Fiscal Policy and CO2 Emissions of New Passenger Cars in the EU

To what extent have national fiscal policies contributed to the decarbonisation of newly sold passenger cars? We construct a simple model that generates predictions regarding the effect of fiscal policies on average CO2 emissions of new cars, and then test the model empirically. Our empirical strategy combines a diverse series of data. First, we use a large database of vehicle-specific taxes in 15 EU countries over 2001-2010 to construct a measure for the vehicle registration and annual road tax levels, and separately, for the CO2 sensitivity of these taxes. We find that for many countries the fiscal policies have become more sensitive to CO2 emissions of new cars. We then use these constructed measures to estimate the effect of fiscal policies on the CO2 emissions of the new car fleet. The increased CO2-sensitivity of registration taxes have reduced the CO2 emission intensity of the average new car by 1,3 percent, partly through an induced increase of the share of diesel-fuelled cars by 6,5 percentage points. Higher fuel taxes lead to the purchase of more fuel efficient cars, but higher annual road taxes have no or an adverse effect.


Introduction
, or to convergence to the EU average, whereas Denmark's move from being average to becoming one of the most fuelefficient countries might be the consequence of its aggressive car tax policies. Overall, we can sort these explanations into five categories. The first category relates to the EU's first pillar, which requires manufacturers to sell more fuel efficient cars. For this reason, the portfolio of cars available for purchase is expected to become more fuel-efficient over time.
Second, the type of cars bought and their fuel efficiency may partly be explained by trends in consumer preferences. The EU second pillar is related to this cateogory. Third is fiscal policy on registration and road taxes, related to the third pillar. Countries have developed widely different fiscal policies aimed at promoting fuel-efficient cars. Some countries have much more aggressive policies vis-a-vis other countries, and countries moving in the same direction still implemented their policies in different years. Fourth are fuel taxes, which differ substantially between countries, and many countries have seen changes in fuel taxes.
Fifth are income and the economic crisis. Higher incomes are associated with larger cars, and lower incomes with smaller, more fuel-efficient cars. In addition, the economic crisis hit the countries very differently. Those countries hit hardest by the crisis are expected to see -all else equal -the largest drop.
In this paper, we focus on the third cause, associated with the third pillar. That is, we address the following research question: to what extent have national fiscal policies contributed to the decarbonisation of newly sold passenger cars? We construct a simple model of a representative agent to generate predictions regarding the effect of fiscal policies on average CO2 emissions of new cars. We study changes at the aggregate level and are interested in differences between countries and changes over time within countries.
That is, the model and our econometric analysis do not provide a detailed micro foundation of consumers' decisions; see Berry et al. (1995) or Meerkerk et al. (2014) for such an analysis. After presenting the model, we build a dataset in which we compare vehicle tax systems across 15 countries over the years 2001-2010. We use a dataset of vehicle-specific taxes, and use these data to characterize each country's tax system at year t with regards to the average registration and road tax, and the sensitivity of the taxes with respect to the car CO2 emissions. We differentiate taxes by petrol and diesel, so that we construct 8 variables to provide an elaborate characterization of a country's vehicle tax system for a given year. Both the construction of the multiple tax proxies and the multi- The constructed variable are used to empirically study the effect of the fiscal treatment, especially the car purchase tax, on the fuel efficiency of newly sold cars. We identify the effect by considering dynamic differences between countries in car taxes and in emission intensities. We control for static differences between countries through country fixed effects, control for income and for common dynamic patters (e.g. EU policies) through time fixed effects. We can identify the effect of fiscal policies on car sales as some countries have consistent low purchase taxes (<30%) that are not very sensitivity to CO2 emissions (Belgium, France, Germany, Italy, Luxembourg, Sweden, United Kingdom), while Spain has low purchase taxes but these have become substantially more CO2 sensitive over the period 2001-2010. Greece has high purchase taxes (>30%) but these became less CO2 sensitive over the years, and the remaining countries (Austria, Denmark, Finland, Ireland, Netherlands, Portugal) have relatively high purchase taxes (30%), with a CO2 component that substantially increased over the years (>10 €/(gCO2/km)), though the countries differ substantially. Our empirical strategy is based on the correlation between the uneven developments in taxes and patterns in the emission intensities for these countries.

