CER-ETH – Center of Economic Research at ETH Zurich The Impact of Policy Awareness: Evidence from Vehicle Choices Response to Fiscal Incentives

Isolating the role of limited knowledge, psychological frictions and policy characteristics is key when evaluating a public program and designing future policies. This paper explores the role of awareness about the presence of fiscal programs in determining their impact on individual choices. Our identification strategy exploits quasi-experimental variation in the introduction of fiscal incentives aimed at promoting the purchase of energy efficient vehicles, and a direct measure of policy awareness at the individual level. We find an important impact of awareness on consumers’ vehicle choices, highlighting that limited awareness may represent a critical barrier to the effectiveness of public programs.


Introduction
The proper design of policy measures crucially relies on the understanding of the reasons why some programs work while others do not. The effectiveness of a public program might be limited for several reasons, ranging from flaws in the incentive scheme to barriers related to the individuals' decision making process such as limited information, low program awareness, stigma, inattention, and other behavioral failures. The role of these factors have been discussed, for instance, in the context of the take-up of tax credits, subsidies for saving accounts, food stamps, Social Security, health insurance and environmental policies (Benartzi and Thaler, 2007, Congdon et al., 2009, DellaVigna, 2009, Gillingham et al., 2009, Chetty et al., 2014, Madrian, 2014, Bhargava and Manoli, 2015, Chetty, 2015, Allcott, 2016. One key challenge for academics and policy makers is to isolate the impact of a single potential factor on the (lack-of) individuals' response to economic incentives.
In this paper, we aim at identifying the impact of awareness about the presence of a fiscal policy in determining its impact on individual choices. Clearly, only those individuals in the target population that are aware about the existence of a specific public program may incorporate its incentives in their decision making process and possibly respond to its introduction. However, the mere knowledge of the existence of the fiscal measure is not sufficient to guarantee that the policy reaches its desired goals in the presence of inadequate incentives, low salience, or behavioral or psychological biases in the individuals' decision making process. Nonetheless, failing to properly consider the role of low policy awareness might have consequences for the evaluation of a program and the design of future policies. Policy makers might indeed conclude that the low effectiveness of a program is due to limitations in its design or other behavioral failures, even though its limited outcomes are only due to the lack of adequate knowledge among the target population. Moreover, corrective non-price interventions can have an equivalent effect than sizable price changes at a fraction of their cost (Allcott andMullainathan, 2010, Bertrand et al., 2010).
To assess the consequences of limited policy awareness on individuals' responses to fiscal incentives, we consider the case of vehicle taxes on consumers' choices in Switzerland. As in many other European countries, Swiss drivers have to pay each year a tax on car ownership -also known as registration tax. The local administrative areas (cantons) are free to set the structure of the tax. In addition, some cantons introduced a Bonus/Malus system based on vehicle energy efficiency or CO 2 emissions. A bonus (malus) applies to very energy efficient (inefficient) cars and provides a sizable percentage discount (increase) to the baseline registration tax. Because the incentives are applied automatically to the baseline vehicle tax, this policy is particularly well suited to study the role of policy awareness as we can rule out the role of transaction costs to access the fiscal benefits.
The Bonus/Malus system provides consumers with incentives for the purchase of efficient vehicles. 1 To estimate the causal effect of individuals' awareness on vehicle choices, we exploit the variation in the introduction of the Bonus/Malus system across cantons and over time, as well as the availability of a direct measure of policy awareness both in the presence and in the absence of the policy. Isolating the role of awareness about the features of the vehicle registration tax scheme in Switzerland provides insights beyond the specific context because it sheds light into the relative importance of different potential factors as barriers to individuals responses to fiscal incentives.
We use data from the Swiss Household Energy Demand Survey (SHEDS), held in 2018 with a representative sample of Swiss individuals. The data include detailed information on respondents' characteristics, their households, as well as their main vehicle. We also ask respondents whether they knew, at the time they bought their main car, if the registration tax in their canton of residence was based on fuel efficiency rating and/or CO 2 emission rate, and define as "policy aware" respondents who answer correctly. Around 42% of respondents do so.
Our main finding is that policy awareness plays a prominent role to understand consumers' response to the fiscal incentives set by the Bonus/Malus system. In fact, while ignoring policy awareness would lead to conclude that such measures are ineffective, we find that awareness about the presence of the incentives has a relevant impact on individuals' vehicle choices when the incentives are in place. Following the introduction of the Bonus/Malus system, policy awareness induces individuals to buy more fuel efficient and newer vehicles. This result is robust to a series of different specification checks, to the addressing of compositional differences between aware individuals in treated and control cantons, and is confirmed by an IV strategy that we exploit to deal with the potential endogeneity of awareness. These results inform policy makers about the potential boost in individuals' response from increased program awareness, providing an upper bound of the impact of information campaigns complementing the introduction of future policies.
Starting from Simon (1955), a large literature in economics has attempted to relax the traditional assumption of individuals taking decisions under full information. Less attention though has been devoted to considering the role of lack-of knowledge in the context of the evaluation of public programs. Our paper is related to a recent literature that considers the role of limited information on the individuals' responses to public policies (Mastrobuoni, 2011, Kling et al., 2012, Bhargava and Manoli, 2015, Liebman and Luttmer, 2015. This literature typically uses information treatments to study the impact of limited knowledge about the characteristics of a policy measure on individuals' choices. However, the effectiveness of an information treatment in influencing behavioral responses to the incentives of a public program depends on whether the intervention affects the actual level of individuals' knowledge about the policy. The process through which knowledge about a policy measure is shaped is largely unknown (as discussed, e.g., by . In fact, several factors might hinder this process, including the initial level of understanding about the policy, the program's complexity, the financial literacy of treatment recipients, and the type and framing of the information provided. We abstract from using a specific information treatment, while exploiting an explicit measure of knowledge about the presence of a program that allows to rule out other factors influencing the process of knowledge accumulation. Heterogeneity in knowledge about the presence of the public program might reflect into substantial improvements in the effectiveness of the existing policy framework given an appropriate information treatment. Using a descriptive analysis of the survey data carried out following their main experiment, Bhargava and Manoli (2015) have already suggested that low program awareness may be one of the possible barriers to the take-up of fiscal benefits. Our main contribution in this paper is to isolate the causal effect of awareness about the presence of a fiscal program on individuals' responses from that of other potential behavioral factors or policy characteristics influencing individual choices.
This work is also largely complementary to the paper by , who estimate the impacts of the Earned Income Tax Credit in the US on labor supply exploiting variation of knowledge at the local level. Although they recognize that, ideally, one would want to use a direct measure of knowledge at the individual level, they proxy for knowledge using an indirect measure at the local level of individuals' residence. In contrast, we exploit a direct measure of policy awareness at the individual level to estimate the impact of fiscal incentives on vehicle choices. To our knowledge, we are the first to use an explicit, direct measure of awareness about a specific public program to study the implications for its effectiveness. We argue that the concept of awareness is distinct from other behavioral anomalies, like salience. 2 Finally, we contribute to the growing literature on the effects of environmental taxation 2 The difference between awareness and salience is clear when considering contexts where low salience has been found to play a role as, for instance, automatic road tolls and sales taxes analyzed by Finkelstein (2009), andChetty et al. (2009), respectively. While individuals might process partially the information about the existence of a toll or a sales tax when taking a decision due to their low visibility (low salience), it is very likely that they would still recognize their existence when explicitly asked about it. In our context, the concept of salience would apply only to the fraction of aware individuals that acknowledges the presence of the policy measure. and information on vehicle choices (D'Haultfoeuille et al., 2014, Klier and Linn, 2015, Alberini and Bareit, 2017, Alberini et al., 2018, Grigolon et al., 2018, Allcott and Knittel, 2019, Cerruti et al., 2019. We show a large lack of awareness by consumers about the presence of tax incentives for the purchase of an energy efficient vehicle. We highlight that limited knowledge might represent a substantial barrier to the effectiveness of vehicle taxes in influencing vehicle purchases.
The rest of the paper is organized as follows. Section 2 sketches a simple model to consider the role of lack-of program awareness in the consumers' decision making process. In Section 3 we describe the Bonus/Malus system in Switzerland. Section 4 presents the data, the measurement of awareness and descriptive evidence. In Section 5 we discuss the identification strategy we exploit for the impact of awareness and present the empirical results. Section 6 concludes.

