Yellow Vests, Carbon Tax Aversion, and Biased Beliefs

Abstract

This paper helps to understand how beliefs form and determine attitudes towards policies. Using a new survey and official households’ survey data, we investigate the case of carbon taxation in France in the context of the Yellow Vests movement that started against it. We find that French people would largely reject a Tax & Dividend policy, i.e. a carbon tax whose revenues are redistributed uniformly to each adult. However, they also overestimate the negative impact of the scheme on their purchasing power, wrongly think it is regressive, and do not perceive it as environmentally effective. Using information about the scheme as instruments to robustly identify causal effects, our econometric analysis shows that if we could rectify these three biased beliefs, it would suffice to generate majority approval. Yet, only a small minority can be convinced by new information and revisions are biased towards pessimism. Finally, if overly pessimistic beliefs cause tax rejection, they also result from it through motivated reasoning, which manifests what we define as “tax aversion”.

Full text

Yellow Vests, Carbon Tax Aversion,
and Biased Beliefs∗
Thomas Douenne and Adrien Fabre†
Link to most recent version
Abstract
Using a representative survey, we find that after the Yellow Vests movement, French people would
largely reject a Tax & Dividend policy, i.e. a carbon tax whose revenues are redistributed uniformly
to each adult. However, they overestimate their net monetary loss, wrongly think the policy is
regressive, and do not perceive it as environmentally effective. We predict widespread approval if
these three biases could be corrected. Yet, only a small minority can be convinced. Indeed, if overly
pessimistic beliefs cause tax rejection, they also result from it through motivated reasoning, which
manifests what we define as “tax aversion”.
JEL classification: D72; D91; H23; H31; Q58
Keywords: Climate Policy; Carbon tax; Bias; Beliefs; Preferences; Tax aversion
∗Acknowledgments: We are grateful to Mouez Fodha, Fanny Henriet and Katheline Schubert for their comments and their help
to get funding. We also thank Stefano Carattini, Linus Mattauch, Joseph Stiglitz, Thierry Verdier, as well as seminar participants
at the Paris School of Economics, Columbia University, University of Pennsylvenia, University of Amsterdam, University of Quebec
at Montréal, KU Leuven, Erasmus University Rotterdam, Environmental Defense Fund, EIEE-CMCC (Milan), OECD, OFCE,
and conference participants at FSR Climate Annual Conference (Florence), ADRES (Lyon), FAERE (Rennes). We are thankful to
Christina Hobbs for the proof-reading. We acknowledge financial support from the Cepremap, EUR PGSE (ANR-17-EURE-0001)
ANR (ANR16-CE03-0011), and Université Paris 1 Panthéon-Sorbonne economics doctoral school (ED 465).
†Douenne: Paris School of Economics, Université Paris 1 Panthéon-Sorbonne, 48 Boulevard Jourdan, 75014, Paris, France
(email: thomas.douenne@psemail.eu); Fabre: Paris School of Economics, Université Paris 1 Panthéon-Sorbonne, 48 Boulevard
Jourdan, 75014, Paris, France (email: adrien.fabre@psemail.eu)
1

1 Introduction
The French government had initially committed to an ambitious trajectory for the price of carbon.1
Initiated in 2014 at 7e/tCO2, the French carbon tax reached 44.6e/tCO2in 2018 and was supposed
to continue growing to hit 86.2e/tCO2by 2022. Yet, at the end of 2018, the same government that
had accelerated the price trajectory decided to abandon it and froze the tax at its current level for
an undetermined period. This turnaround in French climate policy is the direct consequence of the
popular protest of the “Yellow Vests”, which started against the carbon tax.2Among several factors,
the negative impact of the tax on households’ purchasing power has certainly been a key driver
of public’s discontent. The increasing revenues from the carbon tax were mostly used to fund the
budget rather than redistributed to households, raising concerns over the distributive effects of the
policy. In order to tackle the negative impact of carbon taxation on households’ purchasing power,
economists have proposed a mechanism known as “Tax & Dividend”, i.e. a carbon tax whose revenue
is redistributed uniformly to each adult. This strategy has recently been supported by 3,354 American
economists in The Wall Street Journal, “To maximize the fairness and political viability of a rising
carbon tax”. Implicitly, it is therefore assumed that with a design that ensures that the properties of
the tax are aligned with people’s preferences one should be able to generate support for it. But is
it really sufficient? In this paper, we argue that to understand the link between the properties of a
policy and its support, one has to account for a critical ingredient: beliefs.
The objective of this paper is to understand how beliefs about a policy form and then determine
attitudes towards it. The recent events undoubtedly make the French carbon tax an interesting case
study. In order to explain French attitudes towards carbon taxation, we conducted a survey on a
representative sample of 3,002 French households. We focus on a “Tax & Dividend” carbon tax with
uniform lump sum compensation, which allows one to specify clearly the distributive effects of the
policy, in contrast to the policy abandoned by the government. The reform is approved by only 10%
of respondents and disapproved by 70% (the rest do not know or do not want to answer). We analyze
the perceptions of three well-known determinants of acceptance of the carbon tax: the impact on
one’s purchasing power, the progressivity of the scheme, and its environmental effectiveness. We
1More precisely, the “Contribution Climat-Énergie” is a sectoral carbon tax specific to fossil fuels.
2Following a massive petition against rising gasoline prices in November 2018, hundreds of thousands of people
started protesting. They would wear their recognizable fluorescent clothing and gather on roundabouts and tolls every
day, and demonstrate in Paris each Saturday. The Yellow Vests express a general concern for their purchasing power
as well as discontent for French elites and institutions.
2

compare subjective beliefs elicited from our new survey to objective impacts on people’s purchasing
power computed using official households’ survey data. This comparison shows that people largely
overestimate the negative impact of carbon taxation on their purchasing power. For instance, while
70% of households are expected to win from this policy, only 14% think they would. Similarly,
while the scheme proposed in our survey is progressive, a large majority of individuals perceive it
as regressive. In addition, a majority of respondents do not believe that such a policy would reduce
pollution and fight climate change. Using information reported over their energy equipment and
usage, we are able to compute a respondent-specific estimation of the tax incidence. This estimation
enables us to look at the heterogeneity in biases. We find that the people most opposed to the policy,
and in particular those supportive of the Yellow Vests, are the most biased, i.e. the most inclined to
over-estimate their losses. Thus, one may wonder whether biased beliefs lead to policy rejection or if
the causality goes in the other direction.
To disentangle the effect of initial biases on attitudes towards the policy from the reversed effect
of attitudes on perceptions, we investigate the effect of providing new information to respondents
through random treatments. Respondents randomly receive (or not) a piece of information about the
progressivity and/or about the effectiveness of the policy, as well as the customized information —
derived from our respondent-specific estimation — that their household is expected to win or lose from
the policy. We also specify that this latter information is correct in five cases out of six, a probability
that we carefully estimated out-of-sample. A first observation is that our treatments fail to rectify
pessimistic beliefs. For example, among those advantaged by the reform who wrongly believe they
would lose, only 12% are convinced that they would gain when we disclose our estimation to them.
Worse, respondents revise their beliefs in a biased way, giving more weight to new information when
it shows they would lose from the reform, i.e. when it provides them with arguments against the
tax. We also find robust evidence of motivated reasoning in the formation of beliefs, as those who
already approved of the reform are less prone to biases in the revision, while those most opposed to it
such as supporters of the Yellow Vests tend to discard new information unless it goes against the tax.
Moreover, we find that motivated reasoning is accentuated among highly educated people, suggesting
that it stems from an adaptive advantage rather than a cognitive deficiency. The information displayed
being random we can then use it as instruments to measure the causal effect of holding certain beliefs
on acceptance. In the case of self-interest (taken as one’s beliefs about winning or losing purchasing
power from the policy), we supplement these treatments by testing the approval for a different Tax
& Dividend whose compensation is targeted to people with incomes below a threshold that varies
3

between respondents to create exogenous variations in eligibility. The method we use in this case is
noteworthy, as it embeds a regression discontinuity design inside an instrumental variables regression
in order to estimate the causal effect of eligibility at different levels of income. Overall, we find
that beliefs in self-interest, environmental effectiveness, and progressivity all have a large effect on
acceptance: about 40 percentage points (p.p.) for the first two and 27 p.p. for the latter, all else
equal. These motives also appear more as complements than substitutes: in a hypothetical scenario
where all biased beliefs could be corrected, we estimate an approval rate as high as 90%. This result
suggests that rejection of carbon taxation does not commonly result from clashing principles, such
as a disinterest in climate or a dislike of price instruments, but rather from biased beliefs about the
properties of the reform. Since beliefs are formed endogenously in a motivated way, people’s biases
gain inertia, so that new information might only push their attitude in one direction.3
The contribution of this paper is two-fold. First, it contributes to a recent literature that has
emerged to understand the political economy of climate policies, as this issue is becoming critical in
the public debate. For a thorough review of this literature, we refer the reader to Carattini et al.
(2018), and also suggest the more synthetic Klenert et al. (2018), as well as Millner & Ollivier (2016)
for a review of the political obstacles to environmental policies. Stern et al. (1993) is an early work
proposing and testing a model of attitudes for environmental quality aimed at disentangling egoistic
from altruistic motives on the one hand, and beliefs from values on the other hand. Among all
possible attitudes, they show that beliefs about consequences on self-interest are the only predictor
of the willingness to pay Pigouvian taxes. Using a post-electoral survey in Switzerland, Thalmann
(2004) also finds a correlation between carbon tax acceptance and self-interest, proxied by the number
of cars owned. In surveys on British and Swedish respondents respectively, Bristow et al. (2010) and
Brannlund & Persson (2012) document a higher approval rate when the reform addresses distributional
issues. Baranzini & Carattini (2017) report that a majority of the people they interviewed in Geneva
do not believe the tax would be effective, which confirms what Dresner et al. (2006b) find with focus
groups in the UK. Surveying Norwegian people, Kallbekken & Sælen (2011) show that self-interest
matters for acceptance, but less than concerns for environmental effectiveness or distributional effects.
3The “campaign effect” documented by Anderson et al. (2019) (in the case of referenda in the US state of Washing-
ton) is an example of how support for a carbon tax can decrease substantially after it enters the public debate. It may
explain why acceptance of an increase in the carbon tax plummeted with the Yellow Vests movement, down from a level
of 48% (ADEME,2018) in the middle range of other countries’ (Brechin,2010). This effect confirms that the French
carbon tax may be an insightful case study to understand what could happen in other countries when a controversial
policy is publicly debated.
4

On US data, Anderson et al. (2019) argue that ideology explains most of the support for carbon
taxation, and suggest that this effect would dominate that of self-interest.
In the present paper, we also study how acceptance depends on these three motives (i.e. self-
interest, perceived environmental effectiveness and progressivity). We contribute to the literature by
providing robust evidence for causal effects where past studies essentially show correlations, often
relying on proxies such as fuel consumption to proxy self-interest (e.g. Thalmann,2004;Kallbekken
& Sælen,2011;Anderson et al.,2019). In contrast to such shortcuts, we do not assume that people
are fully rational nor have perfect information. Thus, our methodology offers a novel look at the
political economy of climate policies, as it allows one to disentangle erroneous beliefs from pure
effects of preferences.4The paper also quantifies biases regarding the costs of the carbon tax. To
our knowledge, it is the first study that compares subjective beliefs and objective data about the
private costs that arise from carbon taxation. Given the intense public debate over the incidence of
such a policy, identifying and measuring the discrepancy between actual impacts and their subjective
perception is critical.
Beyond the case of carbon pricing, our paper contributes to the literature on biased beliefs and
how they relate to attitudes over policies. Recent research has shown how biased beliefs on inequality
and social mobility affect people’s attitudes regarding distributive policies (e.g. Cruces et al.,2013;
Kuziemko et al.,2015;Alesina et al.,2018). Our paper adds to this literature by investigating the
relationship between beliefs and attitudes on climate policies. It also goes further than previous studies
by identifying a bi-directional relationship as we show that not only do beliefs determine attitudes,
but attitudes over policies in turn shape beliefs. Indeed, using a representative survey, our paper
brings evidence consistent with theories of motivated reasoning (Kunda (1990), see Bénabou & Tirole
(2016) for a recent review) that have so far been mostly tested in the lab (e.g. Redlawsk,2002;Thaler,
2019). In particular, our results support the recent theory of Little (2019) who formalizes motivated
reasoning as a way to reconcile an auxiliary belief (one’s self-interest in the reform) to a core belief
(here, the policy rejection). We believe our results apply beyond the case of carbon taxation, and
illustrate more generally the determinants and consequences of tax aversion. Indeed, the few previous
definitions of tax aversion (Sussman & Olivola,2011) are hardly exploitable empirically, as they do
not relate the concept to an observable phenomenon. This may contribute to the limited number of
4We take preferences over policies as the mapping from beliefs (on facts) to attitudes (on policies), i.e. how attitudes
are determined as a function of beliefs. Conversely, motivated reasoning represents the feedback loop from attitudes to
beliefs.
5

papers on this topic (Kallbekken et al.,2011;Kessler & Norton,2016). Building upon our results, we
can define tax aversion as a gut rejection of a tax (or taxation in general) that influences beliefs about
the tax properties such as its effectiveness, fairness, or sameness with an equivalent measure labeled
differently. Our work then shows that tax aversion can be identified through motivated reasoning, by
observing that the initial tax rejection impacts how one integrates new information into one’s beliefs.
The rest of the paper is organized as follows. In section 2, we describe our survey and other data
sources. In section 3, we compare subjective perceptions to objective data, and measure the bias
regarding the impacts of carbon taxation. In section 4, we highlight new biases in the way individuals
revise these beliefs. In section 5, we estimate the effects on acceptance of rectifying biased beliefs.
Section 6concludes. Further results and methodological complements are reported in the Appendix,
such as an estimation of the willingness to pay for the carbon tax in Appendix G.
2 Survey and data
2.1 Our survey, “Beliefs & climate policies”
2.1.1 Survey data collection
The 3,002 responses were collected in February and March 2019 through the survey company Bilendi.
This company maintains a panel of French respondents whom they can email with survey links.
Respondents are paid 3eif they fully complete the survey. The respondents who choose to respond
are first channeled through some screening questions that ensure that the final sample is representative
along six socio-demographic characteristics: gender, age (5 brackets), education (4), socio-professional
category (8), size of town (5) and region (9). The quotas are relaxed by 5% to 10% relative to
actual proportions. Table A.1 in Appendix Ashows that our sample is still extremely representative.
Nonetheless, observations are weighted to correct for small differences between sample and population
proportions. The median time for completion of the survey was 19 minutes. We made sure that all
questions requiring some concentration were in the first half of the survey.
2.1.2 The survey
The full survey in French can be seen online, and the translated questionnaire is detailed in Appendix
B.5It contains several random branches and treatments that are independent from one another:
5preferences-pol.fr/doc_q.php#_e
6

