Legal Compliance and Detection Avoidance: Results on the Impact of Different Law-Enforcement Designs

1 Introduction

The punishment of violations is often understood as pricing non-compliance (e.g. Cooter 1984). Much of the law & economics literature builds on the hypothesis that potential offenders do not care about how fine revenue is used or why sanctions are imposed (e.g. Polinsky and Shavell 2001).^[1] For example, in the literature on corruption, an individual is usually portrayed as indifferent between paying a bribe and paying a fine as long as these payments are equally high (e.g. Buccirossi and Spagnolo 2006). This reasoning culminates in the argument that corruption is not detrimental to legal compliance as long as the expected bribe does not undercut the expected sanction (e.g. Polinsky and Shavell 2001).

Moreover, in the literature on optimal law enforcement, potential offenders are assumed to be indifferent regarding the enforcement agent’s identity. A strand of the literature, including Becker and Stigler (1974), Polinsky (1980), and Garoupa (1997), considers the possibility of private law enforcement and its desirability relative to public law enforcement. In private law enforcement, profit-maximizing firms detect suspects and may be rewarded by the fine convicts pay. A standard assumption in this literature is that the potential offenders’ violation decisions are not influenced by the identity of the person who receives the fine revenue or influences the enforcement effort.

We present evidence from three controlled laboratory experiments on whether potential offenders behave similarly in different law-enforcement designs. Concerning the law-enforcement designs, we vary (i) whether the fine is used to compensate the victim or as a transfer to the enforcer and (ii) whether an offender’s detection is a purely random and exogenous event or the result of a process somehow endogenously influenced by the enforcer. A scenario where the fine does not benefit the enforcer and the offenders’ detection is perceived as an exogenous event may be interpreted as a proxy of the traditional public law-enforcement model (e.g. Polinsky and Shavell 2006). In contrast, private law enforcement may be understood as a scenario where the enforcer receives the fine revenue and saliently influences the detection outcome.

Concerning choices, one of our studies focuses on the potential offender’s violation choice which consists of taking points from another participant in all of our experiments. This is the traditional focus in the experimental literature on criminal decision-making (Engel 2018). In contrast, two other studies additionally consider the willingness to invest in detection avoidance. While practically very relevant (e.g. Sanchirico 2006), detection avoidance has been neglected in the experimental law & economics literature to date.^[2]

Building on the traditional optimal law-enforcement framework, we predict that potential offenders’ behavior is comparable in our different law-enforcement designs. However, the various law-enforcement designs represent substantial differences. For example, transferring fine revenue to the victim emphasizes restorative instead of retributive justice (e.g. Canton 2017). Evidence suggests that retributive justice often guides peoples’ punishment intuitions (e.g. Darley 2009). As a result, a compensatory sanction may be perceived as less severe than a retributive sanction of the same magnitude. Relatedly, Mulder (2018) asserts that compensatory sanctions are less capable of conveying moral norms than retributive sanctions. The lack of moral activation would suggest that a compensatory sanction should be less deterring than a retributive sanction of equal magnitude. Moreover, people usually find compensatory payments unsatisfactory as a response to harm imposed intentionally instead of carelessly (e.g. Darley and Pittman 2003). However, a regime that transfers the fine to the victim may increase the salience of the victim’s harm, thereby creating a moral suasion effect (e.g. Deffains, Espinosa, and Fluet 2019). This would suggest that the compensatory sanction should have a more significant deterrence effect.

The distinction between the scenario in which fine revenue is received by the victim and the one where the enforcer receives it could interact with the distinction regarding whether the enforcer is perceived to influence the offender’s detection. After all, if the enforcer eventually receives the fine paid by the offender, trying to achieve the offender’s detection is more easily classified as a purely selfish, rent-seeking act (e.g. Baumann et al. 2023). Chen, Zeng, and Ma (2020) find that third-party punishment better expresses a social norm than second-party punishment. It may be argued that an enforcer with financial stakes in the investigation is more like a second than a third party in assessing the offender’s punishment. In this regard, Xiao (2013) emphasizes that a profit-seeking punishment undermines punishment’s ability to influence behavior. Importantly, the law-enforcement design aspects of who receives the fine and how the enforcer influences the offender’s detection could imply differences in perceived legitimacy and procedural justice, which are also relevant to compliance decisions and the extent to which enforcement authority is accepted (e.g. Tyler 2003).