Policy awareness and fiscal policy
In this work we wish to analyze the role of limited policy awareness in the evaluation of policy measures. We consider the introduction of fiscal incentives to promote the adoption of energy efficient cars. This section sketches a simple theoretical framework to highlight how limited awareness about the presence of tax incentives to promote the purchase of energy efficient vehicles can affect consumers' decisions.

Vehicles purchase and awareness about fiscal incentives
We consider the decisions of consumers with respect to the purchase of a new vehicle in the presence of limited awareness in a framework that is similar to DellaVigna (2009) and Allcott et al. (2014). The consumer draws utility from the mileage she is going to drive each year m and other cars characteristics X (such as car engine, number of doors etc.), and disutility from fuel costs (which depend on the vehicle's fuel efficiency f and the fuel price c), vehicle registration tax -to be paid each year a car is owned -τ and upfront costs P .
The consumer then chooses the vehicle i that maximizes her utility over the car's lifetime L. Assuming separability in utility between vehicle's characteristics and vehicle's costs, we can write the problem of the consumer as follows: Assume for simplicity that there are only two types of cars in the consumer's choice set, characterized by different levels of fuel efficiency A and B, with f A < f B . We also assume that the two cars provide the same flow of utility (u(m A , X A ) = u(m B , X B )) such that the decision of purchase only depends on the comparison of total life-cycle costs. 3 In the absence of incentives towards the adoption of fuel efficient cars (τ A = τ B ), it is trivial to show that an optimizing consumer will choose the car with higher fuel efficiency A only if , that is if the savings in the driving costs over the vehicle's lifetime associated with greater fuel efficiency (what Allcott et al. 2014 call "gross utility gains" from energy efficiency) more than compensate the larger upfront costs.
We consider a policy maker introducing a discount D A to the registration tax of a car with fuel efficiency f A . In the presence of the fiscal incentive, an optimizing consumer will choose the fuel efficient car A only if: Several behavioral and psychological biases can influence individuals' valuation of the savings coming from fuel efficiency, and then the effectiveness of the discount to the registration tax to increase the adoption of efficient vehicles. In particular, individuals can correctly evaluate the savings coming from adopting a vehicle with fuel efficiency A if: (i) they have the energy-related knowledge (i.e. they can correctly evaluate fuel efficiency -f A = f B -and fuel prices) and skills to perform an investment calculation (Blasch et al., 2018); (ii) they have no present bias (such that they do not value the future stream of fuel cost and registration tax savings); (iii) they have no limited attention due to salience bias (DellaVigna, 2009); (iv) they are aware about the existence of the policy measure.
In the presence of the discount, a misoptimizing consumer will then choose the efficient vehicle A only if: where Γ is the standard valuation weight in the presence of behavioral and psychological biases (i) to (iii), and η P is an indicator for whether a consumer is aware about the presence of the discount (η P = 1), or not (η P = 0).
Although only the consumers that are aware about the existence of the discount (η P = 1) might respond to its introduction, separating the effect of limited policy awareness from that of other potential behavioral failures (i) to (iii) is key in the design and evaluation of this policy measure. In this work we use a direct individual measure of consumers' knowledge about a specific feature of the fiscal scheme to disentangle the role of policy awareness from that of alternative psychological biases.
Clearly, the presence of policy awareness may interact with other behavioral anomalies. Two remarks are worth making. First, although (lack-of) policy awareness and limited attention due to salience bias are observationally equivalent, they are two separate concepts with different policy implications. In fact, in the presence of salience bias, a consumer will is a function of the salience of the discount and indicates the degree of limited attention, as in DellaVigna (2009). While policy unaware consumers (η P = 0) do not know about the existence of the discount D A , consumers with limited attention see the discount but then process the information only partially (DellaVigna, 2009). Second, we do not exclude that, in the presence of other behavioral biases that keep consumers from making an investment calculation, awareness about the policy measure might still enter the individuals decision making process through the usage of heuristics.
In the next sections we describe how we exploit heterogeneity in registration taxes in Switzerland and information on knowledge about fiscal incentives for the purchase of energy efficient cars at the individual level to identify the impact of awareness on the effect of these fiscal policy measures.

Institutional context
As in many other European countries, car owners in Switzerland have to pay a vehicle registration tax each year. The amount of these taxes is substantial. For instance, the median annual tax on a vehicle bought in 2015 is around 362 CHF, which compares to the median vehicle sale price of about 33'000 CHF. The amount to pay typically depends on vehicle weight, engine size, and engine power, so that larger and more powerful cars pay more. Registration tax rates are not set by the central government: each of the 26 administrative areas of Switzerland (known as cantons) are free to introduce their own scheme.
Between 2005 and 2013, some cantons have introduced a Bonus/Malus incentive system in addition to the regular vehicle tax: generally, while driving a fuel efficient or low CO 2 emission vehicle might guarantee a percentage discount to the baseline registration tax (bonus), driving a fuel inefficient or high emission car might increase the registration tax by a certain percentage (malus). 4 Among the 26 cantons in Switzerland, seventeen introduced some vehicle registration tax incentives based on fuel efficiency, CO 2 emissions, or 4 All cantons that applied a malus system to the registration tax, also applied discount for efficient or low emission cars. Three cantons introduced a Bonus/Malus system based on fixed monetary amounts instead of percentage discounts/penalties. both. Appendix A provides details about the incentive scheme, and its differences across local admninistrative areas.
The registration tax incentive is based on the energy efficiency rating of the car in seven cantons, on the CO 2 emission rate in other seven, and on both criteria in other three cantons. Two cantons adopted a bonus system and then abolished it. While, in most cases, the bonus component of the tax applies only for three or four years after the purchase of the car, the malus component applies for the entire lifetime of the vehicle. The incentives associated with the Bonus/Malus system can be considerable. The discounts to the baseline vehicle tax range from 40% to 100%, while the penalties range from 10% to 50%.
As mentioned earlier, there are two possible criteria used to define which cars are affected by the Bonus/Malus. The first is the CO 2 emission rate, expressed in grams per 100 km, while the second is the energy efficiency rating. There are seven categories of energy efficiency rating, from "A" (most efficient) to "G" (least efficient), and each vehicle is assigned to one rating based on its fuel consumption per 100 km and its weight. Both CO 2 emissions and the energy efficiency rating of a vehicle are listed in its energy label (see Figure 5 in Appendix A). The energy label is assigned to each vehicle on sale in Switzerland and is accessible to consumers before purchasing a car. Contrarily to the registration tax, the information displayed in the energy label and the criteria used to assign the efficiency rating of a vehicle are set by the Swiss federal government and thus are the same in all cantons. The criteria for the application of the Bonus/Malus incentives are different in each canton. However, most cantonal incentive schemes share some common characteristics. For instance, the application of the bonus or malus based on CO 2 emissions to the registration tax depends on whether the vehicle emission rate is below or above a certain threshold, respectively. In the case of the efficiency bonus, a vehicle must have a rating equal to A or B to be eligible. Finally, in most cantons, the age of the car cannot exceed a certain limit (3 or 4 years old) to benefit from a registration tax discount. Importantly, any discount or penalty to the registration tax is applied automatically when the consumers pay the registration tax every year. Therefore the scheme does not require an application by the consumer, who is affected by the discounts and the penalties regardless of her level of knowledge about the features of the registration tax. This characteristic of the fiscal policy allows to rule out the role of transaction costs in the consumers' decision making process when studying the impact of policy awareness.

