CER-ETH – Center of Economic Research at ETH Zurich

We provide an empirical analysis on the relation between culture and revealed environmental economic preferences. Switzerland’s citizens share all major institutions but belong to multiple population groups which differ by culture and language across distinct geographical locations. This unique setting allows us to disentangle the effect of culture on individual consumer preferences from institutional characteristics. We use a spatial fuzzy Regression Discontinuity Design (RDD) at the internal French/German language border on municipality level data to estimate the effect of culture on preferences for energy efficient vehicle design. Our results indicate that French speaking municipalities have a 5.5 percentage points higher share of energy efficient vehicles compared to their German counterparts. In addition, we find that popular votes on environmental issues receive a substantially higher share of approval in French speaking regions. This indicates that they place a higher value on the environment, which may be due to a higher sense of collectivism and altruism.


Introduction
In the context of growing attention to energy efficiency in environmental economics, there is evidence that economic incentives are not the only motivation for consumers to invest in energy saving. A growing literature suggests that culture is an important factor in economic decision making, for instance in explaining saving rates (Guiso et al., 2006) or pursuing business relations (Ahern et al., 2015) 1 . We believe there is strong evidence that culture also influences environmental decisions. In this study, we document differentiated patterns of energy efficient vehicle ownership in Switzerland and propose that culture is a major explanatory factor. Switzerland's citizens share all major institutions but belong to multiple population groups which differ by culture across distinct geographical locations. This unique setting allows us to separate the effect of culture on institutions and on individual consumer preferences.
The role of culture is increasingly present in the economics literature. Bisin and Verdier (2001) suggest that culture has a direct impact on individual preferences. Those findings are supported by both Fehr and Hoff (2011) and Hoff and Stiglitz (2016) who advocate the use of culture as an endogenous determinant of economic preferences. The application of culture can be observed in major economic fields: Guiso et al. (2006) suggest that "cultural variables are as important as economic variables in understanding cross-country differences in national savings rates"; Gould et al. (2011) show how culture shapes various individual economic outcomes at the example of Yemenite migrants in Israel. Ahern et al. (2015) find that "the volume of cross-border mergers is lower when countries are more culturally distant" and Guiso et al. (2009) propose that cultural similarities between countries affect trust which increases bilateral trade and direct investments. Giannetti and Yafeh (2012) argue that banks require higher interest rates to culturally distant borrowers and require more frequently third-party guarantees. Atkin (2016) suggests that cultural preferences for traditional foods can constrain caloric intake and contribute to malnutrition among migrants.
The objective of this paper is to analyze the impact of culture on consumer preferences for energy efficient cars. To do this, we can hardly rely on a cross-country comparison. In fact, with the use of cross-country data it is difficult to identify the effect of culture, since culture is often used to explain the structure of institutions, which, in return, determine economic differences between societies. For instance, Alesina and Giuliano (2015) argue that culture and institutions are both complements and interact in their impact on economic development.
From an empirical point of view, it is important to use an identification strategy that allows us to disentangle the impact of culture from the impact of institutions on the adoption of energy efficient vehicles. Such an identification strategy would be to analyze the behavior of consumers with different cultural backgrounds that live in the same region. This paper will develop an analysis of several cultural groups within one country, where most political and other institutions are homogeneous and therefore the variation in their impact is limited. A within-country analysis provides a natural experiment because institutions relevant to economic development are shared by the whole population, whereas culture is heterogeneous. The analysis will also take into account the impact of local (cantonal 2 ) norms and institution on fuel efficient vehicles and provide an estimator for the impact of culture, robust to regional specific effects.
In the environmental economics literature Bezin (2015) and Schumacher (2015) introduce culture in theoretical models in order to explain the transgression and formation of environmental preferences over generations. Empirical evidence of the effect of culture on environmental decisions mainly relies on stated preferences: Wilhite et al. (1996) propose with a series of interviews that differences in household energy consumption patterns between Norway and Japan are due to cultural differences. Shultz et al. (1998) and Hoyos et al. (2009) suggest using CVM that culture influences stated WTP for environmental goods. An identifica-tion strategy using revealed individual consumer preferences to show the effect of culture on environmental consumption, similar to the more general literature on culture, to our knowledge has not been used until now.
