Dishonesty and Risk-Taking: Compliance Decisions of Individuals and Groups

Unethical behavior in organizations is usually associated with the risk of negative consequences for the organization and for the involved managers if being detected. The existing experimental literature in economics has so far mainly focused on the analysis of unethical behavior in environments that involve no fines or similar monetary consequences. In the current paper, we use a tax compliance framework to study (un-)ethical behavior of individuals and small groups. Our results show that groups are clearly less compliant than individuals. The risk of being detected is the most important aspect in the group communication process when deciding on compliance.


Introduction
Consider an organization that decides on the implementation of provisions of labor laws or on the treatment of tax-relevant circumstances. Such decisions often involve a tradeoff between following the provisions tightly, or deviating from the provisions in relevant aspects to increase profits at the risk of getting detected and having to pay fines or face other forms of punishment. Three elements of such decisions in organizations are crucial: compliance with a (moral) norm, risk of detection, and joint decision making in a group or team.
Several recent examples of behaviors and decisions in organizations that produced massive media attention fall under this definition. Some car producers have allegedly pushed (over the) legal limits of measuring emissions for their diesel engines. Sport organizations have not fully ruled out unethical behavior of their officials. There are many pertinent cases of illegal collusive behavior between firms, cases of companies exploiting their dominant market position, cases of widespread tax fraud of companies (e.g., in connection with the so-called Lux Leaks or the Panama Papers), and cases of financial accounting fraud. The most severe cases make it to the public, but there is of course a continuum of norm violations in terms of severity and impact, meaning that norm-violating behavior in organization is a problem on many levels.
In this paper, we analyze the foundations of group decisions when there is a tradeoff between following a moral norm, resulting in earning smaller profits, and violating the norm, leading to higher profits, with a chance that the norm violation is detected and causing punishment. Building on recent work by, among others, Sutter (2009), Conrads et al. (2013, Gino et al. (2013), Chytilova and Korbel (2014), Muehlheusser et al. (2015), Weisel and Shalvi (2015), and Kocher et al. (2018), we implement a laboratory experiment that uses, without loss of generality, a tax compliance context. Our main innovation is the introduction of a detection probability and a penalty in case of non-compliance to a norm compliance setup that has, thus far, been studied mainly without fines or penalties, when comparing individual and group decisions.
More specifically, each decision maker -an individual or a small group -is a member (set of members) of a tax department that is responsible to file a tax declaration for the organization. Declaring less income than actually earned saves taxes and thus potentially increases profits. When non-compliance is disclosed by an audit, the organization has to pay the evaded taxes plus a penalty. In our individual (group) setting, the tax department consists 2 of one member (three members), but we keep the monetary payoff for each member the same in both conditions, given the same choices. Hence, decisions are directly comparable. Group decisions are the most straightforward implementation of an organizational setup, even though, in their simple form, they have to abstract, e.g., from hierarchies within organizations, to keep the design parsimonious. In order to retain as much experimental control as possible we implement group decisions with communication among group members that take place as anonymous real-time chats.
Our main results are as follows. First, confirming existing evidence in setups without fines or penalties for norm violations, we observe that compliance is significantly lower in the group than in the individual setting, i.e. we confirm what has been termed the individualgroup dishonesty shift (Kocher et al., 2018).
Second, arguments regarding risk-taking become very focal in the group communication. Not surprisingly, these arguments are most predictive for the outcome of the group communication, i.e. the final choice taken by the group. The focus of the chat on detection is ex-post proof for the relevance of the risk dimension when studying unethical or dishonest behavior.
Third, in line with the importance of the risk dimension, the mechanisms behind the dishonesty shift here is mainly a shift in risk tolerance of group members, in contrast to the shift in norm perception in Kocher et al. (2018). In both studies, however, it seems that common knowledge of attitudes (towards unethical behavior and towards risk) and mutual encouragement in the pursuit of non-compliant behavior are important drivers of group shifts.
Fourth, we find that group interaction induces a spill-over effect on subsequent individual compliance. Part of the shift in compliance behavior seems permanent, when former group members are asked to take a subsequent individual decision. However, we still observe that compliance is significantly higher in the individual setting after group interaction than in the group setting, suggesting that the shift in norm perception is not the only driver for the difference in behavior between groups and individuals, but that the mutual encouragement in the group (not present in the individual setting) in the non-compliance decision matters as well.
The remainder of this paper is structured as follows: In Section 2, we discuss the related literature and develop our main hypothesis. The focus will be on the economics literature on group versus individual decision making with regard to risk (e.g., Rockenbach et al., 2007;Masclet et al., 2009;Harrison et al., 2013) and with regard to unethical behavior as well as on the experimental literature on tax evasion (e.g., Torgler, 2002;Hofmann et al., 2008;Alm, 3 2012, Dulleck et al., 2016. Section 3 describes the details of our experimental design. We empirically analyze compliance behavior and treatment differences in section 4. In section 5, we study different types of decision makers and analyze the influence of individual preferences on group compliance. Arguments communicated in the group chats are examined in section 6. Section 7 concludes the paper and draws implications for situations outside the laboratory.

Related Literature and Hypothesis
The general literature on differences between individuals and small groups as decision makers is enormous. Most of the experimental literature in social psychology and economics focuses on so-called unitary groups, i.e. groups whose members have to come up with a joint decision after some form of deliberation and do not face any internal conflict in terms of monetary payoffs. However, there might be differences in preferences and attitudes. In the following, we discuss selected work that is relevant to our setup. Recent surveys that cover a broader spectrum are provided by Charness and Sutter (2012) as well as Kugler et al. (2012).

Risk-Taking Behavior of Individual versus Group Decision Makers
In the 1960s, social psychologists started to investigate the decision behavior of groups and individuals and observed a risky shift in groups, meaning that unitary groups tend to take more risk than individual decision makers (see Isenberg, 1986, for an early review). More recent papers, however, report no differences (Harrison et al., 2013) or even provide evidence for a cautious shift, implying that group decisions are more risk averse than individual decisions (Masclet et al., 2009, Bolton et al., 2015. Studies using the risk elicitation task of Holt and Laury (2002) often find that groups show both risky and cautious shifts in particular domains (Baker et al., 2008, Shupp and Williams, 2008, He et al., 2012. Compared to individuals, groups seem to be more risk averse in lotteries with a low probability of winning the largest payoff (a high risk lottery), but they are less risk averse when this probability is high (a low risk lottery). Studies investigating risky investment decisions outside the laboratory also fail to provide confirmatory evidence for a risky shift in groups. For example, Bliss et al. (2008) and Bär et al. (2011) observe that team-managed mutual funds are less riskexposed than individually-managed funds. When looking at risk-adjusted performance, Bliss et al. (2008) as well as Prather and Middleton (2002) do not provide evidence for any 4 differences between individually-managed and group/team-managed funds. In contrast, Rockenbach et al. (2007) observe that groups accumulate more expected value at lower risk, i.e. they are better at optimizing. Results are also mixed regarding the level of behavioral biases in risky decisions. Whereas Cheung and Palan (2012) and Sutter (2007) show that behavioral biases are reduced in teams, Whyte (1993) and Rau (2015) observe stronger distortions.
In the literature, three main reasons are discussed why risk-taking can differ between groups and individuals. However, their influence on risk-taking is not unambiguous, which could explain why studies fail to find a general tendency in terms of risk-taking. First, there is plenty of evidence that groups take more rational decisions than individuals in both strategic and non-strategic tasks (e.g., Bornstein and Yaniv, 1998;Bornstein et al., 2004;Sutter, 2005;Feri et al., 2010). Reasons, for example, are that groups are better at learning (Kocher and Sutter, 2005;Cooper and Kagel;2005, Fahr andIrlenbusch, 2011), reducing behavioral biases (Sutter, 2007;Cheung and Palan, 2012), avoiding extreme decisions (Bär et al., 2011), forming statistical assessments (Blinder and Morgan, 2005), allocating risk (Rockenbach et al., 2007), and they are more correct in Bayesian updating (Charness et al., 2007). Related to risk-taking decisions, more rational decision making induces less noise, but not a general shift in risky decisions.
The second argument for risk-taking differences between groups and individuals is that social responsibility might lead to more conservative risk-taking. A variety of studies observe that subjects whose risk decisions affect the payoff of others reveal a reduced willingness to take risks (e.g., Charness and Jackson, 2009;Reynolds et al., 2009;Ertac and Gurdal, 2012;Pahlke et al., 2015). Bolton et al. (2015) argue that social responsibility can operate through two channels: "either because decision makers look to avoid blame for bad outcomes or because social responsibility is equated with caution." (p. 110) Third, conformism can cause differences in risk-taking. Conformism refers to the phenomenon that individuals change their behavior to match the behavior of others (Janis, 1972;Cialdini and Goldstein, 2004;Bolton et al., 2015). Recent studies observe that individual decisions under risk can be influenced by the risk preferences of peers such as other group members (Cooper and Rege, 2011;Kocher et al., 2013;Lahno and Serra-Garcia, 2015).
Related to conformism, group polarization refers to the phenomenon that the outcome of group decision making is more extreme than the average initial tendency of the group members (Isenberg, 1986). Obviously, conformism and group polarization can increase or decrease risk-taking in groups compared to individual decision making, depending potentially 5 on whether the average initial tendencies of the group members leaned towards the risky side or the cautious side.