Figure 2.1 presents in a diagram the sequence of treatments (represented by ellipses) and questions
(boxes).
Figure 2.1: Diagram of the sequence of treatments (ellipses) and questions (boxes).
Priming on environmental issues Two blocks of information are randomly displayed or not:
one on climate change and the other on particulate matter. This priming divides the sample in
four groups, who receive either one block of information, the other, none, or both of them. Climate
change information includes temperature trends for the long run future, concerning facts on current
and expected impacts, and a claim that keeping global warming below 2 ℃is technically feasible.
Particulates information consists of the estimated impact on French mortality (48,000 deaths per
year), life expectancy (9 months less), and the assertion that reducing fuel consumption would improve
health. The time spent on each block is saved, and links to scientific references are provided to support
the information.
Household characteristics In addition to the quotas strata, socio-demographic characteristics
include zip code, household structure, income of the respondent and of their household. A block
on energy usage contains questions that allow us to estimate the impact of a carbon tax increase
on housing expenditures: size of accommodation, heating type (collective or individual) and energy
source; as well as on transport expenditures: number of vehicles, type(s) of fuel, distance travelled last
year, and average fuel economy. The distributions of answers are much in-line with official statistics,
7

as shown in Table A.2 in Appendix A.
Partial tax reform One partial tax reform is randomly allocated to the respondent: it consists
of an increase in the carbon tax by 50e/tCO2specific either to heating fuel and gas, or to gasoline
and diesel. Partial reforms on housing and transport feature the same string of questions. We first
ask whether the respondent thinks their household would lose more or less than average in case of
an unspecified increase in indirect taxation, both for the VAT tax and for the energy considered
(housing or transport). Then, we specify in a new block that the revenues of the tax increase would
be distributed equally to each adult, entailing a yearly transfer of 50e(resp. 60e) for the partial
tax on housing (resp. transport).6We also provide the price increases implied by the tax: +13%
(resp. +15%) for gas (resp. domestic fuel) on the one hand, and +0.11e(resp. +0.13e) for a liter of
gasoline (resp. diesel) on the other. We make clear at the beginning that this survey is conducted by
two researchers in social sciences, but we present the policy starting with “The government studies...”
to capture the effect of distrust in government that could arise in the actual political process. Then,
we ask the respondent whether their household would win, lose, or be unaffected by the reform.
Depending on their answer, we further ask them to estimate their expected gain (or loss) among 5
(or 6) intervals. The interval thresholds are tailored to each respondent, as they are computed in
proportion of the number of consumption units of their household (as defined by Eurostat). The
questions were not incentivized. Indeed, Sapienza & Zingales (2013) show that people think that
economic experts are too optimistic regarding the carbon tax, so incentivizing the answers could
have led respondents to misreport their true beliefs and shift them towards what they think the
researchers expect. Finally, respondents are asked to estimate their own elasticity as well as that of
French people. To this end, we borrow the phrasing of Baranzini & Carattini (2017), and ask for
the expected decrease in consumption that would follow a 30% increase in the price of heating (or
equivalently, an increase of 0.50e/L in fuel prices), among 5 brackets.
6We chose to redistribute per adult instead of per consumption unit to make the scheme more understandable. We
limited the number of beneficiaries to two per household to better align with current welfare benefits that depend on
the number of consumption units.
8

Tax & Dividend
Initial perceptions Our main reform of interest is an increase by 50e/tCO2of the French
carbon tax, that concerns both housing and transport.7The revenues generated are redistributed
equally, so that each adult receives a yearly lump sum compensation of 110e(this figure was obtained
using typical elasticities, see 2.2.1). After describing the reform, a first block of questions elicits the
respondent’s perceptions. Their subjective gain is asked in the same manner as for the partial tax.
The priming that “scientists agree that a carbon tax would be effective in reducing pollution” is
randomly displayed or not before asking whether the reform would be effective in reducing pollution
and fighting climate change. Then, respondents are asked to pick the categories of losers and winners
with the reform. Finally, we ask: “Would you approve this reform?” and let the respondent choose
between “Yes”, “No” and “PNR (I don’t know, I don’t want to answer)”. In the following, we say that
a respondent approves a reform if they respond “Yes”, and that they accept the reform if they do
not respond “No”. Given the low rates of approval, we study primarily the acceptance to get tighter
confidence intervals. We also provide results for approval, and obtain similar findings.
Opinion after knowledge To assess how beliefs are formed and measure the importance of
self-interest and fairness motives in the acceptance of the reform, we provide some information on
the effect of the reform. To a random half of the sample, we explain that “this reform would increase
the purchasing power of the poorest households and decrease that of the richest, who consume more
energy”. To two-thirds of the respondents (the remaining half plus one-third of the respondents with
that priming on progressivity), we provide customized information explaining that: “In five cases out
of six, a household with your characteristics would [win/lose] through the reform. (The characteristics
taken into account are: heating using [energy source] for an accommodation of [surface] m2; [distance]
km travelled with an average consumption of [fuel economy] L for 100 km.)”. Indeed, section 2.2.2
shows that we estimate correctly if a household wins or lose in 83% of cases.
As most respondents do not believe the claim that the reform is progressive, we ask the second
half of the sample whether they think so, to analyze the effects of priming on compliant respondents.
We also ask again for the winning category (i.e. if the respondent’s household would win, lose or
be unaffected by the reform) and for the approval to the reform. Finally, we let the respondent
choose the reasons why this reform seems beneficial, and undesirable. We report these results in a
7Electricity is exempt from the French carbon tax “Contribution Climat Energie” (CCE) as it is already taxed by
the EU-ETS, and has a low carbon content due to the large share of nuclear power.
9

companion paper, Douenne & Fabre (2020), where we also study the socio-demographic determinants
of approval.
Tax & Targeted Dividend In this block, we ask for the winning category and for the approval of
four alternative reforms. Each respondent is given just one of them, which differs from the main reform
only in the way revenues are recycled. Here, the payments, still equal among recipients, are targeted
to adults whose income is below some threshold. The four thresholds correspond to the bottom 20%,
30%, 40%, and 50% of the income distribution. They are computed using inflated deciles of individual
income from the Enquête sur les Revenus Socio-Fiscaux (ERFS 2014) produced by Insee (the French
national statistics bureau).8Respondents whose income lies between two thresholds are allocated
randomly to a reform with one of them. When the income is close to only one threshold (i.e. when
its percentile in the distribution is below 20 or within [50; 70]), the allocated reform corresponds to
that one. When the respondent’s income is above 2220e/month (which is the 70th. percentile), the
reform they face is determined by the income of the household’s second adult. Finally, when both
(or the only one) adults in the household earn more than 2220e/month, their reform is allocated
randomly between the four variants. Table 2.1 describes the targeted reforms and the proportion of
respondents allocated to each of them, along with the proportion one would expect from the ERFS.
The two sets of figures match almost perfectly, indicating that our sample is representative along the
income dimension.
Table 2.1: Characteristic of the targeted reform by target of the payment.
Targeted percentiles ≤20 ≤30 ≤40 ≤50
Income threshold (e/month) 780 1140 1430 1670
Payment to recipients (e/year) 550 360 270 220
Proportion of respondents .356 .152 .163 .329
Expected proportion of respondents .349 .156 .156 .339
Other questions We do not detail the other questions of the survey, because we analyze them in
a companion paper, Douenne & Fabre (2020). In that paper, we examine opinions on environmental
policies, including other ways to recycle the revenues of a carbon tax. We measure the knowledge
and perceptions of climate change; ask some specific questions on shale gas, the influence of climate
change on the choice to give birth, and one’s willingness to change their lifestyle. We study the use,
8Incomes entitled to the household rather to its members, such as certain welfare benefits, are divided equally
among the two oldest adults of the household.
10

availability, and satisfaction with public transportation and active mobility. We also ask for political
preferences, including position in relation to the Yellow Vests. Finally, we let the respondent express
any comment in a text box.
2.1.3 Ensuring data quality
We took several steps to ensure the best possible data quality. We excluded the 4% of respondents
who spent less than 7 minutes on the full survey. We confirm that response time is not significantly
correlated with our variables of interest (such as approval or subjective gain). In order to screen
out inattentive respondents, a test of quality of the responses was inserted, which asked to select
“A little” on a Likert scale. The 9% of respondents who failed the test were also excluded, which
yields a final sample of 3002 respondents. Also, when the questions about a reform were spread over
different pages, we recalled the details of the reform on each new page. We checked for careless or
strange answers on numerical questions, such as income or the size of the household. We flagged 10
respondents with aberrant answers to the size of the household (and capped it to 12) and up to 273
respondents with inconsistent answers, such as a household income smaller than individual income,
or a fuel economy higher than 90 liters per 100 km. An examination of these answers shows no
significant correlation with our variables of interest, and suggests that these respondents have simply
mistaken the question. Among these inconsistent answers, 58 respondents have answered more than
10,000eas their monthly income (despite the word “monthly” being in bold and underlined), with
answers in the typical range of French annual incomes. We have divided these figures by 12.
2.1.4 Notations
To improve the understanding of our specifications in the regression Tables, we adopt consistent
notations throughout the paper. For questions where possible answers are “Yes”/“No”/“PNR”, we
define two kinds of dummy variables: the default ones correspond to not “No” answers, while we
put a dot on dummy variables for “Yes”. For example, acceptance is denoted Awhile approval is
denoted ˙
A. Furthermore, for questions that are asked several times, namely acceptance and winning
category, an exponent is added to specify the step at which the question is asked. Table 2.2 describes
these exponents as well as the notations corresponding to the different notions of gain that we use.
Uppercase is used for binary and lowercase for continuous variables, Greek letters denote objective
notions, with a hat for our estimation of gains and without for the true (unknown) ones. To give
another example, the broad notion of self-interest at the initial step, i.e. the belief that one does not
11

lose, is denoted GI, and the strict belief that one wins at Tax & Targeted dividend is denoted ˙
GT.
Table 2.2: Notations for the different reforms and for gain notions.
Step Initial After knowledge with Targeting
Variants −Progressivity Feedback −
Exponent I P F T
Gain Subjective True Estimated
Numeric g γ bγ
Binary ˙
G(g > 0), G(g≥0)Γb
Γ
2.2 Official households surveys
2.2.1 Eliciting objective aggregates and distributions
One of the goals of the paper is to compare respondents’ perceptions with the actual impact of a carbon
tax on households’ purchasing power. For this purpose, we use the database constructed by Douenne
(2020) whose objective was to estimate the distributive effects of a carbon tax for French households.
This database matches two households surveys produced by Insee: the consumer survey Budget de
Famille (BdF 2011) and the transport survey Enquête Nationale Transports et Déplacements (ENTD
2008).
Consumer survey “Budget de Famille” The consumer survey Budget de Famille (BdF 2011)
is a household survey providing information over all households’ revenues and expenditures, together
with many socio-demographic characteristics. It was conducted in several waves from October 2010
to September 2011, over a representative sample of 10,342 French households. The main advantage
of BdF when studying the incidence of carbon taxation is that expenditures in both housing and
transportation energies are reported. Consumption of housing energies is taken from households’ bills,
and for most other goods respondents report their expenditures over the past week. As explained
in Douenne (2020), this data collection is problematic when looking at the incidence of a tax on
transportation energies, as short-run fluctuations in consumption lead to overestimate the variation
in expenditures.
12

Transport survey “Enquête Nationale Transports et Déplacements” To overcome this lim-
itation, BdF is matched with the transport survey Enquête Nationale Transports et Déplacements
(ENTD 2008). ENTD was conducted in several waves from April 2007 to April 2008, over a repre-
sentative sample of 20,178 French households. It provides information on households characteristics,
their vehicle fleet and use over the past week, but most importantly it gives information on annual
distances travelled with these vehicles. This last information enables us to recover the distribution of
transport fuel expenditures without over-estimating its spread. Such matching is not necessary for
housing energies as these already represent consumption over long periods in BdF. Finally, data from
National Accounts is used to make the data representative of the year 2017.
Computing tax incidence and revenues From this dataset, we are able to compute the dis-
tribution of households’ objective net gains in purchasing power after the policies proposed, i.e. the
dividend received minus the increase in fossil fuels expenditures. We also use it to determine the total
tax revenue to be redistributed lump sum. The formulas used for these computations are given in
Appendix C.1. Our computations use typical elasticities found in the literature on French households:
−0.4for transport and −0.2for housing, as well as an incidence borne at 80% by consumers.9
2.2.2 Computing households’ expected net gains
Simulating expected gains for the feedback Households are asked about yearly distance trav-
elled and average fuel consumption of their private vehicles. From this information, it is then possible
to compute the expected cost of a carbon tax on transport fuels. For housing energies, the collected
information does not enable such simple calculations. Instead, we use the housing survey Enquête
Logement (EL 2013) that provides information on household expenditures in housing energies as well
as many demographic and energy characteristics. The survey was conducted between June 2013 and
June 2014 over a sample of 27,137 households in metropolitan France. This enables us to compute the
expected cost from the carbon tax on housing energies, and regress it on household characteristics.
The results are provided in Appendix C.2, where they are shown to be as least as accurate as the ones
obtained from alternative prediction methods and specifications. The distribution of housing energy
expenditures is highly comparable to that of BdF. Adding this estimated cost to the one simulated
for transportation energies, subtracting the amount received in lump sum transfers, and dividing by
9These values correspond to the short run uncompensated price elasticities estimated by Douenne (2020), and are
in line with previous findings on French households (e.g. Clerc & Marcus,2009;Bureau,2011).
13