The mechanism by which culture influences environmental friendly consumption generally passes through positive attitudes towards the environment. Both, consumption and attitudes, are influenced by underlying norms and values, the core constituants of a culture. Kahneman et al. (1999) argue that actions and decisions are direct expressions of underlying attitudes, at the example of environmental protection. Welsch andKühling (2009) andVideras et al. (2012) show that proenvironmental consumption is influenced by the attitudes shared by an individual's social group. Brounen and Kok (2011) find that green party voters, an indicator for environmental attitudes, are more likely to adopt energy labels for the housing market. Kahn (2007) shows that the sales of hybrid and fuel efficient vehicles depend on the share of green party members per community. Gallagher and Muehlegger (2011) confirm those results by taking Sierra Club 3 membership as a proxy for environmental attitudes. According to the social psychology literature, both environmental attitudes and environmental behavior are influenced by underlying norms and values, such as individualism or altruism (Gifford and Nilsson, 2014). Norms and values in return are used to define a culture by Guiso et al. (2006). Empirical studies using revealed preferences to show the underlying mechanism of cultural influence on both environmental attitudes and consumption by connecting both of the to core values has received little attention up to now.
Our analysis has several contibutions to the literature: we provide an empirical analysis on the relation between culture and environmental economic preferences using revealed preferences and a clear identification strategy. Moreover we connect the existing literature on the influence of environmental attitudes on consumption to the impact of culture by associating values, the core element of a culture, to both environmental attitudes and consumption.
The empirical identification of the impact of culture on the adoption of energy efficient cars is based on a regression discontinuity approach. Specifically, to estimate the effect of culture on vehicle choice and on environmentalism, independently of institutions, we will conduct a spatial Fuzzy Regression Discontinuity Design (RDD) at the internal Swiss language border. Several studies have used language as a proxy for variation in culture and as a determinant of economic outcomes, e.g. Fearon (2003), Desmet et al. (2012), Chen (2013) or Falk et al. (2015). In a similar setting, and using a RDD approach, Eugster et al. (2011) describe the effect of culture on the demand for social insurance in Switzerland via differences in popular voting results, Egger and Lassmann (2015) show the effect of culture on international trade at the internal Swiss language border, and Gentili et al. (2017) analyze the effect of culture on decision making concerning elderly care.
Our results show that the population living in the French speaking parts of Switzerland owns a higher share of energy efficient cars than the German speaking population. The difference in energy efficiency is the result of different vehicle characteristics, mainly curb weight 4 , share of diesel engines and replacement rate. Moreover, popular votes on environmental issues receive a higher share of approval in the French speaking part of Switzerland, indicating a different level of concern for the environment.
The paper is organized as follows: the next section explains how language can be used as a proxy for culture in Switzerland. Section 3 provides the empirical strategy used in our analysis and evaluates the necessary conditions. In section 4 we explain the data used for vehicle and municipality characteristics. Our main results are presented and analyzed in section 5. In the final section we will present the conclusions, followed by an appendix containing robustness checks and graphs.

Languages as a proxy for culture
The literature strongly suggests that language can be used as a proxy for culture: Parsons et al. (1965) define culture as a "set of symbols of communication" which means that language is a medium for culture and at the same time an identifier of a specific culture. In linguistics, the meaning of language goes beyond that of a medium for culture: it also shapes the speaker's cognition of the world, making it an active consituent of a culture according to the concept of "Linguistic Relativity" (Whorf, 1997). In the economics literature, Chen (2013) explains differences in saving rates in Switzerland by the linguistic structure of the German and French language.
Switzerland consists of four language groups (German, French, Italian and Romansh) each of which occupy a distinct geographical area of the Swiss territory. According to the 2000 census, 72.5% of the population is German speaking, 21% French, 4.3% Italian and 0.6% Romansh, with a total population of ∼ 8 million 5 . Figure 1 depicts the distribution of the main language per municipality according to the 2000 census (data from the Swiss Federal Statistical Office (BFS)). The map shows a clear spatial separation of the German and French speaking municipalities, where the red line depicts the language border, the so-called "Röstigraben". Italian and Romansh-speaking areas are separated by the Alps from the German speaking part of Switzerland. In the empirical analysis of this paper, we consider the language border between the German speaking and the French speaking municipalities. We exclude the language border between the Italian/Romansh speaking areas and the German speaking regions because the Alps represent an important natural barrier between the language regions. In contrast, the German-French language border is not defined by geographical obstacles: the language border runs 5 A comparison with language data collected and merged from four microcensuses from 2011-2015 shows no significant change in the spatial distribution of the native language. Therefore, the spatial distribution of people speaking Italian, French, German or Romansh remains constant over time. We choose to use the language data from the census of the year 2000, because the microcensuses omit a small number of very small municipalities.