We present evidence from three different studies. The first study considers two extreme treatments as proxies of the public and private law-enforcement models. More specifically, in one regime, the fine is transferred to the victim, and the enforcement is purely random. In contrast, in the other regime, the enforcer receives the fine and makes a choice in the enforcement process. Considering offending and avoidance choices, we find some evidence that people invest more in detection avoidance in the public law-enforcement regime. In the second study, we abstract from avoidance and consider a 2 × 2 design where the first treatment arm concerns the fine revenue’s recipient and the second treatment arm the endogeneity of detection. We find no significant differences regarding the likelihood of taking. Our third study considers detection avoidance and taking, focusing only on the treatment variation regarding who receives the fine revenue. This experiment also elicits social norms following Krupka and Weber (2013) to compare treatments along these lines. Our choice and social norms data suggest that the treatment distinction regarding whether the victim receives the fine or the enforcer does not significantly impact offending or avoidance behavior on average. However, when we analyze the choice data separated by gender, we find that the taking rates of men and women differ by treatment. The null effect for the entire sample is explained by the fact that the men and women react in opposing ways to the treatment variation.

Our paper is related to different strands of the literature. The taking game we consider is similar to designs used by Schildberg-Hörisch and Strassmair (2012) and Rizzolli and Stanca (2012), for example. In a related contribution, Baumann et al. (2023) consider a regime where fine revenue is received by society at large (represented by a donation to a charity in the experiment) or the enforcer. In contrast to our study, they allow the offender to choose any taking amount and, more importantly, have the enforcer invest in creating the detection probability in a stage before the potential offender makes her violation choice. This sequence of choices can mean that the enforcer’s investment signals how society’s members disapprove of the violation to the potential violator. The signaling value of investment in enforcement may be strongly affected by whether society or the enforcer receives a fine revenue. Baumann et al. (2023) find that the deterrent effect of enforcement (i.e. the reduction in crime due to the enforcer’s investment in the first stage) is larger when enforcers receive fine revenues.

Whereas Baumann et al. (2023) and we do not allow for a framing of individuals to extort fine payments (i.e. legal errors), Xiao (2013) focuses on this aspect and shows that punishment is less effective when enforcers have rent-seeking motives. In all of our law-enforcement designs, potential offenders know the level of the sanction and the detection probability before they choose to take points from their peer. In contrast, Tan and Xiao (2018), for example, consider a law-enforcement design where the punishment is not known before the potential offender decides on the crime.

In another domain of law, Baumann et al. (2023) explore whether a compensation payment after an accident induces different care incentives than a fine of the same magnitude, finding that the difference in terms of who receives the tortfeasor’s payment is often not decisive for the care investment. This relates to our findings presented below. Guerra and Parisi (2022) and Guerra and Parisi (2024) also test whether two conditions in the domain of tort law that theoretically produce symmetric incentives with standard preferences induce uniform behavior in the laboratory. Our paper provides a further contribution to this strand of the literature.^[3]

Our empirical analysis includes the participant’s gender as a covariate. Several previous studies have shown the role of gender effects. In an experimental study exploring the implications of self-control for crime, Friehe and Schildberg-Hörisch (2017) find that women take significantly less than men and show more substantial treatment differences. Similarly, in a design testing for the effects of a loss as compared to a gain frame, Baumann, Benndorf, and Friese (2019) find that women take overall less than men and that genders respond asymmetrically to the loss frame. While men’s taking increases, women’s taking decreases in the loss frame when compared to the gain frame. Baumann et al. (2023) find that female subjects are less likely to take points away from a charity. These domain-specific results are consistent with gender differences more generally. Regarding social preferences, Niederle (2016) states in her survey on gender differences that women seem more concerned than men with equalizing payoffs, and Engel (2011) shows that women tend to give more than men in dictator games. Croson and Gneezy (2009) emphasize that women are more responsive to the context than men, which can contribute to an explanation of asymmetric responses to treatment variations. In two of our studies, we also find that the taking rate for women is smaller than that for men. In addition, our Study 3 shows another circumstance in which genders react asymmetrically to a treatment variation.

