet footprint domain to be valid.

Limitations and future research potentials
The limitations of our study can be divided into three main categories: those arising from the assumptions and methods used in our statistical analysis, the scope and temporality of the survey, and the representativeness and calculation of carbon footprints.
First, standard errors are underestimated and biases may arise due to our log-of-0 fix (Bellégo et al 2022). Furthermore, the violation of the normality assumption-occurring in the domains of vehicle use, public transport, leisure travel, pets, and second homes-may not affect the point estimate itself (Schmidt and Finan 2018). However, if our sample size is not large enough, it may reduce the accuracy of our standard errors (and then also confidence intervals and p-values). Therefore, since both of these reasons might lead to the underestimation of standard errors, we should be less confident in the statistical significance of the estimates in these domains, and it is especially important to focus on estimates that are significant at the alpha = 0.01 level.
Various other methods are regularly used when the normality assumption is violated. As alluded to in the previous section (4.2), some of these do not apply to our situation, but others might provide interesting avenues for future research into the relationships between climate concern, carbon emissions, and affluence. The most promising alternative methods from this list are rank-based non-parametric and bootstrap approaches. Knief and Forstmeier (2021) note that rank-based non-parametric approaches may be advisable to avoid the negative consequences of strong deviations from normality. This research would need to investigate how to interpret the results of this approach, as it is less straightforward than the well-known linear regression. However, bootstrap approaches may improve inference and be more easily interpretable than rank-based non-parametric approaches while also circumventing the non-normality issue (Pek et al 2017). The pitfall of these two alternative methods is that they diverge from previous literature on income elasticities of carbon footprints and have more obfuscated interpretations. Violations of non-normality are sometimes interpreted as an indication of a nonlinear relationship. We investigated a few non-linear regressions when deciding on what regression to use for this study and found no convincing evidence of non-linear relationships. However, these alternative methods might be better to uncover non-linear relationships for the domains in question. It is possible that the effect of climate concern has mostly to do with participation than with emissions if already participating. Therefore, one final option for future research may be to look at the effect of concern or income on (non-)participation, e.g., by creating an indicator variable where 1 is assigned if the respondent participates in the activity and 0 if not. This question of (non-)action may provide interesting insights. Furthermore, as we mentioned in section 4.2, the relationship among participants seems to be linear, but the relationship between participants and nonparticipants is unknown. Therefore, a future investigation of (non-)participation may inform a piece-wise model for the relationship between concern and carbon emissions.
As mentioned in section 2.3.2, adding more controls-in particular, age, gender, and urbanity-had no significant effects on the elasticity estimates. Therefore, we chose not to include them in our regression for easier legibility. However, although some studies do not include controls and argue that characteristics other than income and household size are irrelevant (Wier et al 2001, Büchs andSchnepf 2013), some previous studies have had regression coefficients change considerably when controls were added (e.g., Büchs and Schnepf (2013), found a change from 0.6 to 0.43 when adding controls). Including important socio-demographic controls in regressions is standard and generally advisable, and doing so may provide a richer analysis. Despite their negligible effect in our study, we suggest that future research capture and use important socio-demographic variables as controls to reduce the risk of possible confounding effects. In particular, some studies have found that employment status, education, urbanity, household composition, and age are related to emissions (Büchs andSchnepf 2013, Gough et al 2011), indicating these elements as potentially interesting variables to collect and control for.
Adding more controls does have the beneficial effect of increasing the R 2 of each of our regressions, so for future purposes, including more controls may also help increase the goodness-of-fit of the models. For a more in-depth analysis of our R 2 values, please refer to Supplementary section G.
Second, our intended scope of the Nordic countries makes any results only directly applicable within the Nordic countries. Other research must confirm our results and investigate how they differ in other areas. Third, survey data has well-known limitations. Respondents may have reported false values, deliberately or not.