Literature
There is an emerging empirical literature on the effects of fiscal policies on the fuelefficiency of newly sold cars. The general finding is that fiscal policies are an effective tool to influence car purchase decisions. In addition, the literature establishes that purchase taxes are more effective than annual (road) taxes, and that tax reform can cause sizable petrol-diesel substitution.
A strong example of the responsiveness of car purchases to fiscal policies is provided by D' Haultfoeuille et al. (2014). They assess the effect of the system of fees and rebates that existed in France from December 2007 to December 2009. In this system, owners of fuel efficient cars could receive a tax rebate of up to 1000 euros whereas fuel inefficient car owners had to pay a fee of up to 2600 euros. The precise rebate and fee thresholds showed up remarkably in the sales for different car types, with large sales increases just below and drops just above the thresholds.
The effectiveness of car taxes can depend on the subtle features of the policy adopted. The theory of rational choice for car purchases assumes that consumers fully internalize both the expected cost in terms of annual road and fuel taxes, as well as purchases taxes. Empirical evidence however suggests bounded rationality. Consumers do not exploit all available information equally, and tend to give more weight to short-term costs and benefits, known as "consumer myopia" or nearsightedness (DellaVigna, 2009).
For example, when deciding on whether to purchase a more fuel efficient car, consumers tend to calculate the expected savings in fuel costs only for about three years (see Greene et al., 2005;Kilian and Sims, 2006;Greene et al., 2013). This nearsightedness is considered a main reason why, compared to annual taxes, vehicle acquisition taxes are more effective in directing consumers' buying decisions (Brand et al., 2013;Gallagher and Muehlegger, 2011;Klier and Linn, 2012;van Meerkerk et al., 2014).
Another phenomenon identified by the literature is the substitution between petrol and diesel cars. When Ireland differentiated its purchase and annual road taxes according to CO2 emission intensities, sales of smaller cars did not go up. Instead, sales of diesel cars increased at the expense of large petrol cars (Hennesy andTol, 2011, Leinert et al., 2013). This unanticipated shift towards diesel cars reduced the average CO2 emissions by 13 percent in the first year after the tax reform (Rogan et al., 2011). Less advantageous, it also raised NOx emissions (Leinert et al., 2013). The vehicle acquisition tax reform in Norway in 2007 resulted in a drop in CO2 emissions of newly sold cars by 6 gCO2/km in the short run, mainly caused by an increase of diesel market share by almost 23 percentage points (Ciccone, 2014). The tax reform in Denmark in 2007 contributed to the sales of more fuel-efficient cars in the years thereafter. Yet, Mabit (2014) argues that the biggest contribution to the sales of fuel-efficient cars is probably not the tax reform, but technological improvements.
All research discussed above analyses the effect of specific vehicle tax policies in a single country. Hence, these papers cannot control for year-specific effects and the results are not easily generalizable. Our empirical analysis does not consider a single-event in one country, yet studies more broadly the fiscal treatment of car purchases and ownership in relation to car emissions. There are some previous cross-country and panel-data studies on the effect of fuel prices on fuel efficiency (Burke and Nishitateno 2013, Klier and Linn 2013). The effect of the registration and road tax level on car purchases is previously studied in Ryan et al. (2009), who use a panel structure for EU countries. They conclude that vehicle taxes, notably registration taxes, are likely to have significantly contributed to reducing CO2 emission intensities of new passenger cars. Ryan et al. (2009) focus on the average level of registration taxes in a country. 10 We take this analysis one step further by constructing measures of the CO2 sensitivity in addition to the level of registration and road taxes. This allows us to exploit differences between EU countries in the stringency and timing of climate-related vehicle fiscal policies. An important part of our study is thus a 10 Note that Ryan et al. (2009) weigh the registration tax measure by vehicle sales, so that in their analysis the right-hand-side variable depends on policy outcomes. To prevent dependency of right-hand variables on policy outcomes, we construct tax measures that do not use sales for weighing; see footnote 13. more comprehensive characterization of the vehicle tax system that can be used to compare differences across countries and changes over time, based on a large dataset of country-year-vehicle specific prices inclusive and exclusive of taxes.