This paper proceeds as follows. Section 2 explains the common design and procedural features of the three experiments we conducted to answer our research questions. In Section 3, we derive hypotheses on treatment differences. In Sections 4–6, we present the specific design aspects and results of our Studies 1–3. We conclude in Section 7.

2 Basic Experimental Setup

3 Behavioral Hypotheses

To understand how the law-enforcement designs may be relevant to Player B’s decision-making, we briefly consider a possible payoff function and its implications for Player B’s behavior. Our presentation in this section focuses on the distinction between Rent and Comp, but we will discuss the potential implications of the Random and Active treatments at the end.

Assume inequity-averse participants following (Fehr and Schmidt 1999). (Dis)Advantageous inequity concerning the payoff of Player C is considered by using β _C (α _C ) ≥ 0, and likewise for Player A using β _A (α _A ) ≥ 0. Similar to (Feess et al. 2018), we assume that taking points and investing in avoidance can create a moral cost, which we represent with the type variable m. This variable is distributed according to some cumulative distribution function. Variables λ _C ≥ 0 and λ _A ≥ 0 represent how Player B altruistically internalizes payoff consequences of own choices for the enforcer and the victim.

Players B can choose not to take points for a payoff

In our design π _B = π _A and π _B − π _C = 5 hold. Alternatively, Players B can take points and thereby generate an expected payoff of

where a denotes the avoidance decreasing the detection probability p(a) at a diminishing rate,

is Player B’s monetary payoff in the detection state that results after deducting the fine amounting to π _A from Player B’s pre-detection payoff

and

give the non-monetary payoff implications in the detection state in the respective treatments with Π _A and Π _C as Player A’s and Player C’s payoffs after receipt of the transfer T subsequent to Player B’s fine payment in treatments Comp and Rent, and I_COMP is an indicator variable equal to one when treatment Comp applies and zero otherwise. When δ > 0, the moral cost of taking is higher when the fine is transferred to the victim. This is consistent with the moral suasion effect of compensatory sanctions (e.g. Deffains, Espinosa, and Fluet 2019). When γ > 0, the moral cost of investing in avoidance is greater when avoidance seeks to lower the probability that the victim will be compensated.

Assuming that Player B is more altruistic towards the victim and less concerned about disadvantageous payoff comparisons with the victim (i.e. λ _A ≥ λ _C and α _A ≤ α _C ), we get that D_COMP ≥ D_RENT. This means the detection state is more tolerable in treatment Comp than in Rent.^[5] The intuition is that Player B thinks that, if detected, at least the fine payment would compensate the victim. This can have implications for the taking decision and avoidance incentives, where the latter can be derived from the marginal effects:

Avoidance has a direct monetary cost of one, reduces advantageous inequity in the no-detection state, and aggravates disadvantageous inequity in the detection state. The last term in the first line represents moral cost considerations. The second line shows the marginal avoidance benefit via the reduction of the detection probability.

If D_COMP > D_RENT and γ > 0, marginal avoidance incentives are smaller in treatment Comp for a given morality m. However, detection avoidance is selected conditionally on taking points from Player A. This means that the ranking of marginal avoidance incentives could be reversed if the people taking in treatment Comp have a significantly different morality than those in Rent.

