WPS4434 Policy ReseaRch WoRking PaPeR 4434 Litigation and Settlement: New Evidence from Labor Courts in Mexico David S. Kaplan Joyce Sadka Jorge Luis Silva-Mendez The World Bank Financial Private Sector Development Department Enterprise Analysis Unit December 2007 Policy ReseaRch WoRking PaPeR 4434 Abstract Using a newly assembled data set on procedures filed in is also found that cases with multiple claimants against Mexican labor tribunals, the authors of this paper study a single firm are less likely to be settled, which partially the determinants of final awards to workers. On average, explains why workers involved in these procedures receive workers recover less than 30 percent of their claim. The lower percentages of their claims. Finally, the authors find strongest result is that workers receive higher percentages evidence that a worker who exaggerates his or her claim is of their claims in settlements than in trial judgments. It less likely to settle. This paper--a product of the Enterprise Analysis Unit, Financial Private Sector Development Department--is part of a larger effort in the Financial Private Sector Development VPU. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at dkaplan@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Litigation and Settlement: New Evidence from Labor Courts in Mexico David S. Kaplan World Bank and Instituto Tecnológico Autónomo de México dkaplan@worldbank.org Joyce Sadkay Centro de Investigación Económica Department of Economics Instituto Tecnológico Autónomo de México jsadka@itam.mx Jorge Luis Silva-Mendezz Stanford Law School jlmendez@stanford.edu August 2007 Abstract Using a newly assembled data set on procedures ...led in Mexican labor tribunals, the authors of this paper study the determinants of ...nal awards to workers. On average, workers recover less than 30 percent of their claim. The strongest result is that workers receive higher percentages of their claims in settlements than in trial judgments. It is also found that cases with multiple claimants against a single ...rm are less likely to be settled, which partially explains why workers involved in these procedures receive lower percentages of their claims. Finally, the authors ...nd evidence that a worker who exaggerates his or her claim is less likely to settle. We thank Tatiana Carrera, Fernando Espino, Helena Garcia, Oscar Sanchez, Pedro Struck, Antonina Tarassiouk, Lorelen Urioste, and Carlos Zafra for research assistance. Financial sup- port from the Asociación Mexicana de Cultura, the Conferencia Interamericana de Seguridad Social, and the Inter-American Development Bank is gratefully acknowledged. y Corresponding Author. Camino a Sta. Teresa #930. México, D.F., C.P. 10700. Mexico. z J.S.D. Candidate, Stanford Law School. 1 I Introduction One of the major tenets of the ...eld of law and economics is that the legal environment a¤ects economic behavior and outcomes. In studying the legal environment, it is crucial to distinguish between the letter of legal rules and their application or enforcement. Much of the evidence we have on how private law is enforced comes from legal disputes between private parties. The empirical analysis of legal disputes has focused disproportionately on cases litigated at trial, since almost all available data is on trial outcomes. However, ...led cases are not a random sample of the universe of underlying legal disputes, and cases litigated at trial are not a random sample of those ...led. The literatures on law and economics have examined many areas of dispute resolution, including such topics as litigation, arbitration, settlement, and the selection of cases for trial. A key limitation of the empirical tests of theories in this literature has been the lack of data on settled legal disputes. We exploit the fact that Mexican labor law obliges parties in employment disputes to seek rati...cation of settlements, and also mandates that courts approve and record the details of settlements of ...led lawsuits. We use a data set from labor tribunals in Mexico that provides extensive information about settled cases as well as tried cases in order to address a few of the major testable implications of arbitration and litigation theory. Our data analysis has yielded several interesting ...ndings. First, lawsuits that go to trial receive signi...cantly lower ...nal payments. Second, ...nal-payment amounts are lower when several workers are grouped in a case against a single ...rm, and these cases settle with lower probabilities. Finally, we ...nd evidence that workers who exaggerate their claims settle less often, and may be punished in terms of ...nal-payment amounts. These results have important implications for many of the theoretical models proposed in the literature. The paper is organized as follows. Section 2 reviews previous theoretical and empirical work in the areas of arbitration, litigation, and settlement. Section 3 explains the relevant details of the legal environment, namely Mexican labor law relating to the type of lawsuits we examine, as well as rules governing the labor courts in Mexico. Section 4 describes the information available in the data set we use. Section 5 presents our statistical analyses. Section 6 concludes and relates our results to the broader literature. II Previous Work This paper is related generally to the literature on bargaining and dispute res- olution, and more speci...cally to the work on arbitration and litigation. The usual framework in this area is a game theoretic model in which parties decide whether to settle a dispute privately or bring it before an arbitrator or a court. When adjudication is costly, parties have complete information, and they are assumed to behave according to standard rational decision theory, all disputes terminate in private settlement. Hence, incomplete information is the predom- 2 inant explanation in the literature for the fact that a signi...cant proportion of disputes do not settle. In other words, assuming that settlement negotiations are less expensive than litigating the dispute to a ...nal resolution, parties will settle all disputes privately unless they hold su¢ ciently di¤erent beliefs about the expected bene...ts of going to court. Di¤ering beliefs can arise in two ways. On the one hand, the parties may have ex-ante asymmetric information, i.e., one of the parties has superior knowledge about the merits of the dispute and the likelihood of prevailing in an adjudicated award. On the other hand, parties may have ex-ante symmetric information about the merits, but after the case is ...led they may observe di¤erent signals of the probability of prevailing in court. This gives rise to di¤ering posterior beliefs about the expected bene...ts of continuing the litigation rather than set- tling. If based on these posterior beliefs opposing parties are each relatively too optimistic about the expected bene...t of going to court, they will fail to reach a settlement. Both information structures, ex-ante asymmetric information and ex-ante symmetric information with di¤ering posterior beliefs, can result in sys- tematic di¤erences between the average underlying merits of disputes that are settled out of court and disputes that proceed to binding adjudication. A sizeable theoretical literature has focused on explaining why settlement negotiations fail in a signi...cant proportion of legal disputes and on character- izing the selection e¤ect of going to court. However, the lack of information on settlement amounts in most litigation data sets has been a major obsta- cle to testing and measuring the di¤erences in success rates and compensation amounts between settled and tried disputes. Our data set provides an almost unique opportunity to test and measure di¤erences between the outcomes of settled and tried lawsuits that result from ...ring disputes. To provide background for what we do, in this section we review some of the relevant literature. We consider ...rst the implications of models that assume ex-ante asymmetric information as well as the empirical tests of these models. Second, we discuss the theoretical and empirical literature that assumes ex- ante symmetric information. Finally we discuss other empirical work that has measured di¤erences between cases that settle and those that go to court. II.A Asymmetric Information Models P'ng (1983) proposes one of the early models of litigation under asymmetric information. He assumes risk-neutral parties and one-sided asymmetric infor- mation, along with an exogenously set settlement amount. In the Nash equilib- rium of this game the informed party cannot reveal its private information in its settlement o¤er, because this would allow the uninformed party to perfectly predict the potential trial outcome.1 Therefore the average quality of cases that settle need not di¤er from that of tried cases. 1The introduction of judicial error P'ng's model generates some equilibria in which the informed party's information is revealed, but the equilibrium with no information revelation continues to exist. 3 Bebchuk (1984) extends P'ng's analysis by allowing the settlement amount to be chosen by the parties. In this model the uninformed party (the plainti¤) makes a take-it-or-leave-it settlement o¤er to the defendant. In equilibrium the settlement o¤er made by the plainti¤ induces defendants who are relatively likely to lose in court to settle, while defendants who are relatively likely to win refuse to settle and go to court. Hence the quality of cases that settle is higher, on average, than that of cases that go to court. Bebchuk also studies the e¤ects of increasing the stakes in the case, i.e., the potential judgment against the defendant. He ...nds that as the stakes increase, the probability of settlement falls and settlement amounts increase. Nalebu¤ (1987) points out that Bebchuk's results depend on an assumption that the plainti¤'s threat of going to court is always credible. Once this assumption is relaxed, Bebchuk's result on the e¤ects of increasing the stakes can be reversed. As the likely judgment at trial increases (stakes increase) the credibility constraint of the plainti¤ may be relaxed, making her act less aggressively in settlement negotiations, and increasing the likelihood of a settlement. Spier (1992) also assumes risk-neutrality and one-sided asymmetric informa- tion, but models the dynamics of settlement explicitly. Given an exogenously set trial date and a ...nite number of rounds of settlement o¤ers up to that date, she ...nds a U-shaped pattern of settlement, so that the probability of settlement is higher at the beginning of negotiations and close to the trial date. Fenn and Rickman (1999) use a data set containing negligence claims against NHS trusts in England to test Spier's model of delay in settlement. They estimate the con- ditional probability of case settlement and test the e¤ects of changes in legal costs and levels of uncertainty on the probability of settlement. Sieg (2000) uses data on medical malpractice suits in Florida to estimate an asymmetric information model similar to Nalebu¤'s. The data provide in- formation about whether the dispute settles or not, as well as the amounts of compensation paid in court and in a settlement. He ...nds that the asymmetric information model can explain most observed empirical facts, in particular the fact that on average plainti¤s who settle receive a higher level of compensation, while those who do not settle receive a higher level of compensation if they win the case, but are very unlikely to win in court. Farmer and Pecorino (2004) use a data set on ...nal o¤er arbitration in Major League Baseball that contains information on o¤ers from both parties, actual salary earned by the player after the arbitration process, and various statistics pertaining to the player's market value. They ...nd that players who exaggerate their claims, by demanding a salary above their predicted market value, are more likely to reach the ...nal stage of arbitration, and more likely to earn a lower ex-post salary. They conclude that these results are inconsistent with a bargaining model in which the player has private information about some unobservable characteristic related to risk preferences or future market value. In such a model, players who exaggerated their salary demands would tend to be those whose true market value exceeded the predicted market value from publicly available information, and therefore would tend to earn higher salaries after the conclusion of the arbitration process. 