Prosecutorial Discretion and Criminal Deterrence

Franzoni, Luigi Alberto Working Paper Prosecutorial Discretion and Criminal Deterrence Quaderni - Working Paper DSE, No. 234 Provided in Cooperation with: University of Bologna, Department of Economics Suggested Citation: Franzoni, Luigi Alberto (1995) : Prosecutorial Discretion and Criminal Deterrence, Quaderni - Working Paper DSE, No. 234, Alma Mater Studiorum - Università di Bologna, Dipartimento di Scienze Economiche (DSE), Bologna, https://doi.org/10.6092/unibo/amsacta/5079 This Version is available at: https://hdl.handle.net/10419/159077 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-nc/3.0/ Prosecutorial discretion and criminal deterrence Luigi Alberto Franzoni ¤y University of Bologna, Italy September 1995 Abstract. This paper develops a model of law enforcement in which the indicted and the enforcer can negotiate the level of the penalty by means of a settlement. The emphasis of the analysis is on the credibility of the settlement o®er: the enforcer cannot threaten an incredibly large conviction rate if the negotiation fails. The introduction of the negotiation stage brings about several novel features of the optimal enforcement policy, one of which is the possibility that a ¯nite penalty is optimal (globally or locally). We show that the skimming process associated with the negotiation stage reduces the incentives for the enforcer to carry out thorough investigations and increases the rate of noncompliance. JEL Nos. K42 (Illegal Behaviour and the Enforcement of Law) and K41 (Litigation Process) Keywords: settlements, plea bargains, optimal penalty, enforcement, deterrence. ¤ Correspondence to : Luigi Alberto Franzoni, Dept. of Economics, Strada Maggiore 45, 40125 Bologna, Italy. Fax: IT-51-6402664. E-mail: [email protected]. y I am grateful to Chris Harris, Jim Mirrlees, Philippe Aghion, Carlo Scarpa, Vincenzo Denicolo', Carla Marchese and seminar participants at the ESWC in Tokyo for helpful comments and discussions. I would also like to thank Suzanne Scotchmer for having pointed out to me the discrepancy between the ex-post perspective taken by prosecutors and the ex-ante perspective usually taken by the legislature and the economists. 1 Prosecutorial discretion and criminal deterrence 2 1. Introduction In many legal settings, negotiations over the level of the penalty are allowed to take place between the law enforcer and indicted individuals. The negotiation typically ends up with a guilty plea for criminal cases and a settlement for civil cases: in both cases, the o®ender agrees to be in°icted a certain sanction so as to avoid a regular trial. The negotiated enforcement of law is an extremely widespread practice, and targets the objective of preserving the economic resources that would be wasted for costly investigations. Its social desirability, however, is quite controversial, and has been objected by many legal scholars. 1 A strong argument in favour of pre-trial negotiations comes from the literature on law and economics, which recognizes that they reduce the conviction costs and produce an e±ciency gain. The core of this argument can be found in Grossman and Katz (1983) and can be stated in the following terms: settlements, as well as plea bargains, allow guilty agents (who accept the settlement o®er) to separate themselves from innocent ones (who reject the o®er), with the e®ect of reducing the overall conviction costs. In other words, settlements act as a screening device, providing a quick and almost free conviction of the defendants who would otherwise be bound to lose their case. On the negative side, it has been pointed out that settlements may provide indicted agents with unnecessary bargaining power, which allows them to net a discount on the due penalty. In turn, this lowers the expected sanction for o®enders and induces more frequent violations. The latter argument depends strongly on the assumption that the indicted has some bargaining power at the settlement stage, and that he can compel the enforcer to agree on a settlement amount which is less than the statutory penalty for the infraction at hand. In this paper, we provide a di®erent explanation to the fact that settlements usually allow guilty o®ender to get away with a signi¯cantly reduced payment. We assume, as it is often the case, that the indicted has no bargaining power, and that the settlement o®er is made on a take-it-or-leave-it base by the enforcer. The reason why settlements involve amounts lower than the statutory penalty has to be identi- ¯ed with the fact that the enforcer cannot commit itself to put full e®ort into the case after a failure to settle. This, in turn, occurs since the enforcer su®ers from imperfect information at the settlement stage (he does not know whether the indicted is innocent or guilty) and the elimination of this uncertainty (by means of a thorough investigation) is costly. Hence, the enforcer may not be able to threaten a sure conviction to those who refuse to accept his o®er, and the di®erent types of agents (guilty and innocent) do not fully separate at the settlement stage. In general, if the enforcer wants his o®er to be accepted with a larger probability, he will have to 1 A good introduction to the debate on the impact of pre-trial bargains on the legal system can be found in the 1991 issue of the Yale Law Journal . Prosecutorial discretion and criminal deterrence 3 content himself with a low settlement amount. 2 In our model, the terms of the enforcement process depend on the will of di®erent institutions/individuals involved in it at di®erent times. We assume that the enforcement agency is two-tiered; on the ¯rst tier, we have a legislature which decides the level of the penalty for convicted o®enders, on the second, we have an independent investigation branch (\the enforcer"), whose aim is to discover whether the indicted is guilty or innocent with respect to a speci¯c infraction. The two branches of the agency enter the enforcement process at di®erent stages; the legislature decides the penalty at the outset, the enforcer decides which e®ort to devote to the investigation after the violation, if any, has been perpetrated. As in many real situations, the viewpoint of the enforcer is a typical ex-post perspective, in which the attention is focused on the cost-e®ective conviction of guilty o®enders rather than the deterrence e®ect on crime. The legislature, on the other hand, takes an ex-ante perspective, and selects the penalty level that minimizes the social loss (expected harm minus net recovery) associated with the potential o®ense. The main conclusion of this paper is that pre-investigation settlements play a double role in the enforcement process. On the one hand, they allow the enforcer to e®ectively reduce enforcement costs: if the settlement o®er is accepted, the enforcer collects a