Data
We use data from the Swiss Household Energy Demand Survey (SHEDS), an annual survey collecting data on a sample of about 5000 households representative of the French and German speaking Swiss population of 25 cantons (excluding the Italian-speaking canton of Ticino). 5 Details about the information used in the main analysis, and the construction of the final sample are provided in Appendix C.
Most of the information we use comes from the 2018 wave of the SHEDS survey, while some additional data comes from waves 2016 and 2017. 6 The 2018 sample contains 5011 households. Among those, 3621 (72.26%) owned at least one car. Car ownership distribution in our data looks very similar to what has been found by official statistics from the Swiss government. 7 The survey collects detailed information on respondents' socio-economic characteristics and their main vehicle. Data on standard socio-economic characteristics, such as age, education, language and household income, are complemented by a rich set of information on environmental attitudes, values and social norms, life values, trust regarding advice on energy saving provided by various subjects (neighbours, government institutions, environmental organizations), membership to environmental clubs, voting preference for the green party, energy literacy and financial literacy (as in Blasch et al. 2018 andLusardi andMitchell 2014). 8 Information on canton of residence and living area are also available.
Moreover, the survey asks respondents to report information on their vehicle fuel consumption per 100 km, its energy efficiency rating (from A to G), year of purchase and year of first registration. Most of the respondents (94.53%) were able to provide information on fuel consumption per 100 km. On the other hand, only a small fraction of the respondents with at least one car (41.62%) provided information on the energy efficiency category of their 5 The full text of the questionnaire, and information on how to get access to the data, can be found on https://www.sccer-crest.ch/research/swiss-household-energy-demand-survey-sheds/.
6 Some questions are asked only the first time an individual participates to the survey, and are not asked again in the following waves. Regardless, the most relevant information for our analysis and baseline socioeconomic characteristics have been collected in the 2018 wave. In particular, because the key question about policy awareness has been asked only in 2018, we restrict the sample to individuals that have been interviewed in 2018.
7 In our sample, 27.74% of households do not have a car, 45.52% have one, 22.13% have two and 4.61% have three or more. A 2015 survey by the Swiss Federal Statistical Office found these shares to be 22%, 49%, 23% and 6% respectively (Swiss Federal Statistical Office, 2017, p. 11).
8 Some of the questions on environmental attitudes and literacy are asked only to new respondents. Thus, those questions might have been asked in the 2016 or 2017 wave for people who participated to the questionnaire more than once. Questions on baseline socio-economic characteristics, vehicle fuel economy, and policy awareness were all asked in the 2018 wave. main vehicle. 9 We use fuel consumption per 100 km as main outcome variable in the empirical analysis. In fact, fuel consumption rate is strongly correlated with the efficiency rating and, within fuel type, exactly proportional to CO 2 emissions per km. In Appendix B, using the population of Swiss cars, we show that the distribution of actual fuel consumption is matched well by that of the self-reported fuel consumption, and provide descriptive evidence about the association of fuel consumption with efficiency rating and CO 2 emissions.
In the paper we also use the available information on efficiency rating to confirm our main findings. Because of the low response rate on this question, substantial selection might arise from using this outcome variable and thus we prefer to use the self-reported fuel consumption as main outcome.
From the 3621 respondents of our sample with at least one vehicle, we further drop 30.84% of the observations with missing values for any of the variables used in the empirical analysis. 10 The final sample used in the analysis includes 2504 observations.

Measuring policy awareness
We elicit information about respondents' policy awareness about the presence of the registration tax incentives for efficient cars using a survey question. We ask respondents whether they knew, at the time they bought their main car, if the registration tax in their canton of residence was based on fuel efficiency rating and/or CO 2 emission rate. The exact phrasing of the question is "At the time you bought your main car, did you know if in your canton the annual registration tax depended on the level of fuel efficiency and/or on CO 2 emissions of the cars?". The possible answers are "Yes, it depended on the fuel efficiency or on CO 2 emissions", "No, it did not depend on the fuel efficiency or on CO 2 emissions" and "I do not know".
The policy awareness question does not ask about specific characteristics of the Bonus/Malus, or even mentions the terms "bonus" or "malus". 11 Our aim is to capture even vague awareness about the presence of the tax incentives, rather than the respondents' knowledge of the 9 The share of missing answers decreases when the car was bought in recent years, suggesting that information about the efficiency rating might be only available at the time of purchase. Nevertheless, only 50.0% of respondents who bought a car in 2017 reported the efficiency rating.
10 In Appendix C are reported details about the sample selection. The number of missing values is especially relevant for household income (14.94 %) and financial literacy (11.18%). Financial literacy presents a substantial number of missing values because these questions were not asked to respondents surveyed in both 2016 and 2018, but not in 2017. All the results in this paper hold when omitting household income and financial literacy from the set of controls or when adding a "missing answer" category and including the missing observations in the analysis.
11 Registration taxes can be based on engine size, vehicle power, or weight, but only the Bonus/Malus is an explicit and direct link between the tax amount and the efficiency rating or the CO 2 emission rate. details of the registration tax scheme. This allows us to capture the potential role of policy awareness without excluding some potential mechanisms through which this knowledge affects the vehicle choice. Furthermore, asking explicitly about the presence or the absence of the incentives allows us to distinguish between policy aware and not aware individuals also in cantons that never introduced a Bonus/Malus scheme on top of their registration tax.
Our measure of policy awareness combines information coming from the respondents' answer to the awareness question and the presence of the Bonus/Malus in their canton of residence at the time of purchase of the car. We classify as "policy aware" those respondents who answered "yes" and who bought a car in a canton that has introduced some registration tax incentives, when those incentives were in place. Respondents who bought a car when the incentives were not in place and who answered "no" have also been classified as "policy aware". All other respondents were classified as not aware.
Policy awareness in the data A substantial share of respondents could not answer correctly to our policy awareness question. We find that about 42.21% of the individuals in our sample are classified as "policy aware" according to our definition. About 37.81% of respondents answered "don't know", while 19.96% gave the wrong answer. Figure 1 shows the evolution of answers to policy awareness question and the share of policy aware individuals, based on the year in which respondents bought their main car and whether a Bonus/Malus was in place. 12 While answers do not seem to change dramatically over purchase years, we observe that respondents are more likely to answer "no" if they bought the car in absence of the incentives, and "yes" if they were present.
Policy awareness and vehicle characteristics To illustrate differences in vehicle fuel consumption among our groups of interest, in figure 2 we compare its distribution between cars bought when a Bonus/Malus scheme was in place and when not, distinguishing between aware and non-aware respondents, and between incentives based on CO 2 emissions and efficiency rating. When a Bonus/Malus is in place, we observe a shift towards the left of fuel consumption distribution for the group of aware respondents.
We also want to see if there is a correlation between awareness under a bonus and vehicle age at purchase time. Because in many cantons eligibility to discounts to the registration tax depends also on the age of the car, there are two mechanisms why a person responding to the incentives might buy a newer car. First, newer cars are on average more efficient and with lower emissions and thus are more likely to be eligible for a bonus or not affected by a malus. Second, newer cars can become eligible to the bonus, or remain eligible for a longer period of time, compared to older cars with the same efficiency rating and CO 2 emissions. Similarly to figure 2, in figure 3 we show the different distribution of car age at purchase time. 13 Vehicle age distributions between aware and non aware individuals in the absence of fiscal incentives are almost overlapping while, in the presence of a Bonus/Malus system, we observe a much higher frequency of new cars being purchased by policy aware individuals.