4 II.B Symmetric Information Models Models with symmetric information can be divided into two broad categories. The ...rst category includes models of arbitration that typically assume symmet- ric information between the parties to a dispute and model the behavior of an arbitrator who has less information about the dispute than the parties. The second category includes models of litigation in which excessive optimism is the rationale for some disputes going to court. Gibbons (1988) models the decision of an arbitrator under both conven- tional and ...nal-o¤er arbitration structures. The parties to the dispute observe equally noisy signals about the underlying merits of their dispute. The arbitra- tor observes a di¤erent and noisier signal; both parties know the distribution of the error in the arbitrator's signal but do not observe the signal directly. In equilibrium, the arbitrator learns from parties'o¤ers and computes her ideal arbitration award, which is a function of the signal and of the average of parties' o¤ers. When parties'o¤ers di¤er more, these o¤ers provide less precise informa- tion to the arbitrator and their average is assigned relatively lower weight in the arbitrator's ideal award.2 Also, as the parties'information about the case be- comes relatively more important than publicly available information, the award places relatively higher weight on the average o¤er. This decision rule is consistent with empirical evidence on the e¤ects of parties'o¤ers and arbitrators'awards. Farber and Bazerman (1986) study the arbitration decisions of the National Academy of Arbitrators in 25 cases. They ...nd that arbitrators placed relatively more weight on the known facts of the cases than on the o¤ers made by the parties to the dispute, and placed less weight on parties' o¤ers as they diverged more. This is consistent with the notion that an arbitrator who mechanically splits the di¤erence between parties' o¤ers will provide incentives for the parties to take unreasonable or extreme positions. Rather than mechanically splitting the di¤erence, if an arbitrator responds to extreme positions by placing less weight on parties'o¤ers, parties have incentives to submit more reasonable o¤ers. Boden (1992) works with data from 204 disability claims in Maryland, and estimates the adjudicated award as a function of the disability ratings proposed by the physicians of the claimant and the defense, as well as other known facts about the claimant and the injury. He ...nds that few facts relating to the case add explanatory power to the prediction of the adjudicated award, after including the parties' positions. Using data on settled claims, he ...nds that when parties settle they place even less weight on the facts of the case than an arbitrator would, so that the compensation paid is practically always the average of the parties'o¤ers. Moreover, Boden ...nds that as parties'positions become more disparate, both adjudicated awards and settlements continue to place a large weight on the o¤ers. Nevertheless, the data imply that physicians 2Given a signal of the underlying facts received by the arbitrator and a position taken by the opposing party, the more aggressive a party's o¤er is, the lower the weight it will receive in computing the arbitrator's award, and the less likely it is to be chosen as the award under ...nal-o¤er arbitration. 5 tend not to submit wildly di¤erent positions, perhaps because of reputation considerations, or because adjudicators would ignore physician's ratings that clearly contradicted known facts about the injury. Priest and Klein (1984) propose a non-strategic model of the decision to settle a case, and focus on the selection e¤ect of settlement. They show that under several assumptions the win rate for plainti¤s and defendants should tend to 50%, regardless of whether the law favors plainti¤s or defendants. They assume risk neutral parties that sue to decide a discrete issue3 with equal stakes. The parties have incomplete but symmetric information as to the merits of the case, and once a case goes to court, the judge observes the true merits and applies clear legal rules at judgment. Each party observes a noisy signal of the merits of the case, estimates her probability of winning at trial, and decides on a range of settlement amounts that would keep her out of court. Parties settle when their respective ranges of settlement amounts overlap. Priest and Klein show that as parties'signals become more accurate, cases that are litigated have fact patterns arbitrarily close to the legal cuto¤ for liability, and the plainti¤ win rate therefore tends to 50%. In addition, as parties observe almost perfect signals of the merits of the case, the settlement rate should increase so that very few cases are litigated. Hence, the smaller the percentage of litigated cases, the closer to 50% is the win rate at trial. Many studies attempting to test the Priest-Klein result have been unable to verify the 50% win rate for plainti¤s.4 However, the assumptions of the model are quite restrictive, and some are impossible to verify empirically. Hence it is di¢ cult to tell what causes a deviation from the 50% win rate, and such evidence cannot be used to assert that the original case selection theory is incorrect.5 Also, testing settlement behavior is generally di¢ cult because very few data sets have information on pretrial negotiations and on settlement amounts. In fact, in many legal systems out-of-court settlements are generally con...dential and this may be an important reason for parties to reach a settlement.6 Given that directly testing the di¤erences between compensation obtained in settled and tried cases is usually not an option, much of the work on case selection has extended the Priest-Klein framework to derive implications that relate the probability of reaching a settlement - which is generally observable - to variables that can be estimated or observed, such as litigation costs, asymmetry in the stakes of litigation, or risk aversion. Schwab and Eisenberg (1988) compare the success rates between constitu- tional tort plainti¤s and other non-civil rights plainti¤s. The main di¤erence 3For example, the parties in the original case selection model do not litigate the amount of damages that should be paid, but rather a single issue such as whether the defendant is negligent or not. 4Kessler, et. al. (1996) provide an overview of several articles that found plainti¤ success rates signi...cantly di¤erent from 50%. 5This argument is made by Gross and Syverud (1991). 6Gross and Syverud (1991), for example, study cases that proceeded to trial, but are able to obtain data on the best o¤er received by the plainti¤ in pre-trial negotiations. See Daughety and Reinganum (1999) for an extensive discussion and analysis of the con...dentiality of settlements. 6 between the control group of lawsuits and the group of constitutional tort cases is the presence of the government as a defendant in the latter. The government may have an informational advantage in predicting the outcome of ...led suits, may have higher stakes in a given lawsuit since it must potentially defend it- self against many similar suits, and may also be less risk-averse than individual plainti¤s. As a result, the model developed by Schwab and Eisenberg predicts that the government will tend to settle less often and will win more often at trial. The empirical evidence indeed indicates that plainti¤s are signi...cantly more successful in non-civil rights lawsuits than when they ...le a constitutional tort suit. However, since settlement amounts are not observed and settlement is grouped together with winning at trial as a plainti¤ success, the data cannot be used to compare success rates between settled and tried cases. Fournier & Zuehlke (1989) use data from civil cases ...led between 1979 and 1981, and observe whether the case settles or not (but not settlement amounts) as well as some characteristics of trial awards and plainti¤s'claims. They ...nd that a higher variance of trial awards increases the probability of reaching a settlement. Although a higher variance of trial awards should increase the like- lihood of the parties disagreeing on trial award predictions, it also increases the incentives of risk averse parties to settle. This empirical evidence suggests that the latter e¤ect is stronger, so that more uncertainty about trial outcomes promotes a higher settlement rate. Gross & Syverud (1991) study 529 California civil jury trial cases in various areas of law. They observe settlement o¤ers and judgments for cases that go to trial, but have no information on cases that settle. They ...nd that the plainti¤ win rate varies greatly across case types (e.g., personal injury vs. commercial transactions). When plainti¤s pay their own litigation costs, the settlement rate increases and the plainti¤ win rate at trial increases, indicating that higher quality suits are brought on average when plainti¤s must shoulder costs of ...l- ing and trial. Asymmetric stakes (such as repeat players on one side) cause the settlement rate to decrease and the win rate for the high-stakes player to increase. This could imply that high stakes litigants seek to establish a tough reputation by going to trial more frequently and exerting more e¤ort to win cases at trial. Finally, they ...nd that in cases with high potential damages and rich defendants, the plainti¤success rate at trial decreases; this may be evidence that lower quality suits are brought when defendants are richer and potential damage awards higher. Siegelman & Donohue (1995) examine employment discrimination cases and ...nd a plainti¤ win rate well below 50%. They use the business cycle to verify implications of the case selection hypotheses. As macroeconomic conditions become more adverse, they show that the number of lawsuits ...led increases, there is a higher settlement rate, and plainti¤s tend to lose more often at trial. This indicates that on average the quality of cases ...led worsens with economic conditions, and that the selection e¤ect weeds out many but not all of the low-quality cases. Waldfogel (1995) develops the implications of the Priest-Klein model to show that the relationship between the settlement rate and the rate of plainti¤ wins 7 at trial depends on the relative cost of going to trial as opposed to settling, the position of the decision standard used by the court with respect to the distribution of parties'behavior, and the variance of the error in the signals parties