settlement amount without having to prove the guilt of the defendant. On the other hand, however, the screening out of guilty defendants operated at the negotiation stage induces an overall weakening of the deterrent power of enforcement policy, because it reduces the incentives for the enforcer to carry out thorough investigations after a failure to settle. As a consequence, the introduction of the settlement stage increases both the ex-post net recovery from possibly guilty agents and the level of crime. These two factors a®ect social welfare in opposing ways. Which of the two e®ects dominates depends on the ratio between the level of the harm associated with the crime and the amount of resources that can be extracted from the o®ender at the settlement stage. An in-depth account of the credibility constraint hanging over the enforcement process allows us to identify new factors relevant to determining the optimal penalty. In particular, we will provide an alternative justi¯cation to the commonly observed lack of the maximal sanctions which are recommended by deterrence theory. While the \Principle of Maximum Deterrence" states that the optimal penalty is maximal whenever penalties can be in°icted without costs, in the real world, sanctions are rarely set to their maximum level, even when this is ¯nite, i.e., when limited liability applies. 3 Our way of reconciling theory and evidence is rather intuitive; when the level 2 The problem of the credibility of settlement o®ers has been forcefully raised by Nalebu® (1987) in a model involving civil suits. In Nalebu®'s model, the plainti® has imperfect information about the true liability of the defendant and makes a settlement o®er prior to the trial. The credibility problem arises because the plainti® may be tempted to drop the case if a settlement is not reached. Nalebu® limits his analysis to the settlement stage, and does not consider the impact of the negotiation on the production of harm. 3 The optimality of a maximal penalty is derived and discussed in the early works of Becker Prosecutorial discretion and criminal deterrence 4 of harm associated with the crime is relatively small relative to the private bene¯t that the criminal gets from it, it may be socially desirable to let the agent engage in crime so as to reduce the uncertainty su®ered by the enforcer at the settlement stage and obtain the largest bene¯ts form the negotiation. The explicit consideration of the ex-ante e®ects of pre-trial negotiations on the production of crime distinguishes this contribution to most of the other works on settlements. The standard view is to take the probability that a crime has been committed as given, and to study the e®ects of settlements on the distribution of the legal and investigative costs. 4 A work related to ours is Polinsky and Rubinfeld (1988), which analyses the desirability of settlements in a simple model of harm prevention. In their model, the parties enjoy perfect information about the level of harm caused by accidents. As a consequence, trials entail pure \transaction costs." They assume that both parties held some bargaining power at the negotiation stage and that settlements reduce the expected cost of an accident for the injurer. Under these conditions, trials are socially desirable only if settlements do not provide the potential injurer with adequate incentives to care. In Franzoni (1995), a model of tax enforcement is developed in which the tax agency can precommit itself to the audit policy, and uses it to extract a settlement amount from taxpayers. Settlements represent here a form of \renegotiation" of the original enforcement policy, which proves to increase the payo® to the agency when the original policy is not optimally chosen. The paper is organized as follows. In section 2, we introduce the model and derive the equilibrium of the game between the agent and the enforcer. The game is solved backwards from the settlement stage. This allows us to derive a theory of settlements, which is of interest in itself. In section 3, we characterize the optimal penalty and discuss the relation between our result and the Principle of Maximum Deterrence. In section 4, we discuss the desirability of the settlement stage by contrasting the outcome of our game with the outcome of a game in which negotiations cannot take place between the enforcer and the individual. In section 5, the model is extended so as to give account of a variable probability of indictment, a variable bene¯t from noncompliance and a variable penalty. Finally, section 6 provides a brief summary and some concluding remarks. 2. The enforcement game The problem we are going to consider relates to the enforcement of a speci¯c law or regulation, which requires the agent to behave in a speci¯ed manner or to refrain from taking a certain action. For example, environmental regulations might require that ¯rms refrain from using polluting production processes, ¯scal regulations that (1968), Kolm (1973), and Mirrlees (1974). Other contributions include Polinsky and Shavell (1979), Nalebu® and Sharfstein (1987), Shavell (1991) and Boadway, Marceau and Marchand (1993). 4 See Cooter and Rubinfeld (1989) for a good survey of the literature. Recent contributions include, among others, Reinganum and Wilde (1987), Reinganum (1988) and Daughety and Reinganum (1993). Prosecutorial discretion and criminal deterrence 5 taxpayers keep their accounts in a speci¯c way, etc. If the agent does not abide by his obligation, (social) harm of amount H is produced. Our emphasis is on crimes of economic nature, where rationality of both parties, the agent and the enforcer, can be reasonably assumed. We assume that the agent can be of two types: honest or opportunistic. The honest type abides by his obligation out of respect for his ethical code, while the opportunistic type rationally assesses the costs and bene¯ts of compliance, and complies only if the expected penalty is larger than his individual gain from noncompliance. The opportunistic agent maximizes a von Neumann-Morgenstern utility function and is neutral towards risk. For the honest type, the bene¯t from noncompliance is 0 (or negative), while for the opportunistic it is equal to T: The probability that the agent is honest is p: The reason why the opportunistic agent may decide to comply is that he may incur the risk of an investigation. This investigation can take the form of statistical testing in the case of pollution control, an audit in the case of tax infractions, and the reconstruction of the events in the case of generic crimes and frauds. If the investigation shows that the agent has defected, he has to pay a penalty in the amount F: We assume that the investigation process is triggered by a random event (signal), which might be tied up with the behaviour of the agent. For the time being, however, we assume that this