«The Strength of the Symbol: Are we. Willing to Punish Evaders?» Aurélie BONEIN Cécile BAZART. DR n

Transcription

1 «The Strength of the Symbol: Are we Willing to Punish Evaders?» Aurélie BONEIN Cécile BAZART DR n

2 The Strength of the Symbol: Are we Willing to Punish Evaders? Cécile Bazart Aurélie Bonein January 4, 2017 Abstract Starting from the observation of taxpayers heterogeneity in terms of honesty and of the existence of a perverse social dynamic leading the most honest to evade, we propose a simple mechanism to restore honesty. In a laboratory experiment, we test the willingness of subjects to voluntarily contribute to a fund dedicated to finance the fight against tax evasion. This voluntary contribution mechanism, without redistribution, is analyzed thanks to several treatments that differ depending on the type and the level of the required contribution. We show: (i) that a substantial number of subjects contribute and that the frequency of contribution decreases with the increase in the level of contribution; (ii) that the most honest taxpayers are those who contribute and (iii) that risk aversion deters subjects to contribute, while inequality aversion encourages them. Finally, in line with the literature on tax evasion, we confirm that the increase in the audit probability, which is observed only in groups where the collective contribution is sufficient, reduces evasion in these groups. Keywords: Behavioral economics; Tax evasion; Punishment JEL Classification: H26 ; C91 ; C72 ; D03 ; D6 1 Introduction Social dynamics has an important role in experimental economics, especially in terms of tax evasion. Recent works have emphasized the influence of individual s decisions on those of his peers through the frequency or the extent of tax evasion (see Alm, 2012; Cullis, Jones, and Savoia, 2012; Bazart and Bonein, 2014; Lefebvre, Pestieau, Riedl, and We acknowledge financial support from the University of Montpellier 1 under its Bonus Quality Research Program (BURS). Experiments have been conducted at the Labex-EM University Rennes 1. We acknowledge members of our laboratories for their help in the conduct of the experiment, especially Elven Priour. Finally, we acknowledge members of the LEEM (Laboratoire d Economie Expérimentale de Montpellier) for their valuable comments as well as participants at the 6th meeting of the French Economic Association of Experimental Economics (ASFEE), Paris, June 15-16, 2015, at the 32th Applied Microeconomics meeting (JMA), Montpellier, June 4-5, 2015 and at the 25th Annual Tax Research Network Conference, Hull, September 9-10, The usual disclaimer applies. LAMETA, University Montpellier 1, F Montpellier, France CREM UMR CNRS 6211, University Rennes 1, F Rennes, France 1

3 Villeval, 2015, for example). Because taxpayers differ in their degree of honesty, observing the behavior of others may induce a positive or negative dynamic in terms of declaration, whether the others are less or more honest. Justifications either in terms of conformism, social norms or reciprocity have attempted to explain this finding. In a time where the flow of information is constantly increasing and communication channels gain higher capacity, the dissemination of information about the evading behavior of others is true. Therefore, it appears crucial to seek to support the existing honesty and to avoid any spreading of evasion. The present work focuses on this goal by offering honest taxpayers an alternative to tax evasion by expressing their disapproval. In addition, the proposed mechanism could, with a constant administrative budget, 1 help to improve the efficiency of the tax administration. To this end, we conduct a laboratory experiment based on the experiment of Bazart and Bonein (2014) and the Horizontal Inequity treatment. In this experimental treatment, groups of 6 subjects played a pure declaration game without redistribution through the provision of public goods. At the end of each declaration period, players learned the average declaration made by their other group members. The authors find that providing such information may induce a positive or negative dynamics in future declarations. In order to break the negative dynamic through an increase in tax evasion, compared to the original experimental treatment, in the present experiment we add the opportunity for subjects to contribute, voluntarily, to a special fund dedicated to fight against tax evasion. When the level of contributions is high enough, controls become more frequent. So far, the question of the budget dedicated to controls has often been ignored or assumed implicit in experimental literature, except in the recent work of Hsu (2013). 2 The present experiment incorporates this concern and aims to give insights on the following points: To what extent individuals informed of the evading behavior of their peers are more willing to pay - voluntarily - to detect evasion? How honest behaviors are affected by these opportunities of punishment? To answer these questions, we propose a decentralized mechanism, simple, costless that can limit the negative dynamic in declarations, signal a disapproval and have an effective impact on the detection of evasion. Specifically, we study various types of contribution: either the subject is free to determine the amount of contribution he wants or this amount is fixed by the tax administration. In this latter case, we vary the level of contributions. 3 1 We consider that the budget dedicated to controls, that is the budget of the tax administration, is stable from one year to another. 2 Two main differences exist between the experiment of Hsu (2013) and ours: (1) Hsu (2013) studies an endogenous mechanism of audit in which subjects allocate a part of their taxes to controls repeatedly; Hsu (2013) introduce a public good financed by the redistribution of taxes. 3 Non monetary mechanisms of disapproval such as those proposed by Masclet, Noussair, Tucker, and Villeval (2003) or Coricelli, Rusconi, and Villeval (2014) could also be tested. However, they have so far been rarely implemented; for example, the publication of the name of evaders in Switzerland in 2015 (see or in the United Kingdom in 2013 ( 2