The taking decision is made anticipating the privately optimal detection avoidance. When the effect from δ > 0 and γ > 0 dominates the effect from D_COMP > D_RENT, the critical type m ̄ COMP that solves U T a COMP * = U N in treatment Comp is smaller than the critical type m ̄ RENT that solves U T a RENT * = U N in treatment Rent. This argument about a critical type is based on the idea that taking is strictly preferred if m = 0 and that U _T decreases with m, and leads to:

Hypothesis 1: The share of Players B who take points from Player A is larger in Rent treatments than in Comp treatments.

If only Players B with m < m ̄ COMP < m ̄ RENT take points in Comp, this means that their marginal avoidance cost will be on average lower as a result. If this effect is strong enough, avoidance will be smaller in treatment Rent than in Comp.

Hypothesis 2: The average investment into detection avoidance in Comp treatments exceeds the one in Rent treatments.

Our previous discussion has shown that these hypotheses follow only subject to some assumptions about the relative effect sizes. Only if the different opposing effects exactly offset each other would we expect that the taking and the avoidance decisions will not be affected by the treatment variation.

In addition to the distinction between Rent and Comp, we consider the implications from the enforcer’s active rather than passive role in the investigation process. By design, this distinction is, in objective terms, irrelevant to the different possible payoffs of the potential offender and other group members and the likelihood of these different payoffs. However, if Player B perceives an influence of Player C’s decision, treatments Active can provide a different frame than treatments Random. The different frame can change Player B’s beliefs, which can modify Player B’s incentives (e.g. Dufwenberg, Gächter, and Hennig-Schmidt 2011). Comparing conditions Rent × Random and Rent × Active, it is possible that the influence of the enforcer on the investigation process will lead Player B to interpret the interaction as a simple redistribution game instead of as a game allowing for a morally wrong taking. After all, one potential reading of the setting in Rent × Active is that Player B seeks enrichment at the cost of Player A, and Player C seeks enrichment at Player B’s cost. Such an interpretation of the difference between conditions is consistent with smaller possible levels of m in condition Rent × Active when compared to Rent × Random, which tends to induce more taking and avoidance in the former condition. Comparing conditions Comp × Random and Comp × Active instead, Player C’s enforcement effort can be interpreted as a prosocial act highlighting the moral dimension of Player B’s taking. This interpretation is consistent with higher possible levels of m in condition Comp × Active than in Comp × Random, which tends to induce less taking and avoidance.^[6]

Hypothesis 3: The enforcer’s influence on the investigation process in Active treatments induces weakly more (less) taking and avoidance than in Random treatments when the enforcer (victim) receives the fine revenue.

4 Study 1: Taking and Avoidance Choices in Polar Designs

In Study 1, we contrast extreme configurations: a scenario where the enforcer receives the fine revenue and is perceived to influence the offender’s detection with a scenario where the victim receives the fine and the detection is purely random.

4.2 Study 1: Results

In Comp × Random, 50 percent of potential offenders steal points from their potential victim, while 68 percent do so in Rent × Active (see Table 2). This difference is substantial but not statistically significant (p = 0.2581) in a two-sided Fisher Exact Test.

Table 2:

Taking rates and average avoidance investments in Study 1.

	Comp × Random	Rent × Active
Taking rate	0.50	0.68
Avoidance	1.66	1.34
	(0.675)	(0.647)

The results of a linear probability regression in Table 3 also show no treatment effect. However, the regression results show that the taking rate among female participants is smaller than that of their male counterparts.

Table 3:

Linear probability regression: taking decision in Study 1.

	(1)	(2)
Rent	−0.111	0.0107
	(−0.49)	(0.05)
Female	−0.425**	−0.362*
	(−2.14)	(−1.71)
Rent × female	0.446	0.344
	(1.58)	(1.21)
Age		−0.0528
		(−1.67)
Studies econ		0.155
		(0.91)
Constant	0.778***	1.830**
	(4.85)	(2.49)
N	51	51

1. Notes: Results from ordinary least squares regressions where the dependent variable is a dummy variable equal to one if Player B took points from Player A. t statistics in parentheses ^*p < 0.10, ^**p < 0.05, ^***p < 0.01.