along the North-South axis of Switzerland, whereas the major geographical obstacle in Switzerland, the Alps, follow the East-West axis. The Swiss Confederation was founded in 1291 when the Central Swiss Cantons seceded from the German Empire and was successively enlarged over the following centuries. Modern Switzerland dates back to 1848, when the Swiss Constitution established a Federal State composed of 26 Cantons: 18 German speaking, four French speaking, three bilingual (German-French), one trilingual (Graubünden with German, Italian and Romansh).
In order to visualize the spatial distribution of French speakers in Switzerland, we plot the share of French speakers per municipality and the distance per municipality to the language border in Figure 2: as expected, municipalities located in the French regions show a high share of French speakers. Although there are no natural obstacles at the German-French language border, the share of native French speakers drops sharply at the border, which results in a clear cut language-border (see Figure 2). Assimilation and integration processes ensure a dominance of the respective culture in its realm. The persistence of the respective native language in each region indicates that aside from the native population, immigrants either come from the same language area (France or Germany) or adopt the dominant native language in the second generation. In addition, Novembre et al. (2008) found that genetic markers between people living in the French and Germanic area of Switzerland show a higher variation between than within those regions, which indicates that individuals are more likely to start a family with a partner of their own language group. In this context, we can assume that language is a proxy for culture: values, beliefs and attitudes are transmitted within groups, which are defined by their common native language.

Empirical Strategy
From the econometric point of view, the impact of culture on the adoption of energy efficient cars is analyzed using two approaches, a panel regression and a RDD. First, we conduct a linear panel regression analysis using the basic version of the Mundlak correction (Mundlak, 1978) to correct for time-invariant unobserved heterogeneity. We use the following model: where Y is the share of energy efficient cars in municipality i in year t. "language" is the share of French speakers in the year 2000, X it is a set of socio-economic variables at the municipality level containing the share of cars by French producers, income per capita, population density, elevation and a dummy for urban areas.X i is the average over time of X it (naturally, elevation is constant over time) Second, we adopt a spatial fuzzy RDD using the distance of each municipality to the language border. Of course, the RDD is our preferred identification strategy.
In this setting at the proximity of the language border, a regression on efficient vehicle ownership and language will show the average treatment effect of culture. We use a fuzzy RDD because unlike in a sharp RDD, and as shown in Figure 2, the probability of treatment (speaking French) does not change from zero to one but rather from 0.04 to 0.85, with small differences in each municipality.
An important step in the application of this approach is to first visually analyze the level of discontinuity of the dependent variable, the share of energy efficient cars in the proximity to the language border. The level of energy efficiency of a car can be measured by different indicators, such as CO 2 emissions, fuel efficiency or a system of energy labels. In Switzerland, the government has introduced energy efficiency labels for cars (ranging from A to G) that are relative to curb weight, and where A-and B-labels represent the most energy efficient cars. In our study, we consider this label system as the most informative variable for the Swiss consumers. Therefore, the dependent variable in eq (1) is the share of A-and B-label vehicles. In Figure 3, we represent the distribution of municipalities with their share of cars with an A-or B-label and can see a discontinuity at the language border. This observation is confirmed in Figure 6 (Appendix), where we represent the distributions of municipalities with their values for average CO 2 emmisions and average fuel consumption per car. The RDD is based on the estimation of a regression that uses only data within a selected bandwidth above or below the cut-point. In this context, the choice of bandwidth is crucial and has to trade off between opposing criteria: a larger bandwidth contains more precision-increasing observations, whereas a smaller bandwidth minimizes bias.