the number of consumption units, we estimate the impact of the policy on the purchasing power of
each household.
Assessing feedback’s accuracy The previous estimation could have also been conducted with
BdF data. Still, running this estimation on the housing survey enables us to test the accuracy of our
prediction out-of-sample: for each household in BdF, we compute the impact of the policy on their
purchasing power and compare it to the prediction from our simulations. Figure D.1 in Appendix
D.2 shows how the probability that our prediction is correct depends on objective gains. For five
households out of six, we correctly predict whether their purchasing power would increase or decrease
through the policy. We make this ratio symmetrical: among households in BdF predicted to win,
83.4% were actual winners, while among those predicted to lose, 83.4% were actual losers. Since
the simulation and its test are done on different households, there is no concern for over-fitting:
the probability that our feedback is correct is not overestimated. Assuming that the characteristics
reported by our respondents are correct, there is no reason to believe that the probability of error is
higher or lower when simulations are applied to our survey respondents.
3 Biased beliefs
3.1 Self-interest
Over-estimation of policy costs Figure 3.1 plots the kernel density of expected net gains for
objective data from BdF, and subjective beliefs from our survey. Figure 3.2 compares the CDF of
objective vs. subjective net gains.10 It is evident from these figures that on average, respondents
overestimate the cost of the policy, even in the extreme case of perfectly inelastic expenditures. This
result holds both for the carbon tax and for partial carbon taxes on transport and housing energies.
The average net gains from the carbon tax on transport, housing, and both, are respectively 18eper
consumption unit (c.u.), 6eper c.u., and 24eper c.u. from BdF data. Extrapolating from our survey,
we instead find average subjective net gains of respectively −61e,−43e, and −89e. The median gap
of 116ebetween objective and subjective gains indicates a substantial bias towards loss from typical
10The subjective intervals are translated into numerical values, assuming that the distribution within each interval is
the same as that of BdF data. Within each bin, we draw values that match the actual distribution for the PDF, while
we simply take the actual average for the CDF. Among the several methods that we tried to assign numerical values,
all realistic ones yield identical results, and we find an overestimation of policy costs even in the most conservative one
(taking the maximal bounds of intervals).
14

respondents. This bias is widespread, as we find that 89% of respondents underestimate their gain of
purchasing power using our household-specific estimation. This proportion remains as high as 77 %
when assuming inelastic expenditures, which provides a lower bound on the share who underestimate
their net gain in utility. In other words, while 70% of households should benefit (in monetary terms)
from the compensated carbon tax, only 14% think they would (and 22% see themselves unaffected).11
(a) Transport fuels (b) Housing energies (c) Both
Figure 3.1: Distribution of objective (dark blue) vs. subjective (orange) net gains from our Tax &
Dividend.
(a) Transport fuels (b) Housing energies (c) Both
Note: Dashed blue lines represent distributions of objective gains in the extreme case of totally inelastic
expenditures. Vertical dotted orange lines show the limits of intervals answers of subjective gains.
Figure 3.2: CDF of objective (dark blue) vs. subjective (orange) net gains from our Tax & Dividend.
Heterogeneity in bias In order to characterize profiles of individuals more likely to mis-perceive
their gains, we regress mis-perception over many respondents’ characteristics. Mis-perception is
11For transport and housing energy taxes, the objective proportions of winners are very similar at respectively 74%
and 67%, while the subjective shares are 16% and 17% (with 22% and 30% of unaffected).
15

Table 3.1: Determinants of bias in subjective gains
Large bias (|bγ−g|>110)
OLS logit OLS
Initial tax: PNR (I don’t know) −0.179∗∗∗
(0.023)
Initial tax: Approves −0.284∗∗∗
(0.031)
Sex: Female 0.036∗0.030 0.042∗∗
(0.020) (0.020) (0.019)
Ecologist −0.064∗∗ −0.061∗∗ −0.025
(0.026) (0.026) (0.026)
Yellow Vests: PNR 0.039 0.035 0.024
(0.036) (0.035) (0.036)
Yellow Vests: understands 0.081∗∗∗ 0.062∗∗∗ 0.041∗
(0.025) (0.024) (0.025)
Yellow Vests: supports 0.108∗∗∗ 0.103∗∗∗ 0.051∗
(0.026) (0.025) (0.026)
Yellow Vests: is part of 0.202∗∗∗ 0.193∗∗∗ 0.147∗∗∗
(0.048) (0.040) (0.047)
Controls: Socio-demo, political leaning X X X
Observations 3,002 3,002 3,002
R20.061 0.098
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Note: Standard errors are reported in parentheses. For logit, average marginal effects are reported and not
coefficients. The list of controls can be found in Appendix H.
defined as a gap between objectively estimated and subjective net gains beyond 110eper c.u., because
our estimation differs from true objective gain by more than 110ein only 5% of cases. This definition
ensures that the 55% of respondents with a mis-perception have in fact a large bias. Other definitions
for the bias yield very similar results. The results given in Table 3.1 show that mis-perception is
largely idiosyncratic: controlling for a large set of variables12 (column 1), the R2remains small
(0.06). Still, we identify several variables having a significant effect on mis-perception. In particular,
females appear to mis-perceive their gain 4 percentage points (p.p.) more on average, all else equal.
Ecologists on the other hand are about 6 p.p. less likely to display a large bias. Interestingly, while
12The control variables used throughout the paper are described in Appendix H.
16

the standard left/right political leaning has no significant effect, the position towards the Yellow Vests
appears to be the most critical determinant in mis-perception. Relative to respondents who declared
to be opposed to the movement, those who declared to “understand”, “support”, or “be part” of it are
more likely to mis-perceive their gains. This effect is increasing with the degree of adhesion, up to 20
p.p. for individuals who declared to be part of the movement. Column (3) additionally includes one’s
position towards the policy as a covariate: we see that people who approve the policy are 28 p.p. less
likely to mis-perceive their gains relative to those who do not accept it, and 10 p.p. less likely relative
to those who do not know. We can think that the degree of support of the policy is what determines
most of the bias (explaining e.g. why ecologist loses its explanatory power when we control for the
support), and that the Yellow Vests variables remain significant only because they capture different
degrees of rejection of the tax (which our Yes/No question cannot do).
Overall, typical biases are large and closely related to one’s convictions. However, the direction(s)
of causality between beliefs and rejection is not resolved at this stage. Section 4provides robust
evidence that some people think they lose because they oppose the tax, while section 5shows that
perceived outcomes causally influence approval.
3.2 Environmental effectiveness
A well established result in the literature on the acceptability of climate policies is the perceived
ineffectiveness of Pigouvian instruments (e.g. Dresner et al.,2006a;Kallbekken et al.,2011;Baranzini
& Carattini,2017). In particular, people do not see carbon taxes as effective to fight climate change.
Our findings confirm this result: among our survey respondents, only 17% answered “Yes” when asked
whether our Tax & Dividend would be effective in reducing pollution and fighting climate change,
66% answered “No”, 18% that they did not know.
An explanation sometimes encountered to explain perceptions of ineffectiveness is that most peo-
ple believe that energy consumption is quite inelastic (Kallbekken & Sælen,2011;Carattini et al.,
2018). To test this hypothesis, we regress a binary variable Eequal to 0 if the respondent does
not perceive the policy as environmentally effective and 1 otherwise, on their subjective price elas-
ticity for French people. As respondents were randomly assigned to transport or housing, we run a
separate regression for both types of energies. Table D.1 in Appendix D.3 reports results with and
without control variables. They all consistently indicate that a respondent anticipating an elasticity
of −1 is (on average) 6 p.p. more likely to perceive the policy as effective than one anticipating no
elasticity. Although significant, the magnitude of the effect is modest, showing that the perceived in-
17

effectiveness of tax instruments should not be reduced to small subjective elasticities. Indeed, among
respondents who perceive the policy as environmentally ineffective, almost half anticipate responses
to price changes larger than the literature.13
A more plausible explanation for perceived ineffectiveness is that people do not believe that the
policy would be sufficient to substantially affect pollution and climate change. Taking respondents’
average anticipated elasticities for transport and housing energies (that are fairly accurate13), the tax
should reduce French greenhouse gas (GhG) emissions by 5.7 Mt of CO2equivalent (CO2e) each year,
according to the simulation from BdF data. This reduction corresponds to 0.8% of French annual
emissions, 0.01% of global ones, and does not suffice to reach the official objective of carbon neutrality
in 2050.14 Thus, although respondents do anticipate responses to price incentives, our results suggest
that they do not perceive a 50e/tCO2national carbon tax as a proportionate reaction to climate
change.
3.3 Progressivity
It is often argued that a critical barrier to accept carbon taxation is its perceived distributional
impact, in particular the higher burden imposed on lower income households (Bristow et al.,2010;
Brannlund & Persson,2012;Gevrek & Uyduranoglu,2015). A broad literature has shown that carbon
taxation alone is regressive (Poterba,1991;Metcalf,1999;Grainger & Kolstad,2010), meaning that
it is more costly for poorer households as a share of their resources. However, it has also been shown
that redistributing its revenue through uniform lump sum transfers — a mechanism known as flat-
recycling — can make the policy progressive (West & Williams,2004;Bento et al.,2009;Williams
et al.,2015), including for France (Bureau,2011;Douenne,2020). Figure C.2 in Appendix C.3 displays
the average net gain by income decile for our Tax & Dividend. It clearly appears from this figure that
lower income households would gain more than richer households, both in relative and in absolute
terms. Yet, only 19% of respondents think the policy would benefit the poorest households, compared
to 60% who declare it would not, and 21% who do not know. These results show that beliefs about
the distributive effects of carbon taxation are biased, which has not been shown in previous studies.
13Overall, average subjective elasticities are close to these estimates for transport (at −0.45) and somewhat overes-
timated for housing (−0.43). Among those who declared that the policy was not effective, 45% (resp. 43%) anticipated
an aggregate elasticity at or below −0.5 for housing (resp. for transport), while elasticities obtained from the literature
are around −0.2 for housing and −0.4 for transport.
14The computations are based on households’ carbon emissions. In 2014, French GhG consumption based emissions
were equal to 712 MtCO2e (CGDD,2019). 2017 global emissions were 53.5 GtCO2e (UNEP,2018).
18

4 How attitudes shape beliefs
The previous section has shown that people’s low acceptance of our Tax & Dividend coincides with
biased beliefs about the properties of the scheme. As knowledge about these properties has been shown
to be decisive for acceptance (Carattini et al.,2018), it is important to assess how beliefs are formed.
In the following, we test respondents’ reactions to information about their gains, environmental
effectiveness, and progressivity. If their biased views simply reflect a lack of knowledge, we should
expect them to revise their beliefs after new information is provided, what we refer to as “update”.
4.1 Self-interest
Conservatism and pessimism After telling respondents that given their characteristics, they
have 5 chances over 6 to “win” or “lose” from the policy, we observe that only 39% of them agree with
our feedback (GF=b
Γ), far from the 83% expected for a correct update. This result is consistent with
findings of the literature that people tend to update like “conservative” Bayesians (Edwards,1968).
The full transition matrices of people’s beliefs are given in Tables D.2 and D.3 in Appendix D.2. More
concisely, Table 4.1 reports the share of respondents whose beliefs after being informed are aligned
with our feedback, with the corresponding 95% binomial confidence intervals. It shows that, for the
24% of individuals who receive a “lose” feedback (b
Γ = 0), the ex post belief is on average consistent
with Bayesian updating. If anything, these people would rather tend to agree too much with our noisy
signal, especially when excluding people who initially consider themselves as unaffected (i.e. focusing
on gI6= 0), for whom 83% does not lie in the 95% confidence interval.15 The observed conservatism
on the whole sample therefore comes from respondents who receive a “win” feedback (b
Γ = 1). Among
the 60% of respondents who initially thought they would lose in this group, only 12% endorse the
“win” feedback. This is in sharp contrast with the respondents who initially thought they would win
and receive a “lose” feedback, since 82% of them endorse our prediction. Thus, pessimistic beliefs are
persistent to our treatment, but optimistic ones are not.
Table D.4 in Appendix D.2 conducts the same analysis for the 28% of respondents whose gain is
largely positive or largely negative, i.e. above 110eper c.u. in absolute terms. For such respondents,
our out-of-sample prediction of the winning category is correct in 99% of cases, as can be seen in
the Figure D.1 in Appendix D.2. The alignments with our feedback are similar for the whole sample
and for these respondents for whom we are sure to make a correct prediction. The similarity of
15We observe that respondents revise their beliefs less when they feel “unaffected”, indicating that some of them may
have misunderstood this category for “I don’t know”.
19