event is independent of the agent's choice and occurs with a constant probability ¼: In section 5, we will extend our model to the more realistic case in which the probability of an investigation is larger when the agent is noncompliant. Before the investigation is actually began, the agent and the enforcer (i.e. the administrative branch in charge of the discovery of the truth and the conviction of guilty agents) can strike a deal; the agent can plead guilty and accept to pay the settlement amount decided by the enforcer. If the settlement o®er made by the enforcer is rejected, the investigation takes place. Depending on how much e®ort the enforcer puts into the process, the investigation may reach di®erent levels of accuracy. We assume that conviction can be obtained only in the presence of hard incriminatory evidence, i.e. evidence which proves without doubt the guilt of the defendant. Let a be the probability of conviction of a guilty agent subject to the investigation ( a measures the accuracy of the investigation). Since a larger investigative e®ort produces a greater accuracy, we can express the enforcer's problem as the choice of the appropriate level of accuracy for the case at hand. The enforcer's goal is not to produce deterrence, but to recover the penalty with the lowest e®ort from a possibly guilty agent. In other words, the enforcer maximizes the expected net transfer from the agent. For simplicity, we assume that investigation costs are a quadratic function of the probability of conviction ;C ( a )= 1 2 ca 2 : The enforcer maximizes the net recovery taking the penalty level and the noncompliance probability as given. The penalty level is decided by the \the principal" (the legislature), whose objective is ex-ante optimality. It will choose F so as to minimize the social loss from noncompliance, i.e. the di®erence between the expected level of harm and the expected net recovery from the agent. The principal decides the penalty at the outset, Prosecutorial discretion and criminal deterrence 6 and is aware of the way in which it a®ects the game between the enforcer and the agent. The move sequence of the game is the following; 1. the principal decides the penalty level, 2. the agent decides whether to comply or not, 3. depending on an random signal, the agent may be selected for an investigation, 4. the indicted agent is o®ered a take-it-or-leave-it settlement o®er by the enforcer, 5. the agent chooses whether to accept the o®er or not, 6. if the o®er is rejected, the enforcer carries out the investigation. It chooses the e®ort to devote to the case and the resulting conviction probability. We solve the game by backward induction using the concept of Perfect Bayesian Equilibrium. In simple terms, a PBE is a set of strategies and beliefs such that, at any stage of the game, strategies are optimal given the beliefs, and the beliefs are obtained from equilibrium strategies and observed actions using Bayes' rule. We solve the game backwards, considering the optimal settlement o®er ¯rst. The optimal settlement o®er In this section, we derive the optimal settlement o®er for any given probability of guilt of the agent. We are therefore solving the last three stages of the game. Our objective is to show that the optimal settlement o®er will depend on the conviction rate that the enforcer can credibly threaten. Given that the payment of the agent is F in case of conviction and 0 in case of acquittal, the optimal settlement o®er will be q ¤ = a ¤ F; with a ¤ equal to the credible conviction threat. In order to derive the optimal settlement amount, the enforcer has to calculate the continuation equilibrium following each possible settlement amount q: We start our analysis by considering the equilibrium of the continuation game (following the settlement o®er). We derive the optimal behaviour of the defendant ¯rst. The agent only knows whether he is innocent or guilty, while the enforcer believes the agent to be guilty with probability x: The strategy of the agent will depend on whether he is guilty or not. Clearly, an innocent defendant will reject all settlement o®ers with q> 0 : A guilty defendant will instead reject the o®er only if the expected payment in case of rejection (probability of conviction £ penalty) is less that q: Let ½ ( a j q ) be the probability that the guilty defendant rejects an o®er of amount q when the expected conviction probability is a: We have 8 > < > : ½ ( a j q )=1 if q>aF; ½ ( a j q ) 2 [0 ; 1] if q = aF; ½ ( a j q )=0 if q<aF; (1) Consider now the enforcer's optimal choice in case of rejection. It will choose the conviction probability a (through her investigation e®ort) which maximizes the expected net recovery. The latter is calculated through her revised belief about the probability Prosecutorial discretion and criminal deterrence 7 of guilt of the agent, R ( a )= ½x ½x +1 ¡ xaF ¡ 1 2 ca 2 : The enforcer knows that only the guilty defendant may be willing to accept the settlement o®er. If the o®er is rejected, the enforcer will attach a larger probability to the defendant being innocent rather than guilty. This in turn reduces the incentives to carry out thorough investigations. The optimal conviction probability is a ( ½ j x )= 8 > > > < > > > : ½x [ ½x +1 ¡ x ] F c if ½x [ ½x +1 ¡ x ] F c · 1 ; 1 otherwise. (2) The conviction probability is larger if the likelihood that the agent is guilty is larger. If the recoverable ¯ne is far greater than the investigation costs, the enforcer can credibly threaten a ( ½ j x )=1 : Figure 1 depicts the best reply functions of the enforcer and the guilty agent. Figure 1: Optimal rejection probability for the guilty agent and optimal conviction probability. Let us assume that F<c; so that a ( ½ j x ) is always less than one at the equilibrium,and leave the other case to the appendix. The equilibrium following the settlement o®er q is obtained combining eqs. (1) and (2), and depends on the Prosecutorial discretion and criminal deterrence 8 level of q as follows; continuation equilibrium, 8 > > > > < > > > > : a ( q j x )= q F; and ½ ( q j x )= 1 ¡ x x 1 ( F cF q ¡ 1 ) ; if q · x F 2 c ; a ( q j x )= xF c; and ½ ( q j x )=1 ; if q>x F 2 c : (3) If the settlement amount is not too high, 0 <q · x F 2 c ; then at the equilibrium the guilty agent rejects the o®er with a probability ½ ( q j x ) which is larger if the enforcer demands a larger settlement amount; @½ @q > 0. 