Lying and Cheating Behavior of Individual versus Group Decision Makers
The number of studies examining unethical behavior has recently been growing quickly.
Researchers looked at deception, lying, cheating, tax evasion, corruption, promise breaking, etc., and the vast majority of these studies use either laboratory experiments or field experiments, because field data are not easily available. Investigating differences in unethical behavior between individual decision makers and small unitary groups has attracted attention among researchers only very recently. Thus far, the focus has been on settings in which lying and cheating behavior involve no risk of being caught and punished, in contrast to the setting in the current study. When unethical behavior has no consequences, several papers on lying, cheating and deception provide evidence for a dishonesty shift in groups, meaning that groups have a stronger inclination to choose unethically than individuals (e.g., Conrads et al.;2013;Chytilová and Korbel, 2014;Bäker and Mechtel, 2015;Weisel and Shalvi, 2015;Kocher et al., 2018). However, not all studies find differences in unethical behavior between groups and individuals (Sutter, 2009;Azar et al., 2013;Muehlheusser et al., 2015), but we are not aware of any paper that provides evidence for an honesty shift from individuals to groups.
Mainly four reasons for the inclination of groups to behave more unethically than individuals have been discussed. First, groups tend to be more strategical than individuals, i.e. they figure out payoff-maximizing strategies in challenging environments more easily than individuals. This is a typical 'wisdom of the crowd' argument. Second, group membership implies the possibility of hiding behind other group members, when it comes to decision making. Hence, observability of individual actions within a group is potentially reduced compared to an individual decision making situation. As a consequence, group members might feel less individual responsibility or accountability for their actions (Mazar and Aggarwal, 2011;Conrads et al., 2013).
The third reason is that communication within a group can influence the inclination to behave unethically. Communication allows group members to exchange arguments in favor of or against certain actions. The literature suggests that learning about the preferences and attitudes of others might change norm perception. Changes in the norm (perception) might be a consequence of conformism, learning (finding arguments), or 'moral' support by other group members. Although a change in norm perception might increase or decrease unethical behavior in groups, the literature provides evidence that there is a tendency towards more 6 unethical behavior after communication, at least as long as norm violations do not have severe consequences (Gino et al., 2009;Chytilová and Korbel, 2014;Kocher et al., 2018). Fourth, recent studies suggest that groups may have a stronger inclination to behave unethically, when other people benefit from their dishonest behavior (Schweitzer and Hsee, 2002;Wiltermuth, 2011;Erat and Gneezy 2012;Gino et al., 2013;Weisel and Shalvi, 2015). When a group member's unethical behavior increases not only her own payoff but also the payoff of other group members (automatically), this might serve as justification or even a motivation for behaving unethically. 1 Others-serving unethical behavior might be judged as less immoral and seen in a more positive way than purely self-serving unethical behavior (Gino et al., 2013;Weisel and Shalvi, 2015).
An argument against a stronger inclination to behave unethically in groups is that unethical behavior in a group with communication naturally raises image concerns. The intention to behave unethically is usually observable by other group members, which might lead to social image (reputational) concerns (Bénabou and Tirole 2006;Falk and Tirole, 2016;Dufwenberg and Dufwenberg, 2018). It should be noted that it seems difficult to sustain a positive selfimage in terms of honesty, once the positive social imagine has been lost (Gino et al., 2009;Falk and Tirole, 2016).

Empirical hypothesis
The innovation of our study is the combination of the honesty dimension and the risk dimension in a context in which individuals or groups have to decide whether to behave honestly or not. The literature on lying and cheating provides rather conclusive evidence for a dishonesty shift in groups. We expect to replicate this finding. Given mixed results on a risky shift for groups, we remain agnostic with regard to the effect of a risk of being detected and punished. However, it is unclear whether the two dimensions -unethical behavior and behavior under risk -are independent or whether they might interact with each other, potentially to a different extent in individuals than in groups. For instance, the fact that group members can support each other 'morally' could re-inforce the dishonesty shift when a risk of detection is present.
We thus expect, in line with the literature, groups to be less compliant than individuals.
Hypothesis: Compliance is lower in the group than in the individual setting. 1 Such behavior is reminiscent of so-called 'white lies'. 7

Decision Task and Payoff Functions
In our experiment, each decision maker -an individual or a group -faces the decision of an employee (of a group of employees) who has to declare the income of her (their) company.
The actual income of the company, known by the employee(s), is fixed and amounts to 1,000 Lab-points. 2 The decision maker decides on how much of the actual income should be reported, and all integer values from 0 to 1,000 are allowed. The company is requested to pay a corporate tax of 25% on the reported income. 3 With a probability of 30%, the report is audited. If the audit reveals that the reported income is less than the actual income of 1,000, the company has to pay a penalty that is equivalent to twice the evaded tax, i.e. the company has to repay the evaded tax plus a fine which is equal to the tax amount evaded. If the company is not caught misreporting, there are no consequences. The company's after tax profit is… …if no audit occurs: Company's after tax profit = 1,000 -0.25 * reported income; …if an audit occurs: Company's after tax profit = 1,000 -0.25 * reported income -2 * 0.25 * (1,000 -reported income).
The decision maker's payoff is determined by a fixed remuneration of 20 Lab-points and a variable remuneration that amounts to 20% of the company's after tax profit, i.e.: Decision maker's payoff = 20 + 0.2 * company's after tax profit

Individual and Group Setting
In the individual setting, each individual provides an independent tax report. In the group setting, three participants are randomly assigned to a group. The task of the group is the same as in the individual setting (i.e. reporting the company's income). Each group member enters the amount individually. The median of the three proposals determines the income reported 2 The conversion rate from Lab-points to euros is fixed and announced at the beginning: 1 Lab-point corresponds to 0.10 Euro. 8 by the group. Group members only see the median outcome, but not the individual proposals of the others in their group. Before individual decisions are made, group members are allowed to communicate within the other two group members by sending text messages in an anonymous, five-minute chat on the computer screen, without being allowed to reveal their identity, without making side-payments, and without threatening other group members. Each decision maker in the individual setting and in the group setting receives a fixed remuneration of 20 Lab-points and a variable remuneration of 20% of the company's after tax profit; thus, the monetary incentives in the two settings are equivalent.
We implement our settings in three treatments. Each treatment consists of three parts, and each part consists of three consecutive income reporting decisions (i.e. nine decisions in total). Figure 1 provides an overview of the treatments. In treatment I-I-I, the individual (I) setting is applied in all three parts. In treatment G-G-G, the group (G) setting is applied in all three parts. In treatment I-G-I, the individual setting is applied in the first part, the group setting in the second, and the individual setting again in the third part. All subjects participate in only one of the three treatments.
The setup enables us to analyze the differences between individual and group settings in different ways. Comparing treatment I-I-I with G-G-G allows for a between-subject analysis.
Comparing the three parts in treatment I-G-I provides a within-subject analysis. As we have three income reporting decisions per part, we are able to analyze behavior over time (e.g., potential learning effects) within each part and across different parts.