Policy adoption, awareness, and respondents' characteristics
We are also interested in whether respondents' characteristics are correlated with policy adoption and policy awareness. The presence of compositional differences between groups of individuals defined by their policy treatment and awareness status might indeed inform about possible selection patterns. Appendix D contains various tables reporting mean values between different groups of respondents for the variables used in the main analysis.
First, we want to check how different are respondents living in cantons that have introduced the fiscal incentives to promote efficient cars at some point in time, and respondents living in cantons that never had (Appendix D,tables 16,18,20,22,24,26,28,30,32,34).  We find no statistically significant differences between those two groups in terms of baseline characteristics like gender, education, occupation or household size. Instead, we do find that respondents in cantons that have introduced a Bonus/Malus scheme are more likely to live in urban areas, are more likely to be part of a household earning more than 9000 CHF per month, are less likely to be German speakers, and are older. We do not find systematic differences in other characteristics linked to environmental attitudes or life values. For instance, rate of voting preferences for the Green Party or membership in an environmental club are very similar among the two groups. People in cantons adopting the incentives perform worse on average in terms of financial literacy and of certain aspects of energy literacy. 14 Policy aware and policy unaware respondents appear to be balanced in most of the characteristics considered (Appendix D, tables 16, 18,20,22,24,26,28,30,32,34). We do find however that policy aware respondents are more likely to be male, with a university education, older and self-employed, and less likely to be German speakers and unemployed or students. They are also more likely to provide the correct answer to our financial literacy questions. We do not find systematic differences in environmental attitudes or behaviors between policy aware and unaware respondents. Finally, we compare characteristics between respondents who bought their car when the fiscal incentives were in place and those who bought it when there were no incentives (Appendix D,tables 15,17,19,21,23,25,27,29,31,33). In doing so, we consider separately aware and unaware respondents. We do find both pair of groups to be well balanced. This is especially the case for unaware respondents living in a canton that have introduced some fiscal incentives to promote the adoption of efficient cars, and unaware respondents that live in a canton with no such policy. We find some differences between policy aware individuals. Aware respondents who bought their car in a canton that has introduced a Bonus/Malus scheme are older, less likely to be German speakers and less likely to answer correctly to certain financial and energy literacy question.

Identification and Empirical Strategy
The Bonus/Malus system introduced in some administrative areas in Switzerland decreases the vehicle registration tax on high efficiency cars that consumers need to pay each year, while increasing the amount to be paid for the least efficient vehicles. It thus provides incentives for the purchase of vehicles with higher fuel efficiency. This paper aims at estimating the causal effect of individuals' awareness about the presence of the Bonus/Malus system for the vehicle registration tax on their vehicle choices.
The ideal setting to address the question we are asking in this paper would be one where awareness about the presence of the incentive scheme had been randomly distributed to a group of individuals in cantons where the Bonus/Malus system is in place, with no possibility of information spillovers to the other consumers in the same areas. If individuals' awareness was determined exogenously, we could estimate the causal effect of being aware about the presence of the policy by simply comparing the fuel consumption of the vehicle purchased by aware consumers to that of unaware consumers only when the Bonus/Malus system is in place.
In our context, the identification of the causal effect of awareness about the presence of the Bonus/Malus system on vehicle choices is challenging because individuals' policy awareness is endogenous. For instance, prior to the vehicle choice, consumers might determine their level of awareness about the characteristics of the vehicles in their choice set as well as the associated taxation system, depending on the individual incentives that they face and that are unobservable to the researchers.
Our main strategy to identify the causal impact of policy awareness exploits the variation in the Bonus/Malus system across cantons and over time, as well as the availability of a direct measure of policy awareness both in the presence and in the absence of the policy. Hence, we use aware individuals in cantons without the Bonus/Malus system as a comparison group in a difference in differences framework with multiple treatment groups and treatment periods. In this context, the key identifying assumption is that the evolution of vehicle choices over time for policy aware consumers in cantons where the Bonus/Malus system has been introduced would have been the same, in the absence of the policy, as that of policy aware consumers in cantons where the Bonus/Malus system has not been introduced. Importantly for the credibility of this assumption, all cantons in Switzerland require the payment of a registration tax with different characteristics. Therefore, consumers in both treated and control cantons have an incentive to accumulate knowledge about the structure of the registration tax.
We perform standard pre-treatment parallel trend tests to provide evidence that the necessary condition for the validity of the key identifying assumption is satisfied. Moreover, because we exploit data in the form of a repeated cross-section, where the time dimension refers to the year of vehicle purchase, while measuring individual characteristics only at the time of the interview, we need to emphasize that the validity of this strategy relies on the assumption of exogeneity of the control variables. In Section 5.2 we discuss possible threats to the validity of this identification strategy and how we address them.

Main specification
Our basic empirical specification takes the form: (4) where y ict is an indicator of the vehicle choice by individual i living in canton c in year t, Aware is a dummy for whether the respondent is policy aware, BM is a dummy for whether a canton has adopted some Bonus/Malus incentive at any point in time, BM P is a dummy that indicates if a Bonus/Malus system was in place in the canton and year in which the car was purchased, X is a set of respondents' characteristics, and η and ξ denote canton of residence dummies and year of purchase dummies, respectively. As discussed in Secton 4 we use log fuel consumption per 100 km of the vehicle purchased as main indicator of vehicles' energy efficiency, and then perform robustness checks exploiting information on the vehicles' efficiency rating on the subsample for which this information is available. We then use the age of the vehicle at the time of purchase to study whether the incentives set by the Bonus/Malus system induced aware individuals to purchase newer cars. We cluster standard errors at the canton by year of purchase level. 15 The coefficient of main interest β indicates the reduced form effect of consumer's policy awareness on vehicle choices in the presence of the fiscal incentives. The coefficient θ gives the effect of the Bonus/Malus system for the unaware consumers. In the absence of supply side effects of the policy, we would expect the estimated θ to be equal to zero. Finally, ψ captures the time-invariant heterogeneity in vehicle choices within treatment group by awareness status.
To investigate the importance of considering policy awareness when studying the consequences of the Bonus/Malus system on vehicle choices, we also estimate equation 4 setting β, ψ and γ equal to zero. In this case, θ indicates the average effect of the Bonus/Malus system on the treated.
We include a large set of covariates to control for compositional differences among policy aware respondents in cantons with and without the policy. These include standard sociodemographics as well as investment literacy and environmental attitudes that can potentially influence both the decision of purchase of an efficient vehicle and the probability to be aware about the presence of the fiscal incentives. 16 We conduct some robustness checks on the main specification. First, to control for possible unobservables that might affect the choice of purchase of efficient cars at the same time as the introduction of the Bonus/Malus incentive scheme, we also estimate equation 4 including canton by year of purchase fixed effects. Second, one could be worried that the time gap between the interview and the decision of purchase might be correlated with the probability to remember about whether the registration tax depended on the vehicle efficiency at the time of purchase, thus determining a mechanical association between the probability to be aware according to our definition and the presence of the policy. To avoid this potential issue, we conduct a robustness check using only information from vehicles purchased after 2010. 17 Moreover, as the average vehicle fuel efficiency in the consumers' choice set has increased over time, individuals that are about to purchase a new vehicle might have different incentives to accumulate information about vehicle in their choice set (and then become policy aware) over time. To address this possible issue, we estimate equation 4 including the interaction of the awareness dummy with year of purchase dummies. 18 Finally, estimation is conducted also including the interaction of the policy awareness variable Aware with three dummies that indicate the Bonus/Malus type (based on CO 2 , on efficiency rating, on both) instead of the treatment variable BM . We do this to control for potential unobserved factors that might be correlated with vehicle choices and influence how awareness status is determined depending on the type of incentives in place in one's canton of residence. 19