Table 4.1: Share of respondents with new beliefs aligned with feedback.
Aligned with feedback: GF=b
Γ
b
Γ=1 b
Γ = 0
(75.8%) (24.2%)
Initial belief winner (gI>0) 78.8% 81.5%
(14.0%) [73.2%; 83.4%] [65.0%; 91.3%]
Initial belief unaffected (gI= 0) 21.6% 44.9%
(21.7%) [17.6%; 26.2%] [33.5%; 56.8%]
Initial belief loser (gI<0) 12.2% 93.9%
(64.3%) [10.3%; 14.5%] [90.9%; 96.0%]
Initial belief affected (gI6= 0) 26.1% 92.9%
(78.3%) [23.7%; 28.7%] [89.8%; 95.1%]
All 25.1% 85.7%
(100%) [23.0%; 27.3%] [82.2%; 88.7%]
Note: The 95% confidence intervals for binomial probabilities are given in brackets.
alignments for different prediction accuracy rules out the possibility that some respondents do not
update because their private information would be truly more accurate than our prediction.
There are several ways to rationalize the conservatism and pessimism we observe. The simplest
is that respondents perceive our information as biased. They may think we wrongly estimate their
likelihood to win and that we are too optimistic. As a result, they may assign relatively more weight
to pessimistic information coming from us. This is in line with widespread disagreement between
economic experts and the general public about carbon taxation (Sapienza & Zingales,2013), for which
51% of Americans do not trust that the government would actually pay the promised dividends. A
second interpretation is that they give too much value to their private information relative to the
base rate one, a phenomenon known as base rate neglect (Grether,1980). That is to say, pessimistic
winners might be over-confident in seeing themselves as specific so that they partly discard the new
information, e.g. by thinking they are part of the one-sixth for whom our prediction is erroneous.
Another explanation is that uncertainty makes people see their possible gains as a distribution (see
Stiglitz,2019). Then, instead of reporting the average of this distribution, people subject to loss-
20

aversion would reason with conservative estimates for their gains, even more so if they believe that
this survey can influence policy makers. In the end, these interpretations are of course not mutually
exclusive and empirically not distinguishable, as they all lead to too much pessimism in the revision
of their beliefs.
Determinants of correct updating The observed asymmetry in beliefs’ revision echoes a recent
literature that has emphasized how people selectively process information regarding themselves or
specific events, generally in a way that satisfies their ego (Eil & Rao,2011;Mobius et al.,2011;
Sharot et al.,2011). Our results show that in the case of a policy considered as undesirable, people
give more weight to negative news. Thus, both results are consistent with individuals forming their
beliefs in a motivated way. To assess the extent of motivated reasoning, we analyze the heterogeneity
in people’s updating. To handle the notion of correct updating, we define a variable Uwhich equals +1
if the respondent adopts a feedback that invalidates their initial belief, −1if they update against the
feedback that confirms it, or 0if they do not update. Over the sub-sample of invalidated respondents
who should have updated because their initial winning category is not aligned with our feedback, we
regress the correct updating,U, over the initial belief not to lose, GI, and a vector of characteristics,
C:
Ui=δ0+βUGI
i+βCC+ifor i:sgn (gi)6=sgn (ˆγi),(1)
where sgn is the sign function. The high value for βUreported in column (1) of Table 4.2 again proves
that, among those who should have updated, those who initially think they would win (the optimistic
losers) update significantly more correctly than those who do not think so (the pessimistic winners).
Many other specifications have been tested, depending on how the unaffected are treated and whether
controls are included or not, or with the winning category after the feedback as covariate, and they
all confirm a win/lose asymmetry in updating.
Beyond this asymmetry, column (2) shows that some respondents’ characteristics are correlated
with correct updating. Relative to unemployed and inactive people, retired, active, and students
update more correctly, the latter being 25 p.p. more likely to correctly revise their beliefs when
invalidated than unemployed and inactive. Similarly to perceptions of net gains, position towards the
Yellow Vests is significantly correlated with correct updating. The magnitude of this effect is large,
as people who are part of the movement are 17 p.p. less likely to correctly update than people who
oppose it. This can be linked to the Yellow Vests’ higher distrust of the government, documented
21

Table 4.2: Asymmetric updating of winning category
Correct updating (U)
(1) (2) (3) (4)
Constant 0.120∗∗∗ −0.035 −0.146 −0.116
(0.012) (0.179) (0.178) (0.179)
Winner, before feedback ( ˙
G) 0.695∗∗∗ 0.685∗∗∗ 0.646∗∗∗ 0.659∗∗∗
(0.078) (0.080) (0.080) (0.080)
Initial tax: PNR (I don’t know) 0.163∗∗∗ 0.165∗∗
(0.031) (0.067)
Initial tax: Approves 0.158∗∗∗ −0.056
(0.046) (0.115)
Diploma (1 to 4) 0.015 0.016 0.011
(0.013) (0.013) (0.014)
Diploma ×Initial tax: PNR −0.001
(0.025)
Diploma ×Initial tax: Approves 0.074∗∗
(0.037)
Retired 0.143∗0.146∗0.142∗
(0.080) (0.079) (0.079)
Active 0.165∗∗∗ 0.175∗∗∗ 0.175∗∗∗
(0.055) (0.054) (0.054)
Student 0.249∗∗∗ 0.234∗∗∗ 0.239∗∗∗
(0.076) (0.075) (0.075)
Yellow Vests: PNR −0.048 −0.043 −0.044
(0.047) (0.047) (0.047)
Yellow Vests: understands −0.090∗∗∗ −0.063∗−0.064∗
(0.034) (0.034) (0.034)
Yellow Vests: supports −0.101∗∗∗ −0.059∗−0.060∗
(0.035) (0.036) (0.036)
Yellow Vests: is part −0.172∗∗∗ −0.137∗∗ −0.138∗∗
(0.062) (0.062) (0.062)
Among invalidated X X X X
Includes controls X X X
Observations 1,365 1,365 1,365 1,365
R20.055 0.111 0.133 0.136
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Note: Omitted variables are Unemployed/Inactive and Yellow Vests: opposes. The list of controls can be found in
Appendix H.
22

in Algan et al. (2019), that could apply to information provided by researchers regarding policies.
Column (3) includes disapproval as a covariate: it indicates that, as for the initial bias in subjective
gains, disapproving the reform is associated with a less correct update by 16 p.p. and partly explains
the effect of the Yellow Vests. Appendix Econfirms that the result holds true when initial subjective
gains are accounted for, ruling out the explanation that this less correct updating is entirely due to
specific priors of those who disapprove the reform.
Motivated reasoning The previous result suggests that conservatism in belief revision does not
simply follow from people’s cognitive difficulties when dealing with Bayes’ rule. The higher like-
lihood to update correctly of those who approve the reform is robust evidence that ideology and
political identity shape belief formation. Such directional motivated reasoning has been proposed
by Druckman & McGrath (2019) to explain polarization around beliefs on climate change. Kahan
(2013) provides empirical evidence that political motivated reasoning about climate change is not a
reasoning deficiency but rather a reasoning adaptation following the interest that individuals have in
conveying “their membership in and loyalty to affinity groups central to their personal well-being”.
In our case, the position relative to the Yellow Vests proxies the groups that respondents identify
with, and the differentiated updating along this spectrum can be interpreted as motivated reasoning.
Besides, the hypothesis that motivated reasoning follows from a rational adaptation purpose implies
that better educated people are more prone to motivated reasoning, as they are more able to formu-
late specious reasonings and reconcile antagonistic information and ideas. Column (4) supports this
hypothesis, as the interaction term between diploma and initial tax approval is positive and signif-
icant, even capturing all the effect of initial tax approval. To our knowledge, this result is the first
robust evidence of directional and rational motivated reasoning in the context of climate policies, and
it provides empirical support for various models of endogenous belief formation. For example, Little
(2019) formalizes the idea that directional motives may override accuracy motives and update auxil-
iary beliefs (in our case, the winning category) in order to preserve their consistency with core beliefs
(here, rejection of the tax). Building upon the cognitive and social mechanisms described by Kraft
et al. (2015) and documented by e.g. Redlawsk (2002), we hypothesize the following narrative as one
of the possible channels through which aversion for the carbon tax became entrenched. The Yellow
Vests first gathered to defend their interest (above all their purchasing power), and a side effect of
the daily interactions on roundabouts was to bring material and emotional support to the protesters
(Challier,2019). A group identity soon developed, which crystallized shared beliefs and affects such
23

as a rejection of carbon taxation. This group identity gained support from a large majority of the
population, notably through social networks. Now, due to the loyalty to the group as well as the
affects that have entered their subconscious, Yellow Vests supporters oppose instinctively any carbon
tax, and are prone to find excuses to cope with contradictory messages, e.g. by denying the relia-
bility of these messages (Golman et al.,2016). Admittedly, such a narrative falls short of explaining
the majority rejection among those who oppose the Yellow Vests (which may originate from biased
perceptions more than tax aversion), but it illustrates how biased beliefs can be so persistent among
Yellow Vests supporters.
4.2 Environmental effectiveness
Table D.5 in Appendix D.3 reports the effect of displaying relevant information on the belief that
our Tax & Dividend is environmentally effective. The effect of reporting a scientific consensus on
environmental effectiveness (E) is positive and statistically significant, but its magnitude — around
5 p.p. — seems modest given that the question immediately follows the priming. The effects of
information on climate change (CC ) or particulates (P M ) are smaller, and only CC is significant,
which is understandable as they were displayed at the very beginning of the survey and do not mention
any environmental policy. As suggested by Millner & Ollivier (2016), given the complexity of the
mechanisms at play, drawing a causal link between causes and consequences of environmental problems
requires considerable cognitive effort, making it difficult to convince one about the effectiveness of
policies that decentralize efforts to tackle pollution. Finally, we observe that our primings have no
significant effect on beliefs over causes and consequences of climate change. Overall, these primings
appear insufficient to change most people’s mind about climate change and carbon tax effectiveness.
4.3 Progressivity
Table D.6 in Appendix D.4 shows the absence of effect of explaining that our Tax & Dividend is
progressive on perceived progressivity: the correlation between the two is close to 0 (at −0.006) and
even has an unexpected negative sign. Column (2) of the same table clarifies why our treatment
does not change the overall share of people who think the policy is regressive: those who have a
large bias in their perception of gains are in fact more prone to perceive regressivity once provided
the information, by 13 p.p. This result may be a manifestation of the boomerang effect with people
inclined to motivated reasoning, which has already been documented for Republican attitudes over
climate change in the US (Zhou,2016). Indeed, Hovland et al. (1953) showed that when someone
24

is pressured to make a certain choice, psychological reactance (theorized by Brehm,1966) can cause
them to resist this pressure by adopting an opposite alternative. Although the effect on those without
a large bias is not significant, providing them with information is associated with a lower perceived
regressivity by 5 p.p. A possible explanation for the strong belief in regressivity is that people
view the tax as regressive (relative to income) and the transfer as neutral (in absolute values), and
mistakenly conclude that their combination is regressive. In any case, without a deep explanation of
the underlying mechanisms, the progressivity of the policy remains unintuitive for most people, and
we cannot convince them easily.
5 How beliefs determine attitudes
Our results clearly indicate that, as of today, a carbon tax is unlikely to be accepted in France.
However, we have also shown that people largely display incorrect perceptions about the true effects
of the policy. Most of them overestimate the negative impact on their purchasing power, think that
the policy is regressive, and do not see it as environmentally effective. In this section, we examine to
what extent the low acceptance rate reflects intrinsic preferences or wrong perceptions. The question
we address is whether correcting biased beliefs would be sufficient for a carbon tax to be accepted.
5.1 Self-interest
Identification challenge Among the three-quarters of the respondents expected to win from our
Tax & Dividend, 62% both consider that they would not win and disapprove of the policy. We want to
estimate to what extent knowing they would win would lead them to approve of the reform. Because
respondents thinking they would win might differ in many respects from those thinking they would
not, we cannot simply regress approval on perception of winning.
Main identification strategy In order to identify the effect ceteris paribus of self-interest on
acceptance, we exploit exogenous variations in gains and losses. To do so, we consider a Tax &
Targeted Dividend, where respondents are randomly assigned to a compensation scheme to which
they are eligible or not (see section 2.1.2). Our methodology mixes a random discontinuity design
(RDD) and an instrumental variable (IV) strategy. We denote by I1,i respondent’s iincome, and by
I2,i the income of the second adult of their household if there is one, which we control for with a
dummy variable Siequal to 1 when there is a single adult. Let T1,i and T2,i be two binary variables
25

indicating whether these individuals would be eligible for the transfer.16 We also denote by GT
ia
dummy variable equal to 0 if respondent ithinks they would lose from the Tax & Targeted Dividend,
and 1 otherwise. Similarly, AT
iis a dummy variable equal to 0 if respondent idisapproves of this
policy and 1 otherwise. We can then write a two-stage least square model, with the following first
stage equation:
GT
i=α0+α1T1,i +α2T2,i +αcci+αSSi+
2
X
j=1 α1,j Ij
1,i +α2,j Ij
2,i+ηi(2)
where second order terms for income bring more flexibility. We precise that eligibility is defined using
income thresholds cithat are randomly allocated to households (see section 2.1.2) and introduced
as fixed effects to control for threshold-specific preferences. Formally, we define eligibility of adult
k∈ {1; 2}as:
Tk,i =




0,if Ik,i > ci
1,otherwise
(3)
Finally, the second stage writes:
AT
i=β0+β1b
GT
i+βcci+βSSi+
2
X
j=1 β1,j Ij
1,i +β2,j Ij
2,i+i(4)
where c
Gi
Tdenotes the fitted value of GT
ifrom the first stage regression. As can be seen from first stage
results in Appendix F.1, eligibility of both respondents and households’ second adults are positively
correlated with beliefs of winning, so both instruments are relevant. The exclusion restriction states
that conditional on income, being eligible affects approval solely through beliefs of winning. The RDD
procedure employed in the first stage ensures that this is the case: conditional on income, eligibility
is random, and controlling for the target, it should affect acceptance only through self-interest.
Alternative specifications for robustness To further ensure the robustness of the results, we
estimate several alternative specifications. First, we run the same RDD + IV design adding control
variables (specification 2). In particular, we control for initial acceptance of our Tax & Dividend as
this should explain most of the variation in the dependent variable. Second, we compare our results
with a simple OLS where we control for relevant variables (3). Third, we use a logit model (4) to
16As explained in section 2.1.2, we explicitly limit the number of beneficiaries to two per household.
26

ensure that imposing linearity does not bias the results. Finally, we exploit a methodology similar
to the main specification but applied to the customized feedback, without (5) and with (6) control
variables. Indeed, as our estimation of gains for this feedback is a continuous variable bγ, but the
feedback itself is a binary variable b
Γ, we run an RDD to predict the belief of winning after feedback
GF. We then exploit variations around the threshold of zero net gain (which are random conditional
on net gain) to explain acceptance AF. This alternative two-stage least square writes:
GF
i=α0+α1b
Γi+
2
X
j=1
α1,j (bγi)j+ηi(5)
AF
i=β0+β1b
GF
i+
2
X
j=1
β1,j (bγi)j+i(6)
where b
GF
idenotes the fitted value of GF
ifrom the first stage regression. The identification assumption
of this second IV states that conditional on estimated net gains (bγ), receiving a win feedback (b
Γ = 1)
affects approval solely through self-interest. Finally, we also investigate alternative versions of the
previous models where we estimate the effect to “win” instead of “not to lose”, and on “approval”
instead of “acceptance”.
Results First stage regression results are given in Appendix F.1. The effective F-Statistics (Mon-
tiel Olea & Pflueger,2013) range from 37 to 57, indicating that both targeted transfers and feedback
are strong instruments. Table 5.1 provides the second stage results for the six main specifications, and
additional specifications can be found in Appendix F.2. Overall, the estimated effects of self-interest
indicate that believing not to lose increases acceptance by about 50 p.p. The results obtained for the
local average treatment effect (LATE) on compliers in IV regressions — i.e. on people who recognize
they would not lose because of the treatment — are significantly higher than for the average treatment
effect (ATE) estimated with OLS (57 vs. 44 p.p. for Tax & Targeted Dividend). Given the wide
set of control variables used, and in particular our powerful control AI, one can be rather confident
that the OLS estimate is unbiased, and that the difference is due to the specificity of compliers in
the LATE. This explanation holds for both the targeted scheme and acceptance after feedback and
is supported by the findings of section 4.1 for the latter. Indeed, as respondents most likely to revise
their beliefs after a “win” feedback are also more in favor of the tax, the IV coefficient for compliers
is logically higher than the point estimate found with OLS. The comparisons of columns 1 with 2,
and 5 with 6, show that adding control variables does not significantly affect the coefficient for Tax
27