4 Since the seminal works of Yamagishi (1986), Ostrom, Walker, and Gardner (1992) and Fehr and Gächter (2000), it has been demonstrated that voluntary punishment can be effective to deter free riding in different types of public good game (see, for instance, Carpenter, Bowles, Gintis, and Hwang, 2009; Carpenter, 2007; Bochet, Page, and Putterman, 2006; Masclet, Noussair, Tucker, and Villeval, 2003; Lowen and Schmidt, 2013). However, such mechanisms have been little explored in the tax evasion literature. The aim of the present paper is to contribute to this literature by studying the propensity of individuals to punish evaders. Following the terminology used by Casari and Luini (2012), the observed punishment can be either instrumental or expressive. In the former case, the punisher values having the person targeted enduring a certain amount of punishment. In the latter case, utility from punishment comes from the personal act of punishing much more than the monetary incentives associated to it. We should thus observe symbolic contributions. However, even a symbolic contribution can be interesting because it would support honest taxpayers and its financial impact could be non-negligible if we consider its application on large size groups, as the entire population of taxpayers for example. Our results show first that a non-negligible part of subjects contribute to this fund and that frequency of contribution decreases as the cost of contribution increases. Second, subjects who contribute are those who report an income higher than the average of incomes reported by the other subjects with whom they are paired. Third, risk aversion has a negative impact on the willingness to contribute while inequality aversion has a positive effect. Finally, we confirm that an increase in the audit probability, which is observed only in groups for which the collective contribution is high enough, decreases evasion in these groups. 2 Experimental design We ran a pure declaration game, without redistribution and so without public goods, within fixed groups of 6 players who could not identify their other group members. At the beginning of each period, each participant is provided with a constant income of X = 100 points and have to pay a tax at a rate of 30%. 4 We set the fine rate in case of detected evasion at π = 350% (i.e., participants have to pay evaded taxes plus a penalty evasion-fiscale-le-fisc-britannique-publie-une-liste-de-fraudeurs.html). Although these mechanisms can be deterrent for evaders, in France, they face a strong opposition as demonstrated by the proposition of Xavier Bertrand in 2011 regarding the publication of the names of evaders with the Caisse d Allocation Familiale ( xavier-bertrand-veut-publier-dans-la-presse-le-nom-des-fraudeurs-aux-allocations_ html). 4 In order to isolate the effect of inequalities in taxes or declarations on declaration decisions, we made sure that all subjects are fiscally identical, this means that they had the same income and they faced the same tax parameters. 3

5 of 250% of unpaid taxes in the case of an audit). 5 Audits are assumed to be random, perfect and without retro-action that occur with a fixed and announced probability equal to ρ = 1/3. We used the same random sequence of audit to facilitate data comparisons between experimental sessions. At the time they make their decisions, participants have to determine the amount of income they will self-report to the tax authorities. They can choose to report any integer amount from 0 to 100. At the end of each period, they are informed about whether they have been audited, their net payoff and the average income reported by the other members of their group. If the participant has an under-reported tax liability and is audited, then a fine is imposed. Then, a new period begins. This process is repeated over 20 periods, each representing a tax year. Participants are informed that they will be paid their after-tax earnings, obtained in 5 out of the 20 periods, at the end of the entire experiment. The randomly chosen periods are the same for all participants in the same experimental session. The earned points are converted into euros at the end of the experiment and the conversion rate used was 100 points=3.80 euros. Because groups are fixed, with the repetition of the game, reciprocal behaviors may appear, through a reaction to other group members level of declaration. When other group members declare on average more (respectively less) than an individual i, this may induce this latter to increase (respectively decrease) his future declaration (see Bazart and Bonein, 2014). Negative reciprocal behaviors may therefore undermine tax collection. To break this vicious circle, at the end of the 10th reporting period, but before subjects learn if they were audited or not, they had the opportunity to contribute to a special fund dedicated to fight against tax evasion. 6 The underlying idea is that honest taxpayers may be affected when they learn that their other group members do not share their degree of honesty. By contributing to the special fund, these individuals can simultaneously indicate their type (honest) and their disapproval. Such contributions may support honest taxpayers and stigmatize dishonest taxpayers. We assume that honest taxpayers should contribute to the special fund in order to uphold their norm of honesty. The absence of public goods in the game implies that the contribution cannot be justified by any monetary payoff since in the absence of any counterpart to taxation there is no increase in the final payoff. Conversely, it is expensive because any voluntary contribution is deducted from the final payoff after payment of the potential penalty. Following Casari and Luini (2012), to fill this expressive sight of punishment, the contribution to the special fund can be minimal, in order to signal a social disapproval of tax evasion. This can be seen as the symbolic euro against underreporting. However others may seek actual punishment or effective punishment. 5 Parameters have been set such that a risk neutral subject maximized his expected utility if he declared his entire income. The condition π 1 ρ is satisfied. 6 It is noteworthy that at the beginning of the experiment, subjects are aware that a change may intervene after period 10 but they do not know what it is. They are only informed that some instructions will appear directly on their computer screen at the end of period 10, just before learning their net payoff for this period. This would help minimize the potential impact of audit at period 10 on the willingness of subjects to contribute to this fund. 4