5 At the equilibrium, the probability of conviction of the guilty agent in case of rejection is such as to render the agent indi®erent between accepting and rejecting the o®er; a ( q j x )= q=F: As expected, investigations are less thorough if they follow the rejection of a settlement o®er with a small amount. The rejection of the o®er provides the enforcer with bad news about the probability of getting a conviction: the smaller the settlement o®er which is turned down and the more likely it is that the defendant is innocent. Note also that the probability that the guilty defendant accepts the settlement o®er is larger if the probability of guilt, x; is larger. This because a larger probability of guilt relaxes the credibility constraint on the conviction threat by making investigations more pro¯table. If x goes to one (the agent is guilty with certainty), the rejection probability goes to zero. This is an interesting result, which will have important consequences for the optimal enforcement policy. One can start to see that if one of the objectives of the principal is to limit enforcement costs, then a large defection rate may have desirable consequences, since it reduces the credibility constraint on the conviction threat and makes a settlement more likely. If the settlement amount is very large, q>x F 2 c ; then both types reject the offer with probability one, and the probability of conviction of the guilty agent is the largest credible one; a ( q j x )= x F c : In this case, the rejection does not convey any information to the enforcer and the continuation equilibrium is independent of q: We can now turn to the problem of the optimal settlement o®er. The expected net recovery following an o®er of amount q is R ( q j x )= x (1 ¡ ½ ( q j x )) q | {z } exp. settlement intake + x½ ( q j x ) a ( q j x ) F | {z } exp. penalty intake ¡ 1 2 [1 ¡ x (1 ¡ ½ ( q j x ))] c [ a ( q j x )] 2 | {z } exp. investigation costs An increase in the settlement amount a®ects the net recovery in several ways: ¯rst, it increases the amount collected in case of acceptance; second, it reduces the probability 5 Note that this result is brought about by the convexity of the investigation cost function. In a linear model, with C ( a )= ca; the rejection probability would be independent of the settlement amount, since only one level of ½ would allow the inspector to select a conviction probability di®erent from 0 or 1. Here, instead, di®erent rejection rates lead to di®erent conviction probabilities; large rejection probabilities are associated with large conviction rates and, consequently, large settlement amounts. Prosecutorial discretion and criminal deterrence 15 the level of the penalty. The principal chooses the level of the penalty taking into account the e®ects that it will have on the process of enforcement and the level of compliance. Following Reinganum (1993), we assume that the objective of the principal is the minimization of the (uncompensated) social loss resulting from the enforcement activity, measured as the di®erence between the expected harm produced by the o®ender and the expected amount of resources recovered by the enforcer. The level of harm associated with noncompliance is denominated in terms of social utility and is equal to H: 6 The social loss can then be expressed as L ( F )=(1 ¡ p ) ¯ ¤ H ¡ ¼R ¤ ; (4) where (1 ¡ p ) ¯ ¤ is the equilibrium noncompliance probability and ¼R ¤ the equilibrium expected net recovery. We have L ( F )= 8 > > > > > > < > > > > > > : (1 ¡ p ) H ¡ ¼F 2 c (1 ¡ P ) 2 1+ P 2 ; for ¿> 1 ¡ P (small penalty) ; 1 ¡ (1 ¡ ¿ ) 2 1+(1 ¡ ¿ ) 2 H ¡ ¿ 1+(1 ¡ ¿ ) 2 T; for ¿ · 1 ¡ P (large penalty) : (5) For ¿> 1 ¡ P; i.e. for F< ^ F ´ q Tc ¼ (1 ¡ P ) ; the opportunistic agent is noncompliant with probability one. Here, an increase in the penalty a®ects only the expected net recovery from the investigation and provides the enforcer with an incentive to devote more e®ort to the case. The net recovery is larger and the social loss smaller if the penalty is larger. For ¿ =1 ¡ P; we have L ³ ^ F ´ =(1 ¡ p ) H ¡ 1 ¡ P 1+ P 2 T: Note that L ³ ^ F ´ ¸ 0ifand only if the social harm H associated with noncompliance is greater than 1 1 ¡ p 1 ¡ P 1+ P 2 T = 1 1+ P T: 7 For p ! 0 ;L ³ ^ F ´ = H ¡ T; which is positive only if the level of harm is greater than the private bene¯t from noncompliance. For ¿ · 1 ¡ P; i.e. for F> ^ F , an increase in ¿ (i.e. either a reduction in F and ¼ or an increase in c and T ) has two e®ects; it raises the probability that harm is produced and increases the amount of resources recovered through the enforcement system. These e®ects clearly pull in opposite directions for the social loss. We have @L @F < 0 () H> 2 ¡ ¿ 2 4(1 ¡ ¿ ) T: (6) 6 Note that if the victim of crime and the enforcement agency do not coincide (they do coincide, for example, when the crime at hand is tax evasion) and if direct compensation (in expected terms) does not take place, then the enforcement process produces a transfer of `welfare' from the victim of crime to the enforcer. The more undesirable this transfer (because of lack of indirect compensation), the larger will be the social harm H associated with crime. 7 The last inequality stems from the fact that p = 2 P 1+ P 2 : Prosecutorial discretion and criminal deterrence 16 An increase in the penalty level reduces the social loss (at the internal equilibrium) if and only if H is su±ciently large. Note that the coe±cient of T in the previous inequality is strictly increasing in ¿: For ¿ ! 0 (in¯nite penalty), we have L 0 < 0ifand only if H> 1 = 2 T: For ¿ ! 1 ¡ P ,wehave L 0 < 0 if and only if H> ³ 1 2 + (1 ¡ P ) 2 4 P ´ T: If the probability of ethical behavior is negligible, p ! 0 ; then the latter condition cannot be satis¯ed since lim p ! 0 ³ 1 2 + (1 ¡ P ) 2 4 P ´ = 1 : As expected, lim F !1 L ( F )= lim ¿ ! 0 L ( F )=0 : Therefore, ^ F is a local minimum of L ( F )if H · ³ 1 2 + 1 ¡ P 2 4 P ´ T; and is a global minimum if H< 1 1+ P T: Figure 3 plots the loss function for the case with no ethical behaviour ( p =0) : Figure 3: Social loss as a function of the penalty level. Let us focus on ¯gure 3 and consider how a variation in the penalty a®ects social loss when p = 0. For small levels of the penalty, F< ^ F = q cT=¼; the agent is noncompliant with probability one. The enforcer enjoys full information at the settlement stage and his o®er is accepted with certainty. An increase in the penalty decreases the social loss since it increases the settlement amount without a®ecting the probability of a settlement. At F =^ F; the settlement amount is equal to T=¼ and the agent is fully expropriated from his private bene¯t from noncompliance. The settlement o®er is accepted with certainty and no investigation costs are incurred in by the enforcer. For penalty levels greater than ^ F; the credibility constraint on the conviction threat starts to bite. When F is greater than ^ F; but less than a threshold level ~ F; social loss is strictly increasing in the penalty; an increase in F reduces the noncompliance probability and tightens the credibility constraint. The guilty agent ¯nds it more pro¯table to mimic the innocent one and the settlement o®er is more likely to be rejected. Due to the larger probability of a (costly) investigation, the net recovery is smaller and the