Although we had a large sample size, it was insufficient to analyze some values for Denmark, which had fewer respondents than the other countries. In other countries, however, the large sample size and our data cleaning will have reduced the risk and effects of any misreporting. The survey was also conducted in 2021 and 2022 during the COVID-19 pandemic. However, since the calculated carbon footprints align with previous average estimates, we believe this has not caused a consequential bias. Furthermore, there are a varying number of respondents within each income decile, but each income decile was constructed using country averages. Therefore, the overall carbon footprints and incomes are not necessarily in line with the country averages, but the average in each decile should be reasonably well in line with the country averages. The most likely source of bias due to the COVID-19 pandemic is long-distance or leisure travel. Nonetheless, a large number of respondents reported air travel. Compared to Czepkiewicz et al (2018Czepkiewicz et al ( , 2019, the emissions in this domain in our data are likely somewhat downward biased. Furthermore, the cleaned data reduces average footprints in all countries compared to the original data, with only a handful of apparent outliers removed. Third, our analysis may suffer from a lack of representativeness as well as from leaving out the production of some major durable goods, such as the production of homes and cars. The survey group's lack of representativeness, which may arise due to non-probability sampling, may cause a sampling bias. However, it does not necessarily bias our results for our investigated variables. In future efforts, investigating the effects of post-stratification may be of value. Our goal was to introduce a framework for looking at the connections between climate concern or income and emissions, but future efforts should aim to gain representative samples. The production of vehicles can add as much as 1 ton to the carbon footprint of the most affluent (Heinonen et al 2020). This could cause downward bias in both the vehicle use carbon footprint estimate and the income elasticity estimate. Constructing homes is also a significant component of carbon footprints (e.g. Ottelin et al (2015)). However, it is commonly left out and included in a component of capital formation (e.g. Mach et al 2018).
Lastly, our domain-specific elasticities identify the domains where climate concern impacts may be strong. It may be interesting to see whether attitudes toward the importance of personal action or climate literacy strongly affect this impact. Furthermore, it may interesting to investigate longitudinal data for the development of climate concern and carbon footprint. Furthermore, it might also be valuable to investigate whether different constructions of the climate concern variable may provide different results.

Implications
Our small income elasticity estimate is a sign that in the common (typically I/O-driven) methods used to estimate income elasticity, there is inherently a very strong income-emissions relationship that may misconstrue the importance of households' individual income to carbon emissions. Therefore, future research must diverge from this methodology. The smaller income elasticity may indicate that harder policy measures (e.g., bans or taxes) relating to consumption may be more effective than policy measures aimed at changing individual income for the purpose of reducing carbon emissions. That is, policies aimed at affecting production or other systematic elements may be more effective.
Our domain-specific climate concern elasticities are statistically robust, and in the same order of size as income elasticity. This indicates that consumption preferences arising from climate concern may be just as important as households' budgets. The areas that have a large negative elasticity may need fewer hard measures to restrict more polluting consumption and would benefit from soft policy measures, such as information campaigns. However, since overall climate concern elasticity is far below one, it is likely not the best area for incentivizing or enabling sustainable lifestyles. Systematic measures to alter consumption choices are likely to be more effective. Our domain-specific climate concern elasticity estimates indicate areas where more ambitious policies to alter consumption behavior may have the largest effects.
Lastly, we hope that the varied domain-specific climate concern elasticity values provide a guide for domains where behavioral studies can investigate the psychological and economic reasons behind the values of these two elasticities.

Conclusion
In conclusion, we were able to show that income elasticity is lower than previously thought when the mechanical relationship between income and carbon footprints that occurs in popular methods is avoided and that the climate concern elasticity is an approximately equally large relationship. We showed that the relationship between income or climate concern with different domain-related carbon footprints is vastly different between domains, thereby providing a deep framework for future study and potential policies.