Model
We illustrate the effect of vehicle purchase taxes on the average emission intensity with a simple model. We consider two car types. A representative consumer maximises (expected future) utility u dependent on the current purchase of cars, q1 and q2, and income m net of purchase expenditures x: where are costs per quantity, including registration taxes as well as future variable costs and annual taxes. The utility function satisfies the standard assumptions on continuity, differentiability, positive derivatives, and concavity. We also assume that both types are normal goods (increasing consumption with increasing income, decreasing consumption with increasing prices) and that the total budget for cars, , increases in total income, .
We do not model consumers' care about the environmental performance of cars as such (see Achtnicht 2012 for an analysis along those lines), but focus on the effects of government instruments geared to direct consumers' choices. We assume that the tax is fully shifted to consumers, 11 so that the consumer price of cars is where is a type-specific ad valorem tax and is the producer price.
The tax consists of a uniform component and an environmental component, where is a relative weight of the environmental component. The two car types have different emission intensity, say grams of CO2 per km, which we denote by . Without loss of generality, let , for example because car type 2 is more spacious, has more weight, or is more fancy. The type-specific tax becomes: . ( We are interested in the effect of changes in car taxes on the average CO2 intensity of the car fleet, which we define as .
Policy can change the uniform component of the car tax, , the environmental component, θ, or both. We define the average car-tax, given by .
so that we can study shifts in the tax structure while keeping a constant overall tax rate. It is intuitive that an increase in the weight of car-feature θ, while keeping the average tax rate T constant, will decrease the average emission-intensity of the cars.
Proposition 1. An increase in the weight of environmental performance in taxes, θ, while keeping average total taxes T constant, will decrease the average CO2 intensity B: Proof. The policy in the proposition increases the price of the relatively emission-intensive car and decreases the price of the more fuel-efficient car. The result follows immediately from the assumption that both car types are normal goods. Q.E.D.
Thus tilting the car taxes to become more CO2-dependent will make the car fleet more CO2efficient. The effect of an overall car tax increase depends on the comparative income elasticity of the two car types.
Proposition 2. If the environmental tax component is sufficiently small, then feature B decreases with an overall tax increase (or equivalently an increase in T) if and only if the less fuel-efficient car type has higher income elasticity: Proof. Consider    .
When 0, an increase in is equivalent to a decrease in the budget for cars. Because type 2 has a larger income-and budget elasticity , the average CO2intensity B goes down. By continuity, the result also holds for sufficiently small. Q.E.D.
The typical hypothesis asserts that demand for luxurious cars is more income-elastic (Mannering and Winston, 1985). Larger cars, which are also emission-intensive, tend to be more comfortable. For example because they offer more storage and lower occupant fatality rates in vehicle-to-vehicle crashes -attributes that are more easily dispensable than a car's basic transportation service. It thus seems plausible that demand for spacious cars will react more strongly to an equiproportional price increase. The proposition then predicts a decrease in the average pollution intensity if the uniform tax increases.
For high environmental taxes the effect may be reverted, as an increase in the uniform tax rate can then represent a fall in the relative price of less fuel-efficient cars. As we will see however, the relative importance of the environmental component in total car taxes is modest in European countries, so that the proposition's condition seems to apply. their CO2 sensitivity. Here, we define the vehicle registration tax as all one-off taxes paid at the time the vehicle is registered, which is usually the time of acquisition. For road taxes, we include all annual recurrent taxes of vehicle ownership. We construct these data for each country, year and fuel type in our sample using a detailed database with vehicle registration taxes and road taxes at vehicle-country-year level.