receive about the merits of the case. Using data from federal civil cases in the Southern District of New York, which are assigned randomly to judges in the jurisdiction, he is able to identify the e¤ect that the judge (who a¤ects the decision standard and the level of uncertainty faced by the parties) has on the relationship between the rate of settlement and the likelihood of plainti¤ prevailing in court. He ...nds a strong relationship between the rate of settlement and plainti¤'s win rate, and signi...cant variation across judges in both, especially in disputes for which the law is perceived to be less clear. Eisenberg and Farber (1997) assume an underlying distribution of litigation costs for potential litigants, and consider the e¤ects of changes in this distribu- tion on rates of settlement and success at trial. They show that as the variance of potential plainti¤'s trial costs increases, the rate of settlement decreases and the success of plainti¤s at trial falls. They test this result using data from civil suits in federal district courts, and ...nd that individuals, who have a more vari- able distribution of litigation costs than corporations, are less likely to settle as plainti¤s and also less likely to win at trial. II.C Other Empirical Evidence Some evidence about the di¤erences in underlying characteristics and success rates between tried and settled cases can be found in a number of studies that examine litigation trends and costs in particular areas of law. In general, this literature ...nds that average compensation in tried cases is higher than in settled cases. Our study ...nds that average compensation in tried cases is slightly higher in one tribunal, and lower in the other tribunal. However, we ...nd that recovery as a percentage of the original claim is signi...cantly lower for tried cases. Unfortunately, by and large the empirical literature to date has not controlled for the amount of the original claim; therefore the measure of success used in most of the previous work is not directly comparable with ours. There are, however, a few studies that ...nd a much higher compensation ratio of trial to settlement than our data displays. After reviewing several relevant studies, we will attempt to provide some explanations of the di¤erences between their results and ours. Kakalik, et. al. (1984) study the determinants of compensation received in a sample of 513 asbestos injury related lawsuits between January 1980 and August 1982. The authors conducted a survey of plainti¤s'attorney, defendants, and insurers for all tried cases during this period, in addition to a random sample of claims that closed before trial either by settlement or by being dropped. They ...nd that on average trials resulted in ...nal payments that were more than two times greater than the compensation received by plainti¤s in cases that settled before trial. No data on the amount of compensation initially requested by the plainti¤ is given, although the study controls for type of asbestos-related illness, job type, age, sex, marital status, and whether or not the plainti¤ was a smoker. 8 Danzon (1984) analyzes aggregate data at the state level on medical mal- practice claims. She ...nds that legal rules favoring plainti¤s had an important impact in increasing the frequency of medical malpractice claims but not the average payment made on such claims. Also, increased use of medical services and urbanization explain a large portion of the growth in medical malpractice litigation, perhaps indicating that larger plainti¤ verdicts in urban areas af- fected expectations of potential plainti¤s, shifting the demand for litigation in these areas. With lawsuit level data, the shift in legal rules over time could be used as a test of case selection hypotheses. Unfortunately Danzon's data set does not di¤erentiate between claims with respect to case disposition, so that the di¤erences in quality or success across tried and settled cases cannot be examined. Kakalik, et. al. (1988) measure compensation amounts in aviation accident lawsuits for over 2000 cases ...led between 1970 and 1984. They ...nd that claims settled without a suit had an average compensation of $256,200. Claims that resulted in a lawsuit but were settled before trial averaged $387,600, while those that went to trial averaged $599,000. Although case ...le data were available for their study, the authors do not take the amount requested initially in the claim or lawsuit into consideration as a predictor of ...nal payment. As in the case of medical malpractice, which motivates much of the empirical work on litigation, aviation accident plainti¤s experienced a shift in legal rules in their favor dur- ing the time period studied. However, aviation litigation is somewhat peculiar in that it rarely consists of a dispute over liability, but rather focuses almost exclusively on damages. In addition, the underlying informational structure of the bargaining process between plainti¤ and defendant may di¤er in important ways between aviation litigation and medical malpractice suits. Dertouzos, et. al. (1988) use data on 120 jury trials of wrongful termination cases in California between 1980 and 1986 to study the consequences of major shifts away from the employment at will doctrine in many states. The informa- tion from each jury trial includes the award, subsequent alterations of the award through appeal or post-trial settlement, as well settlement o¤ers made before the initial trial. The amount of compensation claimed initially by the plainti¤ is not observed. The average award over all cases was $436,627, while the average settlement o¤er made by the defendant was $30,000. Since every case in this data set went to trial, the settlement o¤ers observed constitute failures to reach agreement. Hence these data do not allow measurement of the selection e¤ect of going to court. When defense costs are considered, straightforward calculations show that for cases with initial demands in the top three quartiles, defendants would have been better o¤ on average accepting settlement demands, which averaged $207,710. Barring irrational behavior by ...rms or severe informational asymmetries, this may indicate that ...rms decide to go to trial to build a rep- utation that would deter future litigation in general or low-quality litigation in particular. Black, et. al. (2005) investigate trends in medical malpractice litigation in Texas using a database including all closed medical liability insurance claims over 15 years. Claims in this data set may be ...led by insured medical profession- 9 als or institutions in anticipation of a lawsuit that never materializes, or entail ...led medical malpractice lawsuits that later settle or go in court. If there is a lawsuit related to the claim, the data do not contain detailed information about the suit, such as the amount of compensation claimed initially or the process of negotiating a settlement. The authors show that when population growth and overall spending on health care are taken into account, there is no strong positive time trend in the amount of medical malpractice payouts. The average payout ranged between $303,000 and $410,000 during the time period studied, while the average verdict on tried cases was $1,528,525. The distribution of trial verdicts is highly skewed, implying a low probability of a high amount of compensation. This seems to indicate that, on average, plainti¤s who settle recover a fraction of trial awards. However, average payouts include claims that never became lawsuits, as well as lawsuits that were dropped. In addition, with- out any information from case ...les, it is di¢ cult to account for di¤erences in underlying characteristics of disputes that jointly determine case outcome and ...nal payout. Another recent study of malpractice litigation was performed by Chandra, et. al. (2005). They use the National Practitioner Data Bank, con- taining information on payments in settled and tried malpractice claims against physicians. The 4% of all claims that result in a trial produce an average award approximately twice as large as the average settlement. Once again, detailed information from the case ...le, including the original amount of the claim, is not available. Several factors limit the comparison between our study and the empirical literature discussed above. The empirical literature tends to focus on shifts in common or statutory law that alter expected recovery at trial for plainti¤s, while such shifts in legal standards are not present in the data we study. In the area of medical malpractice, where much of the empirical work has been done, plainti¤ verdicts can result in extremely high payments. Along with di¤erent relative stakes due to important repeated play features of the environment, this means that defendants may be more willing to go to court for cases with high expected recovery amounts. Most importantly, detailed information such as that o¤ered by our data set is rare. Absence of data on the amount of the initial claim means that we cannot compare the percentage of recovery achieved by the plainti¤ across empirical studies. Information on payments made in settlements is usually unavailable, so empirical work has focused on the probability of settlement or on failed settlement o¤ers. As shown by Waldfogel, a standard case selection model can explain di¤erent relationships between settlement rates and measures of plainti¤ success as features of the model such as litigation costs or relative accuracy of parties'signals about expected recovery change. Clearly, across the di¤erent data sets, legal areas, and time periods studied in the empirical literature, such features of the case selection model can vary signi...cantly, causing di¤erences in the evidence found on case selection, particularly the relative success rates for settled vs. tried cases. Although this paper does not test a speci...c economic model rigorously, we do feel it is useful to summarize some of the important implications of the 10 theoretical models described above, and which the empirical literature to date has not been able to test adequately. The following three issues form the focus of our empirical exercises: 1. Settlement amounts may di¤er on average from judgment amounts. 2. Repeat players (those with higher stakes) may act di¤erently in the bar- gaining and trial process, possibly generating di¤erences in settlement rates and ...nal awards. 3. The "quality"of cases that are settled may di¤er systematically from the "quality"of cases that go to trial. Despite the fact that these three hypotheses have played prominent roles in the theoretical literature, data constraints have made them quite di¢ cult to address empirically. As described in the next section, institutional features of the labor-courts system in Mexico provide us with a unique opportunity to examine these issues empirically. III Legal Environment Mexican labor law regulates many aspects of the employment relationship.7 For the purposes of this paper, the most relevant rules concern the provision of fringe bene...ts, overtime, and the mechanics of ...ring. Fringe bene...ts are mainly composed of vacation time and pay and an end-of-year bonus. Each employee is entitled to a certain number of days of paid vacation depending on tenure at the ...rm. The worker must also be given a vacation bonus, so that she earns 125% of her daily salary during each day of vacation.8 Also, every employee is entitled to an end-of-year bonus of at least 15 days'wages.9 A normal work-week cannot exceed 48 hours. If an employee works more than 48 hours, she is entitled to overtime pay. The law mandates double pay for up to 9 hours of overtime, and triple pay for any hours above 57 per week.10 Firing is classi...ed under the law as justi...ed or unjusti...ed. Justi...ed ...ring is limited to wrongdoing on the part of the worker. For example, an employer may justi...ably ...re a worker for three unexplained absences from work during one month,11 or for deliberately or negligently damaging the employer's machinery. Firing for other reasons, such as low worker productivity, or layo¤s during a recession, is considered unjusti...ed and implies a much higher ...ring cost.12 7All regulations discussed here apply primarily to workers in the formal sector, which covers only around 60% of the Mexican work force. Informal workers can obtain some bene...ts from the labor law, but must be able to prove the existence of an employment relationship as well as facts about the employment contract. 