The relationships between climate concern or income to carbon footprint seem largely disconnected, which may support its use to create complete distributions of carbon footprint where data is scarce. Future research and policy can benefit from our results in a few ways. As described in section 4.4, our income and climate concern elasticity estimates both support the need for systematic policies aimed at altering the production chain of goods and services over policies aimed at altering individual income or climate concern. However, our domain-specific climate concern elasticity estimates indicate areas where sustainable consumption choices are possible, thereby highlighting areas where policies to push consumption choices in a certain direction may have the largest effect.
The study's aim was to destabilize the methods used for income elasticity and generate the climate concern elasticity variable for the first time. However, as our Limitations section (4.3) discusses, future research may make several improvements; in particular, using a different methodology, including more controls, and using a more representative dataset may all be worthwhile efforts. The income elasticity of carbon footprints has been studied for decades and can still be improved in its methodology and interpretation. Similarly, the climate concern elasticity will also require substantial further research. Further research should diverge from previous methods for income elasticity to confirm or deny our income and climate concern elasticity results in the Nordic countries as well as compare our results to other areas. Only if our results are confirmed can they be used for policy purposes. Comparisons with other geographic areas may highlight differences in how carbon footprints come to be in different societies.

Data availability statement
The data cannot be made publicly available upon publication due to legal restrictions preventing unrestricted public distribution. The data that support the findings of this study are available upon reasonable request from the authors.
Appendix A. Relations between carbon footprint, climate concern, and income Figure A1 shows us the same as figure 3 in section 3.1, except that this figure is cut into countries. This highlights differences between countries in the relationships between carbon footprint, income, and climate concern. Although the general trends are roughly similar in each country, they are more extreme in some. Table B1 provides us with the demographics of our sample population and in the population, as estimated by the national statistics offices in each of the countries. Figure A1. Relationships between carbon footprint, climate concern, and income in each country plotted against each other.  Table C1 shows regression ( ** ) with added interaction terms between countries and income or climate concern. Such interaction terms indicate whether the income or climate concern effects vary between countries. When terms are significant, they indicate some difference in that effect in that given country compared to the standard case across countries (which is left out as a regressor to avoid multicollinearity). There does not seem to be much variation in the effects between countries. However, there are fairly many significant terms for income elasticity in the housing energy and goods and services domains. Appendix D. In-Decile climate concern elasticities Figure D1 graphs the in-decile climate concern elasticities for all countries and their 95% confidence intervals for all deciles. Each figure constitutes either the elasticities of total carbon footprint or the elasticities for a specific domain. Generally, due to decile-cut samples being smaller than the original one, the confidence intervals are considerably larger than for the across-decile estimates. This can be confirmed by sight in figures D1 and D2, which include across-decile estimates. The confidence intervals are so large that they provide us with little information. We exclude Denmark from figure D1 since the low number of observations leads to very large confidence intervals, reducing our confidence in those estimates. Note that we have missing estimates and confidence intervals in some places. This is due to low or even no variability of income in the sample of that decile and country. Figure D2 shows the income and climate concern elasticities from our basis regression ( * ) with the total or domain-specific carbon footprint graphed separately by country. This figure focuses on the differences between domains rather than countries and, as opposed to figure D1, includes income elasticities. Indecile income elasticities are discussed in greater detail in the next supplementary section (E). We see in figure  D2 that the income elasticities have massive confidence intervals and quite irrational estimates. However, the indecile climate concern elasticities have considerably smaller confidence intervals, providing more information. Figure D1. Climate concern elasticities as a function of income deciles, separated by domains.

Appendix B. Sample and population demographics
Appendix E. In-decile income elasticities Figure E1 shows the in-decile income elasticity results for different domains, and table E1 shows income variances by country and decile. Due to low variance when we restrict to deciles, the in-decile income elasticity estimates have very low statistical confidence. Furthermore, it is unlikely that the real elasticities vary as much as the estimates and are as significant anywhere as it seems, e.g., in local and leisure travel, pets, and second homes. Their only value is that the confidence intervals largely overlap with the across-decile estimate's confidence intervals. When these do not intersect, it indicates that the true in-decile elasticity might be higher or lower than the across-decile elasticity. A specific description of each such case would be desultory. We instead include these figures in the case that they are helpful for researchers looking at in-decile income elasticities in similar areas in the future. Figure D2. Climate concern elasticities as a function of income deciles, separated by countries.