Experimental Protocol
At the beginning of each part, participants receive written instructions in which all partrelated information are presented. The instructions are available in appendix A1. In the instructions for the first part, subjects are informed that the entire experiment consists of three parts in total and that each part consists of three decisions. Furthermore, participants are informed that, at the end of the experiment, for each subject one out of the nine decision situations will randomly be chosen to determine their individual payoffs.
At the beginning of the experiment, we elicit subjects' willingness to take risks with the Holt and Laury (2002) task (in euro). We use the total number of high risk lottery choices (out of 10) as our proxy for risk attitude. Consequently, subject's willingness to take risk is measured on an 11-point scale, where 0 = not willing to take risk at all, and 10 = strongly 9 willing to take risk. The lottery is resolved at the end of the experiment, and subjects learn the amount that they earned in the lottery after the main part. We obtain further information about individual characteristics of our participants (e.g., gender, age, tax morale, etc.) in a postexperimental questionnaire. At the end of the experiment, each participant receives her total payoff from the experiment plus a show-up fee of 4 euro in cash. In line with most of the tax evasion literature, the instructions are framed in terms of tax decisions. The tax frame should also add to the moral component in decision making.
In the group setting, three subjects are randomly assigned to one group. This assignment was fixed for the rest of the experiment, whenever the group setting is applied. This implies for treatment G-G-G that a subject is in the same group for the entire experiment. In treatment I-G-I, a subject stays in the same group for the three decisions of the second part. All messages sent in the chat are received by all group members, and each group member can independently decide to leave the chat. The number of messages sent is not restricted, but the chat automatically ends after five minutes. At the end of each decision situation, in all treatments, each subject is informed about the reported company income, the resulting amount of taxes, whether an audit has been carried out (including a potential penalty), and about the company's after tax profit and the subject's individual profit.
Although we use a simple setting, each participant receives a pocket calculator and a computerized "what if" calculator for her own calculations. The latter allows subjects to automatically calculate, for example, the company's after tax profit and her payoff for the outcome with or without an audit. In both the individual and group settings, the "what if" calculator is displayed when subjects decide on the reported income. In the group setting, the calculator is, in addition, also displayed during the chat stage. The experimental software was programmed with z-Tree (Fischbacher, 2007). Participants were recruited with ORSEE (Greiner, 2015).

Sample and Data
The experiment was conducted at the computerized experimental laboratory of the University of Cologne (CLER) in March and April 2017. In total, 189 subjects (mainly undergraduate students, 97 females and 92 males) participated and earned, on average, 24.46 euros in approximately 105 minutes (i.e., approximately 14 euros per hour). 48 participants were randomly assigned to treatment I-I-I, 72 to treatment G-G-G, and 69 to treatment I-G-I. Table   1 provides an overview of all our variables and presents descriptive results. [

Results: Treatment Differences
Our compliance measure is the income declared by a subject in a given period. Since the actual income was kept constant across periods and treatments (at 1,000), we can use the absolute values of declared income as our variable of interest. We are interested in how the individual willingness to report income truthfully varies across treatments; hence, we use the income each subject declared in the following analyses (if not stated differently). For our nonparametric analyses, we calculated an average per subject in the individual setting (i.e., one independent observation per subject) and an average per group in the group setting (i.e., one independent observation per group). Compliance is significantly lower in the group setting than in the individual setting.

Result 1: Compliance levels are significantly lower in the group than in the individual setting.
A similar pattern is observed in each single part. In part 1, mean declared income is 463 (N=48) in treatment I-I-I, 291 (N=24) in G-G-G, and 392 (N=69) in I-G-I. Whereas the difference between I-I-I and I-G-I is not statistically significant in the first part, the difference between I-I-I and G-G-G is significant at the 10%-level (Mann-Whitney U test; two-tailed; p=0.066), bearing in mind that we use a very conservative test. The difference between I-G-I and G-G-G is not significant (p=0.417).
In part 2, means are 462 (N=48) in I-I-I, 238 (N=24) in G-G-G, and 161 (N=23) in I-G-I. The differences between the individual and group settings are statistically significant (p=0.022 for I-I-I vs. G-G-G and p=0.004 for I-I-I vs. I-G-I). The difference between treatments I-G-I and G-G-G is not significant (p=0.552).
In part 3, mean declared income is 479 (N=48) in I-I-I, 226 (N=24) in G-G-G and 286 (N=69) in I-G-I. Again, the difference between I-I-I and G-G-G is significant (p=0.014).
However, we now also observe a significant difference between I-I-I and I-G-I (p=0.006) and no significant difference between I-G-I and G-G-G (p=0.929).
Treatment I-G-I deserves special attention. Starting out from an average declared income level of 392 in part 1 (the I-part), the level drops to 161 in part 2 (the G-part) (Wilcoxon signed-rank test; two-tailed; p<0.001; N=69). In part 3 of I-G-I (the final I-part), the level increases to 286, but stays significantly below the first I-part. 4 The level is significantly different from part 3 in I-I-I (Mann-Whitney U-test; two-tailed; p=0.006), but not compared to part 3 in G-G-G (p=0.929). Our data provide evidence for a spillover effect in treatment I-G-I from the G-part to the final I-part.
Result 2: Group interaction induces a negative spillover effect on subsequent individual compliance: individual compliance is significantly lower after a group interaction.
[ Figure 2] Figure 3 shows the histograms for treatments I-I-I and G-G-G (pooled over all parts), and Figure 4 shows the histograms for each part of treatment I-G-I. As standard in tax compliance experiments, we observe that a relatively high number of subjects chose either to report their income truthfully or to report zero income. Furthermore, we observe spikes for round values (i.e. 100, 200, etc.) in all treatments. Coordination in groups is extremely high, despite the fact that it is not required in our design, since the median proposal is implemented.
Nonetheless, almost 90% of the proposals within a group are, on average, the same.

[Figures 3 and 4]
In the following we corroborate our results by running linear regressions that take background variables and the natural correlation structure of data from group interactions into account. We use the declared income by each subject in every period as the dependent variable. As subjects face repeated decisions over several periods, we run multi-level mixed effects linear regressions to capture more than one level of dependence. 5 To account for heterogeneity across individuals and across groups, subject-specific effects, group-specific effects, and the conventional equation error term are included in the estimated equations.
Consequently, this allows us to cluster at the group and at the individual level. 6 Table 2 reports the outcome for the comparison of treatment I-I-I and G-G-G (regression coefficients, standard errors in parentheses). In model 1, we only regress on the treatment dummy "Treatment G-G-G". Since the treatment I-I-I serves as our reference, the coefficient of the treatment dummy measures the difference between treatments I-I-I and G-G-G. We 4 The differences between part 2 (161) and part 3 (286) as well as between part 1 (392) and part 3 are statistically significant (p<0.001 for part 2 vs. part 3 and p=0.011 for part 1 vs. part 3; Wilcoxon signed-rank test; two-tailed; N=69). 5 A detailed description of multi-level modelling is, for example, provided in Moffatt (2015). 6 As robustness tests, we rerun all regressions as random-effects panel regressions (panel variable: subject ID, time variable: period) with standard errors clustered on the group level. All results are robust to this variation. 12 observe a significant lower level of declared income in treatment G-G-G and therefore confirm our result 1.
To control for differences between our three parts, we additionally regress on the dummies "Part 2" and "Part 3" (which take the value of 1 if the decision was made in the respective part, 0 otherwise) in model 2. Coefficients of the interaction terms "Part 2 X Treatment G-G-G" and "Part 3 X Treatment G-G-G" measure any additional difference between treatments I-I-I and G-G-G in parts 2 and 3, respectively. Statistical significance between our two part dummies and our two interaction terms was checked by Wald tests, and the resulting p-values are reported in the bottom of the table. Again we observe a negative and significant effect of the treatment dummy as in model 1, but do not find any significant effect for the additionally included variables. The only exception is the interaction term "Part 3 X Treatment G-G-G". The coefficient is negative and significant at the 5%-level. This implies that in addition to the (negative) main treatment effect, declared income is even further decreased in the third part of treatment G-G-G compared to treatment I-I-I. This is supported by our graphical analysis. In figure 2, we show that reported income decreases over the three parts in treatment G-G-G, whereas it is almost constant in treatment I-I-I. Thus, we can conclude that reported income is generally lower in the group than in the individual setting and that this effect is even more pronounced in the third part of the experiment.
In models 3 and 4, we use the same specifications as in model 1 and 2, but further include the dummy variable "last period audit" (which takes the value of 1 if an income declaration had been audited in the previous period, and 0 otherwise) and "period (1 to 3) within part" (which denotes in which period within a respective part the decision was made, values from 1 to 3) as well as individual-specific variables such as gender, age, etc. We incorporate all 19 individual variables reported in table 1. We show the coefficient of the dummy variable "female subject" in table 2 (which takes the value of 1 if the decision was made by a female, and 0 otherwise). All other individual variables are not displayed. 7 Again, we observe the very similar results as in models 1 and 2. In line with the literature on tax compliance, we observe that women are significantly more compliant than men and that individuals are significantly less compliant if they were audited in the previous period. 8 [ Table 2] 7 The complete set of all regression results are presented in appendix A2. 8 The last result is in line with the "bomb crater effect" first observed by Mittone (2006) and further analyzed by, for example, Maciejovsky et al. (2007) and Kastlunger et al. (2009). This effect describes the tendency of subjects to decrease their compliance rates immediately after they have been audited.