6 We have set a threshold and when contributions, individual or collective, reach 6 points, then the administration is supposed to have enough additional funds to increase the audit probability of the group to 1/2 until the end of the game. 7 Thus the contribution may lead to a collective punishment, which can be deterrent and costly only for future tax evaders. Once all group members have determined their contribution, they learn whether the audit probability is increased or not for the 10 remaining periods; then period 11 began. From period 11, the game restarts with an audit probability of 1/3 or 1/2 if the threshold of 6 points has been reached. Other tax policy parameters as well as composition of experimental groups remain unchanged. To study the propensity of individuals to voluntary contribute, we implemented 4 experimental treatments, knowing that each subject has participated in only one of them. In the first treatment (Free) subjects were free to contribute the amount they wanted. In the three remaining treatments, the level of contributions were set at 2, 4 and 6 points (treatments Low, Medium and High, respectively). Except in the High treatment and in the Free treatment in case of a contribution equal to 6 points, the threshold could be achieved only if several subjects contributed. Therefore, the belief about the contribution of their other group members may be extremely important in the contribution decision. That is why, once all subjects had made their contribution decision, we asked them to indicate their beliefs about their other group members contribution level. 8 In order to disentangle the different motivations in the contributing decision, an experimental session included the following three experiments: (i) the pure declaration game, (ii) elicitation of individual inequality aversion (Blanco, Engelmann, and Normann, 2011), and (iii) a measure of individual risk aversion (Dohmen, Falk, Huffman, Sunde, Schupp, and Wagner, 2011). In order to avoid any order effect, the order of the three experiments has been modified in each experimental treatment. All experiments have been conducted at the LABEX-EM, University Rennes 1, using the Z-TREE software (Fischbacher, 2007). Participants have been recruited by means of the ORSEE system (Greiner, 2015). Overall 192 subjects (48 subjects per treatment) have participated. Each session lasted one hour and half and the average payoff was euros. 9 7 Because there were 6 subjects per group, a contribution such as the symbolic euro would amount to a total contribution of 6 points. So we set the threshold for an increase in audit probability at 6 points. By setting an audit probability equal to 1/2, it remains optimal for a risk-neutral taxpayer to report his entire income. 8 The elicitation of beliefs was implemented after the contributing decision and before individuals learn whether the audit probability increases or not. Furthermore, if the beliefs of a subject matched the exact contribution level of his other group members, he gets an additional payoff of 2 points. 9 A summary of individual sociodemographic characteristics of participants are presented in Appendix 1. The final payoff for a subject included the final payoff in 5 out of the 20 periods of the tax declaration game (conversion rate: 100 points = 3.80 euros), plus his payoff in the inequality aversion experiment (conversion rate: 5 points = 1 euro) and his payoff in the risk aversion experiment if the subject is randomly selected (one chance out of eight; conversion rate: 3.7 points = 1 euro). 5