Addressing possible threats to the validity of our main strategy
The validity of our main identification strategy for the impact of policy awareness on vehicle choices relies on whether consumers that are aware about the features of the registration tax system, living in cantons that have not introduced a Bonus/Malus system are comparable to aware consumers in cantons where a Bonus/Malus system has been introduced. Although the key identifying assumption in this framework is milder than that it would take to simply compare aware and unaware respondents in the presence of the Bonus/Malus system, some potential issues can undermine its validity: (i) selective introduction of the Bonus/Malus system across administrative areas; (ii) lack of control for unobserved individual-specific characteristics that influence the process of information accumulation in the presence and in the absence of the policy; (iii) the presence of measurement error in awareness.
Are aware consumers in the treated group comparable to aware consumers in the control group? Although the descriptive evidence presented in Section 4 shows that respondents living in treated and control cantons, as well as aware and unaware individuals, are comparable with respect to several observable dimensions, some compositional differences emerge. For instance, administrative areas that have introduced a Bonus/Malus system are more urban. Moreover, the share of French speakers is higher in cantons that have introduced some fiscal incentives to promote the adoption of efficient vehicles. This might be problematic for our main identification strategy because heterogeneity in the incentives to foster the adoption of efficient vehicles and/or cultural differences may drive selection into the Bonus/Malus awareness status. Generally, in the presence of selection into treatment or awareness status based on observables that may affect the evolution of vehicle choices over time, the identifying assumption in our basic empirical strategy is not valid.
To overcome this potential issue, we combine our main strategy with a propensity score matching approach that addresses the 'common support problem' (Heckman et al., 1998). The aim is to construct comparison groups that reproduce as close as possible which vehicles the treated consumers would have purchased in the absence of the Bonus/Malus system (or which vehicles aware consumers would have purchased if they did not know about the fiscal incentives), so that the change in the vehicle choices of aware consumers in the presence of the Bonus/Malus system is only due to awareness about the fiscal incentives and not by endogenous selection into the treatment or awareness status.
We perform a regression version of the propensity score matching difference in differences exploiting the rich set of individual characteristics available in our dataset. The analysis is carried out in two steps. In a first step, the matching is performed using the propensity score method (Rosenbaum andRubin, 1983, 1984), separately for the presence of the Bonus/Malus system and awareness status. In our application, the propensity score P (X) is the probability that an individual lives in areas with a Bonus/Malus system, or that an individual is aware about the presence of the fiscal incentives, conditional on the matching characteristics X. We estimate the propensity score using Probit models that include a large number of respondents' and cantons' characteristics and then perform a kernel matching. 20 In a second step, we run regression 4 applying the kernel-based weights and restricting the estimation sample to the common support defined by the propensity score.
In this context, the standard matching assumption of conditional independence must be valid in terms of the before-after evolution instead of levels. Hence, the identifying assumption is that there are no unobservable characteristics that affect both the change in vehicle choices and the treatment or awareness status (i.e., any potential selection occurs through observable individual characteristics). While the matching procedure takes care of the selection into treatment (or awareness status) based on observables, the difference in differences deals with selection on unobservables under the assumption that the bias is time-invariant, conditional on the set of controls. The credibility of this approach relies on how well the set of covariates used in the matching procedure explains the selection process. We perform the propensity score matching using a long list of individual characteristics that might influence both selection and vehicle choices. Importantly, in addition to a rich set of socio-demographics, we include measures of energy and investment literacy, as well as proxies for environmental values and attitudes.
How is policy awareness determined? A crucial assumption we take to identify the effect of policy awareness exploiting the difference in differences strategy is that all control variables X are exogenous (that is, using the potential outcome notation: X 1 = X 0 = X). Some concerns about the validity of this assumption might arise for policy awareness.
The exogeneity assumption would not be met for policy awareness (Aware 1 = Aware 0 ) if the process determining individuals' awareness differed in the presence and in the absence of the Bonus/Malus system. For instance, in the presence of the Bonus/Malus system, individuals might be exposed to information about the introduction of the policy that, based on individual-specific factors unobserved to the researchers (e.g., attention to fuel efficiency, opportunity cost of the time spent in accumulating information), might or might not be incorporated in the determination of the knowledge about the features of the incentive scheme for the vehicle registration tax. In this case, also the conditional independence assumption of the matching difference in differences is not met. This threat might be severe in this context because we do not control nor observe the process that determines individuals' awareness about the features of the vehicle registration tax scheme.
Clearly, the presence of unobserved heterogeneity affecting both vehicle choices and policy awareness is the reason why we cannot simply compare the vehicle choices of aware and unaware consumers in the presence of the Bonus/Malus system to estimate the impact of awareness.
We address the potential issue of endogeneity of policy awareness adopting an instrumental variable (IV) approach. This is implemented both to: (a) simply compare vehicle choices of aware and unaware respondents in treated cantons, after the introduction of a Bonus/Malus scheme, and (b) in the difference in differences framework. In the first case (a), a good instrument must be a good predictor of individuals' awareness about the presence of the Bonus/Malus system, and influence vehicle choices only through individuals' awareness, conditional on the set of controls. In case (b), we require a milder identifying assumption: an instrument must influence the evolution over time of vehicle choices between aware individuals in cantons with a Bonus/Malus system and aware consumers living in cantons that did not implemented such incentives only through individuals' awareness, conditional on the set of controls.
We use two instruments for policy awareness: (i) the distance in years from the introduction of the incentives in a specific canton and (ii) the voting participation rates to eight national referendums on energy, environmental, and transportation topics at the municipality of respondents' residence level, interacted by canton of residence dummies. 21 Because in our difference in differences framework not only policy awareness need to be instrumented, but also its interaction with the treatment variable BM and the indicator for the presence of the Bonus/Malus system BM P , in the set of instruments we also include the interaction of the treatment variable BM with (i) and (ii), as well as the interaction of the indicator for the presence of the Bonus/Malus system BM P with (i) and (ii).
The distance between the introduction of the policy and the purchase of the car as an instrument for policy awareness exploits the idea that advertisement and promotion of the Bonus/Malus system in the local media might be higher immediately after its introduction. Hence, the less time has passed from the introduction of the fiscal measure, the higher the probability for consumers to be aware about the incentive scheme when they take the decision of purchase. The exogeneity of the timing of introduction of the fiscal incentives to individuals' decisions of purchase ensures the validity of the exclusion restriction for this instrument.
At the municipality level, higher voter turnout might be related to higher individual policy awareness through the presence of higher social capital. 22 Clearly, the validity of the exclusion restriction for this instrument depends on the absence of unobserved heterogeneity influencing both voter turnout and the probability to purchase a more energy efficient 21 The eight referendums were held from 2013 to 2017 and concerned the following issues: increase of highway tolls to fund road infrastructure; replacement of the value added tax with a tax on non-renewable energy; restoration of the Gotthard tunnel; earmarking of the revenue from the mineral oil tax exclusively for road circulation purposes; setting goals for the reduction of natural resources consumption; phasing out nuclear power; creation of a permanent government fund for national roads; approval of the new Swiss energy policy (Energy Strategy 2050).
22 Previous literature has indeed used voter turnout as a proxy for social capital at the local level (see Guiso et al. 2004). vehicle. Because we also control for several factors at the individual level related to environmental preferences and attitudes, including environmental activism, we believe the exclusion restriction is valid.

Measurement error in awareness
Individuals' awareness about the presence of Bonus/Malus system is measured at the time of the interview but refers to the time at which the current vehicle was purchased. This might reflect into measurement error and attenuation bias in the estimated effect of awareness on vehicle choices. 23 The IV approach also aims at addressing the issue from the potential measurement error in individuals' awareness. 24 Table 1 reports the main results of the empirical analysis. To underline the importance of considering individuals' policy awareness to evaluate the consequences of the Bonus/Malus system, we first estimate the average effect of these fiscal incentives without taking into account whether respondents were policy aware or not. 25 Column (1) shows the results of this analysis. When we do not distinguish between aware and unaware respondents, we estimate an average effect of the Bonus/Malus system on vehicle fuel consumption of about −0.05%, not statistically significant.

Main specification
To gain insights about the association between individuals' awareness and vehicle choices, we first consider only individuals that purchased a vehicle in a treated canton, after the introduction of the incentives. We then simply regress the log of fuel consumption on awareness status and a large number of respondents' characteristics, canton and year of purchase dummies. 26 Results reported in Column (2) of Table 1 show that individuals living in treated cantons, that are aware of the presence of the Bonus/Malus system and purchased a vehicle after its introduction, own vehicles that consume on average around 6.8 percent less than unaware individuals.
We then start our investigation of the impact of individuals' awareness on vehicle choices by estimating our basic difference in differences specification (4) using OLS. The results are reported in Column (3) to (8) (3), obtained omitting individuals' characteristics, show that the introduction of the Bonus/Malus system has a negative and highly significant impact of 10.9% percent on the fuel consumption of the vehicles purchased by individuals that are aware about the features of the registration tax 23 On the topic of attenuation bias in case of measurement error, see Levi (1973). 24 Besides instrumenting awareness, to address potential concerns related to its measurement, we conduct additional robustness checks modifying the way we characterize unaware individuals. scheme. We find no significant effect of the Bonus/Malus system on the vehicle choices of unaware consumers. The results are almost unaffected when we include a large set of respondents' characteristics (Column 4), include canton by year of purchase fixed effects (Column 5), consider only vehicles purchased in 2010 or later (Column 6), include the interaction of awareness status and treatment period dummies (Column 7), or allowing for Bonus/Malus type specific effects by awareness status (Column 8).