social loss is larger. For F> ~ F; the increase in the Prosecutorial discretion and criminal deterrence 17 rejection probability associated with an increase in the penalty is outweighed by the increase in the probability of harm, and social loss is decreasing with F: If the penalty is in¯nite, no harm is committed, and the net recovery and the social loss are nil. The special case just considered, with p = 0, displays two interesting properties. The ¯rst is that when the penalty is set to ^ F; social loss is equal to H ¡ T; i.e. social harm minus private bene¯t from noncompliance. In this case, a settlement is reached with certainty and the enforcer can extract from the noncompliant agent his private bene¯t in full at zero cost . The second is that social loss is strictly increasing in the penalty for some values of F: Here, an increase in the penalty level is detrimental, since it worsen the credibility constraint on the conviction threat and increases the probability of a (costly) investigation. These facts have straight consequences in terms of the optimal penalty. First, obviously, if the level of harm associated with noncompliance is less than the private bene¯t to the agent, then social loss is minimized at F =^ F: Second, if the penalty is constrained by an exogenous upper bound ¹ F 2 ³ ^ F; ~ F ´ , which may for instance derive from the limited liability of the agent, then the optimal penalty is ^ F: When F =^ F; the enforcer \sells the shop" to the agent, who buys the \right to be noncompliant" at an (expected) price equal to his private bene¯t from it. The settlement e®ectively serves the purpose of legalising the harmful act, taxing individuals by an amount equal to their private bene¯t from the harmful activity. 8 The interesting fact is that the enforcement system can be used to extract the private bene¯ts from individuals at no costs only if two conditions are satis¯ed: i) the penalty is small, and ii) all individuals are interested in buying the right to be noncompliant (i.e. there are no ethical agents). When p 6 =0 ; the previous results have to be quali¯ed. The presence of unconditionally compliant agents reduces the likelihood of a cost-free conclusion of the enforcement process, since they will always refuse to settle. In turn, this leads noncompliant agents to mimic the innocent ones and reject the settlement o®er with positive probability. This means that there are no ways for the enforcer to avoid the investigation costs and to recover T without e®ort. At F =^ F; the probability that a settlement is agreed upon indictment reaches its largest level and is equal to (1 ¡ P ) 2 1+ P : The probability that a costly investigation is carried out and the social loss will therefore be larger if p is larger. Note also that in proximity of F =^ F + ; we have L 0 > 0 if and only if H< ³ 1 2 + (1 ¡ P ) 2 4 P ´ T . This means that when p 6 =0 ;F =~ F is local minimum only if H is su±ciently small. Actually, H has to be smaller if p is larger. The previous arguments suggest that the optimal penalty will depend on the ratio between H and T: 9 8 Note that this argumet in favour of legalisation is slightly di®erent from the standard utilitarian one (which says that an harmful act should not be deterred if the utility that the agent can get from it is larger that the level of harm it produces), since our loss function does not include the violator's utility. See Lewin and Trumbull (1990) for a philosophical defence of this approach. 9 In the appendix, we will show that similar threshold levels can be de¯ned also for the case in Prosecutorial discretion and criminal deterrence 18 Proposition 3. The optimal penalty There are two threshold values h 0 ´ 1 1+ P and h 1 ´ ³ 1 2 + 1 ¡ P 2 4 P ´ such that: 1. If the ratio between the level of harm and the private bene¯t from noncompliance is less than h 0 ; then the optimal penalty is ¯nite, F ¤ =^ F: 2. If the ratio between the level of harm and the private bene¯t from noncompliance lies between h 0 and h 1 ; then an in¯nite penalty is globally optimal and F =^ F is a local optimum; 3. If the ratio between the level of harm and the private bene¯t from noncompliance is greater than h 1 ; then the social loss is monotonically decreasing with F and the optimal penalty is in¯nite. It may be useful to recall that ^ F = q cT ¼ 1 1 ¡ P is the largest penalty compatible with full noncompliance of the opportunistic agent. When H=T is small, the principal is better o® by inducing the opportunistic agent to be noncompliant with probability one and allowing the enforcer to extract from him (in expected terms) his private bene¯t T: This amount can in turn be used to compensate the victim of crime. Note that only opportunistic agents engage in crime and pay the settlement amount, and that the enforcement systems e®ectively screens between the two types of agents (those who value noncompliance and those who do not). This is actually its major advantage with respect to a policy which directly legalises the harmful behaviour. When the ratio between social harm and private bene¯t is larger than h 0 , but less than the threshold h 1 ; then the optimal penalty is in¯nite. However, if limited liability applies and F is bounded above, it may still be optimal for the principal to set F =^ F and maximize the probability of a settlement with the opportunistic agent. Note that the threshold value h 1 = ³ 1 2 + 1 ¡ P 2 4 P ´ is decreasing with the probability of ethical behaviour. As p tends to 0 ;h 1 tends to in¯nity, and ^ F is a local optimum for any (¯nite) level of H=T: Finally, if the ratio between harm and private bene¯t is very large, H=T > h 1 ; then deterrence becomes more important, and the penalty has to be set to its largest feasible level. It is interesting to contrast these results with those deriving from the standard enforcement model, in which no negotiations can take place between the enforcer and the agent (see section 4). In the standard model, the Principle of Maximum Deterrence applies and the optimal penalty is the largest feasible one, as far as social harm is less than the net amount of resources that can be recovered from the agent. Note that the latter amount consists of the penalty in°icted in case of conviction less enforcement costs. When negotiations between the enforcer and the agent take place and the penalty is small, instead, the recovery can take place at no cost, and the condition for the optimality of an in¯nite penalty becomes more stringent. Also, if limited liability puts an upper bound on F; then the optimal penalty may not be the which the constraint \ a · 1" is binding. Prosecutorial discretion and criminal deterrence 19 maximal one. Even in the case with no negotiations, before setting up an enforcement apparatus, the principal has to ask to herself whether enforcement is at all worthwhile, and whether it would