8Articles 76 and 80, Ley Federal del Trabajo (LFT). 9Article 87, LFT. 10Articles 66-68, LFT. 11Nevertheless, "unexplained absence" is not de...ned in the LFT, and anecdotal evidence suggests that is it quite di¢ cult for employers to ...re their workers on this basis alone. 12Article 47, LFT. 11 For either type of ...ring, the ...rm must cover all payments owed to the worker up to the ...ring date, including overtime, unpaid end-of-year bonuses, as well as the percentage of the worker's fringe bene...ts that corresponds to the proportion of the last year in which the worker was employed. Additionally, the worker is entitled to severance pay equivalent to 12 days'wage for each year worked, with wage/day capped at twice the minimum wage.13 At the time of ...ring the ...rm must notify the worker of the exact cause of ...ring as de...ned by the Ley Federal del Trabajo (LFT),14 often leading to a suit in which the worker disputes the ...rm's statement of cause. In all lawsuits related to ...ring, the ...rm carries the burden of proving that it ...red the worker for just cause.15 For unjusti...ed ...rings, which under the LFT constitute the vast majority of worker-job separations, the ...rm incurs much greater costs. To begin with, a worker who proves that she was ...red without justi...cation can ask to be reinstated in her job.16 For the majority of workers, the letter of the law indi- cates that unless the ...rm can prove justi...cation for ...ring, it cannot defeat the worker's plea for reinstatement.17 The ...rm may only refuse to reinstate for cer- tain categories of workers mainly including temporary workers, those with less than one year's tenure, and workers considered to be at-will employees under Mexican law.18 Besides the payments owed to all workers separated from their jobs, all workers ...red unjusti...ably are owed two types of payments. First, they receive back pay including bene...ts covering the period between the date they were ...red and the date at which the court's decision in the lawsuit is executed. Second, they receive three months'salary with bene...ts, with no salary cap. In addition, those workers for whom the ...rm can refuse reinstatement are entitled to 20 days'wage plus bene...ts for each year worked, again with no upper limit on the wage rate.19 Reducing a worker's nominal wage is legally equivalent to an unjusti...ed ...ring.20 A worker whose wage is reduced may force the ...rm to give her full 13Article 162, LFT. 14The worker must be informed in writing of the cause of ...ring. Failure to notify in writing and in a timely fashion implies that the ...ring is considered unjusti...ed under Mexican labor law, regardless of the underlying cause. Article 47, LFT. 15Article 48, LFT. 16In case the worker is reinstated, she receives only back-pay plus fringe bene...ts for the period of time from ...ring to reinstatement. Article 48, LFT. 17Considering that low worker productivity is not a valid cause to for ...ring, the right to demand reinstatement probably constitutes a large ...ring cost for employers, regardless of explicit monetary ...ring costs. Interestingly, in our data we ...nd very few reinstatements. This does not, however, imply that the right to request reinstatement does not a¤ect the bargaining power of workers. 18At-will employees - so called trabajadores de con...anza - include two quite diverse types of employees. On the one hand they include managerial employees, such as supervisors, man- agers, directors, inspectors, and accountants, and on the other hand they include employees whose job implies direct contact with the employer, such as personal sta¤ (for example secre- taries). Article 49, LFT. 19Articles 48 and 50, LFT. 20Article 51-IV, LFT. 12 severance pay, including back pay, three months'salary, and 20 days'wage per year worked, even if the worker is not an at-will employee.21 A ...rm may also avoid having to reinstate workers it ...res without just cause in the case of layo¤s that are warranted given the economic situation of the ...rm.22 A layo¤ is de...ned as a proceeding which the ...rm initiates before the la- bor courts, submitting proof including expert testimony in relation to the ...rm's economic position and the economic situation of the industry. The labor court must then conduct a public hearing in which workers and their representatives, including unions, can participate, as well as the ...rm's experts and experts ap- pointed by the court. After this hearing the labor court declares whether the ...rm can lay o¤ workers. If so, the ...rm avoids having to reinstate any work- ers laid o¤, and need not pay workers the additional 20 days'salary per year worked, although it must still pay three months'wages.23 Finally, a few words about the labor tribunals we study are in order.24 La- bor courts in Mexico (called Juntas de Conciliacion y Arbitraje) are in fact administrative courts that belong to the executive branch and enjoy limited independence from the Secretary of Labor. As their name suggests, these tri- bunals play the role of conciliators as well as adjudicators. Their organic statute mandates at least one conciliation hearing before proceeding to try a case. Fed- eral labor courts have jurisdiction over all labor conicts that involve a certain minimum amount in dispute in a wide range of industries.25 Among the federal tribunals, jurisdiction is determined by industry. The procedure for resolving disputes before the federal labor courts in Mexico is similar to a U.S. style bench trial, preceded by a conciliation process which may terminate the dispute by brokering a settlement between the parties. After the lawsuit is ...led, a conciliation hearing is scheduled.26 The employee must attend this hearing in person, along with his lawyer or legal representative.27 The ...rm may attend via its legal representative, in which case the hearing may result in a settlement. Otherwise, the next step in the process is a hearing similar to a trial, in which each party states its position and presents evidence. Evidence can include oral testimony or written depositions from witnesses, documents, and expert testimony or reports written by experts. This hearing is presided by the judge of the particular labor tribunal with jurisdiction over the ...rm's 21Article 52, LFT. 22Article 434-II, LFT. 23Articles 900-919, LFT. In our sample we do not ...nd any such layo¤ cases initiated by ...rms. However, we do ...nd cases in which ...rms simultaneously ...re large numbers of workers. Given how cumbersome and uncertain the procedure outlined in Articles 900-919 is, it is possible that ...rms almost never make use of the formal layo¤ procedure. 24The following description of the rules governing the operation of the federal labor courts is based on Title 14 of the LFT. 25Article 600-IV, LFT, and Reglamento de la Competencia de las Juntas Especiales que Integran la Junta Federal de Conciliación y Arbitraje. Labor law in Mexico is federal, so that the local juntas are bound by the same substantive law as the federal junta, although they may use simpli...ed procedures. 26Article 876, LFT. 27Legal representatives need not be licensed lawyers for labor disputes under Mexican law. Articles 692-696, LFT. 13 industry. During the hearing itself, no ruling is made by the judge. At the conclusion of the hearing, the judge instructs one of her clerks as to the basic content of the ruling, which the clerk then writes and submits to the judge for comments and revision.28 The decision entails a ruling on all matters of fact, as well as matters of law.29 Once the judge is satis...ed with the ruling, she schedules a meeting of the labor board, consisting of herself along with a representative of each one of the parties.30 In order for a decision to become ...nal, at least one of these representatives must vote along with the judge in favor of the resolution. At least 8 days before voting takes place, the judge must provide the rest of the labor board with a copy of the proposed ruling, and each of these representatives may request an additional hearing in order to allow the parties to present new evidence.31 In the voting session, labor and management representatives can only vote in favor of or against the ruling, but cannot dispute or request changes in any speci...c part of the decision. Parties may resolve their dispute during the conciliation hearing or at any other point in the process before the voting session that ...nalizes the judge's ruling. According to Article 876 of the LFT, settlement agreements that termi- nate a ...led lawsuit must be approved by the judge. All settlements that are not rati...ed by the relevant tribunal are not binding, so that an employee cannot credibly promise not to pursue a suit against his employer unless their settlement is approved by the court. In the data there is no evidence that judges reject settlements, so that the judge's approval appears to serve only as a mechanism for notifying the tribunal and making the agreement binding at law. While Mexican labor law openly promotes the settlement of disputes, it takes an extreme position against the con...dentiality of settlements. Each tri- bunal must record details about settlements it rati...es, such as the date of the settlement and the amount paid to the plainti¤. In addition, given that settle- ment agreements made before a lawsuit is ...led are not legally binding without a labor court's rati...cation, employers and workers very often jointly submit a settlement to the labor court to obtain rati...cation, even though the worker has not sued his employer. Hence, we collect data from ...led settlements as well as 28Article 885, LFT. 29Article 21, Reglamento Interior de la Junta Federal de Conciliación y Arbitraje (Internal Regulations of the Federal Labor Board). 30For each industry in the jurisdiction of the federal labor courts, the Secretary of Labor appoints a labor representative and a ...rm representative. These representatives need not be licensed lawyers. Appointments are often political and usually related to general agreements between the Secretary of Labor and unions or trade associations in the industry. Article 605, LFT. 31Article 886, LFT. According to this article, representatives of workers and ...rms can provide the parties with an additional opportunity to present evidence, thus possibly altering ...ndings of fact made by the judge. The representatives may not challenge ...ndings of law, except by voting against the resolution. Based on interviews with judges and labor lawyers, we believe that representatives'prerogative to request an additional hearing is almost never exercised, and votes are cast simply based on whether the decision generally favors the worker or the ...rm. To the extent that any given ruling must almost always favor one party or the other, the judge obtains the one vote that she needs besides her own, and the resolution stands. 