Data sources
Our first data source is a set of manufacturer price tables as supplied by the European  Ryan et al. (2009), and enables us to differentiate between average taxes and the CO2 sensitivity of vehicle-related taxes. We also take information on fuel taxes from the ACEA tax guides.
Because most cars are petrol or diesel, we restrict our sample to these two fuel types. Sales data play no role in the construction of the country tax data sets.
The next dataset, from Campestrini and Mock (2011), contains information on the CO2 intensity of the newly purchased diesel and petrol cars (CO2 emissions in g/km, weighted by sales, see also Figure 1) and the shares of diesel cars (See Figure 6 in the appendix Section 9.3). We have this information for the EU15 countries, from 2001-2010.
Lastly, data on nominal per capita GDP is taken from Eurostat (2014)

Constructing country average and CO2 sensitivity of car taxes
Countries have widely divergent rules for registration and road taxes. In some countries, vehicle registration taxes are based on CO2 emissions, in others, the cylindrical content is used to compute the tax, or the sales price of the car. In many instances, registration taxes combine multiple variables. Rules for annual road taxes vary even more across Europe.
Some countries base their annual tax on a car's engine power (in kW or hp), while other countries use cylinder capacity, CO2 emissions, weight and exhaust emissions. In addition to the dispersion between countries, for both registration and road taxes, many countries have changed their policies over the period 2001-2010; they adopted (temporary) discounts for fuel efficient cars, or additional charges for cars exceeding specified standards. We compare tax systems across countries by characterizing each country's tax system at year t by the two coefficients used in our model in Section 3. The first coefficient describes the country-year average tax, the second the CO2 sensitivity of the tax. Both variables are computed for both the registration and road tax, and for petrol and diesel. We thus construct 8 variables that characterize a country's vehicle tax system for a given year.
We now provide the details. Let CO2it be the CO2 intensity of car-type i in year t, the (registration or road) (percentage) tax in country c, and let δcit be the index {0,1} identifying whether the data are available for country c. For the sake of exposition, we do not use subscripts for fuel and tax type (registration versus road). We construct the country-specific average CO2 intensity and average tax rate (denoted by bars on top over the variables): 13 13 In the construction of our tax system variables we do not weigh by sales, to prevent our description of the tax system from being contaminated by the subsequent effects of that same tax system. The tax system may of course affect sales, and thereby the average CO2 intensity of new cars. This is discussed in Section 6.
That is, the typical car for a country has emissions 2 and pays a tax rate ̅ . We subsequently calculate the CO2-sensitivity of the tax by comparing how much taxes increase when CO2 emissions increase, on average, and weighted: Where weights are given by the deviation from the average CO2 intensity: The squared weights ensure that the denominator in (10) is strictly positive, and that the CO2 sensitivity is mainly determined by the tax-differences between the fuel-efficient and fuel-intensive cars.
Yet, if we want to determine a country's tax pressure and compare between countries, we should not consider the tax of the typical car for that country, but the tax for a typical car that is the same over all countries. Thus, we construct the (virtual) tax rate that would apply to a car with a CO2-emission profile 2 that is typical for the set of all countries: 1%. If two car types are completely identical (including prices at the factory gate), but one car is 10% more fuel efficient, then the consumer price of the more fuel-efficient car is 0.1* CO2TAX per cent below the consumer price of the more fuel-intensive car. All estimations in the main text are based on the double-log variables. We have reproduced our results for a linear model, which is presented in the appendix, Section 9.2. The appendix also provides the equations with more elaborate references to the details of taking logarithms. to 25% (see Figure 5, right panel), but at the same time reduced the average tax. All other things equal, in 2011, the after-tax price decreases by about 3% if a car is 10% more fuelefficient. The charts in Figure 4 also show that, in the Netherlands, taxes for diesel cars are persistently above those for petrol cars; 14 in our results section, we will come back to the effect of tax differentiation between petrol and diesel cars.