13
In table 3, we use the same approach and specifications to analyze the differences between treatments I-I-I and I-G-I. In models 1 and 3, we observe a significant lower compliance level in treatment I-G-I. However, models 2 and 4 reveal that the treatment difference is only significant in parts 2 and 3, when one would expect it to be. This is indicated by the negative and significant coefficients of both interaction terms "Part 2 X Treatment I-G-I" and "Part 3 X Treatment I-G-I". In contrast, no significant difference is observed in part 1 where the individual setting is applied in both treatments (indicated by the non-significant dummy "Treatment I-G-I" in models 2 and 4). Again, compliance is lower in the group than in the individual setting (result 1).
In both models 2 and 4, we find that compliance increases from part 2 to part 3 in treatment I-G-I. This is indicated by the higher (i.e., less negative) coefficient of the interaction term "Part 3 X Treatment I-G-I" than of "Part 2 X Treatment I-G-I". Wald tests reveal that both coefficients differ significantly (see last row for the corresponding p-values). However, this increase does not compensate the large difference between both treatments occurred in part 2. Consequently, in part 3 compliance is still significantly lower in treatment I-G-I than in I-I-I (indicated by the negative and significant coefficient of the interaction term "Part 3 X Treatment I-G-I"). This provides further evidence for the discussed spill-over effect in treatment I-G-I (result 2). In line with the regression results presented in table 2, we find a positive and significant effect of "female subject" and a negative and significant effect of "last period audit".

Types of Decision Makers in Treatment I-G-I
The design of treatment I-G-I allows us to study an individual's compliance behavior before group interaction (part 1), as a group member (part 2), and after group interaction (part 3). We analyze whether we find patterns of compliance behavior that allow us to identify different types of decision makers. For categorization, we use the individual's mean declared income in each part and analyze the differences across parts. The following patterns are possible, where the first (second) term describes the transition from part 1 to 2 (from part 2 to 3): decreaseincrease, decrease-constant, decrease-decrease, increase-increase, increase-constant, increase-14 decrease, constant-increase, constant-constant, and constant-decrease. We define the transition from part 1 to 2 the following way: If the average declared income of a subject drops (increases) by at least 100 Lab-points from part 1 to 2, then this is classified as a 'decrease' ('increase'), and as 'constant' otherwise. The same logic applies to the transition from part 2 to 3. 9 Examples are shown in appendix A3.
The empirical patterns in our data are very straightforward: 88% (61 out of 69 subjects) can be assigned to three types. The relatively most frequent type is the constant-constant-type (26 subjects). 22 out of these 26 subjects can be further classified as low constant-constanttypes as their reported income did not exceed 100 in any of the three parts. Almost as many subjects are represented by the decrease-constant-type (20 subjects), following an L-shape.
The third type among the most frequent types is decrease-increase, following a V-shape, with 15 subjects being classified as such. The entire distribution of subjects in our categorization of different decision maker types is displayed in Table 4. Interestingly, we observe substantial gender differences. 51% of all men are categorized as low constant-constant-types, whereas the corresponding figure for women is only 12%. Women are predominantly decreaseincrease-(32%) and decrease-constant-types (32%). The distribution of the types of men and women is significantly different (chi-squared test; p<0.05; N=61). Women react to deciding as an individual or as a group member, whereas men do so to a lesser extent. We cannot distinguish whether this is a consequence of women being more responsive to the decision environment or whether this is a floor effect for the in part 1 already less compliant men. [ Table 4]

Group Composition in Treatment I-G-I
In this section we want to examine whether the outcome of a group in treatment I-G-I depends on the group members' individual compliance levels, individual risk preferences, and the sex composition of groups. We use the mean reported income in part 2 in treatment I-G-I (group 9 As robustness checks, we also used transition levels of 150 and 200 Lab-points. All results are qualitatively robust to this variation. 15 decisions) as the dependent variable and regress it on the group members' mean reported income in part 1 (individual decisions) in the same treatment, the mean number of risky choices of the group members in the risk elicitation (the Holt and Laury, 2002, task), and the number of female group members. Table 5 presents the regression results. Both the mean reported income from part 1 and the number of female group members 10 show up significantly in the regressions. The coefficient for risk has the correct sign, but it is far from being significant.

Arguments in the Group Chat
In this section we analyze the communication in the group chats. Two research assistants independently coded all chats using a pre-defined codebook containing all variables of interest. 11 In case of different evaluations by the two coders (which was only the case for 9.5% of all messages), a third research assistant coded the concerned message independently.
For each variable of interest, the median value across all three coders determined the coding that was finally used in our analysis. In total, we have 47 groups engaged in 285 chats. 12 We first categorize the arguments into arguments that are used to encourage compliance and arguments that are used to encourage non-compliance. As our compliance context combines the honesty dimension with the risk dimension, honesty as well as risk can be used as an argument to encourage either compliance or non-compliance. This extends the analysis of Kocher et al. (2018), who lack the risk dimension.
An argument for non-compliance was mentioned (at least once) by all 47 groups, whereas an argument for compliance was only mentioned by 23 groups (49%). We analyze which arguments are used to encourage compliance or non-compliance. The majority of arguments refer to risk, money, honesty, and taxes. We refer to risk if the message of a group member mentions risk as an argument to encourage compliance or non-compliance (e.g., "I support a risky choice, i.e. to declare zero"). Money refers to arguments associated with the 10 A robust finding in the tax compliance literature is that women are more compliant than men (e.g., Kastlunger et al., 2010;Fochmann and Wolf, 2019). At the individual level, we find support for this result (see our linear regressions in tables 3 and 4). Here, we show that group composition matters, in addition to the difference in compliance on the individual level. 11 The complete codebook with all variables of interest can be found in appendix A4. 12 Groups in treatment I-G-I have three separate group chats (three decision situations in part 2; 3 x 23 groups = 69 chats) and groups in treatment G-G-G have nine separate group chats (three parts á three decision situations; 9 x 24 groups = 216 chats). Each chat lasts five minutes.
16 monetary consequences of the compliance decision (e.g., "If we declare zero income, we receive the highest payoff."). We refer to honesty if honesty is mentioned as a norm or value in order to promote a specific behavior (e.g., "Honesty is the best policy."). 'Taxes' refers to (normative) arguments related to taxes or tax collection (e.g., "I think taxes should be paid."). Figure 5 displays the share of chats, in which these arguments are brought forward in the chats. The most frequent type of argument used is associated with risk, for both encouraging and discouraging compliance.