14 from lawsuits that conclude as a settlement or as a court ruling. IV Data We have assembled a data set comprised of a random sample of procedures ...led between 1990 and 1998 in two tribunals in the Mexican federal labor court sys- tem. We sampled from tribunal 15, which covers the pharmaceutical, chemical, paper, automotive and auto parts industries, and from tribunal 6, which covers the textile industry.32 For tribunal 15, we randomly selected 150 case ...les from each year from 1991-1998, with the exception of the year 1992 from which we sampled 215 case ...les. For tribunal 6, we sampled 75 case ...les from each year from 1990-1997.33 There are two main types of procedures: ...led settlements and lawsuits. For ...led settlements, there is only one statement of facts made jointly by the employer and the employee, and resolution of the procedure is always settlement. Lawsuits contain the employee's claim, the employer's answer (if the employer chooses to answer), the terms of settlement reached if the case settles, and the terms of the court's ruling if the case is not settled. Many suits include multiple plainti¤s and are treated as correlated data points in the statistical analysis. In this section we describe the main variables relating to the lawsuit, worker and employer information, and resolution of the conict. For all procedures ...led in our sample, we observe the motive for ...ling,34 the date of ...ling, the geographical location of the dispute, and whether the procedure is a settlement or a lawsuit. With respect to information collected from the worker's ...ling, we have information about the type of job held,35 the date the job started and ended, the salary with and without fringe bene...ts, hours worked per week, the worker's demands,36 as well as worker gender, date of birth, and sometimes the worker's social security ID number.37 With respect 32 These data were obtained by the authors using a new law governing freedom of governmen- tal information in Mexico. Although some of the variables used in this study are considered to be public information under the law, other variables are not public information, and have been obtained under a con...dentiality agreement between the Federal Labor Courts System and the authors. 33 For tribunal 15, the total number of case ...les was 973 in 1991, 951 in 1992, 1020 in 1993, 865 in 1994, 902 in 1995, 722 in 1996, 672 in 1997, and 795 in 1998. For tribunal 6, the total number of case ...les was 728 in 1990, 699 in 1991, 700 in 1992, 860 in 1993, 690 in 1994, 574 in 1995, 414 in 1996, and 403 in 1997. 34 Most procedures in our sample are related to a ...ring. A few suits do not dispute the ...ring decision but claim incomplete severance pay or incomplete payment of fringe bene...ts. There are also a few pension cases. 35 Although the claim speci...es the actual job description, we only use this to classify work- ers as standard employees or as at-will (supervisory) employees, who are entitled to higher severance pay under the labor law. 36 In ...ring law suits, workers generally demand reinstatement, back-pay, overtime, fringe bene...ts, and severance pay. 37 The presence of the ID number allows us to link the data from the lawsuit to con...- dential data on the worker's employment records available from the Mexican social security administration. The latter data tell us the wage reported for the worker since 1985, as well as an identi...er for the worker's employer, the industry and the location of the worker. For 15 to the worker's claims, we collect very detailed data that allow us to construct three variables: the actual amount of money claimed by the worker, an imputed claim assuming the dismissal was unjusti...ed but based only on statements that we believe are easily veri...able, and an imputed claim of what the law would assign to this worker given justi...ed ...ring, again based on statements that we believe are easily veri...able. In order to calculate our imputed claims we assume that the worker is accu- rately reporting certain "easily veri...able"features of the case such as the wage and the dates when the worker was employed. We ignore certain other claims such as having worked an extraordinary amount of overtime or never having re- ceived constitutionally-mandated bene...ts despite the fact that the worker could have demanded these bene...ts prior to the current lawsuit. For the employer we have a ...rm identi...er, the location of the business, and the industry. In lawsuits to which the employer provides an answer, we also have the employer's version of the facts cited by the worker in her claim, such as the worker's job description, salary, and so on. Additionally we code other evidence submitted by the ...rm to establish that the worker was never hired or ...red, was ...red with some justi...cation,38 has received fringe bene...ts payment, or has already accepted a severance package from the ...rm. In terms of the procedures' outcomes, as explained before, a substantial proportion of the procedures ...led arrive to the tribunals as a settlement, and are always rati...ed by the courts as such. For lawsuits, we observe three types of conclusions: dropped suits, settlements, and trials leading to a judgment by the court. We record the date of conclusion of the procedure, the payment received by the worker, and any previous payments recognized by the court. For trials, we observe a trial result stated by the court,39 the votes of the judge and the representatives of labor and management in favor of or against the judgment,40 the facts of the case as recognized by the judge, the number of constitutional appeals ...led,41 and the number of judgments made by the court. the present paper we have only used the social security data to verify wages reported in the lawsuits and to follow up on dropped cases. 38In cases where the ...rm alleges having ...red the worker justi...ably, it provides evidence of one of the causes for justi...ed ...ring described in the law. 39The court states whether its decision is in favor of the worker's claim, the employer, or mixed, in the sense that the judge concedes only part of the claim. 40As explained in the previous section, in order for a ...nal judgment to be valid, at least one of the representatives of the parties must `vote'in favor of it, so that along with the court's vote they constitute a majority. We rarely ...nd both representatives voting in favor of the judge's resolution. 41In cases that proceed to a trial, it is common for one or both parties to ...le constitutional appeals, generally claiming violations of due process. For each successful appeal ...led, the court must issue a new judgment, so that in some cases we observe several decisions by the court. 16 V Statistical Analysis In this paper we do not analyze procedures that arrived to the court as a settle- ment without having initiated as a lawsuit ...led by the worker. In addition, we limit our analysis to lawsuits related to ...ring, which constitute the vast majority of lawsuits in the database. We ...rst report descriptive statistics of lawsuits in our sample. We then present kernel-density estimations that show the distribution of awards and the relationship between awards and claims. Finally we describe our econometric results, which present a story consistent with the intuition provided by the descriptive statistics and by the kernel-density graphs. Table 1 reports several statistics on lawsuits, including statistics on how lawsuits are resolved. Around 70% of lawsuits are settled, and among the 30% that are not, slightly more than half are dropped and slightly less than half go to trial. We ...nd quite similar results in the two tribunals we study. Table 1, like all tables in the paper, is calculated using the inverse of the ex-ante probability of a lawsuit being included in the sample as weights. This is done to approximate what we would have estimated if we had collected data from a census of lawsuits from each tribunal. This weighting procedure essentially adds more weight to lawsuits in years with more total lawsuits, since each sampled lawsuit in these years "represents"a larger number of lawsuits. Table 1 also shows summary statistics for several other variables, including the award received by the employee. All monetary variables are converted to their equivalent value in December 1998 pesos. The employee's claim simply measures the amount of money requested by the plainti¤ in the lawsuit. We also include our two imputed claims, that is, what Mexican labor law would award to the worker based only on facts of the case that are veri...able relatively easily such as dates worked and salary. Based on these easily veri...able facts, the ...rst estimation assumes the worker was ...red without justi...cation, while the second assumes that the worker was ...red with justi...cation. Additionally we report the percentage of the claim obtained by the employee. Note that employees receive substantially less than they ask for, in particular, we do not ...nd that they receive on average half of what they request, as the literature on arbitration might suggest. Bearing in mind that when ...rms answer the lawsuit they often acknowledge some positive amount of money owed to the worker, the amount obtained by workers is far from what "splitting the di¤erence"would suggest.42 Finally, table 1 reports statistics on the percent of litigants who are female, the percent of litigants with "at will"contracts (typically, but not always, white collar workers), tenure at the ...rm, and the percent of litigants with 15 years of tenure or more. These variables will serve as control variables 42Note that average award is less than 10% of average claim in one tribunal, and around 23% in the other. However, the average percentage obtained by the worker is closer to 30%. This di¤erence arises because the percentage statistic computes the average of the percentage that each worker obtains of his claim, rather than the average award divided by the average claim. 