Table 5.1: Effect of self-interest on acceptance.
Targeted Acceptance (AT) Feedback Acceptance (AF)
IV OLS logit IV
(1) (2) (3) (4) (5) (6)
Believes does not lose 0.571∗∗∗ 0.567∗∗∗ 0.443∗∗∗ 0.431∗∗∗ 0.517∗∗∗ 0.434∗∗∗
(0.092) (0.092) (0.014) (0.018) (0.170) (0.135)
Initial tax Acceptance (AI) 0.339∗∗∗ 0.360∗∗∗ 0.342∗∗∗ 0.428∗∗∗
(0.033) (0.026) (0.034) (0.055)
Controls: Incomes X X X X X
Controls: Estimated gain X X X X X
Controls: Target of the tax X X X X
Controls: Socio-demo, other motives X X X X
Observations 3,002 3,002 3,002 3,002 1,968 1,968
R20.033 0.302 0.470 0.044 0.526
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Note: Standard errors are reported in parentheses. For logit, average marginal effects are reported and not
coefficients. The list of controls can be found in Appendix H.
& Targeted Dividend, but does for acceptance after feedback, where the controls seem to capture
part of the specificity of compliers. Overall, both methods yield similar effects: 57 p.p. with targeted
transfers and 52 p.p. with the feedback. Finally, results from the logit regression (4) confirm that
the assumptions of the linear probability model (LPM) do not bias the results.17
By isolating beliefs about self-interest ceteris paribus, we showed that this motive has a large
effect on carbon tax acceptance. This result confirms previous findings of the literature (Stern et al.,
1993;Thalmann,2004;Baranzini & Carattini,2017), but contrasts with the results of Kallbekken &
Sælen (2011) who found that self-interest plays a limited role in Norway. Their different result could
come from their methodological approach that relies on proxies to capture self-interest in regressions,
or on differences between preferences of French and Norwegian people. Our results also qualify the
findings of Anderson et al. (2019) who suggest that ideology better predicts carbon tax acceptance
than self-interest. By distinguishing beliefs from preferences, we find that ideology plays an indirect
17Separation occurs in logistic regressions for self-interest and effectiveness. We made sure this does not lead to
biased coefficients by conducting likelihood ratio tests and by running penalized regressions (Firth,1993;Heinze &
Ploner,2003).
28

role by shaping beliefs about one’s self-interest, and that beliefs directly affect acceptance.
5.2 Environmental effectiveness
Main identification strategy One of the strongest barriers to carbon tax implementation is a
widespread perception of its environmental ineffectiveness. Our objective is therefore to assess to what
extent learning about the environmental benefits of the tax could increase acceptance. To identify
this effect, we estimate a two-stage least squares (2SLS) where the first stage uses random information
to explain beliefs about environmental effectiveness, while the second stage regresses acceptance on
the fitted exogenous variations in these beliefs. Because information on particulate matter (ZP M ) is
poorly correlated with beliefs of effectiveness, we restrict the set of instruments to our primings on the
scientific consensus (ZE) and climate change (ZCC ). As discussed in section 4.2, these instruments
are significantly related to our endogenous variable, yet are potentially weak, as our primings do
not have a large effect on people’s beliefs. If we denote by Ea dummy variable equal to 0 if the
respondents think the policy is not environmentally effective and 1 otherwise, we can write a 2SLS
model as follows:
Ei=α0+α1ZE,i +α2ZC C,i +αCCi+ηi(7)
AI
i=β0+β1b
Ei+βCCi+i(8)
where b
Eidenotes the fitted value of Eifrom the first stage regression, and Ca vector of characteristics.
Alternative specifications for robustness checks To ensure the robustness of our results, we
estimate the previous 2SLS with control variables (1). Acknowledging that our primings could affect
acceptation motives other than effectiveness alone, this potential bias should disappear when con-
trolling for many variables including other motives. In addition, we estimate an OLS (2) model to
compare the LATE of the first specifications with an ATE, and a logit model (3) to check the robust-
ness of assumptions underlying the linear probability model. Finally, we also estimate two modified
versions of our main specifications, where we switch from broad to strict definitions for environmental
effectiveness (4) and for tax approval (5).
Results The first stage regressions results can be found in Appendix F.1. Because of the relatively
modest responses to our primings, the instruments are rather weak (effective F-statistic of 6 in column
29

Table 5.2: Effect of believing in environmental effectiveness on acceptance
Tax Acceptance (AI) Tax Approval ( ˙
AI)
IV OLS logit IV IV
(1) (2) (3) (4) (5)
Environmental effectiveness: not “No” 0.479∗∗ 0.391∗∗∗ 0.370∗∗∗
(0.230) (0.015) (0.018)
Environmental effectiveness: “Yes” 0.505∗∗ 0.416∗∗
(0.242) (0.168)
Instruments: info E.E. & C.C. X X X
Controls: Socio-demo, other motives, X X X X X
incomes, estimated gains
Observations 3,002 3,002 3,002 3,002 3,002
R20.218 0.390 0.218 0.161
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Note: Standard errors are reported in parentheses. For logit, average marginal effects are reported and not
coefficients. The list of controls can be found in Appendix H.
1), a problem that is alleviated in the case of strict definitions (11 in columns 4 and 5). Given the
exogeneity of our instruments, the only concern is a potential bias towards OLS, which would entail
estimates that are too conservative in our case. Table 5.2 reports the results of the second stages
and alternative specifications. They all consistently indicate a strong positive effect of beliefs about
environmental effectiveness on tax acceptance. Overall, the effects are statistically significant, and
their size is comparable to that of self-interest, at 48 p.p. for our main IV regression (column 1).
The LATE is again somewhat higher than the ATE estimated with OLS (2) — 48 vs. 39 p.p. —
which is likely due to differences between compliers and other respondents: people who are most
likely to change their mind following our information might also be more willing to accept the policy.
Still, the coefficients obtained for the ATE remain large and indicate that beliefs over the policy’s
environmental effectiveness are critical for acceptance. Results of the logit model close to the ones of
OLS confirm that the LPM does not induce a bias in the estimation (3). Lastly, the effect is relatively
close with the strict definition of the dependent variable — 42 p.p. in (5), showing that the coefficient
is not driven by a correlation between “PNR” responses to tax effectiveness and approval.
The critical role found for beliefs about environmental effectiveness is in line with findings of
30

the literature (Sælen & Kallbekken,2011;Kallbekken & Sælen,2011;Baranzini & Carattini,2017),
although previous studies do not properly identify a causal effect. Our results thus confirm that con-
vincing people about the environmental effectiveness of a carbon tax would largely increase acceptance
of this instrument.
5.3 Progressivity
Identification challenge and strategies As informing respondents does not convince them that
our Tax & Dividend is progressive (see section 4.3), we cannot identify the causal effect of under-
standing the progressivity on acceptance using an IV estimation. Thus, we estimate how one’s belief
in progressivity correlates with acceptance using simple OLS and logit regressions. Controlling for
many respondents’ characteristics and other acceptance motives, one can be confident that the ef-
fect of progressivity is properly isolated. We focus on the acceptance question after knowledge, i.e.
after asking whether the reform is progressive or not.18 Table 5.3 presents the results of different
regressions, depending on the set of controls and on the choice of variables. Columns (1)-(4) report
regressions of acceptance on the broad definition of motives of acceptance: answers not “No” to pro-
gressivity, effectiveness and not “lose” to winning category. On the contrary, columns (5)-(6) use
strict definitions for both approval and the covariates, where only “Yes” (or “win”) answers activate
the dummy variables.
Results On average, believing that the reform is not regressive is associated with a higher acceptance
rate by 56 p.p. (column 3), while believing it is progressive is associated with a higher approval rate
by 48 p.p. (6). However, when one introduces other motives of acceptance and their interactions as
covariates, with households characteristics as controls, one observes that the effect of progressivity
ceteris paribus is lower. The marginal effect of progressivity at the sample mean — i.e. accounting for
the average marginal effect of interaction terms — is 27 p.p., slightly lower than the marginal effect
obtained for self-interest (40 p.p.) and environmental effectiveness (31 p.p.). The effect obtained for
the latter motive is lower than the OLS estimate found in section 5.2, because here acceptance is taken
at a later step in the survey and not right after asking about environmental effectiveness, making
it less salient. Besides, one might worry that perceived progressivity would be hard to disentangle
from beliefs over net gains, as the latter is influenced by the former for a given income. To address
18As self-interest and effectiveness were made salient from the initial questions, treating progressivity after knowledge
is the only way to make results comparable across motives.
31

Table 5.3: Effect of beliefs over progressivity on acceptance. Covariates refer either to broad (1-4)
or strict (5-6) definitions of the beliefs, where strict dummies do not cover “PNR” or “Unaffected’
answers.
Acceptance (AP) on not “No” Approval ( ˙
AP) on “Yes”
OLS logit OLS
(1) (2) (3) (4) (5) (6)
Progressivity (P)0.223∗∗∗ 0.237∗∗∗ 0.560∗∗∗ 0.544∗∗∗ 0.228∗∗∗ 0.482∗∗∗
(0.038) (0.044) (0.023) (0.019) (0.041) (0.023)
Winner (GP)0.332∗∗∗ 0.332∗∗∗ 0.303∗∗∗
(0.020) (0.020) (0.019)
Effective (E)0.258∗∗∗ 0.259∗∗∗ 0.244∗∗∗
(0.023) (0.023) (0.020)
(GP×E)0.127∗∗∗ 0.127∗∗∗ 0.126∗∗∗
(0.034) (0.034) (0.037)
Interaction: winner (P×GP)0.183∗∗∗ 0.183∗∗∗ 0.098∗∗
(0.050) (0.050) (0.048)
Interaction: effective (P×E)0.172∗∗∗ 0.172∗∗∗ 0.281∗∗∗
(0.057) (0.057) (0.059)
Income (I, in ke/month) 0.017 0.018 0.037∗∗
(0.022) (0.022) (0.018)
Interaction: income (P×I)−0.008 −0.019
(0.013) (0.014)
P×GP×E−0.400∗∗∗ −0.399∗∗∗ −0.314∗∗∗
(0.072) (0.072) (0.083)
Controls: Socio-demo, incomes, X X X
estimated gains
Observations 3,002 3,002 3,002 3,002 3,002 3,002
R20.460 0.460 0.162 0.391 0.130
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Note: Standard errors are reported in parentheses. For logit, average marginal effects are reported and not
coefficients. The list of controls can be found in Appendix H.
32

this dependency, we include the interaction between progressivity and income as a covariate (2,
5). Although the coefficient is negative, in accordance with intuition, the effect is small and not
significant. Finally, using the strict definitions of beliefs and approval yields a smaller correlation (6)
but similar results when accounting for relevant controls (5), showing that the effects are not driven
by a correlation between “PNR” answers.
These results show that progressivity is a motive for acceptance almost as determinant as self-
interest and effectiveness. This concern for distributional effects is consistent with the findings of
Kallbekken & Sælen (2011), Brannlund & Persson (2012), and Gevrek & Uyduranoglu (2015), but
contrasts with the results of Baranzini & Carattini (2017), who found that among Genevan, the
concern exists but does not affect acceptance of environmental policies.
5.4 Complementarity between motives
Table 5.3 (column 1) shows that the combined effect of the three motives together — i.e. adding
up their coefficients and the ones of their interactions — is as high as 90 p.p., confirming that these
motives are the beliefs critically needed to accept the policy, and that they complement (rather
than substitute) one another. Considering approval when respondents do believe that the policy is
effective, progressive, and satisfies their self-interest (5), the results are quite similar, although it
appears that the two altruistic motives are more complementary: their interaction term is high, and
taken together they increase approval by 74 p.p. (marginal effect at the sample mean), against 64 p.p.
for the combination of progressivity and self-interest, and 69 p.p. for self-interest and effectiveness.
The combined effect on approval of the three motives with their strict definition reaches 97 p.p. Thus,
in a scenario where everyone would believe in the effectiveness and progressivity of the policy, and
the 70% of households who win would know it, our estimation indicates that the approval rate would
be as high as 90%. These results suggest that the carbon tax rejection that appears at first glance is
mainly due to biases in beliefs (fostered by tax aversion), and that if theses biases could be corrected
for, carbon pricing could become close to consensual.
6 Conclusion
In this paper, we study how beliefs about a policy form and then determine attitudes towards it.
We investigate this question through the study of carbon taxation in France during the Yellow Vests
movement, that started against fuel price increases. Our analysis is based on a new survey and
33

consumer survey data, enabling one to compare subjective beliefs with objective impacts on French
households. We find that 70% disapprove of a carbon Tax & Dividend policy, which can be explained
by biased beliefs about its properties. 89% of our survey respondents overestimate its negative
impact on their purchasing power, and most of them do not perceive it as environmentally effective,
nor progressive. Biased beliefs appear correlated with people’s support for the scheme: the more they
oppose the mechanism, the more biased they are. We show that causality between beliefs and attitude
towards the policy goes in both directions. People more opposed to the tax are more (pessimistically)
biased in their treatment of new information with respect to it, indicating that beliefs about tax
impacts are partly shaped by motivated reasoning. At the same time, we find that acceptance is
causally determined by beliefs and that if all biased beliefs could be corrected, the approval rate
would reach 90%.
However, our treatments that provide accurate arguments in favor of the scheme never convince
more than 12% of people. We can relate this conservatism to a strong distrust of the government,
documented e.g. in Alesina et al. (2018) and Algan et al. (2019), echoing recent findings that the
ambition of climate policies increases with the level of trust (Rafaty,2018). These results leave us
with three main challenges. First, as it is unlikely that the issue of trust can be resolved in the short
run, it seems necessary to find climate policies that would be accepted by a majority. We address
this question in a companion paper (Douenne & Fabre,2020), in which we assess both knowledge
and beliefs about climate change, and the preferred policies of French people. Second, as trust in
government needs to be restored in the longer run, it is crucial to analyze what causes the distrust
and how it can be overcome. Third, it is important to assess to what extent the mechanisms of belief
formation and their effects on political attitudes we document can be generalized to other policies
and other contexts. Although rejection of the tax may be lower in a different country, biases in
perceptions and political polarization are likely common everywhere. Thus, a lesson must be learned
for policy design and implementation, to avoid another carbon tax debacle à la Française.
34