Result 4:
Arguments for non-compliance are made significantly more frequently than arguments for compliance. Arguments referring to risk are the most frequent arguments to encourage or discourage compliance.
[ Figure 5] Table 6 displays linear regression results with income finally declared by the group in a period as dependent variable. As independent variables, we use dummy variables indicating whether a specific type of argument is mentioned in a chat. 13 Whereas models 1 and 2 consider the general use of arguments in favor of compliance or non-compliance, models 3 and 4 distinguish between the different specific types relating to risk, money, honesty, and taxes to encourage either compliance or non-compliance. Models 2 and 4 further control for differences between treatments I-G-I and G-G-G by using a treatment dummy variable, which is 1 for treatment G-G-G, and 0 otherwise.
Not surprisingly, the use of arguments in favor of compliance significantly increases a group's compliance level, whereas the use of non-compliance arguments significantly reduces the compliance level (p<0.01 in all cases). We further find, that the magnitude of the regression coefficient for compliance is about three times as high as for non-compliance. The regression coefficients differ significantly from each other (Wald-test, p<0.001, two-tailed).
Thus, arguments for compliance have a much greater impact on the reported income by a group than arguments for non-compliance, supposedly because compliance in a group is harder to achieve.
Result 5: Arguments in the group chat used to encourage compliance significantly increase group's compliance, whereas arguments used to encourage noncompliance significantly decrease group's compliance. 13 Our results remain qualitatively unchanged if we use the frequency of each argument (i.e., how often an argument is mentioned in a chat) as independent variable.
Regressing on the different arguments separately (models 3 and 4), we find that the only arguments for non-compliance that has a significant influence on the reported group income are arguments related to risk. When it comes to arguments used to encourage compliance, arguments related to risk and money significantly increase the declared group income.

Result 6:
The influence of communication on the group's compliance behavior is mainly driven by arguments relating to risk.

Summary and Conclusion
The paper analyzes group decisions when there is a tradeoff between following a moral norm, resulting in earning smaller profits, and violating the norm, leading to higher profits. The innovation compared to existing recent work is the introduction of a chance that the norm violation is detected and that norm-violation might be punished. Decisions makers in our experiment are either individuals or groups of three members. They are thought of as a tax department at an organization, responsible for filing a tax declaration. Declaring less income than actually earned saves taxes and thus potentially increases profits. When non-compliance is disclosed by an audit, the organization has to pay the evaded taxes plus a penalty.
We confirm existing evidence from setups without fines or penalties for norm violations: groups declare significantly smaller amounts than individuals, i.e. they behave less honestly. Importantly, the risk dimension is the most important aspect in the decision making procedure within groups. It is most often discussed in the group communications, and it has an effect on the outcome of the declared group income.
We also find conclusive evidence on a spill-over effect of group decision making on subsequent individual compliance. Part of the individual-group dishonesty shift in compliance seems permanent even when former group members are asked to take a subsequent individual decision. However, we still observe that compliance is significantly higher in the final individual setting after the group interaction than in the group setting itself.
First, the replication of previous results in a different frame and with different decision making rules than in previous papers is a comforting outcome. Groups indeed seem inclined to behave more dishonestly than individuals. Second, the hitherto neglected risk domain appears to be relevant. If there is a risk of being detected, it seems that the individual-group         In this table, the results of linear regressions are presented with income reported by the group finally in a period as dependent variable (regression coefficients, standard errors in parentheses). Since groups face repeated decisions, we run linear regression models with random effects, where the period is the time variable and the group's identity number is the cross-sectional variable. *** p ≤ 0.01, ** p ≤ 0.05, * p ≤ 0.1.

A1 Instructions
Appendix A1 includes the translated instructions (from German). All participants received the general instructions in print. Before the actual experiment was executed, subjects participated in the Holt and Laury (2002) task. The instructions for this task (first experiment) were displayed on the computer screen. After that, participants received the specific instructions for each part of the actual (second) experiment in print.

A1.1 General Instructions
Thank you for participating in this experimental study. For your participation, you will receive a participation fee of 4 Euros.
The experimental study consists of 2 experiments in which you have the opportunity to earn money. Before each experiment, you will receive instructions describing each experiment. Then the experiment starts. After completing the second experiment, you will receive a payout (in addition to the participation fee) which depends on the results of both experiments.
The analysis of the experiment will be anonymous. We will never link your name with the data generated in the experiment. You will not learn the identity of any other participant, neither before nor after the experiment. Also the other participants will not learn your identity. At the end of the experiment, you have to sign a receipt to confirm the payments you received. This receipt will only be used for accounting purposes.
Before we start, we would like to draw your attention to a few important points.
• Please note that you are neither allowed to communicate with other participants nor allowed to leave your desk during both experiments. Please do not look at what other participants are doing.
• Please turn off your mobile phone and store it in your bag.
• Please read the instructions thoroughly.
• It is important that you understand the instructions. Therefore, please do not be afraid to ask questions. If you have any questions, please raise your hand. We will then come to you to answer your questions. Please do not ask questions aloud.
• You can write and make markings on the instructions.
• The calculator and the pen that are lying in front of you can be used.
• Please do not take the instructions home, but return them to us at the end of the study.
Before the first experiment starts, we ask you to fill in a short questionnaire on your computer.
After that the instructions for the first experiment will be displayed on your monitor.

A1.2 Instructions for the Holt and Laury (2002) Task
Please choose one of the two lotteries A or B in each of the following 10 decision situations.
You will make a decision for all 10 situations, but your payout from the first experiment is determined only by the one situation that is randomly drawn by the computer after the second experiment.
In each situation, you can either earn 2.00 € or 1.60 € from lottery A and either 3.85 € or 0.10 € from lottery B. The probabilities of winning, however, vary from situation to situation. The further down you move in the table, the higher is the probability of the higher payment and the lower is the probability of the lower payment.
After the first experiment and the second experiment are completed, the computer randomly draws (with the same probability) one of the 10 decision situations. After that, the computer determines your payout from the lottery that you have chosen in this decision situation by a second random draw. For that, the computer uses the probabilities for the higher payment and the lower payment according to the chosen decision situation.

General information
The second experiment consists of 3 parts. The decision situations in the 3 parts are basically identical. Before each part of the experiment, you will receive instructions explaining that part of the experiment. Each part of the experiment consists of 3 periods in which you make one decision each. In total, you make 9 decisions. At the end of the second experiment, one of the 9 decisions will be randomly selected and paid out. How much money you earn depends on your decisions and on chance. These instructions explain to you how to earn money in this experiment. Therefore, read the following paragraphs thoroughly.
For simplification purposes, this experiment does not calculate in euro amounts, but in labpoints. One lab-point is exactly 10 euro cents. That means 100 lab-points are exactly 10 euros.

Corporate employee and corporate income
Imagine you are the employee of a company. Your task is to file the tax return for the company.
As an employee, you receive a fixed remuneration of 20 lab-points. In addition, you receive a variable remuneration, which depends on the company's success. How exactly your personal payout will be calculated is explained below.
In each period, the company has earned a corporate income of 1000 lab-points.

Tax return of the company
In each period, a tax is imposed at a rate of 25 %. The tax revenues will be used to fund future experiments.
The amount of tax to be paid by the company is based on the corporate income declared by you in the tax return of the company. To do this, you simply determine how much of the actual corporate income you want to declare (in the amount of 1000 lab-points). All integer values between 0 and 1000 are possible, whereby the numbers 0 and 1000 can also be chosen. Please note: The declared corporate income can therefore be equal to or less than the actual corporate income, but not higher.
The tax payable amounts to 25 % of the declared corporate income: The declaration of the corporate income is the only decision that you make in a single period.
In the next period, the decision about the declared corporate income is made again.

Audit of tax return and corporate success
With a probability of 30 %, the provided information on the corporate income is audited.
With the counter-probability of 70 %, the information is not audited. If there is an audit and the declared corporate income does not coincide with the actual corporate income, the company has to repay the unpaid tax. In addition, the company must pay a fine equal to the amount of the unpaid tax.