17 and are not the main focus of the paper.43 Tables 2, 3, and 4 show the same summary statistics presented in table 1, but presented separately for lawsuits that are settled, tried, and dropped respectively. Although we will undertake more formal analyses later in this section, we believe it is useful to begin with simple comparisons of means. Comparing tables 2 and 3, we see some interesting di¤erences between law- suits that end up being settled and lawsuits that go to a ...nal judgment. First note that, in both tribunals, workers receive a higher percentage of what they ask for in lawsuits that are eventually settled. In both tribunals, our imputed claims of what the worker is entitled to based on relatively easily veri...able facts is higher in lawsuits that eventually settle than in lawsuits that do not settle. Worker claims as a percentage of our imputed claims, however, are larger in lawsuits that end up going to a ...nal judgment. In fact, we see in the tribunal 15 that, despite the fact that our imputed claims lead us to believe that the law- suits that end up being settled are "stronger"cases for the workers, the workers in this tribunal ask for more in lawsuits that go to ...nal judgment. Taken to- gether, these results suggest that the lawsuits that get settled are the ones in which the worker is asking for a payment that is more in accordance with a conservative reading of Mexican labor laws.44 Table 3, which presents descriptive statistics on tried lawsuits, also presents some statistics on whether the judgment was favorable for the worker or for the ...rm. The judge's ruling, in addition to containing the amount that will be awarded to the worker, contains a summary of the judgment that can have three possible values. The most common value (68% in tribunal 15, 67% in tribunal 6) is "favorable for the ...rm." We will present further evidence that the ...rm typically wins tried cases later in this section. The second most common value (26% in tribunal 15, 32% in tribunal 6) is "mixed", that is, neither completely in favor of the ...rm nor completely in favor of the worker. The least common value is "favorable for the worker." In fact, this value was never observed in tribunal 6 and only observed in 7% of the tried lawsuits in tribunal 15. We see from table 4 that dropped lawsuits typically involve claims that dis- pute signi...cant amounts of money. In both tribunals, for instance, the workers' claims in lawsuits that eventually get dropped are higher than the averages for all cases. In tribunal 15, our imputed claims of what the worker is entitled to based on easily veri...able facts is also higher for dropped lawsuits than for all lawsuits. One might worry that these dropped lawsuits really represent ones in which the ...rm rehires the worker to convince the worker to drop the case. If this 43Perhaps some bivariate correlations of these variables with the main outcome variables would be useful. In both tribunals, women receive signi...cantly lower log ...nal payments. In the tribunal 15, women also receive a lower percentage of their claim and are more likely to go to trial and less likely to settle. In both tribunals, tenure is positively correlated both with ...nal payment received and percent of claim obtained. Results are similar for the dummy for tenure of 15 years of tenure or more. In tribunal 15, at will employees obtain lower percentages of their claims. 44Table 3 also presents evidence that, in tribunal 15, women tend to go to trial more often. This will be supported by econometric analysis later in this section. 18 were true, these dropped lawsuits might cause serious econometric concerns. We would not observe any monetary award for the plainti¤, but in reality the worker might have received substantial compensation. To evaluate the potential severity of this problem, we examined those workers for whom we observe a social security ID number. In tribunal 15, we observed 20 workers who dropped their lawsuits, 4 of whom we observed to be working at the same ...rm after the lawsuit was dropped. For settled lawsuits (again for those workers for whom we observed the social security number) only 2 out of 99 workers were observed at the same ...rm after the lawsuit was settled. For lawsuits that went to trial, 3 workers out of 47 were observed at the same ...rm after the lawsuit ended. Unfortunately, sample sizes were too small to do meaningful comparisons in tribunal 6. Our interpretation is that the majority of dropped lawsuits are workers who simply gave up and received no compensation. Nevertheless, there is some evidence that dropped lawsuits might in some cases be successes for the workers. For this reason, we estimate models in which dropped lawsuits are included and treated as if the worker received nothing and we estimate models in which dropped cases are excluded from the analysis. We now turn to our graphical analyses. Figure 1 shows a kernel-density estimate of the distribution of the log di¤erence between the amount the worker asks for and the amount the worker receives for both tribunals. When the worker receives zero, we set the log of the payment equal to zero. Note ...rst that the majority of the distribution lies in the negative region of the ...gure, indicating that nearly all workers receive less than they demanded. Also note that the distribution is bimodal. We interpret the bimodal feature of the distribution as evidence that the worker either "wins"or "loses." One might suspect that the "worker lose" spike is dominated by lawsuits in which the worker receives no compensation. Figure 2 demonstrates that this is indeed the case by excluding lawsuits in which the worker received no compensation. It is clear that the "worker lose" spike in the distribution has disappeared. Figure 3 does the same exercise as in ...gure 1 for lawsuits that reach a ...nal judgment, that is, for lawsuits that are not settled and are not dropped. Once again we see a bimodal distribution. Note that in this ...gure the "worker wins" spike is smaller compared to ...gure 1 and the "worker loses"spike is bigger com- pared to ...gure 1. These results lend further support to the evidence in table 3 that workers do relatively poorly in lawsuits that reach a ...nal judgment. Fig- ure 4 o¤ers further evidence by examining lawsuits that eventually get settled. Note that this distribution is unimodal and that the spike of the distribution lies approximately where the "worker wins"spike lies in the previous ...gures. One might, however, question our interpretation of the two spikes as "worker wins"and "worker loses."In order to bolster this assertion, ...gure 5 examines tried cases that ended in a favorable judgment for the ...rm. It seems apparent that the "worker loses" spike is dominant for these lawsuits. We continue to analyze tried lawsuits in ...gure 6, but we focus here on judgments that were not favorable to the ...rm, that is, we aggregate into one category judgments that 19 were favorable for the worker and "mixed"rulings. It seems apparent here that the "worker wins"spike now dominates the distribution.45We now turn to the econometric section of the paper where we will further bolster the above claims and undertake additional analyses. Table 5 reports the results of econometric models estimating the determi- nants of the log of the ...nal award. Once again we set the log award equal to zero when the actual award is zero. We present both ordinary least squares models and Tobit models; awards of zero are treated as censored observations in the Tobit models. The key independent variables are the log of the worker's claim as well as a dummy for whether the case ends in a ruling by the judge. We also include a dummy variable for whether the worker is female, a dummy variable for whether the worker has an "at will"contract, a dummy variable for whether the worker has tenure of at least 15 years in the ...rm, a dummy vari- able for the year in which the worker separated from the ...rm (seven dummies for the eight years in the data in each tribunal) and a dummy variable for the quarter in which the worker separated from the ...rm (three dummies for the four quarters). The models for tribunal 15 also include three dummy variables for the indus- try of the ...rm (the four categories are chemical/pharmaceutical, automotive, paper, and no industry information). Since all ...rms in tribunal 6 are textile ...rms, the models for tribunal 6 contain no industry dummies. For ease of ex- position we do not report the coe¢ cients on the year, quarter, and industry dummies although we are happy to provide these results upon request. Our models allow for the possibility of heteroscedasticity and allow for a correlation of outcomes when we observe multiple plainti¤s in the same case ...le against the same defendant. We estimate these models separately for each tribunal and estimate them separately both including and excluding dropped cases from the analysis. Two main results emerge from table 5. First, there is only weak evidence that the worker's claim is correlated with the amount she receives. We only observe a signi...cant coe¢ cient for tribunal 15 when dropped cases are excluded. We will present evidence later in this section that the worker's claim often contains highly dubious assertions that tend to be ignored by the judge. We also see that lawsuits that reach a ...nal judgment pay substantially less than cases that do not reach a ...nal judgment. This is a particularly interesting result when we include dropped lawsuits in the analysis, noting that all dropped lawsuits involve no payments at all. We emphasize the fact that the terms of settlement are typically not observed in data sets of this nature. Few empirical studies can compare the award amounts between lawsuits that reach a ...nal judgment and those that do not.46 45Since there were no rulings in favor of the worker in tribunal 6, in ...gure 6 for this tribunal the distribution is comprised entirely of "mixed"rulings. 46We were surprised to see that the female dummy was insigni...cant in all models. In tribunal 15, both the dummy for having an "at will" contract and the dummy for tenure at the ...rm of at least 15 years have positive and signi...cant coe¢ cients when dropped cases are excluded. In tribunal 6, the dummy for tenure at the ...rm of at least 15 years has a positive 20 Table 6 adds an additional independent variable that we believe is interest- ing. We include a dummy variable indicating whether the plainti¤ is involved in a lawsuit that is grouped together with other plainti¤s against the same ...rm. These estimations may shed light on game-theoretic models involving repeat players. When the ...rm is taking its decision with regard to one plainti¤, the ...rm must take into account any inferences the other plainti¤s might make about the ...rm's willingness to be aggressive. We see from table 6 that, when we include dropped lawsuits in the analysis, lawsuits involving multiple plainti¤s tend to pay less to the workers, although the result is not signi...cant in tribunal 6. We see no such evidence, however, when we exclude dropped lawsuits. One wonders, however, whether it is ap- propriate to control for the mode of termination of the lawsuit. If one believes that a ...rm would act more aggressively in lawsuits involving multiple plainti¤s, the e¤ects on award amounts might work through the di¤erences in settlement probabilities. Table 7 investigates this possibility by estimating similar models, without controlling for mode of termination. The results for tribunal 15 now indicate that lawsuits involving multiple plainti¤s pay less to the worker, whether or not dropped lawsuits are included in the analysis. the pattern looks quite similar in tribunal 6, although the results are not statistically signi...cant. Taken together, tables 6 and 7 present evidence that cases involving multiple plainti¤s end up paying less to the workers, and some of this e¤ect is due to cases involving multiple plainti¤s di¤ering in their modes of termination. This evidence is consistent with the theory that ...rms involved in multiple or repeated lawsuits will settle less and exert more e¤ort to win at trial in order to build a reputation that will deter future lawsuits. Also, these ...ndings may be consistent with a negative selection e¤ect of multiple plainti¤s. A potential plainti¤ may ...nd suing less expensive given that other workers are involved in the same suit, and this may result in a lower average quality of multiple lawsuits. Table 8 presents further evidence on this issue by explicitly examining the determinants of the mode of termination. When dropped lawsuits are included in the analysis, we estimate separate logit models for the three possible modes of termination. When dropped lawsuits are excluded from the analysis, we estimate one logit model of the probability of settlement. We control for the worker's claim at the time of the ...ling of the lawsuit and include a dummy variable indicating whether the lawsuit involved multiple plainti¤s in the same case ...le.47 In tribunal 15, whether or not we exclude dropped lawsuits from the analysis, cases are less likely to be settled when they involve multiple plainti¤s.48 Once again the results for tribunal 6 look similar but with weaker statistical and signi...cant coe¢ cient whether or not dropped cases are included in the analysis. 