be better to legalise the o®ence at hand. Clearly, this is de¯nitively a feasible alternative when all individuals enjoy a private bene¯t from noncompliance, but is certainly more problematic when some individuals bene¯t from committing the o®ense and others do not. One can hardly imagine an auction for the right to commit criminal acts or engage in harmful behaviour. In fact, the easiest way to extract the private bene¯t from the agents who gain from the harmful behaviour is to refrain from setting a large penalty and allow for pre-trial negotiations. People who net a large bene¯t from noncompliance will then engage in crime and pay it out at the settlement stage. Agents who do not attach any value to noncompliance will instead comply with the law and go through the normal enforcement procedure in case they are (wrongfully) indicted. Clearly, this outcome is desirable only if the private bene¯t from noncompliance to the agents is su±ciently greater than the social harm attached to it and the legal system provides e®ective guarantees to innocent defendants. In order to asses the overall e®ect of the settlement stage on the enforcement policy, in the next section we investigate the features of the compliance game assuming that negotiations between the enforcer and the agent cannot take place. 4. Negotiations versus straight investigations One of the controversial issues in the theory of law enforcement and in general public debates is whether negotiations between the enforcer and `suspected' o®enders should be allowed to take place in the course of the enforcement process. The reason for this is that the prospect of a negotiated penalty may provide agents with additional incentives to engage in criminal activity. The increased noncompliance rate may then partially or completely o®set the bene¯ts accruing from the reduction in enforcement costs, which is provided by the introduction of the settlement stage. In this section, we tackle this issue by comparing the equilibrium con¯guration of the game in which negotiations take place between the enforcer and the indicted agent with the equilibrium of a game in which negotiations are ruled out. We assume that the penalty level is given. For sake of simplicity, we limit our analysis to the case in which the probability of ethical behaviour is negligible, p = 0, and the enforcement costs are su±ciently large to rule out equilibria with perfect investigations, c>T=¼ . 10 Let us consider the case in which pre-investigation negotiations are not feasible. The enforcement game has a very simple structure; the agent chooses whether to comply or not, and the enforcer what level of e®ort to devote to the investigation. This model captures the typical situation in which noncompliance occurs only if the 10 When c<T=¼; at the equilibrium we have ¯ = 1 and a =1for F 2 [ c;T=¼ ] : For F belonging to this interval, the introduction of the settlement stage has no e®ect on the noncompliance probability (which is constantly one) and brings enforcement costs down to zero. Prosecutorial discretion and criminal deterrence 20 probability of conviction is su±ciently large, while the probability of conviction is large only if there is a positive probability that the agent is guilty. The best reply functions of the players can be easily derived. The agent chooses to comply only if the prospective punishment is larger than his private bene¯t from noncompliance; 8 > > > > > > < > > > > > > : ¯ I ( a )=1 if a I < T ¼F ; ¯ I ( a ) 2 [0 ; 1] if a I = T ¼F ; ¯ I ( a )=0 if a I > T ¼F : (7) The net recovery for the enforcer is R I ( a I j ¯ I )= a I ¯ I F ¡ 1 2 ca I : For any given ¯ I ; the best reply for the enforcer is a I ( ¯ I )= 8 > < > : ¯ IF c if ¯ IF c · 1 1 otherwise. (8) The optimal accuracy of the investigation is larger if the probability of unearthing an infraction is larger and if the investigation costs are smaller. The equilibrium of the game between the enforcer and the agent is derived from the interception of the two best reply functions, eqs. (7) and (8) ; and is described below. 8 > > > > < > > > > : ¯ ¤ I = ¿; and a ¤ I = T ¼F; for ¿ · 1 ; ¯ ¤ I =1 ; and a ¤ I = F c; for ¿> 1 : For small investigation costs, ¿ = cT ¼F 2 · 1 ; the conviction rate is a ¤ I = T ¼F and the agent is noncompliant with probability ¯ ¤ I = ¿: The defection probability is larger if c and T are larger, and F and ¼ smaller. For large investigation costs, ¿> 1 ; the enforcer is not able to deter the agent; the optimal conviction rate is a ¤ I = F c and the agent is noncompliant with probability one. It can be easily seen that the conviction probability in the game without negotiations, a ¤ I ; is the same as that obtained in the game with negotiations, a ¤ N . This is the unique conviction probability that fully expropriates the agent's bene¯t from noncompliance. As far as the probability of noncompliance is concerned, one can easily see that, for any penalty level, it is larger when negotiations are feasible; ¯ ¤ N ¡ ¯ ¤ I =1 ¡ (1 ¡ ¿ ) 2 1+(1 ¡ ¿ ) 2 ¡ ¿ =(1 ¡ ¿ ) ¿ 2 1+(1 ¡ ¿ ) 2 > 0 : Prosecutorial discretion and criminal deterrence 21 When negotiations between the enforcer and the agent can take place, a fraction of guilty agents settles the penalty; this reduces the pro¯tability of the investigation and lowers the investigative e®ort. As a consequence, the deterrent e®ect of the investigations is decreased and noncompliance occurs with a larger probability. At the equilibrium, the expected net recovery for the enforcer is R ¤ I = 8 > > > > < > > > > : a ¤ I ¯ ¤ I F ¡ 1 2 c ( a ¤ I ) 2 = ¿ 1 2 T; for ¿ · 1 ; a ¤ I F ¡ 1 2 c ( a ¤ I ) 2 =1 2 F 2 c; for ¿> 1 : (9) At the internal equilibrium, ¿ · 1 ; the expected net recovery is decreasing in F and ¼; and is increasing in T and c: Changes in these variables a®ect the probability of noncompliance and, consequently, the probability that the penalty is recovered. At the corner equilibrium, ¿> 1 ; the agent defects with probability one; a larger F and a smaller c increase the pro¯tability of the investigation and increase the net recovery. The social loss in the game with no-negotiations is L ¤ I = 8 > < > : ¯ ¤ I ( H ¡ T )+ ¼ 1 2 c ( a ¤ I ) 2 ; for ¿ · 1 ; H ¡ a ¤ I F + 1 2 c ( a ¤ I ) 2 ; for ¿> 1 ; (10) that is L ¤ I = 8 > > < > > : ¿ ³ H ¡ 1 2 T ´ ; for ¿ · 1 ; H ¡ 1 2 ¼ F 2 c ; for ¿> 1 : By looking at the social loss expression, we can see that here an in¯nite penalty is optimal only if the level of harm, H; exceeds the net recovery, T= 2 : For small levels of harm, H<T= 2 ; the net recovery is larger than the level of harm and a ¯nite penalty is optimal ( F ¤ = q cT=¼ ). At the optimum, the agent is noncompliant with probability one and faces an expected penalty equal to his private bene¯t. It is important to note that in order to collect this amount, the principal has to expend resources on the investigation process, since