14 The Netherlands is atypical in the sense that registration taxes and fuel taxes are used as instruments to segregate the car market. Diesel fuel taxes are low (relative to petrol) while diesel registration taxes are high (relative to petrol). The tax scheme intends to separate longdistance drivers (who buy diesel cars) from short-distance drivers (who buy petrol cars).    Table 5 in the appendix, Section 9.2, for the tax levels and CO2 sensitivity based on the linear model.
Vehicle fiscal measures are correlated, also when we take out country and time fixed effects. Petrol and diesel registration taxes move in tandem, both for the levels and CO2sensitivity. The same applies to the annual taxes, where correlations exceed 80%. 17 Petrol and diesel fuel taxes are also positively correlated. The year fixed effects separate fuel price developments from fuel tax changes. There is almost no correlation between the three groups of tax instruments. For annual taxes, we see a very strong negative correlation between the level of annual taxes and its CO2 sensitivity, implying that the set of annual taxes are strongly multi-collinear, so that we must be careful when interpreting individual coefficients for annual taxes. 18 17 See Table 7 in the appendix for details 18 The negative correlation between the level of annual taxes and its CO2 sensitivity is 'natural' in the following sense. If the level of annual taxes increase, typically they increase less then

Econometric strategy
The benchmark model estimates the dependence of the CO2 intensity of the new car fleet in country c in year t (as in Figure 3), separately for diesel and petrol, 19 on the two dimensions of the registration car taxes: its level and its CO2 sensitivity where and are country and time dummies, and the country-time specific control variables Z include income, the share of diesel cars in total sales, and gasoline taxes. 20 For our preferred logarithmic model, we use logarithms for the dependent variable. In the linear model (see appendix, Section 9.2), the dependent variable is measured in average grams of CO2 emissions per km.
We add convergence patterns through the control variable, through where 2 is the CO2 intensity of the new fleet in the base year 2001. Convergence between countries is measured through a negative coefficient for the interaction term (16).
We assume there is no systematic correlation between observed fiscal vehicle policies and unobserved policies such as vehicle retirement plans that could induce omitted variable bias.
proportional with the car's size, weight and price. Thus, annual taxes have a tendency to be regressive. This is picked up by a negative coefficient for the CO2 sensitivity. 19 All variables are specific for diesel and petrol. We use both petrol and diesel independent variables when we estimate the diesel share as dependent on car taxes. 20 The fuel tax is estimated for each country-year-fuel type by fuel: tax=ln(1+{fuel tax level}/{fuel price}), where we take the fuel price as the average fuel price across the countries.
We estimate the model for both fuel types jointly and separately, with and without the annual taxes. One of the control variables, the share of diesel cars, is itself dependent on a country's tax regime. As discussed in the introduction, diesel cars are typically less CO2 intensive than petrol cars, and a high CO2 intensity of taxes may thus encourage individuals to switch to diesel cars. To assess whether this affects our results, we reestimate the model without the diesel share.

Results
The main results are presented in Tables 2 and 3.

Fuel-type specific effects
Starting with the CO2 intensity of new diesel cars, we find a clear significant effect of registration taxes on CO2 emissions (see Table 2). Especially the CO2 sensitivity is an effective instrument to change the characteristics of newly bought vehicles: a 1% increase in CO2 sensitivity reduces the CO2 intensity by 0.04 to 0.1 percent. The effect is weaker when we control for the diesel share, suggesting that part of the effect goes through the changes in the diesel share. As we see in Table 3, a higher CO2 sensitivity of diesel registration taxes increases the share of diesel cars. Buyers who decide to acquire a diesel car as a substitute for a petrol car typically buy diesel cars that are smaller compared to the average diesel car, while they substitute away from petrol cars that are larger compared to the average petrol car (see the case study of Ireland, Rogan et al., 2011, Hennessy and Tol, 2011, Leinert et al., 2013. These consumers who substitute diesel cars for petrol cars thereby reduce the average emissions from both diesel and petrol cars. The mechanism is confirmed by the negative coefficients for diesel share in the diesel emissions (Table 2, columns 1 and 3) and in the petrol emissions (Table 2, columns 5 and 7). Indeed, a closer look at our data (not shown here) shows that diesel cars are on average 20 percent heavier compared to petrol and the average weight for both diesel and petrol cars decreases with an increase in the diesel share. Part of the emission reduction of new cars in the EU has been achieved by lower registration taxes for diesel cars (Table 1 and Appendix, Table 5), which translated in an increased share of diesel cars (Table 3, column 5 and Appendix, Figure 6), which are typically more fuel efficient than petrol cars (Figure 1), and thus in turn decreases the CO2 intensity of the average car (Table 3, column 1).