Regarding avoidance, we find that Players B who decided to take points in Comp × Random invested significantly more (1.66 points) than Players B in Rent × Active (1.34 points, p < 0.0477, Wilcoxon rank-sum test, one-sided).

Figure 1a shows how frequently the different avoidance levels were chosen. Figure 1b shows the cumulative distribution functions. The maximum investment is the most attractive one. A relatively large number of Players B in Rent × Active choose to invest 1 point. Subjects who invest in treatment Comp × Random choose the maximum avoidance more often than subjects who invest in treatment Rent × Active (p = 0.033 in a two-sided Fisher Exact Test).

Figure 1:

Investment into avoidance in Study 1. (a) Histogram. (b) Cumulative distribution.

In summary, Study 1 suggests that potential offenders, if anything, take less often in Comp × Random than in Rent × Active while it at the same time finds that Players B who take points from their Player A invest significantly more in avoidance in treatment Comp × Random. As this study focused on the two polar cases, we cannot attribute these differences clearly to one of our treatment variables. Study 2 is designed to help disentangle possible effects.

5 Study 2: Taking Choices in a Full Factorial Design

In Study 2, we are more comprehensive by considering all 2 × 2 treatment combinations, but more restrictive in that Player B only has a taking (and no avoidance) choice.

5.2 Study 2: Results

In treatments Comp × Random, Comp × Active, and Rent × Random, 76 percent of Players B decided to take points from Player A (Figure 2). In Rent × Active, 85 percent of the Players B did so. The differences in proportions are not statistically significant according to a Fisher Exact Test: when we compare Comp × Active or Rent × Random to Comp × Random, the one-sided Fisher Exact test p-value is p = 0.6201. For Rent × Active, the p-value is p = 0.2957. Thus, we do not find any treatment effect.

Figure 2:

Taking rates by treatment in Study 2.

The results of linear probability regressions shown in Table 4 confirm the null effect, controlling for several covariates.

Table 4:

Linear probability regression: taking decision in Study 2.

	(1)	(2)
Active	−0.0769	−0.0898
	(−0.56)	(−0.65)
Rent	0.0575	0.0704
	(0.41)	(0.50)
Rent × Active	0.0908	0.0718
	(0.57)	(0.44)
Female	−0.127	−0.116
	(−0.93)	(−0.84)
Active × female	0.132	0.126
	(0.83)	(0.78)
Rent × female	−0.0963	−0.106
	(−0.61)	(−0.67)
Age		−0.00564
		(−0.27)
Studies econ		0.128
		(1.31)
Constant	0.824***	0.924*
	(7.84)	(1.94)
N	113	113

1. Notes: Results from ordinary least squares regressions where the dependent variable is a dummy variable equal to one if Player B took points from Player A. t statistics in parentheses ^*p < 0.10, ^**p < 0.05, ^***p < 0.01.

An important element of Study 2 is the endogenous detection probability resulting from the hide-and-seek game played by the offender and the enforcer. Table 5 summarizes the choice frequencies of the four symbols. Although we carefully pretested the action labels to have Player C randomize, control by Player C occurred only with a probability of 11 percent. This difference in detection probability was, in all likelihood, not expected by our participants. If so, this could have counteracted a possible treatment effect: A lower detection probability could induce more Players B to take points in the Active treatments. This means, if the randomization had worked properly, Players B might have stolen less when being controlled by Player C instead of a random mechanism.

Table 5:

Frequency distribution of symbol choices in the hide-and-seek game in Study 2.

	Symbol 1	Symbol 2	Symbol 3	Symbol 4
Player B	0.23	0.21	0.36	0.20
Player C	0.18	0.21	0.36	0.25

From Study 2, we may conclude that neither of the two treatment variables has any effect on the taking choice by Player B. However, the uneven distribution of detection probabilities resulting from the hide-and-seek game might have counteracted a treatment effect.