only a conviction can result in the application of the penalty. This marks a major di®erence between the present model and that in the previous section, since in the latter the full extraction of the private bene¯t could be obtained with no cost (by a pure threat) when the penalty was optimally set. Note, ¯nally, that when the optimal penalty is in¯nite ( H> 1 = 2 T ) ; the Principle of Maximum Deterrence applies: the principal is better o® setting F to its largest feasible level. This because variations in F do not a®ect the likelihood of a cost-free conviction (settlement) of the o®ender. Finally, we can compare the social loss associated with the game without negotiations ( L ¤ I ) with the social loss associated with the negotiation game( L ¤ N ) : From eqs. Prosecutorial discretion and criminal deterrence 22 (5) and (10), we have L ¤ N ¡ L ¤ I = 8 > < > : ( ¯ ¤ N ¡ ¯ ¤ I )( H ¡ T ) ¡ ¼ Pr(sett.) 1 2 c ( a ¤ ) 2 ; for ¿ · 1 ; ¡ ¼ 1 2 c ( a ¤ ) 2 ; for ¿> 1 : The introduction of the settlement stage increases the probability that the agent is noncompliant and reduces the expected investigation costs by an amount proportional to the probability that the settlement is accepted. The di®erence in the social loss is then composed of two parts: i) the increase in the expected level of harm net of the amount that can be recovered from the noncompliant agent, and ii) the reduction in the enforcement costs. Clearly, when the penalty is so small that ¯ is constantly equal to 1 ; the introduction of the negotiation stage does not a®ect the expected level of harm, and the social loss in the negotiations regime is smaller by an amount equal to the reduction in enforcement costs. When the equilibria of the games with and without negotiations are internal, ¿< 1, which of the e®ects is the dominant one depends on the size of H compared with T: In particular, we have L ¤ N ¡ L ¤ I · 0 () (1 ¡ ¿ ) ¿ 2 1+(1 ¡ ¿ ) 2 H ¡ 1 2 ¿ ³ 1 ¡ (1 ¡ ¿ ) 2 ´ 1+(1 ¡ ¿ ) 2 T · 0 () H< 1 ¡ 1 2 ¿ 1 ¡ ¿T: (11) The next proposition summarizes our ¯ndings. Proposition 4. The optimality of the negotiation The introduction of the settlement stage has two e®ects on the enforcement system: i) it increases the probability that the agent is noncompliant, ii) it reduces the expected investigation costs by an amount which is proportional to the probability that a settlement is agreed upon. Given a penalty level large enough to exert some deterrence on the agent ( ¿< 1) , the introduction of the settlement stage is desirable if and only if the ratio between social harm and private bene¯t from noncompliance is less than a cut-o® value ¹ h ( ¿ ) ´ 1 ¡ 1 2 ¿ 1 ¡ ¿: Note that the threshold ¹ h is increasing in ¿; and is therefore decreasing in ¼ and F; and increasing in T and c: For ¿ ! 1 (small penalty), the condition for the desirability of the negotiation stage collapses to H=T < 1 ; while for ¿ ! 0 (in¯nite penalty), this condition collapses to H=T < 1. 11 11 By comparison of eqs.(6) and (11), one may note that if the social loss (with negotiations) is increasing in the penalty, then the removal of the negotiation stage is de¯nitely detrimental (while the reverse is not necessarily true). Prosecutorial discretion and criminal deterrence 23 Given H and T; with H>T; we have L ¤ N ¡ L ¤ I · 0 , ¿ ¸ H ¡ T H ¡ 1 2 T: The last result can be rephrased as follows. Corollary 5. Given the level of the social harm H and the private bene¯t from noncompliance T; with H>T; the introduction of the settlement stage reduces the social loss if and only if the penalty level is su±ciently small. When the penalty level is large, the agent is noncompliant with a small probability and the possibility to reach a settlement is hampered by the credibility constraint hanging over the conviction threat. This reduces the bene¯ts from the negotiation. On the contrary, when the penalty is small, the agent is noncompliant with a large probability: he will ¯nd it hard to pretend that he is innocent and will accept the settlement o®er with a large probability. The reduction in the enforcement cost will then outweigh the increase in the expected level of harm. One may wonder how a positive probability of ethical behaviour would a®ect our results. It can be easily seen that if the equilibrium is internal both with and without negotiations, then equation (11) still holds. Hence, the introduction of the negotiation stage is still desirable only if the penalty is su±ciently small. If the equilibria are on the corner, then the introduction of the negotiation stage does not a®ect the expected level of harm and yields a net reduction in the social loss. 5. Extensions In this section we look as some ways of extending the model with pre-investigation negotiations and show how some of its limitations can be overcome. Selective indictment The analysis of the previous sections has been carried out under the assumption that the probability of indictment was the same irrespective of whether the crime had been committed or not. We intend now to show how our results can be generalized to the case in which the probability of an investigation depends on the compliance choice of the agent. Let us assume that by not complying the agent increases the probability of the signal s that triggers the investigation. Let ¼ n and ¼ c be the probabilities that this signal occurs when the agent is, respectively, noncompliant and compliant, and let ¼ n >¼ c : One can calculate the probability that the agent is noncompliant when the signal is observed. If x is the (prior) probability of noncompliance, we have ^ x ´ Pr(noncomplianace j s )= Á ( x ) ´ ¼ n x ¼ n x + ¼ c (1 ¡ x ) : Prosecutorial discretion and criminal deterrence 24 Since the indictment process is biased against the noncompliant agent, we have ^ x>x for all x 6 =0 ; 1 : Once an agent has been indicted, the case can be settled before the investigation is started. The problem is the same as that analysed in section 2; the only di®erence is that the belief of the enforcer about the probability of noncompliance is now ^ x instead of x: The optimal settlement o®er is hence that, which is described in section 2. Since ^ x is generally greater than x; the optimal settlement amount will be larger: the selection bias of the indictment stage makes the conviction threat more credible. From an ex-ante perspective, the fact that noncompliance increases the chances of being indicted decreases the incentive to defect. An opportunistic agent will comply only if ¼ n q (^ x ) · T: The noncompliant agent faces now a larger probability of indictment and a tougher enforcer. At the equilibrium (which we assume for simplicity to be internal), the expected payment must equate the private bene¯t from defection, so that q (^ x )= T=¼ n : We have therefore ^ x = q ¡ 1 ( T=¼ n ) and ^ x = Á ( x ) ; so that x = Á ¡ 1 ( q ¡ 1 ( T=¼ n )). Given ¼ n ; the