In order to use a RDD approach, three conditions should be fulfilled (Lee and Lemieux, 2010). First, the running variable to determine the discontinuity must be continuous. In this case the running variable is the distance to the language border in km, which is continuous. Second, agents should not be able to precisely control whether they receive the treatment or not. This condition is fulfilled in our setting, since the native language is not a choice an individual can make and native language is used as a proxy for culture. Third, covariates must be balanced at the discontinuity or vary smoothly. Importantly, there should be no discontinuous change at the language border in order to have no correlation of the covariates with the treatment.
The classical and most important explanatory variables used in the empirical literature to explain the demand for cars are income per capita, population density, topography, fuel prices and policy measures such as fuel economy standards and subsidies. Johansson and Schipper (1997) find that income, fuel price, population density and national fuel economy standards have a positive impact on the adoption of fuel efficient vehicles. Small and Van Dender (2007) obtained similar results, although income shows no statistical significance. Similarly, Klier et al. (2010) and Beresteanu and Li (2011) estimate the volume of hybrid car sales in the US by using standard economic variables which reflect price sensitivity. Both studies find a positive impact of fuel price on fuel intensity. In addition, Beresteanu and Li (2011) suggest a lower price sensitivity of high income consumers, a positive impact of large households with children as well as consumers living in urban areas on the sales volume of hybrid vehicles. More recent findings by Heutel and Muehlegger (2015) confirm earlier results. The literature also includes non-monetary incentives which influence car choice: Goldberg and Verboven (2001) suggest that consumers have a home bias for domestically produced cars and Sexton and Sexton (2014) find evidence for willingness to pay for a hybrid vehicle as a status symbol. Moreover, Allcott and Knittel (2017) conclude after two experimental studies that imperfect information and inattention do not explain consumer choices with respect to fuel economy. For Switzerland, we can also assume that geographical differences such as altitude could also play a role. Some of these proposed variables could show variation at the regional level, whereas others not. For instance, fuel economy standards do not differ within Switzerland, and the price for cars, the price for fuel and the quality of the public transportation system are generally homogeneous at the regional level.
As discussed previously, in a RDD, those factors have to be independent of the treatment variable (culture) in order to isolate its effect. Therefore, we examine, using figures and computing statistics, the variation at the language border of the most important explanatory variables. Figure 7 (Appendix) shows that in 2010 most of the socio-economic covariates do not show a discontinuity except altitude. Even if we can observe a slight mean difference in the population density, the variance in this variable is high, which suggests that the mean difference (and thus the discontinuity) is not significant. The share of cars by French producers does not show a discontinuity at the language border, which implies that the French speaking population does not have a preference for French cars. Altitude does show a difference at the language border, this difference however is not sustained throughout all Switzerland. Table 1 compares the mean difference of outcome variable, treatment and controls at a narrow range of 25km from the language border with a t-test for the year 2010 6 . The results confirm the intuition given by the graphs in figures 2, 3, 6 and 7 (Appendix), that outcome and treatment change discontinuously whereas most of the controls do not. The difference between the French and German speaking regions is evaluated using Welch's t-test. The share of French cars shows a significant difference, however using an alternative test based on a RDD (results shown in the Appendix, Table 8), the difference is non-significant. Moreover, we control for this variable in one of the model specifications. Municipality elevation shows as well a significant difference at a bandwidth of 25km, throughout the entire dataset there are however not two clearcut groups as with language. Moreover, Filippini et al. (2015) showed that elevation is associated with higher fuel consumption, which means in our context that omiting elevation from the model would lead to an overestimation of the effect of culture. The map in Figure 4 highlights the area at a distance of 25km from the language border. Not all municipalities at the language border are highlighted in the map because the relevant measure is the street distance to the border. Some municipalities, even though they are located at the border, are in mountainous regions which means that their driving distance to the next municipality exceeds 25km.

Figure 4: Share of A and B label vehicles, 25 km from language border (in 2010)
The language border is marked in red Source: MOFIS/TARGA dataset In the fuzzy RDD we estimate a standard cross section model and add the distance of each municipality to the language border as well as the interaction term of language and distance. In a second step, we subsequently decrease the bandwidth (i.e. distance to the language border) in order to minimize the effect of structural factors other than language. Both models use a 2SLS approach to implement the fuzzy RDD. In the first stage, we use a dummy indicating whether the municipality is located in the French-speaking region, as an instrument for the percentage of French-speakers. The second stage of the cross section fuzzy RDD is specified as follows: where Y is the share of A and B-labeled vehicles in municipality i. "language" is the share of French-speakers, F i is the distance of the municipality to the lanugage border for French speaking municipalities or for German speaking ones (this specifications allows for different slopes on either side of the language border). F is a j-th degree polynomial of the distance to the language border. X i contains a set of socio-economic variables at the municipality level: the share of cars by French producers, income per capita, population density, elevation and a dummy for urban areas.