Placebo Bonus/Malus
Our basic specification assumes that vehicle choices of aware individuals living in administrative areas where a Bonus/Malus system has been introduced would have evolved over time similarly to those of aware individuals in cantons without incentive schemes to promote energy efficient cars. To provide evidence that such condition is met in the pre-treatment periods, we perform a placebo test where we set the introduction of the Bonus/Malus system in some pre-treatment period in each treated canton, and restrict the sample period to years before the actual introduction of the Bonus/Malus system. Clearly, failing to reject the null of parallel pre-treatment trends between vehicle choices of aware individuals in cantons with and without fiscal measure would raise serious concerns about the validity of this strategy. For instance, aware respondents in cantons that eventually introduced a Bonus/Malus might have a stronger preference for efficient cars -not captured by our observables -than aware respondents in cantons without such policy. Then, the effect we would measure in the main specification would not be due to the introduction of the Bonus/Malus, but rather due to the higher propensity to buy the more energy efficient cars launched on the market over the years.
Results of the placebo difference in differences regressions are reported in Table 2, where Columns (1) to (3) show results when the placebo Bonus/Malus is introduced 2 years, 3 years and 4 years earlier than the actual introduction, respectively. The coefficients associated to Aware * BM P are statistically equal to zero in all three cases, suggesting that there are no differences in the trend of vehicle choices among aware respondents in cantons with and without fiscal incentives.
Propensity score matching difference in differences The first step of our approach is to match respondents along two dimensions separately: (i) Bonus/Malus treatment status and (ii) policy awareness status. We then estimate two separate Probit models for the Bonus/Malus treatment and policy awareness variables. Selected estimated coefficients are reported in Table 35 in Appendix E.
Although the aim of this estimation procedure is to match respondents on all observable characteristics that may influence selection into either Bonus/Malus treatment status or (3)  awareness status and not to estimate a casual model of selection, some relevant associations emerge in the Probit estimates. These largely confirm the descriptive evidence reported in Section 4. Individuals living in cantons that have introduced a Bonus/Malus scheme are less likely to live in rural areas, tend to have lower income, are associated with a lower level of investment literacy and have a lower probability to be German speakers. Without getting rid of these compositional differences, one might fear that, for instance, cultural differences and heterogeneity in the incentives of the two groups, due to the different availability of resources and type of living areas, induced a selective introduction of the Bonus/Malus incentives and influenced the decisions with respect to the efficiency of the newly purchased cars. The propensity score matching procedure based on the Bonus/Malus treatment status identifies 192 individuals in the treatment group that are not comparable in terms of observable characteristics to those in the control group. The probability to be policy aware shows a hump-shape profile over the age of the respondents, is higher for male respondents and individuals that are self-employed or retired, and is positively associated with the level of energy literacy. However, the compositional differences between aware and unaware individuals that emerge as a result of the propensity score matching are less relevant than those resulting from the matching procedure based on the Bonus/Malus treatment status. 27 Column (2) of Table 3 shows the estimation results of our propensity score matching difference in differences approach with the matching based on the Bonus/Malus treatment status, while Column (3) reports the results obtained when the matching is performed based on the awareness status. These estimates for the effect of policy awareness are very close to those estimated with our basic differences in differences estimation. This evidence suggests that issues related to selection on observables into treatment and awareness status do not represent a relevant threat to our main identification strategy. Table 3 show the IV estimates obtained when policy awareness is instrumented to deal with its potential endogeneity.

IV estimates Columns (4) and (5) of
In particular, Column (4) reports the 2SLS estimates of equation 4 when we set β = θ = ψ = 0 and restrict the estimation sample to individuals that purchased a vehicle when fiscal incentives towards the adoption of energy efficient vehicles were in place in a treated canton. The policy awareness variable Aware is instrumented with the distance in years from the introduction of a Bonus/Malus system in a certain canton and the purchase of the vehicle, and the voter turnout in the national referendums at the municipality of residence level, interacted by canton. The results of the first stage regression lead us to reject the null of weak instruments at the 1 percent confidence level. 28 The 2SLS estimate indicates that, in the presence of the Bonus/Malus system, policy awareness induces individuals to purchase vehicles that consume on average 16.3% less than unaware consumers. Notice that this estimate is larger than the corresponding OLS estimate reported in Column (2) of Table  1, suggesting that neglecting the potential endogeneity of awareness in this simple framework might bias its estimated effect downwards.
In Column (5) we report the IV estimates of our main specification when we instrument Aware (and its interaction with the policy variables Aware * BM P and Aware * BM ) with the distance in years from the introduction of a Bonus/Malus system in a certain canton and the purchase of the vehicle, the voter turnout in the national referendums at the municipality of residence level, and the interaction of those with the treatment variable BM and the indicator for the presence of the Bonus/Malus system at the time of purchase BM P . Also in this case, the first stage regression results lead us to reject the null of weak instruments at the 1 percent confidence level. 29 Adopting the 2SLS estimation approach for equation 4, we find that policy awareness induces individuals living in a canton with a Bonus/Malus scheme in place to purchase vehicles that consume on average 30.3% less than unaware individuals living in those cantons. This IV estimate for the impact of policy awareness is substantially larger than the corresponding estimate obtained with our basic difference in differences specification. Hence, our results confirm that the endogeneity issues might tend to attenuate the magnitude of the effect of policy awareness estimated with OLS.
Alternative definitions of unaware respondents So far we defined unaware respondents those who provided a wrong answer to our awareness question, together with those who answered "Don't know". However, individuals who gave a wrong answer might have different expectations towards the fiscal incentives in place in their canton of residence compared to respondents who hold no information on the presence or absence of such incentives. In fact, a wrong answer to our awareness question in the absence of the Bonus/Malus implies that, at the time the respondent purchased her car, she mistakenly believed that a fiscal incentive to promote the purchase of efficient vehicles had been in place. In contrast, individuals that gave a wrong answer in the presence of the policy believed that there was no fiscal incentive to purchase an efficient car when this was instead in place. Hence, these unaware individuals might act on such believes by purchasing vehicles with higher energy efficiency in the absence of the policy compared to those that unaware individuals buy in the presence of the policy.
Moreover, some previous research discuss the role of self-confidence on the probability to give a definite answer vis à vis reporting "Don't know" (see, for instance Lusardi and Mitchell 2014). This could be problematic if this self-confident attitude was correlated with vehicle choices.
In column (6), we include in the sample only unaware respondents who answered "Don't know" to our awareness question. In column (7) only respondents who gave the wrong answer ("yes" in absence of a Bonus/Malus, "no" when it was present) are included in the estimation sample. In both cases, we observe a strong and significant effect of the Bonus/Malus on aware respondents.
However, when we consider only people who provided the wrong answer, we estimate a positive and significant effect of the presence of the Bonus/Malus for the unaware consumers. This suggests consumers that believe that there is no fiscal incentive towards efficient vehicles, in cantons with a Bonus/Malus, buy less efficient cars than consumers that believe such incentives exist in cantons with no such measure.