Appendix G. Interpreting the R 2 results
We note that when more controls are added to our regressions, the R 2 values become higher. Nonetheless, the improvements do not suffice to make the R 2 close to what is commonly accepted. Thus, we must analyze our R 2 results further. The reader can confirm in tables F2-F9 in supplementary chapter F that the R 2 is high when the intercept is left out of the regression. Some might argue that a no-intercept model should therefore be used. In some situations, it makes theoretical sense to assume that if the independent value is 0, the dependent variable ought to be 0. One should not exclude an intercept when the value of the independent variable (such as income) is far removed from 0 unless the data supports the use of such a model. Theoretically, if the reported income is the only source of income that can be used for consumption, then 0 income would mean no consumption and no consumption-based carbon footprint. However, one cannot, for example, live without consuming food. Therefore, this assumption would be fundamentally flawed. It is more likely that even if reported incomes are 0, there is some consumption-based carbon footprint when income and climate concern is 0. Previous papers analyzing income elasticity of carbon footprint have often included an intercept for a linear regression equation.
Only when using a non-linear regression specification have some excluded an intercept. Therefore, a fully specified equation (such as equations * and ** ) is preferable to any regression model excluding an intercept, despite a generally low R 2 . The high R 2 for a non-intercept model is likely a result of a differently specified computation of R 2 . When an intercept is included, R 2 is 1 minus the residual sum of squares (RSS) over the total sum of squares (TSS): Where i is each data point, y î indicates the prediction for i, and y¯indicates the average. Intuitiv ely, this compares the regression model in question to a reference model that merely uses a constant term. When we exclude the intercept, we compare it to noise only. In this case, the total sum of squares goes from y y i i 2 (¯) å to y i i 2 ( ) å . Thus, the denominator becomes larger. The R 2 estimate increases if this increase outweighs the change in RSS. If the data is such that it has a slope not very far from 0, and all of the data is far from the origin, then the correct R 2 should be low since a considerable part of the variation is also due to noise. This is the case for our data; There is a large variance in carbon footprints, and the process seems to have a slope near 0. As tables tables F2-F9 in supplementary section F highlight, the behavior of R 2 is standard since adding climate concern to the regression increases the R 2 .
It is not strange that the R 2 is lower here than in many previous studies of income elasticities. Since carbon footprints are calculated only by allowing people to purchase the "average good" in a given sector, the noise is mechanically reduced.
R 2 is likely low due to our high variance of carbon footprints making the residual sum of squares large. Despite this, it is important to note the often high statistical significance. This indicates that our regression model captures the mean response in each case well. A larger sample should help with this issue, as confirmed by our single regression method ( ** ), since the R 2 is either in the range we get from regression ( * ) or higher.
Furthermore, this variable for explained variance may be low due to a high variability due to the data coming from a survey. On the other hand, the high possible variability of carbon footprint for all incomes is a considerable strength of this dataset compared to the normal household budget survey data used to get income elasticities. A paper by Rössel (2008) is an excellent supplement to understanding and interpreting the R 2 from lifestyle surveys such as ours. It confirms that such surveys often have low R 2 . It argues that situations where behavior can be aestheticized, and in low-cost situations, the explanatory power becomes higher. Since most domains are mostly high-cost situations with restricted options, it is to be expected that R 2 will be low. Our results are not useful to (dis)confirm their results on aestheticizable behavior. Supplementing our results with an analysis of how aesthetic typical behavior in various domains are could provide useful to that end.
Lastly, R 2 has often been criticized for providing little value or insight, and some have argued that the standard error of regression should be used instead as a measure of good fit. We include it due to its widespread use but interpret it with caution. We did not explicitly report standard errors since they can be inferred from confidence intervals, which we deemed more valuable for the reader.