The registration tax level reduces the CO2 intensity of new diesel cars, yet only significantly so when we account for the diesel share. In response to the observed decline of diesel registration tax levels during the sample period, consumers of small diesel cars switch to relatively less fuel-efficient, heavier diesel models, but there is also substitution from petrol cars to small diesel cars. We find no significant effect for road taxes on the emissions by diesel cars. Higher diesel fuel tax rates increase the fuel efficiency of newly acquired vehicles, as expected (Burke and Nishitateno, 2013). In addition, we find higher CO2 intensities with increasing income and a clear convergence pattern between EU countries.
We find a similar pattern for petrol vehicles. This paper is one of the first including annual road taxes, in addition to registration and fuel taxes, in the analysis of car purchase behaviour. We find that an increase in the annual road tax level and CO2 sensitivity increases the CO2 intensity of new petrol cars. We are not sure what causes this finding. It is not obvious that individuals account for future annual tax expenses, as discussed in Section 2. In our regressions, even though the annual tax rates enter significantly, excluding them from the regression has only little effect on the coefficient for the other variables. Hence, we can interpret the other coefficients with 21 The insignificance is not driven by lack of variation as petrol taxes show up significantly in Table 3.
confidence, and conclude that leaving annual taxes unaccounted for probably does not greatly alter our conclusions.

Aggregate effects
Then consider the overall effect of car taxes on the new fleet emission intensity (Table 3). The diesel share is the most important mediating variable: none of the policy variables is significant when this share is included in the regression, though the joint effect of more CO2-sensitive diesel and petrol taxes is significantly different from zero (columns 1 and 2). 22 The results in Table 3 allow us to assess the effect of the changes in registration taxes on the diesel intensity. We subtract the log of taxes in 2001 from those in 2010 (Table 1) and multiply the differences with the coefficients in Table 3 (column 5). We conclude that the changes in registration taxes have increased the diesel share by 6.5 percentage points. Using a similar calculation (but using column 2 from Table 3), we find that the changes in registration taxes have reduced the CO2 intensity of the average new car by 1.3% 23 . 0.9 percentage points of this overall effect is explained by changes in the diesel share. 24 The overall effects are modest; an explanation is that the large countries with a major domestic car industry (France, Germany, Italy, United Kingdom), have relatively low registration taxes that are almost independent of emission intensities.
When leaving out the diesel share, the diesel registration taxes stand out as the most important determinants. Lower overall taxes for diesel cars increase the share of diesel cars and thereby decrease average overall emissions, but also encourage existing diesel drivers to switch to more polluting models. At the same time, a more CO2 sensitive diesel registration tax favours small diesel cars, increasing the diesel share (column 5 and 22 As noted previously and presented in Table 7 in the appendix, policy measures for petrol and diesel vehicles are strongly correlated. This inflates the standard errors of the individual regressors in Table 3, columns 1-4. When testing, we find in column 1 and 2 that the sum of the diesel and petrol coefficients for the CO2 sensitivity of registration taxes is significantly different from zero.