6 Study 3: Taking and Avoidance Choices, and Social Norms When Victims or Enforcers Receive Fines

Using the Active detection from Study 1 in all treatments, our third study focuses on the treatment variation regarding who receives the fine after the offender’s detection, the victim (Comp × Active) or the enforcer (Rent × Active).^[8] It has more statistical power than the two earlier studies because more observations per treatment result from participants making decisions in all roles. Furthermore, this study elicits many control variables, including subjects’ attitudes towards risk, and various measures of their beliefs about the social appropriateness of Player B’s decisions.

6.2 Study 3: Results

Table 6 provides a first glance at the choice data. We report the taking rates and the average investment in detection avoidance jointly with the standard deviation.

Table 6:

Taking rates and average avoidance investments in Study 3.

	Comp × Active	Rent × Active
Taking rate	0.47	0.53
Avoidance	11.71	10.85
	(8.35)	(7.24)

6.2.1 Taking

In Comp × Active, 47 percent of the potential offenders decided to take points from their potential victim. In contrast, 53 percent did so in Rent × Active (Figure 3). This difference is insignificant in a two-sided Fisher Exact Test (p = 0.4267) and a one-sided Fisher Exact Test (p = 0.2133).

Next, we consider the individual characteristics and run linear probability regressions (see Table 7) with a treatment dummy variable equal to one when the fine was transferred to the enforcer. Transferring the fine to the enforcer instead of the victim does not significantly influence the taking decision on average.

Table 7:

Linear probability regression: taking decision in Study 3.

	(1)	(2)	(3)
Rent	−0.193*	−0.196**	−0.208**
	(−1.93)	(−1.99)	(−2.44)
Female	−0.322***	−0.291***	−0.263***
	(−3.33)	(−3.02)	(−3.11)
Rent × female	0.366***	0.366***	0.324***
	(2.81)	(2.85)	(2.93)
Age	−0.0260***	−0.0247***	−0.0205***
	(−3.11)	(−2.97)	(−2.83)
Studies econ	0.124	0.0899	0.0905
	(1.64)	(1.19)	(1.39)
Efficiency type		0.175	0.0854
		(1.16)	(0.66)
Egalitarian type		−0.205**	−0.196**
		(−2.08)	(−2.31)
MaxiMin type		−0.141	−0.0875
		(−1.57)	(−1.13)
Taking belief			0.0112***
			(8.47)
Risk tolerance			0.0160
			(1.41)
Trust			0.00478
			(0.42)
Constant	1.270***	1.350***	0.464**
	(5.80)	(5.76)	(2.03)
N	228	228	228

1. Notes: Results from ordinary least squares regressions where the dependent variable is a dummy variable equal to one if Player B took points from Player A. The second block of covariates includes the distributional preference types. The last block of covariates includes items from Falk et al. (2018) and a belief about how many Players out of 100 take from their Player A. t statistics in parentheses; ^*p < 0.1, ^**p < 0.05, ^***p < 0.01.

However, there is a large gender difference. While females take more often in the Rent × Active than the Comp × Active treatment, the effect for male participants goes in the opposite direction. This asymmetric response to the treatment variation explains the null effect for the full sample. In a two-sided Fisher Exact Test, the difference for men (women) is significant with a p-value of 0.092 (0.024). Thus, we conclude that the taking choices of females align with Hypothesis 1, whereas the decisions by males violate it.

In addition to the gender effect, other characteristics, such as age and an egalitarian distributional fairness preference, significantly affect the taking probability. The significant impact of the belief about the prevalence of taking is interesting. It confirms that the strength of the social norm not to take is undermined if people expect that many people violate the social norm (Funk 2005). Concerning our behavioral hypotheses, we may interpret the beliefs as information about the type m. People with a lower m are more likely to take points from Player A.