14 4 Data Description For the analysis, we use data on individual vehicle registration in Switzerland from 2008 to 2012, the share of French native speakers per municipality, the distance of each municipality to the language border and a set of covariates. Information on the share of energy efficient vehicles (cars with A and B labels, CO 2 emissions and level of fuel efficiency) at the municipality level has been obtained from a data set created at the Center for Energy Policy and Economics at the ETH Zurich. 7 We use income data from the Swiss Federal Tax Administration on the direct federal tax. As income we use the total taxable net revenue per municipality for natural persons, measured in thousand CHF. Population data was obtained from the BFS, using the balance of permanent residents by districts and municipalities, where the population for each year is the mean of the population at the beginning and at the end of the year, measured in 1000s of inhabitants. The area of each municipality was obtained from the area statistics of Switzerland 2004/09 from the BFS and is measured in ha. Altitude is measured in meters above sea level and depicts the median value per municipality, obtained from the BFS. Data on the quality of Public Transportation in 2010 was obtained from the Federal Office for Spatial Development (ARE). The quality of public transportation per municipality ranges from A to D (if the municipality has a public transportation system), where A indicates a high quality, measured by proximity of transportation stops and the frequency of transport. Results from the 2011 federal election were obtained from the BFS and are given in percentage points. The language statistics contain the number of native speakers of each language per municipality, according to the census of the year 2000. The data was obtained from the BFS and municipality mergers were performed retroactively. 8 7 Relevant information about this dataset can be found, together with further information, in Alberini et al. (2015) and Alberini et al. (2018).
8 Between Jan 1. 2000 and Jan 1. 2017, several municipalities were merged in Switzerland. Details about the mergers can be found at the BFS under "Amtliches Gemeindeverzeichnis der Schweiz". In the income, population and area datasets, the mergers were performed retroactively by aggregating the concerned municipalities and assigning them a new municipality code.
In order to measure the distance of each municipality to the language border, we use a matrix of street distances between all Swiss municipalities (from center to center), originating from Search.ch 9 . Similarly to Eugster et al. (2011) the municipalities were attributed to a single language group (French or non-French): we determined a total of 29 French speaking municipalities constituting the border towns to the German speaking region. Since those municipalities constitute the language border, the distance of a border-municipality i to the language border F is F i = 0. In a second step, we calculate the distance of all other municipalities to the nearest border town (center to center) and multiply the distance by (−1) if the municipality is located in the French region. As a result, French municipalities either have a distance of zero if they constitute the border or a negative distance. German speaking municipalities have a positive distance to the language border. As the distance data dates from the year 2010, we took into account municipality mergers until Jan 1st 2017 by taking the average distance of the municipalities concerned by mergers.
As described previously, we omit the Italian speaking cantons of Tessin and Graubünden in order to compare two distinct language groups, French-and Germanspeaking. The final dataset contains data on 1836 municipalities from 2008 to 2012 (9171 observations). 87 municipalities were omitted from the dataset due to missing values. In the RDD we will select all municipalities located within a distance of 25km to the language border: the RDD sample contains 353 municipalities which are in the cantons Bern, Fribourg, Solothurn, Basel-Landschaft, Vaud, Valais, Neuchatel and Jura. Some cantons imposed a label policy in order to promote energy efficient labels. Among the previously mentioned cantons this concerns the canton of Fribourg, however the language border runs through the canton of Fribourg, which means that a possible effect of the label policy would not manifest itself in the magnitude of the language treatment since we can include canton-level fixed effects.