7 3 Results 10,11 This section summarizes the main results of our experiment. Our first result deals with the proportion of individuals who contribute to the special fund. As Fig.2 depicts, a non-negligible part of subjects contribute to the special fund: 29.16% in the Free treatment, 27.08% in the Low treatment, 22.92% in the Middle treatment and 14.58% in the High treatment. A noticeable feature is the decrease in these frequencies when the cost of contribution increases (Chi-square tests: χ 2 = , p = between the Low and Middle treatments, χ 2 = , p < between the Middle and High treatments, χ 2 = , p < between the Low and High treatments). In addition the frequency of contribution is the highest when subjects are free to contribute the amount they want (χ 2 = , p < 0.001). In this treatment, various levels of contributions are observed: 42.86% of contributors give 1 point, 42.86% give 2 points and only 14.28% give 6 points so that their action is accompanied by a sure deterrent impact. The minimal contributions can be seen as either a symbolic signal of disapproval (an expressive motivation to punishment, as defined by Casari and Luini, 2012) or as an individual contribution to a punishment that should be socially implemented. In other words, these individuals wish to signal their disapproval without wanting to bear alone the cost of the punishment. Figure 1: Frequency of contribution by experimental treatment Our starting hypothesis was that honest taxpayers should contribute to the special fund in order to support their norm of honesty. Our results are in line with our hypoth- 10 Results obtained in the inequality aversion experiment and in the risk aversion experiment are reported in Appendices 2 and Because the first 10 periods are similar across our experimental treatments, we naturally observe no significant difference in the average declaration across them. All Mann Whitney U tests are not significant. Subjects declare on average points in the Free treatment, points in the Low treatment, in the Middle treatment and points in the High treatment. This result highlights the robustness of the declaration decisions under the same experimental design. 6

8 esis except that we have only few subjects who are perfectly honest (i.e., a perfect declaration in the first 10 periods): from none in the Free treatment, to 20.83% in the High treatment. In addition, only some of them contribute to the special fund: from 66.66% in the Low treatment to 10% in the High treatment. As depicted by Fig. 2, subjects who contribute are those who declare more on average. And this, regardless of the experimental treatment (Mann Whitney U tests: Free z = 3.187, p = ; Low z = 3.966, p = ; Middle z = 4.139, p = ; High z = 3.181, p = ). In addition, in the Free treatment, a positive correlation between the average declaration over the first 10 periods and the level of contribution is observed: for contributions of 1, 2 and 6 points, average declarations are equal to 63.93, and 90 points respectively (Spearman rank correlation coefficient: ρ = , p = ). Thus, a first explanation in the contributing decision refers to the level of declared income, and potentially, to the gap between the declaration of a given individual and the average declaration of his other group members. Note that this was observed in Bazart and Bonein (2014) where tax evasion was shown to increase by negative reciprocity. Therefore, the special fund may well be seen as an alternative to the increase in tax evasion when honest taxpayers observe the evasion of their peers. Figure 2: Average declaration by experimental group depending on the contribution decision Besides the level of declarations, intrinsic characteristics of subjects may explain their contributing decisions. Thus, individuals who express a strong aversion toward inequality may be more willing to contribute while those who express a strong risk aversion and who are not perfectly honest should not contribute to the special fund in order to 7

9 keep the audit probability constant. To confirm these conjectures, we estimate the determinants of the contribution decisions with a probit analysis. The explanatory variables can be grouped into three blocks. The first block includes the difference in declarations between the declaration of an individual i and the average declaration of his other group members (X i X i ), by distinguishing as recommended by Fehr and Gächter (2000), between advantageous (X i > X i ) and disadvantageous inequalities (X i < X i ). Because individuals have at their disposal the past declarations, we test the existence of a memory effect by including the observed difference in declarations over the last 3 periods. We also include a dummy variable that accounts for past audit experience, and this for the last 2 periods. The second block deals with individual characteristics in terms of inequality aversion and risk aversion using data obtained in the two dedicated experiments. We also include an individual value of risk attitudes that comes from answers provided in the post experimental questionnaire. The last set of independent variables refers to the judgment expressed by taxpayers in the post experimental questionnaire regarding the tax system as well as the behavior of other taxpayers. To control for the experimental conditions, we include fixed effects for the experimental treatments. Finally, more specifically in the Low and Middle treatments, the success of the contribution decision depends on the contribution decision of their other group members since any individual contribution is not sufficient to increase the audit probability. It follows that one can assume that the contribution decision of a taxpayer depends on his belief regarding the contribution decisions of his other group members. To check this assumption, we include as independent variable the belief expressed by taxpayers about the contribution decisions of their other group members. Nevertheless, proceeding in this way may cause an endogeneity bias. In particular, if taxpayers declared beliefs depend on their decisions, then elicited beliefs would be endogenous in our analysis. To address this potential endogeneity problem we conducted a two-stage IV probit regression analysis using the session and the order of the games as instruments for beliefs. Because the session and the order of the games are exogenous to the subjects, these variables should be uncorrelated with the error term in the probit regression model. 12 The results are reported in Table 1. Results show that the current declaration has a positive and significant impact on the contribution decision. As expected, those who contribute more than the average of declarations made by their other group members are more likely to contribute and it is a long lasting effect. Conversely, the disadvantageous difference in declarations has no impact on contribution decisions. These results are strengthened by the stronger impact of individual estimates of advantageous inequality aversion (measured through β) than the one of the disadvantageous inequality aversion (measured through α). We note that taxpayers who express few risk aversion or those who are risk lover are more willing to contribute. Reported estimates confirm that taxpayers will be less likely to contribute in the Middle and High treatments compared to the Free treatment, and this 12 See Bicchieri and Xiao (2009) for a more detailed explanation. 8