6.2.2 Avoidance

Offenders could invest between zero and 20 points to reduce their detection probability from the initial level of 25 percent to a minimum of 5 percent. Figure 4a shows that the corner solutions were again relevant for many subjects. A significant fraction of participants who took points from their potential victim decided to either leave the detection probability at its initial level or reduce it to its minimum.

Figure 3:

Taking rates in Study 3.

Figure 4:

Investment into avoidance in Study 3. (a) Histogram. (b) Cumulative distribution.

Figure 4b shows the cumulative distribution functions for avoidance investments in both treatments. It shows that investments in treatment Comp × Active are relatively more concentrated on very low levels (more than 20 percent invest nothing) and the maximum investment (more than 30 percent invest the maximum amount). In contrast, in treatment Rent × Active, a sizable share of subjects invest 10 points.

Offenders, on average, invested 10.85 points in avoidance in treatment Rent × Active, thereby inducing a detection probability of less than 15 percent. The average investment in treatment Comp × Active amounts to 11.71. This difference in average investment levels is not statistically significant (p = 0.5378; Wilcoxon Rank Sum Test).

Next, we seek to control for subject information. Table 8 presents results for the level of avoidance from Tobit regressions and confirms the absence of a treatment effect. Whereas women take less often than men (Table 7), we find that women who take points from their Player A invest more in avoidance.

Table 8:

Tobit regression: avoidance decisions in Study 3.

	(1)	(2)	(3)
Rent	1.011	0.865	0.572
	(0.50)	(0.44)	(0.29)
Female	5.236**	4.397**	4.350**
	(2.48)	(2.13)	(2.07)
Rent × female	2.294	1.353	0.977
	(1.14)	(0.68)	(0.48)
Age	0.0697	0.0251	0.0271
	(0.32)	(0.12)	(0.12)
Studies econ	0.268	0.939	0.881
	(0.17)	(0.60)	(0.56)
Efficiency type		−2.641	−2.996
		(−1.00)	(−1.11)
Egalitarian type		4.359**	4.117*
		(2.02)	(1.89)
MaxiMin type		1.900	1.897
		(1.04)	(1.02)
Taking belief			0.0423
			(1.03)
Risk tolerance			0.0750
			(0.25)
Trust			0.0444
			(0.13)
Constant	7.490	7.241	3.964
	(1.39)	(1.29)	(0.58)
N	114	114	114

1. Notes: Results from Tobit regressions where the dependent variable is the level of avoidance. The second block of covariates includes the distributional preference types. The last block of covariates includes items from Falk et al. (2018) and a belief about how many Players out of 100 take from their Player A. t statistics in parentheses; ^*p < 0.1, ^**p < 0.05, ^***p < 0.01.

The treatment variation did significantly influence the decision to invest a positive amount in detection avoidance. Women are more likely to invest positively in avoidance than men, and this difference is less pronounced in Rent × Active (see Column (3) in Table A.7). The treatment variation did not significantly impact whether the offender invested the maximal amount.^[10] We find that traits relevant to taking also influence the level of avoidance. With the interpretation that the belief is information about the inverse of m, we find that it influences the chosen level of avoidance consistently with the marginal avoidance incentives shown above.

6.2.3 Social Norms

We elicited information about the social appropriateness of taking and avoidance for the two law-enforcement designs. To study the effect of the law-enforcement design on norms, we follow Krupka and Weber (2013) and transform the ratings into an empirical measure of the norm by converting subjects’ ratings into numerical scores. Specifically, a rating of “very socially inappropriate” receives a score of −1, “socially inappropriate” receives a score of −0.6, “somewhat socially inappropriate” receives a score of −0.2, and likewise for the appropriate categories. Table 9 lists the distribution of social appropriateness ratings for each choice alternative, separated by treatment. Not taking points and not investing in avoidance is clearly most socially appropriate.

Table 9:

Results of the norm elicitation in Study 3.