Results
In Table 2 we report the estimation results from the basic regression model. These indicate a strong relationship between culture and the share of vehicles with A and B labels. All model specifications include year-and canton Fixed Effects. Therefore, this specification should be able to to control for any Cantonal policy measures. Column (1) shows the model without any covariates, results under this specification are less robust than with covariates, as in column (2) and (3), which suggest that the covariates control for a source of unobserved heterogeneity. In column (4) we show the model with the Mundlak correction that controls for timeinvariant unobserved heterogeneity. The estimator for language does not change significantly with the Mundlak extension, which suggests that the treatment effect is neither correlated with the covariates nor any time-invariant unobserved heterogeneity. Furthermore, a reduction of the sample to the three bilingual Cantons in column (5) (Bern, Fribourg and Valais) shows no change in the treatment effect either. Table 3 presents the results of the influence of language on vehicle choice, obtained using a fuzzy RDD. All model specifications include Canton Fixed Effects. Therefore, this specification should be able to to control for any Cantonal policy measures. In column (1) we use the entire sample with a quadratic polynomial, column (2) restricts the estimation to the three bilingual cantons, with little effect on the results. Further, in column (3) to (5), we use different bandwidths and in columns (4) and (5), we use a linear model instead of a second degree polynomial. Column (5) also includes covariates (income per capita, population density, elevation, share of French cars and an urban dummy), which do not significantly influence the result compared to column (4). Importantly, the difference in energy efficiency is not due to a preference for French car brands. The estimation results for each year suggest a strong effect of French-speaking culture on efficient vehicle choice. Throughout the period from 2008 to 2012, the effect of culture appears to increase every year. With increasing distance the effect increases throughout all years and the standard errors of the coefficient decrease. 10 As expected, the language coefficient decreases and the standard error increases as the bandwidth becomes more narrow. Therefore a simple estimation using the entire Swiss dataset would overestimate the effect of language, due to unobserved heterogeneity which is eliminated with the RDD. Overall, the treatment effect of language is of a magnitude of about 5.5 percentage points. Compared to the total mean difference at the language border of 6.2 percentage points 11 , cultural preferences account for 88% of the difference between the two regions.
The magnitude of the results is comparable to several policies on fuel efficiency: Huse and Lucinda (2014) find that in Sweden a rebate of SEK 10000 (about USD 1200) for energy efficient vehicles increases their market share by 5.5 percentage points. Beresteanu and Li (2011) show that a US federal income tax deduction of up to USD 2000 for hybrids explains 5% of sales. Gallagher and Muehlegger (2011) find that a 1000$ tax incentive for hybrid vehicles increases sales by 3% if the incentive is allocated on federal income taxes and a 45% increase in case of a sales tax rebate; Chandra et al. (2010) associates a 34% increase in sales with a CAD 1000 sales tax rebate on hybrids in Canada. Moreover, Gallagher and Mueh-legger (2011) propose that an additional 100$ of annual fuel savings, associated with hybrids, results in 13% more hybrid sales, a 10% gasoline price increase leads to a 8.6% increase in hybrid sales. Bento et al. (2009) suggest that the impact of a 25 cent gasoline price increase has only a marginal impact on fuel economy: 0.08% increase of mileage per gallon in the short-run (1 year) and a 0.16% increase in the long-run (10 years), mostly through the increased scrapping of old and inefficient vehicles.  We believe that the specific element of culture which influences consumer preferences for energy efficient vehicles can be found in people's concern for the environment. Similarly to Eugster et al. (2011), we use popular voting outcomes to depict people's undelying attitudes. Popular votes in Switzerland are held on a regular basis and allow the entire population to vote on propositions initiated by their fellow countrymen. A graphical representation of the voting outcomes shows the same discontinuity at the language border as previously shown for the energy labels. In Figure 5 we present the outcome of the six most recent popular votes on environmental issues in Switzerland. All graphs indicate that voters in the French speaking regions are more favorable to environmental issues than their German speaking counterparts 12 . 12 We use results of six popular votes, obtained from the Swiss Federal Office of Statistics (BFS): • Initiative for non-genetically modified food ("für Lebensmittel aus gentechnikfreier Landwirtschaft"), Nov. 27 th 2005; accepted with 55.7% (participation 42.24%). The initiative bans genetically modified food in Switzerland for the following five years.
• Green Economy ("Grüne Wirtschaft"), Sept. 25 th 2016, rejected with 36.4% (participation 43%). The initiative proposed an amendment to the constitution in order to reduce the Swiss resource and energy consumption until 2050- • Initiative against nuclear energy ("Atomausstiegsinitiative"), Nov. 27 th 2016, rejected with 45.8% (participation 45.38%). The initiative proposed to reduce the lifetime for existing nuclear energy plants and ban new ones.