References
ADEME. Représentations sociales de l’effet de serre. Technical report, 2018. 4
A. Alberini, A. Bigano, M. Ščasný, & I. Zvěřinová. Preferences for Energy Efficiency vs. Renewables:
What Is the Willingness to Pay to Reduce CO2 Emissions? Ecological Economics, 2018. 60
A. Alesina, S. Stantcheva, & E. Teso. Intergenerational Mobility and Preferences for Redistribution.
American Economic Review, 2018. 5,34
Y. Algan, E. Beasley, D. Cohen, M. Foucault, & M. Péron. Qui sont les gilets jaunes et leurs soutiens
? Technical report, CEPREMAP et CEVIPOF, 2019. 23,34
S. Anderson, I. E. Marinescu, & B. Shor. Can Pigou at the Polls Stop US Melting the Poles? Working
Paper 26146, National Bureau of Economic Research, 2019. 4,5,28
A. Baranzini & S. Carattini. Effectiveness, earmarking and labeling: testing the acceptability of
carbon taxes with survey data. Environmental Economics and Policy Studies, 2017. 4,8,17,28,
31,33
R. Bénabou & J. Tirole. Mindful Economics: The Production, Consumption, and Value of Beliefs.
Journal of Economic Perspectives, 2016. 5
A. M. Bento, L. H. Goulder, M. R. Jacobsen, & R. H. von Haefen. Distributional and Efficiency
Impacts of Increased US Gasoline Taxes. American Economic Review, 2009. 18
R. P. Berrens, A. K. Bohara, H. C. Jenkins-Smith, C. L. Silva, & D. L. Weimer. Information and
effort in contingent valuation surveys: application to global climate change using national internet
samples. Journal of Environmental Economics and Management, 2004. 60
R. Brannlund & L. Persson. To tax, or not to tax: preferences for climate policy attributes. Climate
Policy, 2012. 4,18,33
S. Brechin. Public opinion: a cross-national view. In Handbook of Climate Change and Society.
Routledge, lever-tracy, constance edition, 2010. ISBN 978-1-135-99850-9. 4
J. W. Brehm. A theory of psychological reactance. Academic Press, Oxford, England, 1966. 25
A. L. Bristow, M. Wardman, A. M. Zanni, & P. K. Chintakayala. Public acceptability of personal
carbon trading and carbon tax. Ecological Economics, 2010. 4,18
35

B. Bureau. Distributional effects of a carbon tax on car fuels in France. Energy Economics, 2011. 13,
18
T. A. Cameron. Individual option prices for climate change mitigation. Journal of Public Economics,
2005. 60
D. Cantoni, D. Y. Yang, N. Yuchtman, & Y. J. Zhang. Protests as Strategic Games: Experimental
Evidence from Hong Kong’s Antiauthoritarian Movement*. The Quarterly Journal of Economics,
2019. 56
S. Carattini, M. Carvalho, & S. Fankhauser. Overcoming public resistance to carbon taxes. Wiley
Interdisciplinary Reviews: Climate Change, 2018. 4,17,19
CGDD. Chiffres clés du climat France, Europe et Monde. Technical report, 2019. 18
R. Challier. Rencontres aux ronds-points. La Vie des idées, 2019. 23
M. Clerc & V. Marcus. Elasticité-prix des consommations énergétiques des ménages. Technical report,
Insee, 2009. 13
G. Cruces, R. Perez-Truglia, & M. Tetaz. Biased perceptions of income distribution and preferences
for redistribution: Evidence from a survey experiment. Journal of Public Economics, 2013. 5
T. Douenne. The vertical and horizontal distributive effects of energy taxes: A case study of a french
policy. The Energy Journal, 2020. 12,13,18
T. Douenne & A. Fabre. French attitudes on climate change, carbon taxation and other climate
policies. Ecological Economics, 2020. 10,34,48
S. Dresner, L. Dunne, P. Clinch, & C. Beuermann. Social and political responses to ecological tax
reform in Europe: an introduction to the special issue. Energy Policy, 2006. 17
S. Dresner, T. Jackson, & N. Gilbert. History and social responses to environmental tax reform in
the United Kingdom. Energy Policy, 2006. 4
J. N. Druckman & M. C. McGrath. The evidence for motivated reasoning in climate change preference
formation. Nature Climate Change, 2019. 23
W. Edwards. Conservatism in human information processing. In Formal representations of human
judgement. Wiley, New York, b. kleinnmutz edition, 1968. 19
36

D. Eil & J. M. Rao. The Good News-Bad News Effect: Asymmetric Processing of Objective Infor-
mation about Yourself. American Economic Journal: Microeconomics, 2011. 21
D. Firth. Bias reduction of maximum likelihood estimates. Biometrika, 1993. 28
Z. E. Gevrek & A. Uyduranoglu. Public preferences for carbon tax attributes. Ecological Economics,
2015. 18,33
R. Golman, G. Loewenstein, K. O. Moene, & L. Zarri. The Preference for Belief Consonance. Journal
of Economic Perspectives, 2016. 24
C. A. Grainger & C. D. Kolstad. Who Pays a Price on Carbon? Environmental and Resource
Economics, 2010. 18
D. M. Grether. Bayes Rule as a Descriptive Model: The Representativeness Heuristic. The Quarterly
Journal of Economics, 1980. 20
W. M. Hanemann. Welfare Evaluations in Contingent Valuation Experiments with Discrete Re-
sponses. American Journal of Agricultural Economics, 1984. 60
G. Heinze & M. Ploner. Fixing the nonconvergence bug in logistic regression with SPLUS and SAS.
Computer Methods and Programs in Biomedicine, 2003. 28
C. I. Hovland, I. L. Janis, & H. H. Kelley. Communication and persuasion; psychological studies of
opinion change. Yale University Press, New Haven, CT, US, 1953. 24
J. D. Jenkins. Political economy constraints on carbon pricing policies: What are the implications
for economic efficiency, environmental efficacy, and climate policy design? Energy Policy, 2014. 60
D. M. Kahan. Ideology, motivated reasoning, and cognitive reflection. Judgment and Decision Making,
2013. 23
S. Kallbekken & H. Sælen. Public acceptance for environmental taxes: Self-interest, environmental
and distributional concerns. Energy Policy, 2011. 4,5,17,28,31,33
S. Kallbekken, S. Kroll, & T. L. Cherry. Do you not like Pigou, or do you not understand him? Tax
aversion and revenue recycling in the lab. Journal of Environmental Economics and Management,
2011. 6,17
37

J. B. Kessler & M. I. Norton. Tax aversion in labor supply. Journal of Economic Behavior &
Organization, 2016. 6
D. Klenert, L. Mattauch, E. Combet, O. Edenhofer, C. Hepburn, R. Rafaty, & N. Stern. Making
carbon pricing work for citizens. Nature Climate Change, 2018. 4
M. J. Kotchen, K. J. Boyle, & A. A. Leiserowitz. Willingness-to-pay and policy-instrument choice for
climate-change policy in the United States. Energy Policy, 2013. 60
P. W. Kraft, M. Lodge, & C. S. Taber. Why People “Don’t Trust the Evidence”: Motivated Reasoning
and Scientific Beliefs. The ANNALS of the American Academy of Political and Social Science, 2015.
23
Z. Kunda. The Case for Motivated Reasoning. Psychological Bulletin, 1990. 5
I. Kuziemko, M. I. Norton, E. Saez, & S. Stantcheva. How elastic are preferences for redistribution?
evidence from randomized survey experiments. American Economic Review, 2015. 5
A. T. Little. The Distortion of Related Beliefs. American Journal of Political Science, 2019. 5,23
A. Longo, A. Markandya, & M. Petrucci. The internalization of externalities in the production
of electricity: Willingness to pay for the attributes of a policy for renewable energy. Ecological
Economics, 2008. 60
G. E. Metcalf. A Distributional Analysis of Green Tax Reforms. National Tax Journal, 1999. 18
A. Millner & H. Ollivier. Beliefs, Politics, and Environmental Policy. Review of Environmental
Economics and Policy, 2016. 4,24
M. M. Mobius, M. Niederle, P. Niehaus, & T. S. Rosenblat. Managing self-confidence: Theory and
experimental evidence. Working Paper 17014, National Bureau of Economic Research, 2011. 21
J. L. Montiel Olea & C. Pflueger. A robust test for weak instruments. Journal of Business & Economic
Statistics, 2013. 27
J. J. Murphy, P. G. Allen, T. H. Stevens, & D. Weatherhead. A Meta-analysis of Hypothetical Bias
in Stated Preference Valuation. Environmental and Resource Economics, 2005. 60
J. M. Poterba. Is the Gasoline Tax Regressive? Tax Policy and the Economy, 1991. 18
38

R. Rafaty. Perceptions of Corruption, Political Distrust, and the Weakening of Climate Policy. Global
Environmental Politics, 2018. 34
D. P. Redlawsk. Hot Cognition or Cool Consideration? Testing the Effects of Motivated Reasoning
on Political Decision Making. The Journal of Politics, 2002. 5,23
H. Sælen & S. Kallbekken. A choice experiment on fuel taxation and earmarking in Norway. Ecological
Economics, 2011. 31
P. Sapienza & L. Zingales. Economic experts versus average americans. American Economic Review,
2013. 8,20
T. Sharot, C. Korn, & R. Dolan. How unrealistic optimism is maintained in the face of reality. Nature
Neuroscience, 2011. 21
P. C. Stern, T. Dietz, & L. Kalof. Value Orientations, Gender, and Environmental Concern. Envi-
ronment and Behavior, 1993. 4,28
J. E. Stiglitz. Addressing Climate Change through Price and Non-Price Interventions. European
Economic Review, 2019. 20
D. Streimikiene, T. Balezentis, I. Alisauskaite-Seskiene, G. Stankuniene, & Z. Simanaviciene. A
Review of Willingness to Pay Studies for Climate Change Mitigation in the Energy Sector. Energies,
2019. 60
A. B. Sussman & C. Y. Olivola. Axe the Tax: Taxes are Disliked More than Equivalent Costs. Journal
of Marketing Research, 2011. 5
M. Thaler. The “Fake News” Effect: An Experiment on Motivated Reasoning and Trust in News.
2019. 5
P. Thalmann. The Public Acceptance of Green Taxes: 2 Million Voters Express Their Opinion. Public
Choice, 2004. 4,5,28
UNEP. Emissions Gap Report 2018. Technical report, 2018. 18
S. West & R. Williams. Estimates from a consumer demand system: implications for the incidence
of environmental taxes. Journal of Environmental Economics and Management, 2004. 18
39

R. C. Williams, H. Gordon, D. Burtraw, J. C. Carbone, & R. D. Morgenstern. The Initial Incidence
of a Carbon Tax Across Income Groups. National Tax Journal, 2015. 18
J. Zhou. Boomerangs versus Javelins: How Polarization Constrains Communication on Climate
Change. Environmental Politics, 2016. 24
40

Appendices
A Raw data
Table A.1: Sample characteristics: quotas stratas.
Population Sample
gender
woman 0.52 0.53
man 0.48 0.47
age
18-24 0.12 0.11
25-34 0.15 0.11
35-49 0.24 0.24
50-64 0.24 0.26
>65 0.25 0.27
profession
farmer 0.01 0.01
independent 0.03 0.04
executive 0.09 0.09
intermediate 0.14 0.14
employee 0.15 0.16
worker 0.12 0.13
retired 0.33 0.33
inactive 0.12 0.11
education
No diploma or Brevet 0.30 0.24
CAP or BEP 0.25 0.26
Bac 0.17 0.18
Higher 0.29 0.31
size of town
rural 0.22 0.24
<20k 0.17 0.18
20-99k 0.14 0.13
>100k 0.31 0.29
Paris area 0.16 0.15
region
IDF 0.19 0.17
Nord 0.09 0.10
Est 0.13 0.12
SO 0.09 0.09
Centre 0.10 0.12
Ouest 0.10 0.10
Occ 0.09 0.09
ARA 0.12 0.13
PACA 0.09 0.09
Table A.2: Households’ characteristics.
Population Sample
Household composition (mean)
Household size 2.36 2.38
Number of adults 2.03 1.93
c.u. 1.60 1.61
Energy source (share)
Gas 0.42 0.36
Fuel 0.12 0.09
Accomodation surface (m2)
mean 97 96
p25 69 66
p50 90 90
p75 120 115
Distance travelled by car (km/year)
mean 13,735 15,328
p25 4,000 4,000
p50 10,899 10,000
p75 20,000 20,000
Fuel economy (L/100 km)
mean 6.39 7.25
p25 65
p50 6.5 6
p75 7.5 7
Sources: Matched BdF; except for number of adults
(ERFS) and domestic fuel (CEREN).
41

B Questionnaire
Priming
1. [No priming] Welcome to this survey.
It was conceived by two researchers in social science. It lasts about 15-20 minutes.
2. [Info PM] Welcome to this survey.
It was conceived by two researchers in social science. It lasts about 15-20 minutes.
Before starting, please read carefully the information below on particulate matter pollution:
•particulate matter are responsible for 48,000 deaths in France each year;
•particulate matter reduce the life expectancy of French people by 9 months;
•reducing fuel consumption would reduce the health problems associated with particulate matter.
Source: France Public Health Report (2016)
3. [Info CC] Welcome to this survey.
It was conceived by two researchers in social science. It lasts about 15-20 minutes.
Please read carefully the information below on climate change.
•Climate change is already responsible for 150,000 deaths annually.
•If greenhouse gas emissions continue on their current trend, the average global warming will be
+5°C in 2100 and +8°C in 2250.
•A rapid transition to renewable energies is technically possible and would contain global warming
at +2°C.
According to scientists, in the absence of ambitious measures:
•a large proportion of species face an increased risk of extinction;
•natural disasters will intensify (hurricanes, heat waves, droughts, floods, forest fires, etc.);
42