Alternative indicators of vehicle choices
As discussed in Section 4, in the main analysis we use self-reported fuel consumption as dependent variable because almost all our respondents were able to provide it, its distribution matches quite well that of the population of vehicles in Switzerland, and it is strongly correlated with both CO 2 emissions and efficiency rating (see Appendix B). Nevertheless, in this section we use the two vehicles' features that determine the eligibility to the Bonus/Malus system, namely efficiency rating and CO 2 emission rate, as dependent variables.
We construct two indicators for efficient cars: a dummy for vehicles with rating A, and a dummy for vehicles with rating A or B. Equation 4 is estimated for each of these as a linear probability model, where the coefficient of interest β is now the effect of respondents' policy awareness under a Bonus/Malus scheme on the probability of buying a high efficiency rating car. We also construct a measure of CO 2 emissions per 100 km based on self-reported vehicle fuel consumption and fuel type. This is possible because a linear relationship exists between fuel consumption and CO 2 emission rate within fuel type. 30 Results in columns (1) and (2) of table 4 show that the introduction of fiscal incentives for efficient vehicles increased the probability of aware respondents of buying a rating A vehicle (+31.1%) and a vehicle with either rating B or higher (+38.4%). One important 30 We use data from the Swiss vehicle registration database to estimate the linear relationship between fuel consumption and CO 2 . Regressing CO 2 emissions on fuel consumption separately for diesel and gasoline vehicles and omitting the intercept gives a coefficient of 26.44701 and 23.6457 respectively, with virtually perfect fit (R-squares 0.9998 and 0.9997). See Appendix B for more information about the Swiss vehicle registration database.  caveat in interpreting this set of results is that a large number of respondents (58.38%) were not able to indicate the efficiency rating of their car. Thus, it is possible that estimates obtained by using the rating as dependent variable suffer from selection bias. For instance, unobservable factors such as limited attention or cognitive abilities might drive both the selection into the awareness status and the probability to report the efficiency rating of the vehicle. Another possibility is that consumers are more likely to remember the energy rating of a highly efficient car.
The estimated impact of awareness on CO 2 emissions, reported in Column (3), is very similar in magnitude to that on fuel consumption, corroborating the results of the main specification.  (3) is the Log of grams CO 2 emissions per 100 km. Standard errors in parenthesis, clustered at canton by year of purchase level. * / * * / * * * indicate statistical significance at the 10, 5, and 1 percent level, respectively.

Vehicle age and fuel consumption
As discussed in Section 3, age of the vehicle is typically an eligibility criteria under the bonus component of the Bonus/Malus system. Moreover, newer models tend to be more energy efficient.
The fiscal incentives can then induce a behavioral response of consumers through at the least two mechanisms: (i) they might access the fiscal incentives provided by the Bonus/Malus system through the purchase of a more efficient vehicle of the same age (efficiency channel); (ii) they could buy newer cars to receive the registration tax discount for a longer period of time (vehicle age channel). We wish to check whether both or only one of those two channels are driving our main results.
In column (1) of table 5 we report results obtained using vehicle age at the time of purchase as a dependent variable. Aware consumers buying a car under a Bonus/Malus scheme choose vehicles on average around 1.46 years newer. In column (2) we use fuel consumption as dependent variable instead, but we look at the effect of awareness in the presence of fiscal incentives, conditional on the age of the vehicle at the time of purchase. In particular, we distinguish between new cars, second-hand cars still eligible to the Bonus (maximum 4 years old), and second-hand cars not eligible anymore (more than 4 years old). 31 We find that policy awareness has an impact on the fuel economy of the newly purchased vehicles up to four years of age at the time of purchase. In contrast, when the vehicle purchased cannot be eligible anymore due to age limits, we do find any effect of policy awareness. Thus, both the efficiency channel and the vehicle age channel play a role in determining the overall impact of policy awareness on vehicle choices. Importantly, the virtually zero effect for cars not anymore eligible due to their age confirms the robustness of our main findings.
Heterogeneity by respondents' characteristics Finally, we provide illustrative evidence that aware consumers are not homogeneous in terms of their response to fiscal incentives for efficient cars. For instance, perspective buyers with high income might be less prone to switch to a more efficient vehicles even when aware about the presence of the fiscal advantages in doing so. This would be the case, for instance, if the budget constraint was binding for low income earners in the absence of the Bonus/Malus scheme. In addition, aware individuals with low energy and investment literacy might not be able to incorporate the fiscal incentives correctly in their decision making process. Finally, conditional on being aware, individuals with strong pro-environmental values might be more responsive to the fiscal incentives.
In tables 6 and 7 we present results that aim at highlighting what are the most important sources of heterogeneity in the effect of policy awareness. We focus on a selected group of respondent characteristics that might influence vehicle choices, such as monthly house- Results show some considerable heterogeneity in the response of aware individuals over some of those dimensions. 33 In particular, aware respondents react less strongly to a Bonus/Malus scheme when high income earners, or when they indicate "pleasure" or "wealth" as important personal life values. On the opposite, aware respondents who are Green Party voters, are member of an environmental club, or have higher energy literacy, tend to choose even more efficient cars than other aware consumers. We find no heterogeneity in the response between male and female consumers or between consumers belonging to different language groups.
These results reveal that information campaigns complementing the introduction of fiscal policies might have substantially heterogeneous effects across different groups of consumers. Importantly, the heterogeneity in the effect of policy awareness between individuals with low and high energy literacy suggests that policy awareness enters the decision making process mainly through the reduction in the vehicles' life cycle costs and not through heuristics. This prompts policy makers to complement information campaigns with literacy programs that allow policy aware consumers to better understand the role of the fiscal incentives in their decision making process.
While these results allow to gain some insights into how awareness may interact with other individual factors in determining their responses to the fiscal incentives, we want to stress that the present analysis cannot pin down the exact mechanism through these factors induce aware respondent to react differently to the fiscal incentives. 34 32 As in the previous results, such characteristics are also included as controls. For energy literacy here we mean whether a respondent understands that the efficiency rating is a relative measure of fuel efficiency. 33 We also check for heterogeneity for all other individual characteristics we consider as baseline controls. In the majority of cases, the coefficients are not statistically significant, but the sign of the coefficients are in line with the results presented here.
34 For instance, the evidence that aware respondents who value wealth highly tend to react less to the incentives might be due to either personal beliefs or a different financial situation, even after controlling for household income. Notes: Results disaggregating the effect of awareness by selected respondents' characteristics. Dependent variable is the log of fuel consumption per 100 km. Coefficients associated with respondents' characteristics (rows 2-5) show the change in the baseline effect of awareness in the presence of Bonus/Malus (row 1). Standard errors in parenthesis, clustered at canton by year of purchase level. * / * * / * * * indicate statistical significance at the 10, 5, and 1 percent level, respectively.  -5) show the change in the baseline effect of awareness in the presence of Bonus/Malus (row 1). Standard errors in parenthesis, clustered at canton by year of purchase level. * / * * / * * * indicate statistical significance at the 10, 5, and 1 percent level, respectively.

Discussion and Conclusion
When evaluating the impact of a public program, separating the role of policy awareness from that of other behavioral anomalies or possible limitations of the policy design is important because each of these factors have different implications for the understanding of individual behavior and the design of future policies. In this paper we explore the role of awareness about the presence of fiscal incentives for the purchase of energy efficient vehicles on consumers' vehicle choices. To identify the impact of policy awareness, we exploit the variation in the introduction of the policy over time and across administrative areas, as well as a direct measure of awareness at the individual level. While ignoring policy awareness might lead to conclude that the fiscal incentives are ineffective, we find that policy awareness induces individuals to purchase more efficient and newer vehicles. This result is robust to several robustness checks, to the addressing of compositional differences between aware individuals in treated and control cantons and is confirmed by an IV strategy to deal with the potential endogeneity of awareness. Our findings about the impact of policy awareness complement previous descriptive evidence suggesting low program awareness to be an important barrier to the take-up of fiscal benefits (Bhargava and Manoli, 2015).
In the context of fiscal incentives to promote the adoption of efficient vehicles, our results suggest that a simple information treatment (e.g., a mandatory registration tax disclosure), making consumers aware of the existence of the fiscal incentives at the time of purchase, would induce a significant vehicle choice response. More generally, the evidence we present prompts then policy makers to complement the design and introduction of public policies with adequate information campaigns. We show that failing to do so might lead to a low average impact of the policy measure, which could be wrongly blamed on flaws of the policy design or other behavioral factors that would need different interventions to be corrected. Importantly, our estimates for the impact of policy awareness should be interpreted as the potential increase in individuals' response that could be achieved with an information treatment that succeeds in making everyone in the target population aware about the existence of the fiscal program. Another implication of our results is that the tax burden of fiscal measures disproportionally affects unaware individuals. While aware individuals might modify their behavior to benefit from the monetary incentives, unaware consumers will not respond even though they could potentially be better off when doing so.
These findings have important implications for a wide range of existing public programs. The impact of increasing awareness on individual behavior might be especially prominent in contexts where the application of the benefits is automatic and requires little or no addi-tional effort from the recipients (e.g., Medicare, marriage taxes and benefits, student quotas, discounts for public transport subscriptions). Clearly, the presence of other barriers to the access of fiscal benefits, such as transaction costs or stigma, might hinder the response of aware individuals to a public program. Moreover, in contexts with higher policy complexity, possible behavioral failures might play a more prominent role.
A limitation of this study is that we cannot specify the minimum level of knowledge about the fiscal system and individuals' sophistication that are necessary to induce a behavioral response to the policy. In order to identify the effect of awareness, we use a basic indicator capturing an even vague knowledge about the presence of the policy. Although we document some heterogeneity in the response to the policy among the group of aware individuals, we cannot say which level of sophistication is required for the consumers to react. We believe such a question should be addressed by future research. Future work should also explore what type of interventions are best suited to foster the relevant individuals' awareness in different contexts.