22
The literature proposes various explanations on how differences in concern for the environment are formed. Schumacher (2015) proposes that high wealth levels allow societies to develop an environmental culture. Imhof et al. (2016) suggest that countries with a culture that favours future oriented behaviour are more likely to include environmental protection in their constitution. However, both theories do not seem to apply for the case of Switzerland: wealth levels were historically lower in the French part of Switzerland and the German-speaking regions are likely to have a more future oriented culture (Chen, 2013), this would predict a lower concern for the environment in French speaking regions, however we observe the contrary. In social psychology, the Fishbein-model by Ajzen and Fishbein (1980), composed of personal attitudes and social norms, aims to explain social behaviour. Laroche et al. (1996) use the Fishbein-model to explain differences in environmental behaviour and attitudes in Canada between English speaking Ontario and French speaking Quebec. The authors conclude that English speakers have a strong individual-attitudes component which motivates their environmental behaviour, while French speakers have a greater level of collectivism which results in a high importance of social norms. A measure for the degree of collectivism of cultures is given by Hofstede et al. (2010) and was recently used in the economics literature by Gorodnichenko and Roland (2017). Similarly, Gifford and Nilsson (2014) argue that since environmental friendly behavior has only an effect if it is executed collectively, the importance of social relationships and collectivism in society is crucial to pro-environmental actions. Indeed, the French part of Switzerland shows a higher level of collectivism than the German regions (64/100 vs. 69/100, lower scores indicate collectivism, higher scores individualism). This difference in collectivism indicates that social norms for environmental concern, even though those norms might exist in all of Switzerland, have a higher importance to individuals in French speaking regions due to their sense of collectivism. The more recent literature of social psychology also emphasizes the importance of altruism for environmental friendly behavior (Gifford and Nilsson, 2014). Data from the Swiss Household Energy Demand Sruvey (SHEDS) 13 supports this claim by suggesting that the French speaking regions are characterized by a higher level of altruism 14 compared to the German speaking parts (an index of 16.2/20 for French-speaking Regions vs 15.6/20, with a high significance level of a mean difference test).
Moreover there is evidence that the trustworthiness of government shapes environmental attitudes as well (Tjernström and Tietenberg, 2008). In this context, the French speaking regions of Switzerland score significantly higher on the index for the acceptance of government power by Hofstede et al. (2010) (70/100 vs 26/100), which could explain that social norms are especially accepted if they are promoted by the government, for instance through campaigning. Those results support Stephenson et al. (2010) who suggest that policy makers can actively shape social norms through campaigning and the promotion of energy literacy. Furthermore, the European Union project HarmoniCOP proposes a framework which builds on the idea of influencing environmental culture through social learning in order to promote a more sustainable resource management (Pahl-Wostl et al., 2008).

Conclusions
In this paper, we analyze to which extent culture, expressed through native language, affects consumer preferences for fuel efficient vehicles. Our results support theoretical models from the literature which suggest that consumer choices are influenced directly by culture, independently of institutions (see (Atkin, 2016, Bisin and Verdier, 2001, Guiso et al., 2006). In the context of a growing literature on the influence of culture in economics, our study adds empirical evidence that there is a clear effect of culture on environmental decision making and more specifically on efficient vehicle choice. Moreover, our observations on voting outcomes allow us to argue that consumers' concern for the environment is the major channel through which culture affects energy efficient vehicle preferences. Concerns for the environment in turn have been identified by the literature as a factor for energy efficient vehicle choice ( (Gallagher andMuehlegger, 2011, Kahn, 2007)).
Using findings from social psychology, we can argue that the cultural difference in attitudes and environmental behavior is caused by different underlying values, specifically altruism and individualism. In Switzerland, unique conditions allow for a natural experiment to estimate the direct effect of culture on consumer preferences: several culturally distinct groups, expressed via different native languages, live in the same country and share most institutions. In order to estimate the treatment effect of culture, we used a dataset containing all registered vehicles in Switzerland from 2008 to 2012. We aggregated the data at the municipality level and added socio-economic variables, the share of French speakers and the distance from each municipality to the language border. In order to estimate the effect of culture on efficient vehicle choice, we perform a spatial fuzzy RDD at the internal Swiss language border.