•by 2100, 270 million more people would be flooded each year due to sea-level rise;
•violent conflicts and migration flows can be expected to increase.
Sources: Burke et al (2009),Hinkel et al (2014),IPCC Report (2014),Meinshausen et al (2011),Patz
et al (2005)
Socio-demographics
4. What is your postal code?
5. What is your gender (in the sense of civil status)?
Female; Male
6. What is your age group?
18 to 24 years old; 25 to 34 years old; 35 to 49 years old; 50 to 64 years old; 65 years old or more
7. What is your employment status?
Permanent; Temporary contract; Unemployed; Student; Retired; Other active; Inactive
8. What is your socio-professional category? (Remember that the unemployed are active workers).
Farmer; Craftsperson, merchant; Independent; Executive; Intermediate occupation; Employee;
Worker; Retired; Other Inactive
9. What is your highest degree?
No diploma; Brevet des collèges; CAP or BEP [secondary]; Baccalaureate; Bac +2 (BTS, DUT,
DEUG, schools of health and social training...); Bac +3 (licence...) [bachelor]; Bac +5 or more
(master, engineering or business school, doctorate, medicine, master, DEA, DESS...)
10. How many people live in your household? Household includes: you, your family members who live
with you, and your dependents.
11. What is your net monthly income (in euros)? All income (before withholding tax) is included
here: salaries, pensions, allowances, APL [housing allowance], land income, etc.
12. What is the net monthly income (in euros) of your household?All income (before withhold-
ing tax) is included here: salaries, pensions, allowances, APL [housing allowance], land income,
etc.
43

13. In your household how many people are 14 years old or older (including yourself)?
14. In your household, how many people are over the age of majority (including yourself)?
Energy characteristics
15. What is the surface area of your home? (in m²)
16. What is the heating system in your home?
Individual heating; Collective heating; PNR (Don’t know, don’t say)
17. What is the main heating energy source in your home?
Electricity Town gas; Butane, propane, tank gas; Heating oil; Wood, solar, geothermal, aerothermal
(heat pump); Other; PNR (Don’t know, don’t say)
18. How many motor vehicles does your household have?
None; One; Two or more
19. [Without a vehicle] How many kilometers have you driven in the last 12 months?
20. [One vehicle] What type of fuel do you use for this vehicle?
Electric or hybrid; Diesel; Gasoline; Other
21. [One vehicle] What is the average fuel economy of your vehicle? (in Liters per 100 km)
22. [One vehicle] How many kilometers have you driven with your vehicle in the last 12 months?
23. [At least two vehicles] What type of fuel do you use for your main vehicle?
Electric or hybrid; Diesel; Gasoline; Other
24. [At least two vehicles] What type of fuel do you use for your second vehicle?
Electric or hybrid; Diesel; Gasoline; Other
25. [At least two vehicles] What is the average fuel economy of all your vehicles? (in Liters per 100
km)
26. [At least two vehicles] How many kilometers have you driven with all your vehicles in the last 12
months?
44

Partial reforms [transport / housing]
27. Do you think that an increase in VAT would result in a loss of more purchasing power for your
household than for the average French household?
Yes, much more; Yes, a little more; As much as the average; No, a little less; No, a lot less; PNR
(Don’t know, don’t say)
28. Do you think that an increase in [fuel taxes / taxes on gas and heating oil] would cause your
household to lose more purchasing power than an average French household?
Yes, much more; Yes, a little more; As much as the average; No, a little less; No, a lot less; PNR
(Don’t know, don’t say)
29. The government is studying a fuel tax increase, whose revenues would be redistributed to all
households, regardless of their income. This would imply:
•[an increase in the price of gasoline by 11 cents per liter and diesel by 13 cents per liter / a 13%
increase in the price of gas, and a 15% increase in the price of heating oil];
•an annual payment of [60 / 50]eto each adult, or [120 / 100]eper year for a couple.
In terms of purchasing power, would your household be a winner or a loser with
such a measure?
Winner; Unaffected; Loser
30. [Winner selected] According to you, your household’s purchasing power would increase:
From 0 to [10·uc] eper year; From [10·uc] to [20·uc] eper year; From [20·uc] to [30·uc] eper
year; From [30·uc] to [40·uc] eper year; More than [40·uc] eper year
31. [Loser selected] According to you, the purchasing power of your household would de-
crease:
From 0 to [15·uc] eper year; From [15·uc] to [40·uc] eper year; From [40·uc] to [70·uc] eper
year; From [70·uc] to [110·uc] eper year; From [110·uc] to [160·uc] eper year; From more than
[160·uc] eper year
32. If fuel prices increased by 50 cents per liter, by how much would your household reduce its fuel
consumption?
0% - [I already consume almost none /I am already not consuming ]; 0% - [I am constrained on all
45

my trips /I will not reduce it]; From 0% to 10%; From 10% to 20%; From 20% to 30%; More than
30% - [I would change my travel habits significantly /I would change my consumption significantly]
33. In your opinion, if [fuel prices increased by 50 cents per liter / gas and heating oil prices increased
by 30%], by how much would French people reduce their consumption on average?
From 0% to 3%; From 3% to 10%; From 3% to 10%; From 10% to 20%; From 20% to 30%; More
than 30%
Tax & dividend: initial
34. The government is studying an increase in the carbon tax, whose revenues would be redistributed
to all households, regardless of their income. This would imply:
•an increase in the price of gasoline by 11 cents per liter and diesel by 13 cents per liter;
•an increase of 13% in the price of gas, and 15% in the price of heating oil;
•an annual payment of 110eto each adult, or 220eper year for a couple.
In terms of purchasing power, would your household win or loser with such a
measure?
Win; Be unaffected; Lose
35. [Winner selected] According to you, your household’s purchasing power would increase:
From 0 to [20·uc] eper year; From [20·uc] to [40·uc] eper year; From [40·uc] to [60·uc] eper
year; From [60·uc] to [80·uc] eper year; From more than [80·uc] eper year
36. [Loser selected] According to you, the purchasing power of your household would de-
crease:
From 0 to [30·uc] eper year; From [30·uc] to [70·uc] eper year; From [70·uc] to [120·uc] eper
year; From [120·uc] to [190·uc] eper year; From [190·uc] to [280·uc] eper year; From more than
[280·uc] eper year
37. [ [empty] / Scientists agree that a carbon tax would be effective in reducing pollution.] Do you
think that such a measure would reduce pollution and fight climate change?
Yes; No; PNR (Don’t know, don’t say)
46

38. In your opinion, which categories would lose [ [blank] / purchasing power] with such a measure?
(Several answers possible)
No one; The poorest; The middle classes; The richest; All French people; Rural or peri-urban
people; Some French people, but not a particular income category; PNR (Don’t know, don’t say)
39. In your opinion, what categories would gain purchasing power with such a measure? (Several
answers possible)
No one; The poorest; The middle classes; The richest; All French people; Urban dwellers; Some
French people, but not a particular income category; PNR (Don’t know, don’t say)
Tax & dividend: after knowledge
40. [Feedback] We always consider the same measure. As a reminder, it would imply:
•an increase in the price of petrol by 11 cents per liter and diesel by 13 cents per liter;
•an increase of 13% in the price of gas, and 15% in the price of heating oil;
•an annual payment of 110eto each adult, or 220eper year for a couple.
In five out of six cases, a household with the same characteristics as yours would [win / lose]. (The
characteristics taken into account are: heating with [source] for a dwelling of [size] m²; [distance] km
covered with an average consumption of [fuel economy] liters per 100 km).
Based on this estimate, do you now think that your household would be:
Winner; Unaffected; Loser
41. [Info on progressivity] On average, this measure would increase the purchasing power of poorest
households, and decrease that of the richest, who consume more energy.
In view of this new information, do you think this measure would benefit the poorest?
Yes; No; PNR (Don’t know, don’t say)
42. [No info on progressivity] Do you think this measure would benefit the poorest?
Yes; No; PNR (Don’t know, don’t say)
43. In view of the above estimate, would you approve of such a measure?
Yes; No; PNR (Don’t know, don’t say)
47

44. Why do you think this measure is beneficial? (Maximum three responses)
Contributes to the fight climate change; Reduces the harmful effects of pollution on health; Reduces
traffic congestion; Increases my purchasing power; Increases the purchasing power of the poorest;
Fosters France’s independence from fossil energy imports; Prepares the economy for tomorrow’s
challenges; For none of these reasons; Other (specify):
45. Why do you think this measure is unwanted? (Maximum three answers)
Is ineffective in reducing pollution; Alternatives are insufficient or too expensive; Penalizes rural
areas; Decreases my purchasing power; Decreases the purchasing power of some modest households;
Harms the economy and employment; Is a pretext for raising taxes; For none of these reasons;
Other (specify):
Tax & targeted dividend
46. The government is studying an increase in the carbon tax, whose revenues would be redistributed
to the [20 / 30 / 40 / 50]% of the poorest French people only. This would imply:
•an increase in the price of gasoline by 11 cents per liter and diesel by 13 cents per liter;
•an increase of 13% in the price of gas, and 15% in the price of heating oil;
•an annual payment of [550 / 360 / 270 / 220]efor each adult earning less than [780 / 1140 /
1430 / 1670]eper month (welfare benefits included, before withholding tax);
•no compensation for the others.
We estimate that in your household, [number of recipients] persons would receive this payment.
In terms of purchasing power, would your household win or lose with such a measure?
Win; Be unaffected; Lose
47. Would you approve such a measure?
Yes; No; PNR (Don’t know, don’t say)
Other questions The survey is completed by other attitudinal questions, treated in our companion
paper, Douenne & Fabre (2020). Hereafter, we only describe questions that are used in the present
paper.
48

48. Please select “A little” (test to check that you are attentive).
Not at all; A little; A lot; Completely; PNR (Don’t know, don’t say)
49. Do you smoke regularly? Yes; No
50. How much are you interested in politics?
Almost not; A little; A lot
51. How would you define yourself? (Several answers possible)
Extreme left; Left; Center; Right; Extreme right; Liberal; Conservative; Liberal; Humanist; Patriot;
Apolitical; Ecologist
52. How do you keep yourself informed of current events? Mainly through...
Television; Press (written or online); Social networks; Radio; Other
53. What do you think of the Yellow Vests? (Several answers possible)
I am part of them; I support them; I understand them; I oppose them; PNR (Don’t know, don’t
say)
54. The survey is nearing completion. You can now enter any comments, comments or suggestions in
the field below.
C Estimation for feedback
C.1 Formulas to compute monetary effects of carbon tax policy
In order to compute the monetary impact of a carbon tax increase, we decompose current energy
expenditures E(τ)as a product of current price P(τ)and current quantities consumed Q(τ), each
being a function of the excise tax τwithin which the carbon tax is comprised:19
E(τ) = P(τ)Q(τ)
All variations in expenditures can then be expressed as:
dE
E(τ) = dP
P(τ) + dQ
Q(τ)
19The French carbon tax “Contribution Climat Energie” is a component of existing taxes on energetic products:
TICPE for transport and domestic fuels, TICGN for natural gas.
49

from which we can write the effect of a price change on quantities consumed:
Qτ0=Q(τ)1 + edP
P(τ)
where e=dQ
dP ·P
Qis the price elasticity of the energetic good considered, that is here assumed constant.
For all energies, the final price can itself be decomposed as:
P(τ)=(p+iτ) (1 + t)
where tis the value added tax (VAT) rate that applies after excise taxes, ithe incidence of excise
taxes on consumers assumed constant, and p+ (i−1) τthe producer price as a function of τand for
a given value of t.20 When the carbon price changes so that the excise taxes varies from τto some
level τ0, we therefore have:
dP (τ)
P=P(τ0)−P(τ)
P(τ)=(p+iτ0) (1 + t)−(p+iτ) (1 + t)
(p+iτ) (1 + t)=i(τ0−τ)
p+iτ
Thus, following a carbon price increase, one can express the associated increase in expenditures for
each energy as:
Eτ0−E(τ) = E(τ) (1 + e)dP
P=E(τ) (1 + e)i(τ0−τ)
p+iτ
We can replicate similar calculations to obtain the expected variations in tax revenue T. Starting
from its expression — which is the sum of excise taxes and the VAT over this tax — we have:
T(τ) = Q(τ) ((1 + t)τ+t(p+ (i−1) τ))
from which we obtain:
Tτ0−T(τ) = Q(τ)1 + ei(τ0−τ)
p+iτ tp+ (i−1) τ0+ (1 + t)τ0−Q(τ) [t(p+ (i−1) τ) + (1 + t)τ]
Following the literature, we assume price elasticities of −0.4for transport fuels and −0.2for housing
energies. For the tax incidence on consumers, we assume a value of 0.8. These values were used to
compute aggregate variations in tax revenue and determine the level of lump sum transfer per adult
that a budget neutral policy would finance. When asked to estimate the impact of the policy on their
own purchasing power, respondents simply had to make an estimation over:
Eτ0−E(τ) = E(τ) (1 + e)dP
P
where for simplicity dP was given for transport fuels, and dP
Pfor housing energies. Thus, they were
not required to make any specific assumption about existing taxes or tax incidence, but simply to
estimate their consumption and price elasticity.
20Hence pis the producer price for a given value of t, when τ= 0.
50