10 result is all the more pronounced with the increase in the cost of contribution. Finally, and as expected, those who believe that their other group members will contribute will be more likely to contribute too. Table 1: Determinants in the decision to contribute to the special fund Estimates Marginal effects Declaration 0.009** (0.004) Advantageous difference in declarations in period * (0.008) in period *** (0.008) in period ** (0.007) Disadvantageous difference in declarations in period (0.008) in period (0.008) in period * (0.007) Control=1 if Yes in period (0.378) in period (0.458) Risk and inequality aversion Personal risk index (0.096) r a 1.465*** (0.396) α b 0.198** (0.093) β c 2.575*** (0.646) Experimental treatments d Low (0.335) Middle ** (0.290) High ** (0.608) Belief 0.134* (0.077) Constant ** (2.153) Sociodemographic controls Opinions Yes Yes Yes Yes Statistics Observations 192 Pseudo R Wald χ Prob > χ Notes:,, denote statistical significance at the 1%, 5% or 10% level, respectively. Robust standard errors in parentheses, clustered at the group level. Sociodemographic controls include: age, gender (Female = 1) and level of study. a : individual level of risk preferences. b : individual parameter of disadvantageous inequality. c : individual parameter of advantageous inequality. See Appendices A et B for individual estimates for inequality and risk preferences respectively. d : Free treatment is the reference. Marginal effects are computed at the mean of the independent variables. For dummy variables it corresponds to the discrete change from the base level. After having studied the propensity of individuals to contribute to the special fund 9

11 and the determinants of contributions, we analyze the consequences thereof. The cumulative contributions exceed 6 points and the probability of audit increases in 25% of the groups for the Free and Low treatments, 37.5% of groups for the Middle treatment and 62.5% of groups for the High treatment. This result implies that for the last 10 periods, some experimental groups will face an audit probability of 1/2 and others of 1/3. In what follows, we call the experimental groups in which the audit probability remains unchanged the control group and the experimental groups in which the audit probability increases from period 11 the treatment group. Following the theoretical literature, an increase in the audit rate should yield an increase in the level of declarations (or, similarly, a decrease in tax evasion). To test this prediction, we use a difference-in-differences approach, testing how the increase in the audit probability changes the declared incomes. The validity of this methodology rests on the assumption of common trends before the change in audit probability. Figure 3: Average declarations in the control and treatment groups by experimental treatment Fig. 3 shows how the declarations of the control and treatment groups evolve during the 20 periods to provide some evidence that this assumption holds. The graph is striking. Before contribution decisions, the average declaration of the treatment group follows that of the control group reasonably closely. Thus, the control group should be a sensible approximation of how the treatment group would behaved if the audit rate has not been increased. However, after contribution decisions at the end of period 10, the average declaration of the treatment group clearly lies everywhere above that of the 10