47The worker's claim normally increases over time since "lost wages"continue to accumulate. We do not allow for this claim to increase over time in the logit models of mode of termination to avoid endogeneity problems. We also include the same additional controls used in previous tables. 48It may also be worthwhile to note that, for tribunal 15, women are less likely to settle and more likely to go to trial. Results for tribunal 6 are statistically insigni...cant. 21 signi...cance.49Multinomial logit results produce very similar results.50 Thus far we have not exploited our imputed claims of what the worker would be entitled to given the easily veri...able facts of the case. In table 9, we return to our models of award amounts by adding the log of our imputed claim, assuming the dismissal was not justi...ed. We generally ...nd that neither the claim of the worker nor the imputed claim is signi...cant. In tribunal 15, however, the imputed claim is positively and signi...cantly associated with the amount received when dropped cases are excluded. In table 10, we estimate the same models as in table 9, but without control- ling for mode of termination. If an exaggerated claim a¤ects the probability of the mode of termination, excluding mode of termination may be a better way of estimating the overall e¤ect of an exaggerated claim. Now that we are not controlling for mode of termination, the coe¢ cient on the log of the total claim is negative in all models and statistically signi...cant when dropped cases are ex- cluded.51 That is, we ...nd some evidence that workers are actually punished for exaggerating their claims. The coe¢ cient on the imputed claim is positive and signi...cant in all models suggesting that it is a good indicator of the strength of the worker's case. To further address the point on how an exaggerated claim might a¤ect the mode of termination, we return to logit analyses of the mode of termination in table 11. We include an additional control variable of the log of the ratio between the worker's claim and our imputed claim assuming an unjusti...ed dismissal. In both tribunals, whether we include dropped lawsuits in the analysis or not, we ...nd that cases involving "excessive"worker claims tend not to be settled.52 Overall, tables 9-11 present an interesting story. Some workers exaggerate their claims more than others. Additionally, workers who exaggerate their claims settle less often, presumably because they are either aggressive by nature or because they incorrectly estimate the strengths of their cases. In any event, the courts are not fooled by these exaggerated claims. In fact, the courts may punish workers for exaggerating their claims. VI Conclusions Motivated by the previous theoretical and empirical literatures on the resolution of legal disputes, we use data from labor tribunals in Mexico to address three 49When all three modes of termination are considered, we ...nd that cases involving multiple plainti¤s are less likely to be settled and more likely to go to trial, with both results signi...cant at the 10% level. When dropped cases are excluded, we ...nd that cases involving multiple plainti¤ are less likely to settle, but the estimated coe¢ cient is not statistically signi...cant. 50The multinomial logit speci...cation requires the standard "independence of irrelevant al- ternatives" assumption, which we believe is di¢ cult to defend in our context. Since the multinomial logit produces the same results as separate logit speci...cations on each case out- come, we prefer to report the latter results. 51In tribunal 15 when dropped cases are included, the coe¢ cient on total claim is negative and signi...cant at the 10% level. 52Adding the worker's claim to these models yields insigni...cant coe¢ cients and does not change the main results. 22 empirical issues. First, we compare ...nal payments in cases that are settled to those in cases that go to trial, a comparison which is relatively rare in this literature since settlement amounts are rarely observed. Second, we address di¤erences between "high-stakes" players and others by examining di¤erences between ...rms involved in cases involving multiple workers compared to ...rms in cases involving a single worker. Finally, we examine whether tried cases tend to be of higher or lower "quality" by examining workers who are apparently exaggerating their claims. Perhaps our strongest result is that ...nal payments to workers are signif- icantly lower in cases that go to trial compared with cases that settle. We also ...nd that ...rms involved in cases with multiple workers (...rms with higher stakes) tend to settle less and make lower ...nal payments to their workers. Fi- nally, we ...nd that workers who exaggerate their claims settle less often and may be punished for such exaggeration in the ...nal award amounts. We conclude by placing our results in the broader literature. Several theoret- ical models imply that ...nal payments may be di¤erent in settled cases compared with tried cases due to a selection e¤ect of going to trial. We do ...nd substantial di¤erences in ...nal payments between cases that are settled and cases that go to trial. In fact, we ...nd direct evidence of a case-selection e¤ect by showing that workers who exaggerate their claims (and therefore have weaker cases) tend to settle less often. Our results on ...rms involved in cases against multiple workers are also rele- vant for models involving repeat players or "high-stakes"players. These models predict that high-stakes litigants will be concerned about how their current ac- tions a¤ect their reputations. These ...rms would therefore be more aggressive in negotiations and would exert more e¤ort at trial. Our results that ...rms involved in cases with multiple plainti¤s settle less often and end up making lower ...nal payments are consistent with these theoretical predictions. 23 References 1. Bebchuk, Lucian Ayre (1984). Litigation and Settlement under Imperfect Information, The Rand Journal of Economics 15(3): 404-415. 2. Black, Bernard, Charles Silver, David A. Hyman, and William M. Sage (2005). Stability, Not Crisis: Medical Malpractice Claims in Texas, 1988 - 2002. Journal of Empirical Legal Studies 2: 207-245. 3. Boden, Leslie (1992). Dispute Resolution in Workers'Compensation, Re- view of Economics and Statistics, 74(3): 493-502. 4. Chandra, Amitabh, Shantanu Nundy, and Seth A. Seabury (2005). The Growth of Physician Malpractice Payments: Evidence from the National Practitioner Data Bank, Health A¤airs 24:240-50, Chevy Chase. 5. Danzon, Patricia (1984). The Frequency and Severity of Medical Mal- practice Claims. Journal of Law and Economics 27(1): 115-148. 6. Daughety, Andrew F. and Jennifer F. Reinganum (1999). Hush Money, The Rand Journal of Economics 30(4): 661-678. 7. Dertouzos, James N., Elaine Holland, and Patricia Ebener (1988). The Legal and Economic Consequences of Wrongful Termination, Rand, R- 3602-ICJ. 8. Eisenberg, Theodore and Henry S. Farber (1997). The Litigious Plainti¤ Hypothesis: Case Selection and Resolution, Rand Journal of Economics 28: S92-112. 9. Farber, Henry S. and Max H. Bazerman (1986). The General Basis of Arbitrator Behavior: An Empirical Analysis of Conventional and Final- O¤er Arbitration, Econometrica 54(4): 819-844. 10. Farmer, Amy and Paul Pecorino (2004). The Causes of Bargaining Failure: Evidence from Major League Baseball, The Journal of Law and Economics 47: 543-564. 11. Fenn, Paul and Neil Rickman (1999). Delay and Settlement in Litigation, Economic Journal 109(457): 476-491. 12. Fournier, Gary M. and Thomas W. Zuehlke (1989). Litigation and Settle- ment: An Empirical Approach, The Review of Economics and Statistics 71(2): 189-195. 13. Gibbons, Robert (1988). Learning in Equilibrium Models of Arbitration, American Economic Review 78(5): 896-912. 14. Gross, Samuel R. and Kent D. Syverud (1991). Getting to No: A Study of Settlement Negotiations and the Selection of Cases for Trial, Michigan Law Review 90: 319-393. 24 15. Kakalik, James S., Patricia A. Ebener, William L.F. Felstiner, Gus W. Haggstrom, and Michael G. Shanley (1984). Variation in Asbestos Litiga- tion Compensation and Expenses, Rand, R-3132-ICJ. 16. Kakalik, James S., Elizabeth M. King, Michael Traynor, Patricia A. Ebener, and Larry Picus (1988). Costs and Compensation Paid in Aviation Acci- dent Litigation, Rand, R-3421-ICJ. 17. Kessler, Daniel., Thomas Meites, and Geo¤rey Miller (1996). Explain- ing Deviations from the Fifty-Percent Rule, The Journal of Legal Studies 25(1): 233-59. 18. Nalebu¤, Barry (1987). Credible Pretrial Negotiation, The Rand Journal of Economics 18(2): 198-210. 19. P'ng, I.P.L. (1983). Strategic Behavior in Suits Settlements and Trial, The Bell Journal of Economics 14(2): 539-550. 20. Priest, George L. and Benjamin Klein (1984). The Selection of Dispute for Litigation, The Journal of Legal Studies 13(1): 1-55. 21. Schwab, Stewart and Theodore Eisenberg (1988). Explaining Constitu- tional Tort Litigation: The Inuence of the Attorney Fees Statute and the Government as Defendant, Cornell Law Review 73: 719-783. 22. Sieg, Holger (2000). Estimating a Bargaining Model with Asymmetric Information: Evidence from Medical Malpractice Disputes, The Journal of Political Economy 108(5): 1006-1021. 23. Siegelman, Peter and John J. Donohue III (1995). The Selection of Em- ployment Discrimination Disputes for Litigation: Using Business Cycle E¤ects to Test the Priest-Klein Hypothesis, The Journal of Legal Studies 24(2): 427-462. 24. Spier, Kathryn E. (1992). The Dynamics of Pretrial Negotiation, The Review of Economic Studies 59(1): 93-108. 25. Waldfogel, Joel (1995). The Selection Hypothesis and the Relationship between Trial and Plainti¤ Victory, The Journal of Political Economy 103(2): 229-260. 25 Figure 1: Kernel Density Estimate of Log Difference Between Claim and Payment (Log payment set to zero when payment is zero) Tribunal 15 Tribunal 6 Density Density .15 . 