Your personal payout in a period
Your personal payout in a period consists of two components. On the one hand, you receive a fixed remuneration of 20 lab-points. On the other hand, you receive a variable remuneration which depends on the company's success. The variable remuneration amounts to 20 % of the company's success. Your personal payout will be as follows: After each period, you will receive information about whether an audit has been carried out or not. In addition, you will receive an overview of all important data as well as your personal payout.

Payout
Since the second experiment consists of 3 parts, each of which consists of 3 periods, you make decisions in 9 periods. After making decisions in all 9 periods, one period is randomly selected by the computer at the end of the second experiment. The payout of this period is converted into euros and will then be paid out to you in cash.

Final information
When deciding how much corporate income you want to declare, you have the option to run trial calculations on your computer (lower half of the screen). Among other things, this will show you the resulting company's success as well as your personal payout, both in the event that no audit is carried out and that an audit is carried out. In addition, you can use the calculator at your workplace for your own calculations.
Before the second experiment starts, you are asked to answer some questions at your computer. Answering these questions is only a check of your understanding and is not payout relevant.

A1.3.1.2 Instructions for Part 2
The second part of the experiment is identical to the first part of the experiment. This means that you make the same decisions as in the first part. The second part of the experiment again consists of 3 periods in which you make one decision each.

Corporate employees and corporate income
No changes to the first part of the experiment.
Continue to imagine you are an employee of a company. Your task is to file the tax return for the company.
As in the first part of the experiment, you as an employee receive a fixed remuneration of 20 lab-points. In addition, you receive a variable remuneration, which depends on the company's success.
In each period, the company has earned a corporate income of 1000 lab-points.

Tax return of the company
No changes to the first part of the experiment.
Therefore, in each period, a tax is again imposed at a rate of 25 %.
The amount of tax to be paid by the company continues to be based on the corporate income declared by you in the tax return of the company. To do this, you simply determine how much of the actual corporate income (which is 1000 lab-points) you want to declare. All integer values between 0 and 1000 are possible, whereby the numbers 0 and 1000 can also be chosen. Please note: The declared corporate income can therefore be equal to or less than the actual corporate income, but not higher.
The tax payable amounts to 25 % of the declared corporate income: The declaration of the corporate income is the only decision that you make in a single period.
In the next period, the decision about the declared corporate income is made again.

Audit of tax declaration and corporate success
No changes to the first part of the experiment.
It therefore continues to apply that the provided information on the corporate income is audited with a probability of 30 %. If there is an audit and the declared corporate income does not coincide with the actual corporate income, the company has to repay the unpaid tax. In addition, as in the first part of the experiment, the company must pay a fine equal to the unpaid tax.
Therefore, the company's success continues to result in the case of an audit as follows:  actual corporate income company's success 1000 -tax -tax repayment -fine = The company's success results in the case of no audit as follows:  actual corporate income company's success 1000 -tax =

Your personal payout in a period
No changes to the first part of the experiment.
Your personal payout in a period continues to consist of two components. On the one hand, you receive a fixed remuneration of 20 lab-points. On the other hand, you receive a variable remuneration which depends on the company's success. The variable remuneration amounts to 20 % of the company's success. Your personal payout will be as follows: After each period, you will receive information about whether an audit has been carried out or not. In addition, you will receive an overview of all important data as well as your personal payout.

Payout
No changes to the first part of the experiment.
It therefore continues to apply that at the end of the second experiment (after making decisions in all 9 periods), one period is randomly selected by the computer. The payout of this period is converted into euros and will then be paid out to you in cash.

Final information
No changes to the first part of the experiment.
It therefore continues to apply that when deciding how much corporate income you want to declare, you have the option to run trial calculations on your computer (lower half of the screen).

A1.3.1.3 Instructions for Part 3
The third part of the experiment is identical to the first and second part of the experiment. This means that you make the same decisions as in the first and second part. The third part of the experiment again consists of 3 periods in which you make one decision each.

A1.3.2.1 Instructions for Part 1 General information
The second experiment consists of 3 parts. The decision situations in the 3 parts are basically identical. Before each part of the experiment, you will receive instructions explaining that part of the experiment. Each part of the experiment consists of 3 periods in which you make one decision each. In total, you make 9 decisions. At the end of the second experiment, one of the 9 decisions will be randomly selected and paid out. How much money you earn depends on your decisions and on chance. These instructions explain to you how to earn money in this experiment. Therefore, read the following paragraphs thoroughly.
For simplification purposes, this experiment does not calculate in euro amounts, but in labpoints. One lab-point is exactly 10 euro cents. That means 100 lab-points are exactly 10 euros.

Corporate employee and corporate income
Imagine you are the employee of a company. Your task is to file the tax return for the company.
As an employee, you receive a fixed remuneration of 20 lab-points. In addition, you receive a variable remuneration, which depends on the company's success. How exactly your personal payout will be calculated is explained below.
In each period, the company has earned a corporate income of 1000 lab-points.

Tax return of the company
In each period, a tax is imposed at a rate of 25 %. The tax revenues will be used to fund future experiments.
The amount of tax to be paid by the company is based on the corporate income declared by you in the tax return of the company. To do this, you simply determine how much of the actual corporate income you want to declare (in the amount of 1000 lab-points). All integer values between 0 and 1000 are possible, whereby the numbers 0 and 1000 can also be chosen. Please note: The declared corporate income can therefore be equal to or less than the actual corporate income, but not higher.
The tax payable amounts to 25 % of the declared corporate income: The declaration of the corporate income is the only decision that you make in a single period.
In the next period, the decision about the declared corporate income is made again.

Audit of tax return and corporate success
With a probability of 30 %, the provided information on the corporate income is audited.
With the counter-probability of 70 %, the information is not audited. If there is an audit and the declared corporate income does not coincide with the actual corporate income, the company has to repay the unpaid tax. In addition, the company must pay a fine equal to the amount of the unpaid tax.
tax repayment = unpaid tax

Fine = unpaid tax
The unpaid tax is:

Your personal payout in a period
Your personal payout in a period consists of two components. On the one hand, you receive a fixed remuneration of 20 lab-points. On the other hand, you receive a variable remuneration which depends on the company's success. The variable remuneration amounts to 20 % of the company's success. Your personal payout will be as follows: After each period, you will receive information about whether an audit has been carried out or not. In addition, you will receive an overview of all important data as well as your personal payout.

Payout
Since the second experiment consists of 3 parts, each of which consists of 3 periods, you make decisions in 9 periods. After making decisions in all 9 periods, one period is randomly selected by the computer at the end of the second experiment. The payout of this period is converted into euros and will then be paid out to you in cash.

Final information
When deciding how much corporate income you want to declare, you have the option to run trial calculations on your computer (lower half of the screen). Among other things, this will show you the resulting company's success as well as your personal payout, both in the event that no audit is carried out and that an audit is carried out. In addition, you can use the calculator at your workplace for your own calculations.
Before the second experiment starts, you are asked to answer some questions at your computer. Answering these questions is only a check of your understanding and is not payout relevant.

A1.3.2.2 Instructions for Part 2
The second part of the experiment is identical to the first part of the experiment. The only exception is that you now make your decisions in a triad. Your remuneration therefore also depends on the decisions of other participants.
The second part of the experiment again consists of 3 periods in which you make one decision each.

Group
Together with 2 other, randomly selected participants, you form a triad that stays together during the second part of the experiment. Each of these 3 group members makes the same decisions.

Corporate employees and corporate income
Imagine you and the other two members of your group are employees of a company. Your common task is to file the tax return for the company.
As in the first part of the experiment, each employee receives a fixed remuneration of 20 labpoints. In addition, each employee receives a variable remuneration, which depends on the company's success.
In each period, the company has earned a corporate income of 1000 lab-points.