Transmission mechanisms
Finally, we present a brief assessment of the transmission channels through which fiscal car taxes change emissions. We have already seen that consumers switch between petrol and diesel cars, in response to tax measures, but within a fuel type, they can also respond to tax measures by switching to lighter cars with less powerful engines, or alternatively, they can choose for cars with more fuel efficient engines while keeping the preferred car specifications unaffected (Fontanas and Zamaras, 2010).
In Table 4 we present, for diesel and petrol separately, the effect of fiscal measures on the CO2 intensity with and without additional controls for average vehicle mass and engine power. If controlling for mass or power reduces the (absolute) value of the policy coefficient, this can be taken as an indication that part of the policy's effect is transmitted through the car features. Next, we estimate the direct effect of fiscal policies on average vehicle mass and engine power. In all models, we control for convergence, income, time and country fixed effects.
To allow easy comparison, columns 1 and 5 in Table 4 reproduce Table 2 columns 1 and 5 respectively. Column 2 and 6 confirm that larger cars with more powerful engines have higher emission intensities. For diesel cars, registration taxes do not significantly affect average mass or engine power of newly purchased vehicles, although adding these features does slightly reduce the (absolute) coefficient on registration taxes in column 2 compared to column 1. A similar effect is found for the CO2 sensitivity of diesel registration taxes. One possible interpretation of this finding is that higher and more CO2-sensitive diesel registration taxes push the technology frontier for cars, providing the same qualities (mass and horsepower) to the consumers, at lower CO2 emissions. For petrol cars, the effects of registration taxes appear to be transmitted through the car features: higher (CO2 sensitivity of) registration taxes reduce the average mass and horse power of newly purchased vehicles, even when controlling for the diesel share. There is less indication of a technology effect, and more evidence of switch in the type of cars bought by consumers.
We note that the effects of income on CO2 intensities appear to be fully transmitted through car features, both for diesel and petrol cars. Such an outcome is intuitive, as increasing income will be used mainly to increase the level of desirable features. The effect of income on CO2 intensity does not become significantly negative when controlling for mass and horse power, so we find no evidence that consumers use income increases to purchase more environmentally friendly cars. For diesel cars, the effect of diesel fuel taxes is also fully transmitted through the car features.

Discussion
We find empirical evidence that fiscal vehicle policies significantly affect emission intensities of new bought cars. Increasing CO2-sensitivity of registration taxes and higher fuel taxes lead to the purchase of more fuel efficient cars, but higher annual road taxes have no or an adverse effect. The former is consistent with the literature; the latter is counter-intuitive, possibly because annual road taxes are not salient, but the high collinearity between annual road taxes may also play a role. We decomposed the vehicle registration tax rate into two variables, the level and CO2-sensitivity, and found that countries. The principle, sometimes referred to as a 'waterbed-effect', implies that environmental gains from fiscal national policies can leak away as the sale of more fuel-efficient cars in a country with a fiscal regime that puts a large premium on CO2 emissions, is countered by the sale of more fuel-intensive cars in other countries. National fiscal policies, aimed at the demand side, and in line with the third pillar of EU-policies, might thus be less effective conditional on the effectiveness of the first pillar of EU-policy, aimed at the supply of fuel efficient cars throughout the EU.Given an exogenously set ceiling for the EU-wide CO2 emissions, there is no clear economic gain from a diversified fiscal regime between EU countries, while there are social costs (Hoen and Geilenkirchen, 2006). Indeed, a few years ago, the EU proposed to harmonize vehicle taxes in the EU, but the proposal was rejected by the Member States. We also mention a few other potential disadvantages of fiscal support of fuel efficient cars.