12 control group. 13 This is consistent with the theoretical positive relationship between the level of declaration and the audit rate. Based on the results reported above, we are able to examine the changes in the level of declarations before and after contribution decisions for the treatment group relative to the control group using a difference-in-differences methodology. If the before after change for the treated group is significantly different than the average change for the control group, we infer that the increase in the audit probability has an effect on taxpayer compliance behavior. The difference-in-differences approach is a very attractive tool as it controls for factors that simultaneously affect control and treatment groups. In particular, we estimate the following equation: Y it = ηt it + ϕp it + γt it P it + ωx i + µ t + ν i + ɛ i,t (1) where Y it is the level of declaration, for individual i in period t,t is a dummy variable indicating the group to which individual i belongs; T takes value equal to 1 for the treatment group. T captures the potential differences between groups before period 10. P is a dummy variable equal to 1 for post-treatment and γ stands for the difference-indifferences coefficient. X i includes sociodemographic controls and individuals estimates for risk and inequality preferences. µ t and ν i capture the period and individual fixed effects and ɛ it is an error term, which captures any other things not included in the equation, also those we cannot observe. It must be emphasized that this analysis is particularly sensitive to the specification of the control and treatment groups. Taxpayers can endogenously fall into the treatment group due to their choice of contribution, so there is a potential source of endogeneity in eq. 1 as the dummy T can be correlated with the error term. This endogeneity problem is particularly acute in the High treatment where the contribution decision of one taxpayer is sufficient to classify himself and his belonging group in the treatment group. To control for this and as a robustness check of the results obtained for the first specification, we estimate the same model by restricting our sample to taxpayers who did not contribute to the special fund (column 2, Table 2). It results that if these taxpayers fall into the treatment group, this is not due to their own contribution decisions. Results are reported in Table 2. Observations made from Fig. 3 are confirmed since Table 2 reports a strong causal effect: declarations are higher in groups in which the audit probability has increased after period 10. In fact, the difference-in-differences estimate is positive and strongly significant in all specifications. The impact is the lowest in the Low treatment and the highest in the Free treatment. Besides, we find a period effect: all other things being equal, declarations are lower after the contribution decisions, and declarations are, on average, higher in the treatment group. 13 All Mann Whitney U tests are not significant before the opening of the special fund and they are significant at the 5% level for the Free and Low treatments and at the 1% level for the Middle and High treatments after the opening of the special fund. 11

13 Table 2: Difference-in-differences estimates All subjects Non contributors Free treat. Low treat. Middle treat. High treat. η 1.163*** 1.076*** 1.931*** *** 0.697*** 1.334*** (0.102) (0.155) (0.352) (0.168) (0.133) (0.149) ϕ *** *** ** ** (0.279) (0.330) (0.052) (0.558) (0.551) (0.311) γ 1.199*** 1.327*** *** 1.464**** 1.230*** 1.253*** (0.203) (0.309) (0.704) (0.335) (0.265) (0.299) Constant 1.397*** 8.689*** *** *** *** 8.578*** (0.192) (0.219) (0.387) (0.403) (0.369) (0.340) Statistics N R-square Notes:,, denote statistical significance at the 1%, 5% or 10% level, respectively. Standard errors in parentheses, clustered at the group level. Fixed effects for subjects and periods are included in regressions 1 to 6. Regressions 1 and 2 include also fixed effects for experimental treatments. 4 Conclusion Monetary social punishment mechanisms have so far been little explored in the literature on tax evasion, contrarily to non-monetary identification punishment or stigma. To our knowledge, in the absence of public good, only the evolutionary model of Antoci, Russu, and Zarri (2014) identifies the punishment as a supporting instrument of declarative dynamic. In the present study we demonstrate that social punishment through the contribution to a special fund dedicated to the fight against tax evasion is used by our subjects. In particular, we observe that the free and voluntary contribution increases the frequency of contribution, compared to situations in which the level of contribution is fixed by the tax administration. However, in case of free contribution, the amounts contributed are lower, corresponding mainly to minimal contributions with few cases of effective punishment, i.e., contributions that guarantee a hardening of the repressive policy. This result is in line with expressive motivations rather than instrumental ones. As expected, we observe a positive correlation between the degree of honesty and the contribution to the special fund. We also confirm that inequality aversion motivates punishment (Masclet and Villeval, 2008) and we demonstrate its long-lasting effect. Finally, even if individuals cannot observe the contribution decisions of their other group members, their beliefs about their contributing decision impact their own contribution decisions. It follows that we cannot exclude any strategic dimension in punishment. Lastly, in line with the literature on tax evasion, we confirm that an increase in audit rate, which is observed only in groups with a sufficient collective contribution, decreases tax evasion. The mechanism of punishment proposed in the present study is interesting insofar that at a national level, symbolic contribution could: allow an increase in the administrative resources, provide a signal on the acceptance of tax evasion and, as a consequence, increase the efficiency of punishment. The recent work of Xu, Cadsby, Fan, and Song (2013) confirms this point. We have therefore a lot to expect of the symbol. 12