Tax return of the company
Therefore, in each period, a tax is again imposed at a rate of 25 %.
The amount of tax to be paid by the company is based on the corporate income, which your group declares in the tax return of the company. The group decides by vote on the amount of the declared corporate income. For this purpose, each individual group member makes a personal proposal of how much of the actual corporate income (which is 1000 lab-points) should be declared. As a proposal all integer values between 0 and 1000 are possible, whereby the numbers 0 and 1000 can also be chosen. Please note: The declared corporate income can therefore be equal to or less than the actual corporate income, but not higher.
The median of the proposals of the three group members determines the amount of the declared corporate income of your group in this period. The median is the value that stands at the middle (central) location when sorting the values by size (from small to large). This also means that if two or three group members propose the same value, this proposed value is the median. Please note that the median is not the same as the mean. After each member of the group has made his binding proposal, the median is automatically determined and the amount of the declared corporate income specified.
The tax payable is 25 % of the declared corporate income: The declaration of the corporate income is the only decision that you and the two other members of your group make in a single period. In the next period, the decision about the declared corporate income is made again.
Before each member submits his binding proposal, the three group members can communicate in writing for a maximum of 5 minutes in a chat. More information about the chat can be found on the last page of these instructions.

Audit of tax declaration and corporate success
No changes to the first part of the experiment.
It therefore continues to apply that the provided information on the corporate income is audited with a probability of 30 %. If there is an audit and the declared corporate income does not coincide with the actual corporate income, the company has to repay the unpaid tax. In addition, as in the first part of the experiment, the company must pay a fine equal to the unpaid tax.
Therefore, the company's success continues to result in the case of an audit as follows:

Your personal payout in a period
No changes to the first part of the experiment.
Your personal payout in a period continues to consist of two components. On the one hand, you receive a fixed remuneration of 20 lab-points. On the other hand, you receive a variable remuneration which depends on the company's success. The variable remuneration amounts to 20 % of the company's success. Your personal payout will be as follows: • Since your personal payout depends on the company's success, it also depends on the tax and (possible) fine paid by the company.
• Each member of your group receives the same payout.
After each period, each group member will receive information about whether an audit has been carried out or not. In addition, each member receives an overview of all important data as well as the personal payout.

Payout
No changes to the first part of the experiment.
It therefore continues to apply that at the end of the second experiment (after making decisions in all 9 periods), one period is randomly selected by the computer. The payout of this period is converted into euros and will then be paid out to you in cash.

Final information
While chatting and deciding how much corporate income you want to declare, you have the option to run trial calculations on your computer (lower half of the screen).

Information about the chat
You have the option of communicating with the other two members of your group through a chat to discuss the proposal on the amount of declared corporate income that each group member will subsequently enter.
You have 5 minutes to exchange information. The group discussion ends after 5 minutes or as soon as all 3 group members have clicked the button "Leave Chat". If 1 or 2 group members click on the button, the chat will continue until either all group members have clicked on the button or the time has expired. If you have clicked on the button "Leave Chat", but do not want to leave the chat, you can click on the button "Back". After the group discussion, each member makes a binding proposal on the amount of the declared corporate income.
Basically, the content of the communication is open, but it is forbidden to share personal information. Personal data is: name, age, gender (please always use gender-neutral terms), subject (this includes the mentioning of specific lecturers, courses or course descriptions, that allow for identification of the subject) or similar topics that could identify you (e. g. your cabin number or row). Furthermore, it is prohibited to agree on side payments within your group. If you violate these rules, you will be excluded from the experiment and will not receive a payout for the entire experiment.
During the given time each group member can send as many messages as he likes. Each of your messages automatically appears on the screen of the other two group members.
Messages to a single person are not possible.
The screen with the chat will look like this:

40
To send a message, click on the purple box, type in your message and press the "enter" key. After that, your message will appear in the gray box above. This procedure allows you to send as many messages as you want. The other group members see your messages only when you hit the "enter" key, that is, when your message appears in the gray box.

A1.3.2.3 Instructions for Part 3
The third part of the experiment is identical to the first part of the experiment. This means that you make the same decisions as in the first part. Please note, therefore, that you make the decisions on your own and not in a group anymore. The third part of the experiment again consists of 3 periods in which you make one decision each.

General information
The second experiment consists of 3 parts. The decision situations in the 3 parts are basically identical. Before each part of the experiment, you will receive instructions explaining that part of the experiment. Each part of the experiment consists of 3 periods in which you make one decision each. In total, you make 9 decisions. At the end of the second experiment, one of the 9 decisions will be randomly selected and paid out. How much money you earn depends on your decisions, the decisions of other participants, and on chance. These instructions explain to you how to earn money in this experiment. Therefore, read the following paragraphs thoroughly.
For simplification purposes, this experiment does not calculate in euro amounts, but in labpoints. One lab-point is exactly 10 euro cents. That means 100 lab-points are exactly 10 euros.

Group
Together with 2 other, randomly selected participants, you form a triad that stays together during the first part of the experiment. Each of these 3 group members makes the same decisions.

Corporate employees and corporate income
Imagine you and the other two members of your group are employees of a company. Your common task is to file the tax return for the company.
Each employee receives a fixed remuneration of 20 lab-points. In addition, each employee receives a variable remuneration, which depends on the company's success. How exactly your personal payout will be calculated is explained below.
In each period, the company has earned a corporate income of 1000 lab-points.

Tax return of the company
In each period, a tax is imposed at a rate of 25 %. The tax revenues will be used to fund future experiments.
The amount of tax to be paid by the company is based on the corporate income, which your group declares in the tax return of the company. The group decides by vote on the amount of the declared corporate income. For this purpose, each individual group member makes a personal proposal of how much of the actual corporate income (in the amount of 1000 labpoints) should be declared. As a proposal all integer values between 0 and 1000 are possible, whereby the numbers 0 and 1000 can also be chosen. Please note: The declared corporate income can therefore be equal to or less than the actual corporate income, but not higher.
The median of the proposals of the three group members determines the amount of the declared corporate income of your group in this period. The median is the value that stands at the middle (central) location when sorting the values by size (from small to large). This also means that if two or three group members propose the same value, this proposed value is the median. Please note that the median is not the same as the mean. After each member of the group has made his binding proposal, the median is automatically determined and the amount of the declared corporate income specified.
The tax payable is 25 % of the declared corporate income: The declaration of the corporate income is the only decision that you and the two other members of your group make in a single period. In the next period, the decision about the declared corporate income is made again.
Before each member submits his binding proposal, the three group members can communicate in writing for a maximum of 5 minutes in a chat. More information about the chat can be found on the last page of these instructions.

Audit of tax declaration and corporate success
With a probability of 30 %, the provided information on the corporate income is audited.
With the counter-probability of 70 %, the information is not audited. If there is an audit and the declared corporate income does not coincide with the actual corporate income, the company has to repay the unpaid tax. In addition, the company must pay a fine equal to the unpaid tax.
tax repayment = unpaid tax The unpaid tax is:

Your personal payout in a period
Your personal payout in a period consists of two components. On the one hand, you receive a fixed remuneration of 20 lab-points. On the other hand, you receive a variable remuneration which depends on the company's success. The variable remuneration amounts to 20 % of the company's success. Your personal payout will be as follows: • Since your personal payout depends on the company's success, it also depends on the tax and (possible) fine paid by the company. • Each member of your group receives the same payout. After each period, each group member will receive information about whether an audit has been carried out or not. In addition, each member receives an overview of all important data as well as the personal payout.

Payout
Since the second experiment consists of 3 parts, each of which consists of 3 periods, you make decisions in 9 periods. After making decisions in all 9 periods, one period is randomly selected by the computer at the end of the second experiment. The payout of this period is converted into euros and will then be paid out to you in cash.

Final information
While chatting and deciding how much corporate income you want to declare, you have the option to run trial calculations on your computer (lower half of the screen). Among other things, this will show you the resulting corporate success as well as your personal payout, both in the event that no audit is carried out and that an audit is carried out. In addition, you can use the calculator at your workplace for your own calculations.
Before the second experiment starts, you are asked to answer some questions on your computer. Answering these questions is only a check of your understanding and is not payout relevant.