Results indicate that culture accounts for a difference of about 5.5 percentage points, 88% of the total difference, in the share of energy efficient vehicles between the French speaking regions of Switzerland and their German speaking counterparts. Since the efficiency labels control for curb weight, the results imply that French-speaking consumers, given a type of vehicle, would prefer the more efficient vehicles. In addition, the difference in fuel efficiency between the language regions is not due to a preference for German/French car brands as we control for this variable (in addition, we control for all vehicle brands in the individual linear probability model shown in the Appendix). The magnitude of our results corresponds to the effect of standard rebate policies for energy efficient vehicles as documented in the literature. In addition, we find that popular votes on environmental issues receive a significantly higher share of approval in the French speaking regions, indicating that they place a higher value on the environment. Possible explanations for a higher concern for the environment of the French speaking population are given by social psychology theories which emphasize the importance of collectivism and altruism, which is supported by several value surveys in Switzerland.
Our results give insights to both car producers and policy makers. Producers can adapt their marketing strategies by taking into account local cultural preferences for energy efficient design. For policy makers this result has two implications: first, to introduce additional policy measures, such as campaigning, which address culture as a factor. Second, to direct traditional economic policy efforts to groups where the potential outcome is higher. Switzerland is organized as a federal state where the cantons have a large autonomy and can decide on most environmental policies themselves. Since the French speaking population is more sensitive to environmental concern and have a higher acceptance of government power, policy makers could use other channels to target environmental goals, mainly by trying to directly influence consumer preferences via marketing and campaigning. In contrast, the German speaking part of Switzerland, which shows a lower concern for the environment, could be addressed with more traditional monetary incentives. Finally, if policy makers seek to influence environmental consumer-culture (in this case in the Swiss German part), as proposed by Stephenson et al. (2010) and the European Union project HarmoniCOP (Pahl-Wostl et al., 2008), our results show the potential magnitude of energy savings a change in culture could induce.

Robustness Checks
We first check the robustness about our assumptions on the covariates. Table 8 shows a non-parametric, sharp RDD on the covariates used in the analysis (using the driving distance to the language border as running variable). In comparison to the t-test in table 1, the RDD is more restrictive: the share of French cars, as well as the other covariates, show a non-significant difference with the RDD, with exception of municipality elevation.
In order to further analyze the robustnes of our results, we conduct the same fuzzy RDD using as a dependent variable CO 2 emissions and mean fuel consumption. Further, we also apply a non-parametric fuzzy RDD. Finally, we decided to exploit the individual data and estimate a probit model. In this case, we are interested to know the impact of culture on the probability that an energy efficient car is adopted. Of course, one downside of this approach is that the explanatory variables are not measured at the individual but at the municipality level. However, we believe that this model represents a valid robustness check.
In Tables 4 and 5 we present the results of a panel fuzzy RDD using as dependent variables the mean CO 2 emissions and mean fuel consumption of a municipality's car fleet. Results are similar to the results reported in Table 3 and suggest that CO 2 emissions and fuel consumption follow the same trend as the distribution of A-and B-labels, although the effect is less significant than with the labels. The underlying reason is that labels do not measure absulute emissions and consumption but are relative to the curb weight. Table 6 shows the results of a non-parametric fuzzy RDD using a uniform Kernel density estimation with a first order polynomial local linear regression. In order to determine the optimal bandwidth of the RDD, we use the non-parametric procedure by Calonico et al. (2014) and Calonico et al. (2015), which yields an optimal bandwidth of 15.85km. Results confirm the sign of the treatment effect, found using the parametric RDD. However, we believe that in our case, the parametric approach is more appropriate because the error terms of our parametric estimations follow a normal distribution. Table 7 shows a fuzzy RDD using individual data from 2010 (linear probability model). Results indicate that the probability of choosing an energy efficient car is higher in the French speaking regions and persists as we approach the language border. In column (5) of table 4 we introduce control variables, including vehicle brand, which does not have a large effect on the language estimate. This implies that the measured effect of language does not depend on consumer preferences for certain brands.
28     Standard errors in parentheses, * * * : p < 0.01, * * : p < 0.05, * : p < 0.10 The RDD was implemented with the package rdrobust in R, using a sharp, nonparametric design with a bandwidth of 25km and a uniform kernel