	Comp							Rent
Decision	Mean	− − −	− −	−	+	+ +	+ + +	Mean	− − −	− −	−	+	+ +	+ + +	Rank sum
Taking	−0.62	35.1 %	40.5 %	21.6 %	0 %	1.8 %	0.9 %	−0.57	23.1 %	49.6 %	25.6 %	0.9 %	0.9 %	0 %	−1.501
Not taking	0.81	0.9 %	0.9 %	0 %	9 %	21.6 %	67.6 %	0.78	0.9 %	0 %	0 %	10.3 %	29.1 %	58.1 %	1.282
Investing 0	0.3	5.4 %	5.4 %	11.7 %	32.4 %	27 %	18 %	0.35	2.6 %	5.1 %	15.4 %	27.4 %	30.6 %	20.5 %	−0.593
Investing 10	−0.28	8.1 %	29.7 %	43.2 %	12.6 %	5.4 %	0.9 %	−0.25	10.3 %	23.1 %	42.7 %	18 %	5.1 %	0.9 %	−0.718
Investing 20	−0.56	42.3 %	33.3 %	8.1 %	6.3 %	7.2 %	2.7 %	−0.46	35.9 %	30.8 %	16.2 %	3.4 %	7.7 %	6 %	−1.312

1. Notes: Responses are: comprising: “very socially appropriate” (+ + +), “socially appropriate” (+ +), “somewhat socially appropriate” (+), “somewhat socially inappropriate” (−), “socially inappropriate” (− −), and “very socially inappropriate” (− − −). ^*p < 0.1, ^**p < 0.05, ^***p < 0.01.

As shown in the last column, none of the comparisons yields a significant difference. Participants do not believe taking is more or less socially appropriate when the victim receives the fine instead of the enforcer. However, the ranking of mean values is consistent with that of the taking rates, pointing out a tendency that taking may be less appropriate in Comp × Active.

The social norms data also align with choice data regarding avoidance. Many participants opt for the maximum investment in treatment Rent × Active. The social norms data presents a ranking of mean ratings indicating that the maximum investment may be more appropriate in Rent × Active. Importantly, the widespread practice of detection avoidance is considered to be socially inappropriate.

Whereas the taking rates’ reaction to the treatment variation is gender-specific, we find that the social norms are relatively comparable for women and men (see Tables A.9 and A.10 in Appendix E).

7 Discussion and Conclusion

This paper compared the decision-making of potential offenders in different law-enforcement designs. We report data from three experimental studies. The treatment arms are concerned with whether the victim or the enforcer receives the fine revenue and whether an enforcer’s choice influences the detection probability. In contrast to previous experimental studies on criminal decision-making, we allow for investment in detection avoidance in addition to the focus on the violation decision. In this respect, we find that most subjects who steal prefer investing in detection avoidance.

Our findings suggest that the law-enforcement design aspects we examined have, on average, little impact on potential offenders’ decision-making. This comes even though the design differences seem substantial. One study delivers significant differences in avoidance, which we cannot replicate in another, higher-powered study. The taking decision is not significantly different across treatments in any of our three studies. However, in Study 3, we find that men and women react strongly to the treatment variation concerning who receives the fine revenue, but in opposite directions, explaining the null effect on average.

This paper should be viewed as a catalyst for further research. We offer a comparison of behavior in different stylized institutions, laying the groundwork for a more comprehensive understanding of public and private law enforcement. In our designs, we sought to maintain the detection probability before any avoidance at 25 percent for a cleaner comparison of the taking choices in the different regimes. Clearly, in reality, the different regimes might show varying effectiveness (e.g. Helland and Tabarrok 2004). A fuller representation of the regimes may need to consider an office motive, as in Engel and Zhurakhovska (2017), or a possibility of legal errors (e.g. Garoupa 1997), among other avenues for exploration. Moreover, some of the distortions such as “policing for profit” usually associated with private law enforcement may also occur under public law enforcement, depending on the wider institutional framework (e.g. Harvey 2020). Finally, the gender difference in Study 3 regarding how transferring the fine to the enforcer instead of the victim influences the taking rate is interesting and deserves further scrutiny.^[11]