C.2 Predicting gains and losses
We regress the increase in housing energy expenditures on households’ characteristic using EL 2013
survey. Table C.1 presents several specifications for such regression, and its last row shows the
out-of-sample error rate, computed with BdF data. All specifications yield a similar error rate of
15-17%. Fearing that respondents could make mistakes when filling the accommodation size in the
entry field, we used the first specification in our survey, as it does not rely as heavily as the others on
the accommodation size. In order to balance the error rate for losing households that are mistakenly
estimated winners, and winners who are mistakenly estimated losers, we add a constant of 16.1 in
our estimation of yearly net gain, which is thus the sum of 16.1 plus 110 times one or two (depending
on the number of adults) minus increases in transport and housing energy expenditures. We selected
OLS as our prediction method for the estimation of net gain because it compared well with respect
to alternative methods. We also classified winners and losers using a decision tree, and obtained an
very close error rate: 17.4% (see Figure C.1). Finally, statistical matching provided an error rate of
17.7%.
Table C.1: Determinants of housing energy expenditures.
Increase in housing energy expenditures (e/year)
(1) (2) (3)
Constant −55.51∗∗∗ −0.634
(1.237) (1.489)
Housing energy: Gas 124.6∗∗∗ 1.173
(1.037) (2.323)
Housing energy: Fuel oil 221.1∗∗∗ 129.8∗∗∗ 130.4∗∗∗
(1.719) (3.752) (4.002)
Accommodation size (m2) 0.652∗∗∗ 0.024
(0.012) (0.015)
Accommodation size ×Gas 1.425∗∗∗ 1.397∗∗∗
(0.007) (0.024)
Accommodation size ×Fuel oil 0.945∗∗∗ 0.922∗∗∗
(0.029) (0.032)
Observations 26,729 26,729 26,729
R20.545 0.716 0.599
Error rate 0.166 0.155 0.155
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
51

Figure C.1: Decision tree that classifies households between winners and losers.
C.3 Distributive effects
Figure C.2: Average cost of the carbon tax and dividend policy, by income decile.
Note: The definition used for the net gain is the change in fossil fuels expenditures, not the change in taxes and
transfers. Although the latter would sum to zero in aggregate because the reform is budget neutral, the former does
not because fossil fuels expenditures adjust downwards following the increase in the carbon tax.
52

D Beliefs and persistence
D.1 Elasticities
Table D.1: Effect of subjective elasticities on perceived environmental effectiveness.
Environmental effectiveness: not ‘No’
(1) (2) (3) (4)
Price elasticity: Housing −0.062∗−0.055∗
(0.032) (0.032)
Price elasticity: Transports −0.056∗−0.060∗∗
(0.030) (0.030)
Controls: Socio-demo, energy X X
incomes, estimated gains
Observations 1,501 1,501 1,501 1,501
R20.003 0.002 0.089 0.090
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
D.2 Self-interest
Table D.2: Transition matrix after telling respondents they are expected to win (75.8%).
Before \After Winner (25%) Unaffected (28%) Loser (47%)
Winner (16%) 79% 13% 8%
Unaffected (24%) 22% 63% 15%
Loser (60%) 12% 18% 70%
Table D.3: Transition matrix after telling respondents they are expected to lose (24.2%).
Before \After Winner (3%) Unaffected (12%) Loser (86%)
Winner (7%) 16% 3% 81%
Unaffected (15%) 5% 50% 46%
Loser (78%) 1% 5% 94%
53

Figure D.1: Probability that our estimation of net gains correctly predicts the winning category.
Table D.4: Share with new beliefs aligned with feedback, among those with large gain or loss (|bγ|>
110).
Aligned with feedback: GF=b
Γ
b
Γ=1 b
Γ = 0
(81.6%) (18.4%)
Initial belief winner (g > 0) 77.6% 78.4%
(19.4%) [68.5%; 84.7%] [43.2%; 94.5%]
Initial belief unaffected (g= 0) 20.7% 32.7%
(28.2%) [14.8%; 28.1%] [14.7%; 57.7%]
Initial belief loser (g < 0) 10.8% 92.2%
(52.3%) [7.3%; 15.8%] [84.5%; 96.3%]
Initial belief affected (g6= 0) 32.7% 91.1%
(70.8%) [27.7%; 38.1%] [83.5%; 95.4%]
All 28.9% 83.0%
(100%) [24.8%; 33.3%] [74.8%; 88.9%]
Note: The 95% confidence intervals for binomial probabilities are given in brackets.
54

D.3 Environmental effectiveness
Table D.5: Effect of primings on beliefs about environmental effectiveness
Environmental effectiveness
not “No” “Yes”
OLS logit OLS
(1) (2) (3) (4)
Info on Environmental Effectiveness (ZE) 0.043∗∗ 0.063∗∗∗ 0.052∗∗∗ 0.059∗∗∗
(0.017) (0.018) (0.018) (0.014)
Info on Climate Change (ZCC ) 0.044∗0.041∗0.043∗0.029
(0.024) (0.024) (0.024) (0.018)
Info on Particulate Matter (ZP M ) 0.039 0.029 0.037 0.017
(0.024) (0.024) (0.024) (0.019)
ZCC ×ZP M −0.040 −0.033 −0.042 −0.005
(0.035) (0.034) (0.033) (0.027)
Controls: Socio-demographics X X X
Observations 3,002 3,002 3,002 3,002
R20.003 0.047 0.075
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
D.4 Progressivity
Table D.6: Effect of information on perceived progressivity
Progressivity: not No (P)
(1) (2) (3)
Constant 0.419∗∗∗ 0.435∗∗∗ 0.052
(0.022) (0.033) (0.319)
Information on progressivity (ZP)−0.021 0.050 0.051
(0.027) (0.040) (0.041)
Large bias (|bγ−g|>110) −0.028 −0.040
(0.045) (0.045)
Interaction ZP×(|bγ−g|>110) −0.130∗∗ −0.117∗∗
(0.055) (0.055)
Controls: Socio-demo, politics X
Observations 1,444 1,444 1,444
R20.0004 0.018 0.094
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
55

E Robustness of motivated reasoning
There could be a reason why motivated reasoning would come from an accuracy motive (i.e. the
willingness to form a correct representation of reality) rather than a directional motive (i.e. the desire
to preserve one’s prejudices). The opponents to the reform could process new information in the
same Bayesian way as the supporters, but have different (possibly well justified) priors. As argued by
Cantoni et al. (2019), accounting for the heterogeneity in priors is essential to correctly measure the
effect of a treatment on beliefs. In our case, the only prior that matters is arguably the subjective
gain. Indeed, it could be the case that informing a respondent that they would win makes them revise
their subjective gain by, say, 100eupwards, leading most supporters of the reform to discover that
they would win, while opponents would keep thinking they would lose.
Column (2) in Table E.1 somehow validates this hypothesis, as even controlling for the bias, a
higher subjective gain is significantly associated with a higher propensity to update correctly (while
those who feel unaffected update less correctly all else being equal). However, the coefficients rela-
tive to the approval are left practically unchanged (compared to column 1) when subjective gain is
included.21 This analysis provides robust evidence in favor of directional motivated reasoning.
Finally, it is worth noticing that these results on motivated reasoning are entirely driven by the
“pessimistic winners”: we find no effect of motivated reasoning on the “optimistic losers” (columns
3-4). In other terms, the updating of people who wrongly think they win does not depend on their
approval, another indication that motivated reasoning is driven by tax aversion.
21Logically, the coefficient affected by the new specification is the binary version of subjective gain: Winner.
56

Table E.1: Asymmetric updating of winning category (complementary results).
Correct updating (U)
(1) (2) (3) (4)
Constant −0.146 −0.039 −0.110 −0.258
(0.178) (0.179) (0.181) (0.890)
Winner, before feedback ( ˙
G) 0.646∗∗∗ 0.551∗∗∗
(0.080) (0.083)
Initial tax: PNR (I don’t know) 0.163∗∗∗ 0.179∗∗∗ 0.199∗∗∗ 0.113
(0.031) (0.032) (0.033) (0.155)
Initial tax: Approves 0.158∗∗∗ 0.176∗∗∗ 0.216∗∗∗ −0.162
(0.046) (0.046) (0.049) (0.185)
Subjective gain (g) 0.0004∗∗ 0.001∗∗∗ −0.001
(0.0002) (0.0003) (0.004)
Subjective gain: unaffected (g= 0)−0.127∗∗∗ −0.208∗∗∗ −0.331
(0.033) (0.033) (0.219)
Bias about gain (g−ˆγ)−0.00005 −0.001∗−0.0003
(0.0001) (0.0003) (0.0002)
Diploma (1 to 4) 0.016 0.014 −0.001 0.148∗
(0.013) (0.013) (0.013) (0.078)
Retired 0.146∗0.130∗0.108 0.124
(0.079) (0.079) (0.080) (0.435)
Active 0.175∗∗∗ 0.166∗∗∗ 0.160∗∗∗ 0.113
(0.054) (0.054) (0.054) (0.365)
Student 0.234∗∗∗ 0.224∗∗∗ 0.183∗∗ 0.402
(0.075) (0.075) (0.074) (0.526)
Yellow Vests: PNR −0.043 −0.045 −0.031 0.013
(0.047) (0.047) (0.048) (0.246)
Yellow Vests: understands −0.063∗−0.065∗−0.059∗0.141
(0.034) (0.034) (0.034) (0.170)
Yellow Vests: supports −0.059∗−0.063∗−0.050 −0.156
(0.036) (0.036) (0.036) (0.206)
Yellow Vests: is part −0.137∗∗ −0.141∗∗ −0.106∗−0.985∗∗
(0.062) (0.061) (0.063) (0.367)
Includes “pessimistic winners” X X X
Includes “optimistic losers” X X X
Includes controls X X X X
Observations 1,365 1,365 1,265 100
R20.133 0.144 0.115 0.696
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Note: Omitted variables are Unemployed/Inactive;Yellow Vests: opposes. Cf. Appendix Hfor the list of controls.
57

F Estimation of acceptation motives
F.1 Two-stage least squares: first stage results
Table F.1: First stage regressions results for self-interest
Believes does not lose
Targeted tax (GT) After feedback (GF)
(1) (2) (5) (6)
Transfer to respondent (T1) 0.268∗∗∗ 0.227∗∗∗
(0.028) (0.027)
Transfer to spouse (T2) 0.180∗∗∗ 0.174∗∗∗
(0.031) (0.030)
T1×T2−0.190∗∗∗ −0.161∗∗∗
(0.038) (0.037)
Initial tax Acceptance (AI) 0.163∗∗∗ 0.333∗∗∗
(0.033) (0.038)
Simulated winner (b
Γ) 0.217∗∗∗ 0.210∗∗∗
(0.036) (0.035)
Controls: Incomes X X X
Controls: Estimated gain X X X
Controls: Target of the tax X X
Controls: Socio-demo, other motives X X
Effective F-Statistic 44.093 40.834 37.966 57.866
Observations 3,002 3,002 1,968 1,968
R20.082 0.177 0.131 0.319
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
58

Table F.2: First stage regressions results for environmental effectiveness
Environmental effectiveness
not “No” “Yes”
(1) (4,5)
Info on Environmental Effectiveness (ZE) 0.062∗∗∗ 0.059∗∗∗
(0.017) (0.014)
Info on Climate Change (ZCC ) 0.030∗0.028∗∗
(0.017) (0.013)
Controls: Socio-demo, other motives, X X
incomes, estimated gains
Effective F-Statistic 5.886 11.145
Observations 3,002 3,002
R20.121 0.123
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
F.2 Additional specifications
Table F.3: Effect of self-interest on acceptance: second stages of alternative specifications.
Targeted Tax After Feedback
Acceptance Approval Acceptance Approval
(1) (2) (3) (4) (5) (6)
Believes wins 0.636∗∗∗ 0.382∗∗∗ 0.764∗∗∗ 0.612∗∗∗
(0.100) (0.080) (0.251) (0.197)
Believes does not lose 0.351∗∗∗ 0.414∗∗∗
(0.073) (0.133)
Controls: Incomes X X X X X X
Controls: Estimated gain X X X
Controls: Target of the tax X X X
Observations 3,002 3,002 3,002 1,968 1,968 1,968
R20.033 0.018 0.018 0.044 0.017 0.017
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
59

G Willingness to pay
For respondents who believe in effectiveness of our Tax & Dividend, we are able to infer their willing-
ness to pay (WTP) for climate mitigation by studying the acceptance rate in function of subjective
gain. We adopt a common practice in the literature and define the WTP as the monetary loss that
the median agent is willing to incur (Hanemann,1984). Figure G.1 indicates that this WTP is about
60e/year per c.u., as this corresponds to the subjective loss below which a majority accepts the
policy. This WTP is computed only among people who believe that the tax is not ineffective, as it
would make little sense to assume that some people are willing to pay for an instrument that does
not achieve its expected goal. Indeed, Figure G.1 shows that the “WTP” of the whole sample is zero,
meaning that the median person accepts the policy only when they personally gain from it. Our
method has several advantages. First, it can be interpreted as a willingness to accept as much as a
willingness to pay, because our instrument is neither framed as a good to buy nor as a damage to be
compensated for, and net gains do not distinguish cost increases from payments received. Second,
our method is more akin to revealed preferences — and hence probably less biased (Murphy et al.,
2005) — than previous ones, because most studies directly ask respondents to select their preferred
option for climate mitigation, be it in a in a contingent valuation method (Berrens et al.,2004;
Cameron,2005;Kotchen et al.,2013) or in a discrete choice experiment (Longo et al.,2008;Alberini
et al.,2018). Still, our estimation has two notable limitations relative to the literature: it relies on a
non-representative sub-sample, and subjective gains are endogenous with acceptance.
To compare our estimation with those of the literature, expressed per household, we have to
multiply our WTP by the average number of consumption units by households: 1.6. The WTP per
household we get, 96e, lies in the typical range of the literature (Jenkins,2014;Streimikiene et al.,
2019), suggesting that the protests against carbon taxation encountered in France do not reflect
specific preferences for environmental policies.
60

Figure G.1: Acceptance rate by subjective gain, which informs one on the willingness to pay for
climate mitigation.
H Control variables
Socio-demographics: respondent’s income, household’s income, sex, age (5 categories), employ-
ment status (9 categories), socio-professional category (8 categories), region of France (10 categories),
size of town (5 categories), diploma 4 categories, household size, number of people above 14, number
of adults, number of c.u., income per c.u., smokes, favored media for news (5 categories).
Politics: extreme left, left, center, right, extreme right, interest in politics (3 categories), conservative,
liberal, humanist, patriot, ecologist, apolitical.
Political leaning: extreme left, left, center, right, extreme right, indeterminate.
Energy: heating mode (collective vs. indivual), heating energy (7 categories), annual distance trav-
elled, fuel economy, diesel (binary), gasoline (binary), number of vehicles.
Incomes: income of respondent, income of the second adult, income of respondent squared, income
of the second adult squared, dummy for absence of second adult.
Estimated gains: simulated net gain, squared simulated gain.
61