equilibrium probability of noncompliance is smaller if the indictment process is more selective, i.e. if the probability of indictment of the compliant type is smaller. The social loss associated with the enforcement process can be easily derived. Recall that L is the di®erence between expected harm and expected net recovery. Once investigations are started, net recovery is as in section 2, with ^ x in place of x: The expected probability of an investigation is ¼ ( x )= ¼ c (1 ¡ x )+ ¼ n x: Thus, the social loss becomes L ( x )= xH ¡ ¼ ( x ) R (^ x ) : An improvement in the selectivity of the indictment process leads to a reduced level of expected harm (since defection is more easily detected), a smaller probability that the investigation is undertaken (since defection takes place with a smaller probability), and to a larger net recovery if the investigation is carried out (since the probability of noncompliance given indictment is larger, and the credibility problem is relaxed). Since ^ x = x at x =0,1 ; the results on the optimality of a ¯nite penalty for p =0 are con¯rmed. In general, for p> 0 ; the results obtained with a constant probability of indictment have to be adjusted so as to take into account the degree of selectivity in the indictment process. Their qualitative features, however, remain unchanged. Variable private bene¯t from noncompliance In many situations the private bene¯t from noncompliance is not known with certainty by the principal, since it may vary from individual to individual. We model this lack of information by assuming that the principal and the enforcer share a common prior on T; represented by a cumulative probability distribution G ( T ) with support h 0 ; ¹ T i and G 0 > 0 throughout : Since honest agents do not attach any (positive) value to noncompliance, their private gain is T =0 : Prosecutorial discretion and criminal deterrence 31 Figure 5 plots the social loss as a function of the penalty for the case with p =0 : Figure 5: Social loss as a function of the penalty for the case in which \ a · 1" is binding. Recall that the constraint \ a · 1" is binding for F 2 [ c;T=¼ ] : In this region, the agent defects and accepts the settlement o®er with amount q = F with probability one. An increase in the penalty increases the settlement amount without a®ecting the credibility constraint, and reduces the social loss. For F = T=¼; the settlement amount equals the private bene¯t from noncompliance; the game admits of a multiplicity of equilibria characterized by a defection probability ¯ 2 [ c¼=T; 1] and a rejection probability for the noncompliant type equal to ½ 2 [0 ; T ¡ c¼ T ¡ 1 = 2 c¼ ] : The larger the defection probability, the more likely it is that the settlement is accepted. For ( T=¼ ) ¡ ; we have ¯ =1and L = H ¡ T: For F greater than T=¼; the constraint \ a · 1"is not binding, and the social loss is as in ¯gure 3. We can now derive the optimal penalty. Recall that L ( F )= 8 > > > > > > > > > < > > > > > > > > > : (1 ¡ p ) H ¡ ¼ F 2 c (1 ¡ P ) 2 1+ P 2 ; for F · c 1 ¡ P ; (1 ¡ p )( H ¡ ¼F )+ ¼ 1 2 cp F F ¡ c ; for c 1 ¡ P <F< T ¼ ; 1 ¡ (1 ¡ ¿ ) 2 1+(1 ¡ ¿ ) 2 H ¡ ¿ 1+(1 ¡ ¿ ) 2 T; for T ¼ <F: For ¼F < T; we have L 0 ( F ) < 0 ; while for ¼F > T; we have L 0 ( F ) < 0ifandonlyif condition (6) is satis¯ed. For ¼F = T; the game admits a multiplicity of equilibria; the opportunistic agent is indi®erent on whether to defect with probability one and reject the settlement o®er (of amount ¼F = T ) with probability ½ = p 1 ¡ pc F ¡ c ,orto defect with probability ¯ = 1 1 ¡ p 1 ¡ (1 ¡ ¿ ) 2 1+(1 ¡ ¿ ) 2 and reject with probability ½ = 1 ¡ ¿ 1 ¡ 1 = 2 ¿ .We Prosecutorial discretion and criminal deterrence 32 have L ³ T=¼ ¡ ´ =(1 ¡ p )( H ¡ T )+ ¼ 1 2 cp T=¼ T=¼ ¡ c and L ³ T=¼ + ´ = cT [ H (2 T ¡ c¼ ) ¡ T 2 ] T 2 +( T ¡ c¼ ) 2 : With a good deal of algebra, one can establish that 8 > > > > > > > > > > > < > > > > > > > > > > > : L ( T=¼ ¡ ) · L ( T=¼ + ) () H · T 22 T ¡ c¼ T ¡ c¼ ; L ( T=¼ ¡ ) ¸ 0 () H ¸ T T ¡ c¼ ³ T ¡ 2 ¡ p 2 ¡ 2 p c¼ ´ ; L ( T=¼ + ) · 0 ) L 0 ( T=¼ + ) ¸ 0 ; [ L 0 ( T=¼ + ) ¸ 0 _ L ( T=¼ ¡ ) · 0] ) L ( T=¼ ¡ ) · L ( T=¼ + ) : Again, lim F !1 L ( F )=0 : From the previous results one can infer that F = T=¼ is a global minimum if and only if H=T · h 0 ´ 1 T ¡ c¼ ³ T ¡ 2 ¡ p 2 ¡ 2 p c¼ ´ and a local minimum if and only if h 0 <H=T · h 1 ´ 1 22 T ¡ c¼ T ¡ c¼ : Prosecutorial discretion and criminal deterrence 33 REFERENCES Becker, G.S., 1968, Crime and punishment: An economic approach, Journal of Political Economy 76, 169-217. Boadway, R., N. Marceau and M. Marchand, 1993, Time-consistent criminal sanctions, mimeo. Cooter, R.D. and D.L. Rubinfeld, 1989, Economic analysis of legal disputes and their resolution, Journal of Economic Literature 27, 1067-1097. Daughety, A.F. and J. Reinganum, 1993, Endogenous sequencing in models of settlement and litigation, Journal of Law, Economics and Organization 9, 314-348. Franzoni, L.A., 1995, Tax amnesties, plea bargains and the optimal enforcement policy, w.p. 221, University of Bologna. Grossman, G.M. and M.L. Katz, 1983, Plea bargaining and social welfare, American Economic Review 73, 749-757. Lewin, J.L. and W.N. Trumbull, 1990, The social value of crime, International Journal of Law and Economics 10, 271-284. Kaplow, L. and S. Shavell, 1994, Optimal law enforcement with self-reporting of behaviour, Journal of Political Economy 102, 583-606. Kolm, S.-Ch., 1973, A note on optimum tax evasion, Journal of Public Economics 2, 265-270. Mirrlees, J., 1974, Notes on welfare economics, information and uncertainty, in Balch, Mc.Fadden and Wu (eds.), Essays on economic behavior under uncertainty , North Holland, Amsterdam. Nalebu®, B., 1987, Credible pretrial negotiations, RAND Journal of Economics 18, 198-210. Nalebu®, B. and Sharfstein, 1987, Testing in models of asymmetric information, Review of Economic Studies 44, 265-277. Polinsky, A.M. and S. Shavell, 1979, The optimal tradeo® between the probability and magnitude of ¯nes, American Economic Review 69, 880-891. Polinsky, A.M. and D.L. Rubinfeld, 1988, The deterrent e®ect of settlements and trials, International Review of Law and Economics 8, 109-116. Prosecutorial discretion and criminal deterrence 34 Reinganum, J., 1988, Plea bargaining and prosecutorial discretion, American Economic Review 78, 713-728. Reinganum, J. and L. Wilde, 1985, Income tax compliance in a principal-agent framework, Journal of Public Economics 26, 1-18. Reinganum, J. and L. Wilde, 1987, Settlement, litigation, and the allocation of litigation costs, RAND Journal of Economics 17, 557-566. Schulhofer, S.J., 1992, Plea bargaining as disaster, Yale Law Journal 101, 1979-2009. Scott, R.E. and W.J. Stuntz, 1992, Plea bargaining as a contract, Yale Law Journal 101, 1909-1968. Shavell, S., 1991, Speci¯c versus general enforcement of law, Journal of Political Economy 99, 1088-1108.