A Description of cantonal bonus systems
We reconstructed the criteria and the evolution of the vehicle registration tax and the Bonus/Malus system for each canton by looking at the official legislation between 2005 and 2018. We have information on thresholds of CO 2 emissions and/or efficiency rating necessary to be eligible to the bonus, the maximum age of the car, the size of the discounts and the penalties (in most cases a percentage discount on the baseline registration tax), and the years in which the Bonus/Malus was implemented. To be eligible for a bonus, some cantons require a vehicle to be registered for the first time only after a certain date. Seven cantons also implemented a malus -that is, an increase in the baseline tax if a car has low energy rating or high CO 2 emissions. Table 8 and figure 4 show the characteristics of the Bonus/Malus system in each canton.
Some cantons have also benefits for hybrid electric vehicles, plug in battery electric vehicles, or other alternative fuel vehicles. Those vehicles represent a very small share of the SHEDS sample and therefore those incentives are not considered in our analysis.  consumers when buying a car: baseline vehicle characteristics, fuel consumption per 100 km, CO 2 emissions per km, and energy efficiency rating. Therefore the label reports two clear indicators on the level of energy efficiency of a car -fuel consumption and efficiency rating -and an environmental indicator -CO 2 emission rate. Estimates of fuel costs per km or vehicle registration taxes are not shown. The efficiency rating is a relative measure of efficiency, and the the thresholds used to assign the efficiency rating are recalculated each year. All recent vehicle types on the market are ordered according to an evaluation coefficient which depends on both absolute fuel consumption per 100 km and fuel consumption per 100 km over vehicle weight. Based on this coefficient, vehicle types are assigned to seven equally sized groups from A to G.

B Fuel consumption variable
The main outcome of interest in our analysis is self-reported information on vehicle fuel consumption. In this Appendix we provide descriptive evidence on the accuracy of our self-reported data, and then on the strong relationship between actual fuel consumption, efficiency rating, and CO 2 emissions.
To do so, we use data from the Swiss vehicle registration database, containing information on all vehicles registered in Switzerland at the end of 2015. While, due to data confidentiality reasons, we cannot match each individual vehicle in SHEDS with its corresponding vehicle in the registration data, we can compare the fuel consumption distribution between the two databases. Figure 6 plots the percentile of fuel consumption using SHEDS data and the registration data. In both datasets, we limit the sample to cars bought in 2015 or earlier, excluding the canton of Ticino. We observe a good overlap between the self-reported SHEDS fuel consumption and the actual fuel consumption through all percentiles, suggesting a limited presence of measurement error issues. Figure 7 shows fuel consumption distribution by efficiency rating using the registration data. We group vehicles in two categories: ratings A and B (which can benefit from a bonus) and ratings from C to G (which cannot receive a bonus). Despite some overlap, we see that most vehicles belonging to worse efficiency ratings have higher fuel consumption.
Finally, in Figure 8 we illustrate the almost linear relationship between CO 2 emissions and fuel consumption. For visibility sake, we include only vehicles registered for the first time in 2015. The graph clearly shows an almost perfect linear relationship among cars with the same fuel type.

C Information on SHEDS questionnaire
SHEDS is organized in multiple waves, from 2016 to 2020, carried out by the survey company Intervista. Participants are taken from a larger pool of potential respondents, and some individuals are surveyed in multiple waves. Each wave, the pool of respondents is selected to be representative of the French and German Swiss population. People living in the canton of Ticino are not surveyed. Our main dataset comes from the 2018 wave. Baseline respondent characteristics, like income, education, and household size, are collected by Intervista each year outside the questionnaire. In our survey some questions are asked only to new respondents i.e. individuals who were never interviewed in the previous waves. We use data from the 2016 and 2017 waves to collect the answers in case of recurring respondents.
The 2018 wave of SHEDS contains 5011 individuals. Among those, 3621 (72.26%) own at least a car in their household. Due to the presence of missing variables, the final sample used in the baseline analysis contains 2504 respondents. Table 9 shows how many observations have been omitted among respondents with a car. The majority of missing observations comes from two variables: financial literacy and household monthly income. Many values from financial literacy are missing simply because all respondents surveyed in waves 2016 and 2018, but not in 2017, were not asked the two questions that compose the financial literacy indicator used in the analysis. We argue that whether a respondent participated to a specific wave is essentially exogenous to our phenomenon of interest. Some other respondents were reluctant to provide information on their household's income. Nevertheless, repeating our main analysis either excluding these two variables or including the missing variables under a separate index does not change our results.
We also trim the top and bottom 0.5% fuel consumption percentiles (24 observations), to exclude patently wrongly reported values. That means dropping values of fuel consumption (L/100 km) lower than 2 and higher than 40. As in the previous case, including those observations in the analysis does not affect our conclusions.
In the analysis we include a rich set of controls to take into account factors related to the fuel consumption of the car respondents bought. 35 These are summarized in tables 10, 11, 12, 13 and 14. While some of the questions were asked in wave 2018, others were not asked again in case of participation to multiple waves. For what concerns vehicle information, many questions were not asked every wave unless the respondent told she changed vehicle between waves.
Besides baseline questions on respondents and household and vehicle characteristics, we have also questions on energy and financial literacy (table 10). Financial literacy measures whether respondents answered correctly to two standard questions to measure the understanding of interest rates and inflation. For energy literacy, we have indicators for the correct answer to four questions: electricity cost for running a desktop computer for one hour; electricity cost for running a washing machine at 60°C with 5 kg load; knowledge that vehicle energy efficiency rating is a relative measure of efficiency rather than absolute; which has higher energy cost per 100 km between a gasoline car consuming 5 L/100 km and an electric car consuming 20 kWh per 100 km.
Another set of controls represent general attitudes towards life. We ask respondents how important certain values or lifestyles are for them using a scale from 1 (Not at all important) to 5 (Very important), and we create indicators equal to 1 for those who indicated 4 or 5, and zero otherwise. The list of values and lifestyles are summarized in table 11.
The first set of environment-related controls used in the paper are shown in table 12. They represent statements about feelings towards the environment, to which respondents are asked to indicate their degree of agreement with a scale between 1 and 5 (from totally disagree to totally agree). We then create indicators on whether people answered 4 or 5.
We also have a second set controls about environmental activism, and behavior and expectations towards the environment (table 13). The latter are once more based on a scale of agreement between 1 and 5, and they are transformed into binary indicators under the same criteria as before.                         (2) is a dummy that indicates whether the respondent is aware about the presence of the policy. Regressions also control for respondents' education, household size, indicators for energy literacy, preferences towards the environment, life attitudes and values. Standard errors in parentheses. * / * * / * * * indicate statistical significance at the 10, 5, and 1 percent level, respectively.