Information about the chat
You have the option of communicating with the other two members of your group through a chat to discuss the proposal on the amount of declared corporate income that each group member will subsequently enter.
You have 5 minutes to exchange information. The group discussion ends after 5 minutes or as soon as all 3 group members have clicked the button "Leave Chat". If 1 or 2 group members click on the button, the chat will continue until either all group members have clicked on the button or the time has expired. If you have clicked on the button "Leave Chat", but do not want to leave the chat, you can click on the button "Back". After the group discussion, each member makes a binding proposal on the amount of the declared corporate income.
Basically, the content of the communication is open, but it is forbidden to share personal information. Personal data is: name, age, gender (please always use gender-neutral terms), subject (this includes the mentioning of specific lecturers, courses or course descriptions, that allow for identification of the subject) or similar topics that could identify you (e. g. your cabin number or row). Furthermore, it is prohibited to agree on side payments within your group. If you violate these rules, you will be excluded from the experiment and will not receive a payout for the entire experiment.
During the given time each group member can send as many messages as he likes. Each of your messages automatically appears on the screen of the other two group members. Messages to a single person are not possible.
The screen with the chat will look like this:

44
To send a message, click on the purple box, type in your message and press the "enter" key. After that, your message will appear in the gray box above. This procedure allows you to send as many messages as you want. The other group members see your messages only when you hit the "enter" key, that is, when your message appears in the gray box.

A1.3.3.2 Instructions for Part 2
The second part of the experiment is identical to the first part of the experiment. This means that you make the same decisions as in the first part. The second part of the experiment again consists of 3 periods in which you make one decision each.

Group
Please note that your triad consists of the same group members as in the first part of the experiment and that you therefore interact again in the second part of the experiment with the same participants. Each of the 3 group members makes the same decisions again.

Corporate employees and corporate income
No change to the first part of the experiment.
Continue to imagine you and the other two members of your group are employees of a company. Your common task is to file the tax return for the company.
As in the first part of the experiment, each employee receives a fixed remuneration of 20 labpoints. In addition, each employee receives a variable remuneration, which depends on the company's success.
In each period, the company has earned a corporate income of 1000 lab-points.

Tax return of the company
No changes to the first part of the experiment.
Therefore, in each period, a tax is again imposed at a rate of 25 %.
The amount of tax to be paid by the company continues to be based on the corporate income, which your group declares in the tax return of the company. The group decides by vote on the amount of the declared corporate income. For this purpose, each individual group member makes a personal proposal of how much of the actual corporate income (which is 1000 labpoints) should be declared. As a proposal all integer values between 0 and 1000 are possible, whereby the numbers 0 and 1000 can also be chosen. Please note: The declared corporate income can therefore be equal to or less than the actual corporate income, but not higher.
The median of the proposals of the three group members determines the amount of the declared corporate income of your group in this period. The median is the value that stands at the middle (central) location when sorting the values by size (from small to large). This also means that if two or three group members propose the same value, this proposed value is the median. Please note that the median is not the same as the mean. After each member of the group has made his binding proposal, the median is automatically determined and the amount of the declared corporate income specified.
The tax payable is 25 % of the declared corporate income: The declaration of the corporate income is the only decision that you and the two other members of your group make in a single period. In the next period, the decision about the declared corporate income is made again.
Before each member submits his binding proposal, the three group members can communicate in writing for a maximum of 5 minutes in a chat. More information about the chat can be found on the last page of these instructions.

Audit of tax declaration and corporate success
No changes to the first part of the experiment.
It therefore continues to apply that the provided information on the corporate income is audited with a probability of 30 %. If there is an audit and the declared corporate income does not coincide with the actual corporate income, the company has to repay the unpaid tax. In addition, as in the first part of the experiment, the company must pay a fine equal to the unpaid tax.
Therefore, the company's success continues to result in the case of an audit as follows:

Your personal payout in a period
No changes to the first part of the experiment.

46
Your personal payout in a period continues to consist of two components. On the one hand, you receive a fixed remuneration of 20 lab-points. On the other hand, you receive a variable remuneration which depends on the company's success. The variable remuneration amounts to 20 % of the company's success. Your personal payout will be as follows:  fixed remuneration variable remuneration payout in a period 20 20 % of the company's success = +  Please note: • Since your personal payout depends on the company's success, it also depends on the tax and (possible) fine paid by the company. • Each member of your group receives the same payout. After each period, each group member will receive information about whether an audit has been carried out or not. In addition, each member receives an overview of all important data as well as the personal payout.

Payout
No changes to the first part of the experiment.
It therefore continues to apply that at the end of the second experiment (after making decisions in all 9 periods), one period is randomly selected by the computer. The payout of this period is converted into euros and will then be paid out to you in cash.

Final information
No changes to the first part of the experiment.
It therefore continues to apply that while chatting and deciding how much corporate income you want to declare, you have the option to run trial calculations on your computer (lower half of the screen).

Information about the chat
You have the option of communicating with the other two members of your group through a chat to discuss the proposal on the amount of declared corporate income that each group member will subsequently enter.
You have 5 minutes to exchange information. The group discussion ends after 5 minutes or as soon as all 3 group members have clicked the button "Leave Chat". If 1 or 2 group members click on the button, the chat will continue until either all group members have clicked on the button or the time has expired. If you have clicked on the button "Leave Chat", but do not want to leave the chat, you can click on the button "Back". After the group discussion, each member makes a binding proposal on the amount of the declared corporate income.
Basically, the content of the communication is open, but it is forbidden to share personal information. Personal data is: name, age, gender (please always use gender-neutral terms), subject (this includes the mentioning of specific lecturers, courses or course descriptions, that allow for identification of the subject) or similar topics that could identify you (e. g. your cabin number or row). Furthermore, it is prohibited to agree on side payments within your group. If you violate these rules, you will be excluded from the experiment and will not receive a payout for the entire experiment.
During the given time each group member can send as many messages as he likes. Each of your messages automatically appears on the screen of the other two group members. Messages to a single person are not possible.
The screen with the chat will look like this: To send a message, click on the purple box, type in your message and press the "enter" key. After that, your message will appear in the gray box above. This procedure allows you to send as many messages as you want. The other group members see your messages only when you hit the "enter" key, that is, when your message appears in the gray box.

A1.3.3.3 Instructions for Part 3
The third part of the experiment is identical to the first and second part of the experiment. This means that you make the same decisions as in the first and second part. The third part of the experiment again consists of 3 periods in which you make one decision each.
Please note that your triad consists of the same group members as in the first and second part of the experiment and that you therefore interact again in the third part of the experiment with the same participants.

A2 Regressions
This section presents the complete set of regression results. Table A2.1 corresponds to Table  2 and Table A2.2 to Table 3.  Wald test: Part 2 = Part 3 p = 0.5806 p = 0.52492 Part 2 X Treatment I-G-I = Part 3 X Treatment I-G-I p = 0.0061 p = 0.0026 Note: In this table, the results of multi-level mixed effects linear regressions are presented with declared income as dependent variable (regression coefficients, standard errors in parentheses). *** p ≤ 0.01, ** p ≤ 0.05, * p ≤ 0.1.

A4 Codebook
This section presents the codebook that was used by the coders.
Arguments used: • Risk Risk discussed as an argument in the group chat (in general) o Risk_compliance Risk discussed as an argument in favor of compliance Example: "I do not want to take any risks now", "I want to play it safe" o Risk_noncompliance Risk discussed as an argument in favor of non-compliance Example: "I favor to be risky and to declare 0" • Money Money discussed as an argument in the group chat (in general), arguments resting on the monetary consequences of the compliance decision o Money_compliance Money discussed as an argument in favor of compliance Example: "We gain quite a lot if we report honestly" o Money_noncompliance Money discussed as an argument in favor of non-compliance Example: "If we declare 0 income, we receive the highest payoff" • Honesty Honesty discussed as an argument in the group chat (in general), honesty mentioned as a norm or value o Honesty_compliance Honesty discussed as an argument in favor of compliance Example: "In my tax return I'm honest", "Honesty is the best policy" o Honesty_noncompliance Honesty discussed as an argument in favor of non-compliance Example: "Let's deceive" • Taxes Taxes discussed as an argument in the group chat (in general), arguments related to taxes or tax collecting o Taxes_compliance Taxes discussed as an argument in favor of compliance Example: "I think taxes should be paid" o Taxes_noncompliance Taxes discussed as an argument in favor of non-compliance Example: "The state does not receive anything"