14 References ALM, J. (2012): Measuring, explaining, and controlling tax evasion: Lessons from theory, experiments, and field studies, International Tax and Public Finance, 19(1), ANTOCI, A., P. RUSSU, AND L. ZARRI (2014): Tax evasion in a behaviorally heterogeneous society: An evolutionary analysis, Economic Modelling, 42, BAZART, C., AND A. BONEIN (2014): Reciprocal relationships in tax compliance decisions, Journal of Economic Psychology, 40, BERANEK, B., R. CUBITT, AND S. GÄCHTER (2015): Stated and revealed inequality aversion in three subject pools, Journal of the Economic Science Association, 1(1), BICCHIERI, C., AND E. XIAO (2009): Do the right thing : but only if others do so, Journal of Behavioral Decision Making, 22(2, (4)), BLANCO, M., D. ENGELMANN, AND H.-T. NORMANN (2011): A Within-Subject Analysis of Other-Regarding Preferences, Games and Economic Behavior, 72(2), BOCHET, O., T. PAGE, AND PUTTERMAN (2006): Communication and punishment in voluntary contribution experiments, Journal of Economic Behovior and Organization, 60(1), CARPENTER, J. (2007): Punishing free-riders: how group size affects mutual monitoring and the provision of public goods, Game and Economic Behavior, 60(1), CARPENTER, J., S. BOWLES, H. GINTIS, AND S. HWANG (2009): Strong reciprocity and team production: theory and evidence, Journal of Economic Behavior and Organization, 71(2), CASARI, M., AND L. LUINI (2012): Peer punishment in teams: expressive or instrumental, Experimental Economics, 15(2), CORICELLI, G., E. RUSCONI, AND M. C. VILLEVAL (2014): Tax evasion and emotions: An empirical test of re-integrative shaming theory, Journal of Economic Psychology, 40(C), CULLIS, J., P. JONES, AND A. SAVOIA (2012): Social norms and tax compliance: framing the decision to pay tax, The Journal of Socio-Economics, 41(2), DOHMEN, T., A. FALK, D. HUFFMAN, U. SUNDE, J. SCHUPP, AND G. G. WAGNER (2011): Individual Risk Attitudes: Measurement, Determinants, And Behavioral Consequences, Journal of the European Economic Association, 9(3), FEHR, E., AND S. GÄCHTER (2000): Fairness and Retaliation: The Economics of Reciprocity, Journal of Economic Perspectives, 14(3),

15 FEHR, E., AND K. M. SCHMIDT (1999): A Theory of Fairness, Competition, and Cooperation, Quarterly Journal of Economics, 114(3), FISCHBACHER, U. (2007): Z-Tree: Zurich toolbox for ready-made economic experiments, Experimental Economics, 10(2), GREINER, B. (2015): Subject pool recruitment procedures: organizing experiments with ORSEE, Journal of the Economic Science Association, 1(1), HSU, L.-C. (2013): Tax Auditing as a Public Good Game: An Experimental Study on Punishment and Compliance, Pacific Economic Review, 18(4), LEFEBVRE, M., P. PESTIEAU, A. RIEDL, AND M. C. VILLEVAL (2015): Tax evasion and social information: an experiment in Belgium, France, and the Netherlands, International Tax and Public Finance, 22(3), LOWEN, A., AND P. SCHMIDT (2013): Cooperation limitations under a one-time threat of expulsion and punishment, The Journal of Socio-Economics, 44, MASCLET, D., C. NOUSSAIR, S. TUCKER, AND M.-C. VILLEVAL (2003): Monetary and Nonmonetary Punishment in the Voluntary Contributions Mechanism, American Economic Review, 93(1), MASCLET, D., AND M.-C. VILLEVAL (2008): Punishment, inequality, and welfare: a public good experiment, Social Choice Welfare, 31(3), OSTROM, E., J. WALKER, AND R. GARDNER (1992): Covenants with and without a sword: self governance is possible, American Political Science Review, 86, XU, B., C. CADSBY, L. FAN, AND F. SONG (2013): Group size, coordination and effectiveness of punishment in the voluntary contribution mechanism: An experimental investigation, Games, 4(1), YAMAGISHI, T. (1986): The provision of a sanctioning system as a public good., Journal of Personality and Social Psychology, 51(1),

16 A Appendix A. Summary of sociodemographic characteristics of participants Table 3: Summary of sociodemographic characteristics of our sample Treatments All treatments Free treat. Low treat. Middle treat. High treat. Age Gender Men Women Level of study Undergraduate 1st year Undergraduate 2nd year Undergraduate 3rd year Postgraduate 1st year Postgraduate 2nd year Field of study Economics Management Law, Politic sciences Medicine Literature Other Place of study University Other Labor activity No Yes Notes: reported data correspond to the mean of the variable age and to the observed frequencies for the other variables. B. Experiment dedicated to the elicitation of risk aversion estimates We conducted an experiment dedicated to the estimation of individual parameters of risk aversion, following the lotteries task implemented by Dohmen, Falk, Huffman, Sunde, Schupp, and Wagner (2011). In this game, participants had to make 20 successive choices between 2 options. The first option is a safe option that varies from row to row, from 0 to 19 points, by increments of 1 point. The second option is a lottery that remains constant: there is a 50 percent chance of winning 0 point and a 50 percent chance of winning 30 points. The switching point between the two options informs us about the risk attitude of the participants. Because the expected value of the lottery is 15, risk-neutral participants should switch at 15. Risk-loving participants should choose the lottery when the offered safe option is higher than 15, while risk-averse participants should prefer the safe option for safe payments lower than 15. In this way, we are able to compute individual risk preferences parameters and subsequently see whether risk preferences and evasion decisions are related. It is noteworthy that when estimating individual parameters of risk aversion, we have considered as definitive any change from the safe option to the risky option for individuals who have proceeded to several switches. Results are depicted on Fig. 4. Recall that r < 1, = 1 and> 1 represents a risk-averse, a risk-neutral and a risk-lover individual, respectively. We observe a large heterogeneity 15