In this paper, we focus on the average emission intensity of new cars. Reducing taxes for small, fuel-efficient cars can lead to scale effects (i.e. more cars) and intensity-of-use effects (i.e. more kilometres per car). Konishi and Zhao (2014) show that in a green tax reform in Japan, this scale effect offset the composition effect (i.e. a bigger share of fuel-efficient cars) by approximately two third. In addition, there is a rebound effect. Fuel-efficient cars are cheaper to drive, and a portion of the CO2 gains by CO2 -based vehicle purchase tax is lost as the fuel-efficient cars increase car travel demand (Khazoom, 1980). The existence of the effect is undisputed, but its magnitude remains an issue of debate (see e.g. Brookes, 2000, Binswanger, 2001, Sorrell and Dimitropoulos, 2008. Frondel and Vance (2014)  There are also secondary effects of fiscal policies. When consumers choose lighter cars that are more fuel efficient, not only CO2 emissions fall but emissions of NOx and PM10 as well. A weight reduction of 10% results in a decrease of the emission of NOx with 3-4% (Nijland et al., 2012). On the other hand, substituting diesel cars for petrol cars improves CO2 fuel efficiency by about 10-20%, yet increases the emissions of NOx (Nijland et al., 2012). In the case of PM10 the situation is not clear, as modern petrol cars with direct injection might emit more PM10 than modern diesel cars (Köhler, 2013). Lighter cars also reduce fatalities for drivers of other vehicles, pedestrians, bicyclists, and motorcyclists (Gayer 2004, White 2004. The design of the fiscal regime, encouraging lighter cars or encouraging diesel cars, can alter the secondary effects substantially. We used CO2 emission data according to the NEDC guidelines. It is known that the tests typically report lower emissions compared to realistic conditions, especially for cars that score very well at the tests (Ligterink andBos, 2010, Ligterink andEijk, 2014).
Moreover, the gap between test results and realistic estimates for normal use have where weights are given by the deviation from the average CO2 intensity (11): We then construct the (virtual) tax rate LOGTAXct that would apply to a car with a CO2emission profile that is typical for the aggregate of all countries (12) (13): The two constructed variables LOGTAXct and LOGCO2TAXct, are used as independent variables explaining the average emission intensity of the new car fleet (14). Note that the country-average CO2 intensity constructed in (8) or (17) is not the same variable used in the econometric regression, used as independent variable in Section 5 (14). The countryaverage CO2 intensity in (8) or (17) is measured only for those car types for which we have price and tax data, and its purpose is solely to construct the CO2 sensitivity of car taxes in (10) or (19). The country-average CO2 intensity used in Section 5 (14) is from an independent source, and is based on all car sales in a country-year; it is the independent variable that we explain using the country tax variables constructed in Section 4.2.

Linear model
In the main text, we characterized a country's tax system by two coefficients: the average rate, and its CO2 sensitivity, which is defined as elasticity of the tax rate with respect to CO2 emissions. In this appendix, we take a linear approach. Here, the CO2 sensitivity is instead defined as the increase in the tax level for a given increase in CO2 emissions (in grams per km). To decompose the tax in these elements, we estimate 2 2 2 where is the tax paid (in euro's) for vehicle i in country c at time t, is the tax exclusive purchase price, 2 the vehicle CO2 emission in g/km and 2 the average time t CO2 emissions in g/km. We then characterize a tax system by , which is the average tax rate as a percentage of the purchase price, and 2 which is the additional tax, in euro's, per g/km additional CO2 emissions. 26  With this decomposition, we consider the effect of the vehicle registration tax rate, and the CO2 sensitivity of the tax paid on the average CO2 intensity of newly purchased vehicles.
Results are presented in Tables 6 and 7. Since we now take the level of the additional tax on CO2 emissions, and the level of the average CO2 intensity of newly purchased vehicles interpretation is slightly different compared to Tables 2 and 3. Take for example the first column of Table 6. Here, a 10 percentage point increase in the vehicle registration tax rate is expected to reduce the CO2 intensity of diesel cars by 2.2 gCO2/km. Similarly, the (insignificant) coefficient of -0.005 on CO2TAX registration implies that a 10 euro increase in the effective registration tax rate on CO2 emissions for diesel cars, is expected to reduce the average CO2 intensity of diesel cars by 0.05 gCO2/km. In terms of sign and significance, results are in line with the logarithmic model in the main text.