17 in our sample, from risk-averse subjects (r = ) to risk-lover ones (r = ). Figure 4: Distribution of risk aversion parameters C. Experiment dedicated to the elicitation of inequality aversion estimates We conducted an experiment dedicated to the estimation of individual parameters of inequality aversion, following Fehr and Schmidt (1999) s model. This model assumes that the utility of a player i may be written as: U i = x i α i max ( x j x i, 0 ) β i max ( x i x j, 0 ) (2) where x i is the monetary payoff of player i, x j is the monetary payoff of player j, α i is the parameter for disadvantageous inequality of player i and β i is the parameter for advantageous inequality of player i. It is assumed that α i β i. We followed the procedure of Blanco, Engelmann, and Normann (2011), whereby subjects make decisions in two different games: an ultimatum game using the strategy method and a modified dictator game. In each game, participants did not learn their role (for example, proposer or responder in the ultimatum game) until the end of the game. More precisely, the ultimatum game is used to elicit the individual parameter of disadvantageous inequality, α i. In this game, the proposer must divide 20 points between himself and the responder. Next, the responder must decide whether to accept or reject the proposition. In our experiment, all subjects decided first as a proposer and second as a responder. To avoid any feedback and to elicit the complete strategy of responders, we used the strategy method; that is, responders must decide whether to accept or reject any of the 21 possible distributions (ranging from (20, 0) to (0, 20)). The estimation of α i is obtained through the decisions of the responder i and corresponds to the switch point between rejecting and accepting the distribution. Regarding advantageous inequality, we used the modified version of the dictator game in which subjects must make decisions as a proposer by choosing between two 16

18 distributions - a non-egalitarian one (20, 0) and an egalitarian one (x i, x i ), for 21 possibilities (ranging from (0, 0) to (20, 20)). The estimate of the advantageous inequality parameter, β i, corresponds to the switch point from the unfair distribution (20, 0) to the egalitarian one (x i, x i ). The implementation of the two games allows us to determine the joint distribution of the α and β parameters. Fig. 5 depicts both individual parameters, which are found to be widely distributed in our subject pool, a finding indicative of the highly heterogeneous subject pool used in our experiment. We find a positive correlation between α i and β i (Spearman rank correlation coefficient, ρ = , p < ) and 55.21% of subjects decisions are not consistent with the hypothesis that α i β i (the corresponding data points lie below the α = β line in Fig. 5). This last finding is in line with the recent work of Beranek, Cubitt, and Gächter (2015). Figure 5: Joint distribution of inequality aversion parameters D. Results of the first-stage estimation of the IV probit on the determinants of contribution decisions Because the contribution of one individual may depend on his belief about the contribution decisions of others, an endogeneity bias may exist. To address this potential endogeneity problem we conducted a two-stage IV probit regression analysis using the session and the order of the games as instruments for beliefs. Because the session and the order of the games are exogenous to the subjects, these variables should be uncorrelated with the error term in the probit regression model. We report in Table 4 the results of the firststage estimation that aims at estimating beliefs with the session and the order of games 17

19 as independent variables. Table 4: Estimation of belief - First stage of the IV probit estimation All subjects Session *** (0.701) Order of games 1.186*** (0.293) Constant 3.337*** (0.572) Statistics Observations 189 R Prob > F Notes: denotes statistical significance at the 1% level. Robust standard errors in parentheses. 18

20 Documents de Recherche parus en 2017 DR n : Tristan LE COTTY, Elodie MAITRE D HOTEL, Raphaël SOUBEYRAN, Julie SUBERVIE «Linking Risk Aversion, Time Preference and Fertilizer Use in Burkina Faso» DR n : Cécile BAZART, Aurélie BONEIN «The Strength of the Symbol: Are we Willing to Punish Evaders?»

21 DOCUMENT DE RECHECHE LAMETA : Conditions d acceptation : Contact : garnierl@supagro.inra.fr/

22