2 .1 .15 .1 .05 .05 0 0 -20 -15 -10 -5 0 5 -20 -15 -10 -5 0 5 Log difference between claim and payment Log difference between claim and payment Figure 2: Kernel Density Estimate of Log Difference Between Claim and Payment: (Positive payments only) Tribunal 15 Tribunal 6 Density Density .3 .2 .2 .1 .1 0 0 -15 -10 -5 0 5 -20 -15 -10 -5 0 5 Log difference between claim and payment Log difference between claim and payment Figure 3: Kernel Density Estimate of Log Difference Between Claim and Payment: (Judges' Rulings Only, log payment set to zero when payment is zero) Tribunal 15 Tribunal 6 Density Density .1 .08 .08 .06 .06 .04 .04 .02 .02 0 0 -20 -15 -10 -5 0 5 -20 -15 -10 -5 0 5 Log difference between claim and payment Log difference between claim and payment Figure 4: Kernel Density Estimate of Log Difference Between Claim and Payment: (Settlements Only, log payment set to zero when payment is zero) Tribunal 15 Tribunal 6 Density Density .3 .2 .2 .1 .1 0 0 -20 -15 -10 -5 0 5 -20 -15 -10 -5 0 5 Log difference between claim and payment Log difference between claim and payment Figure 5: Kernel Density Estimate of Log Difference Between Claim and Payment: (Judge's ruling in favor of firm, log payment set to zero when payment is zero) Tribunal 15 Tribunal 6 Density Density .2 .4 .15 .3 .1 .2 .05 .1 0 0 -20 -15 -10 -5 0 5 -20 -15 -10 -5 0 5 Log difference between claim and payment Log difference between claim and payment Figure 6: Kernel Density Estimate of Log Difference Between Claim and Payment: (Judge's ruling not in favor of firm, log payment set to zero when payment is zero) Tribunal 15 Tribunal 6 Density Density .2 .15 .2 .1 .1 .05 0 0 -20 -15 -10 -5 0 5 -20 -15 -10 -5 0 5 Log difference between claim and payment Log difference between claim and payment Table 1: Descriptive Statistics: all lawsuits Tribunal 15 Obs Mean Std. Dev. Min Max Percent Settled 1076 68 47 0 100 Percent Tried 1076 14 35 0 100 Percent Dropped 1076 18 38 0 100 Award 1076 23,629 59,626 0 1,001,167 Claim 1076 259,171 703,228 6,747 11,700,000 Imputed award unjustified firing 1076 64,190 226,613 3,820 5,729,232 Imputed award justified firing 1076 38,916 215,297 210 5,718,458 Percent of claim obtained 1076 29 52 0 298 Percent Female 1076 30 46 0 100 Percent of workers who are "at will" 1076 19 39 0 100 Years of tenure at firm 1076 5 6 0 39 Percent with tenure 15 years 1076 8 28 0 100 Tribunal 6 Percent Settled 547 71 46 0 100 Percent Tried 547 14 34 0 100 Percent Dropped 547 16 37 0 100 Award 547 56,387 412,704 0 4,760,639 Claim 547 238,833 882,595 3,043 11,600,000 Imputed award unjustified firing 547 115,958 712,606 2,038 11,600,000 Imputed award justified firing 547 94,474 679,038 348 11,600,000 Percent of claim obtained 547 26 38 0 435 Percent Female 547 26 44 0 100 Percent of workers who are "at will" 547 15 35 0 100 Years of tenure at firm 547 8 9 0 50 Percent with tenure 15 years 547 17 38 0 100 Notes: All monetary values are expressed in December 1998 pesos. Each observation is given the weight of the inverse of its ex-ante probability of being included in the sample. See text for details. Table 2: Descriptive Statistics: settled lawsuits Tribunal 15 Obs Mean Std. Dev. Min Max Award 729 27,133 54,438 0 1,001,167 Claim 729 224,272 673,560 6,747 11,700,000 Imputed award unjustified firing 729 64,264 262,967 3,820 5,729,232 Imputed award justified firing 729 40,167 254,404 210 5,718,458 Percent of claim obtained 729 40 58 0 298 Percent Female 729 27 44 0 100 Percent of workers who are "at will" 729 18 39 0 100 Years of tenure at firm 729 5 6 0 39 Percent with tenure 15 years 729 7 26 0 100 Tribunal 6 Award 388 76,455 490,065 0 4,760,639 Claim 388 243,902 949,655 3,043 11,600,000 Imputed award unjustified firing 388 140,149 844,468 2,038 11,600,000 Imputed award justified firing 388 116,014 805,069 378 11,600,000 Percent of claim obtained 388 33 42 0 435 Percent Female 388 24 43 0 100 Percent of workers who are "at will" 388 15 36 0 100 Years of tenure at firm 388 8 8 0 50 Percent with tenure 15 years 388 18 38 0 100 Notes: All monetary values are expressed in December 1998 pesos. Each observation is given the weight of the inverse of its ex-ante probability of being included in the sample. See text for details. Table 3: Descriptive Statistics: tried lawsuits Tribunal 15 Obs Mean Std. Dev. Min Max Award 153 36,634 99,472 0 657,101 Claim 153 370,782 748,560 12,351 4,778,867 Imputed award unjustified firing 153 54,441 90,441 5,059 874,882 Imputed award justified firing 153 28,700 62,573 494 603,464 Percent of claim obtained 153 16 40 0 218 Percent Female 153 41 49 0 100 Percent of workers who are "at will" 153 21 41 0 100 Years of tenure at firm 153 5 6 0 31 Percent with tenure 15 years 153 10 30 0 100 Percent of rulings for worker 153 7 26 0 100 Percent of rulings for firm 153 67 47 0 100 Percent of "mixed" rulings 153 26 44 0 100 Tribunal 6 Award 69 18,339 32,977 0 186,293 Claim 69 172,608 288,101 32,519 1,894,045 Imputed award unjustified firing 69 56,401 101,758 8,997 710,333 Imputed award justified firing 69 41,088 88,366 1,101 611,824 Percent of claim obtained 69 18 27 0 157 Percent Female 69 24 43 0 100 Percent of workers who are "at will" 69 12 33 0 100 Years of tenure at firm 69 8 10 0 42 Percent with tenure 15 years 69 20 40 0 100 Percent of rulings for worker 69 0 0 0 0 Percent of rulings for firm 69 68 47 0 100 Percent of "mixed" rulings 69 32 47 0 100 Notes: All monetary values are expressed in December 1998 pesos. Each observation is given the weight of the inverse of its ex-ante probability of being included in the sample. See text for details. Table 4: Descriptive Statistics: dropped lawsuits Tribunal 15 Obs Mean Std. Dev. Min Max Award 194 0 0 0 0 Claim 194 300,871 764,719 7,908 6,092,656 Imputed award unjustified firing 194 71,740 138,378 4,724 1,301,103 Imputed award justified firing 194 42,411 106,689 387 1,074,676 Percent of claim obtained 194 0 0 0 0 Percent Female 194 33 47 0 100 Percent of workers who are "at will" 194 20 40 0 100 Years of tenure at firm 194 6 7 0 39 Percent with tenure 15 years 194 11 32 0 100 Tribunal 6 Award 90 0 0 0 0 Claim 90 272,725 913,254 7,216 5,691,863 Imputed award unjustified firing 90 59,626 129,257 4,261 801,570 Imputed award justified firing 90 44,615 121,703 348 758,498 Percent of claim obtained 90 0 0 0 0 Percent Female 90 33 47 0 100 Percent of workers who are "at will" 90 14 35 0 100 Years of tenure at firm 90 8 9 0 37 Percent with tenure 15 years 90 14 35 0 100 Notes: All monetary values are expressed in December 1998 pesos. Each observation is given the weight of the inverse of its ex-ante probability of being included in the sample. See text for details. Table 5: Effects of worker's assertion and mode of termination on final payment (Dependent Variable: Log of final payment) Dropped cases included Tribunal 15: Tribunal 6: Tobit OLS Tobit OLS coef std err coef std err coef std err coef std err ln(worker's claim) 0.17 0.21 0.14 0.16 0.17 0.37 0.19 0.30 trial -5.10 *** 1.13 -3.39 *** 0.70 -3.09*** 1.11 -2.28 *** 0.80 female -0.42 0.61 -0.36 0.45 -0.99 0.82 -0.74 0.62 at-will worker 0.58 0.58 0.49 0.43 0.49 0.78 0.36 0.62 tenure 15 years 1.27 0.87 1.03 0.67 1.51 ** 0.75 1.32 ** 0.61 R2 0.14 0.11 Number of obs 1,076 1,076 547 547 Censored obs 288 122 Dropped cases excluded Tribunal 15: Tribunal 6: Tobit OLS Tobit OLS coef std err coef std err coef std err coef std err ln(worker's claim) 0.31 *** 0.11 0.29 *** 0.10 0.22 0.20 0.23 0.19 trial -6.24 *** 0.75 -5.51 *** 0.60 -4.41*** 0.89 -4.01 *** 0.77 female 0.33 0.29 0.29 0.26 -0.34 0.34 -0.30 0.32 at-will worker 0.73 *** 0.27 0.70 *** 0.25 0.29 0.34 0.26 0.32 tenure 15 years 2.31 *** 0.41 2.17 *** 0.37 1.45 *** 0.48 1.39 *** 0.45 R2 0.53 0.36 Number of obs 882 882 457 457 Censored obs 94 32 Notes: The dependent variable is the log of the amount awarded to the employee in December 1998 pesos. In cases in which the amount awarded was zero, we set the log of the award to zero. Additionally, these cases are treated as censored observations in the Tobit model. Standard errors are calculated allowing for heteroscedasticity and for the possibility that the outcomes in cases that have been grouped into the same proceeding may be correlated. We use the notation of *** to denote significance at the 0.01 level. Similarly ** denotes significance at the 0.05 level and * denotes significance at the 0.10 level. All models include dummy variables for the year when the employee separated from the firm and the quarter when the employee separated from the firm. Industry dummies are included in the models for tribunal 15. Each observation is given the weight of the inverse of its ex-ante probability of being included in the sample. See text for details. Table 6: Effects of worker's assertion, mode of termination, and cases involving multiple workers on final payment (Dependent Variable: Log of final payment) Dropped cases included Tribunal 15: Tribunal 6: Tobit OLS Tobit OLS coef std err coef std err coef std err coef std err ln(worker's claim) 0.09 0.22 0.09 0.16 0.20 0.35 0.21 0.29 trial -4.68 *** 1.09 -3.09 *** 0.69 -2.94*** 1.12 -2.18 *** 0.81 multiple workers -2.03 ** 0.82 -1.42 ** 0.56 -1.30 1.08 -0.91 0.81 female -0.40 0.58 -0.33 0.42 -0.80 0.74 -0.60 0.57 at-will worker 0.60 0.55 0.52 0.40 0.37 0.75 0.29 0.59 tenure 15 years 1.15 0.84 0.95 0.65 1.61 ** 0.74 1.39 ** 0.60 R2 0.16 0.12 Number of obs 1,076 1,076 547 547 Censored obs 288 122 Dropped cases excluded Tribunal 15: Tribunal 6: Tobit OLS Tobit OLS coef std err coef std err coef std err coef std err ln(worker's claim) 0.30 ** 0.12 0.28 *** 0.11 0.22 0.20 0.23 0.19 trial -6.18 *** 0.69 -5.46 *** 0.56 -4.40*** 0.85 -4.00 *** 0.74 multiple workers -0.23 0.38 -0.16 0.33 -0.09 0.63 -0.05 0.58 female 0.32 0.30 0.28 0.26 -0.33 0.33 -0.30 0.31 at-will worker 0.72 *** 0.27 0.69 *** 0.25 0.28 0.33 0.26 0.31 tenure 15 years 2.29 *** 0.40 2.16 *** 0.36 1.45 *** 0.47 1.39 *** 0.45 R2 0.53 0.36 Number of obs 882 882 457 457 Censored obs 94 32 Notes: The dependent variable is the log of the amount awarded to the employee in December 1998 pesos. In cases in which the amount awarded was zero, we set the log of the award to zero. Additionally, these cases are treated as censored observations in the Tobit model. Standard errors are calculated allowing for heteroscedasticity and for the possibility that the outcomes in cases that have been grouped into the same proceeding may be correlated. We use the notation of *** to denote significance at the 0.01 level. Similarly ** denotes significance at the 0.05 level and * denotes significance at the 0.10 level. All models include dummy variables for the year when the employee separated from the firm and the quarter when the employee separated from the firm. Industry dummies are included in the models for tribunal 15. Each observation is given the weight of the inverse of its ex-ante probability of being included in the sample. See text for details. Table 7: Effects of worker's assertion and cases involving multiple workers on final payment (Dependent Variable: Log of final payment) Dropped cases included Tribunal 15: Tribunal 6: Tobit OLS Tobit OLS coef std err coef std err coef std err coef std err ln(worker's claim) -0.15 0.23 -0.07 0.17 0.11 0.36 0.14 0.29 multiple workers -2.61 *** 0.89 -1.82 *** 0.60 -1.51 1.09 -1.08 0.81 female -0.75 0.58 -0.55 0.42 -0.73 0.74 -0.55 0.56 at-will worker 0.80 0.57 0.68 0.42 0.56 0.74 0.44 0.58 tenure 15 years 1.08 0.79 0.95 0.61 1.61 ** 0.74 1.42 ** 0.60 R2 0.10 0.09 Number of obs 1,076 1,076 547 547 Censored obs 288 122 Dropped cases excluded Tribunal 15: Tribunal 6: Tobit OLS Tobit OLS coef std err coef std err coef std err coef std err ln(worker's claim) -0.12 0.16 -0.08 0.14 0.04 0.27 0.06 0.26 multiple workers -1.51 ** 0.70 -1.29 ** 0.59 -0.58 0.74 -0.49 0.67 female -0.33 0.40 -0.29 0.35 -0.25 0.38 -0.23 0.36 at-will worker 0.96 ** 0.38 0.90 *** 0.34 0.62 0.38 0.56 0.35 tenure 15 years 2.17 *** 0.36 2.05 *** 0.33 1.54 *** 0.48 1.49 *** 0.45 R2 0.16 0.10 Number of obs 882 882 457 457 Censored obs 94 32 Notes: The dependent variable is the log of the amount awarded to the employee in December 1998 pesos. In cases in which the amount awarded was zero, we set the log of the award to zero. Additionally, these cases are treated as censored observations in the Tobit model. Standard errors are calculated allowing for heteroscedasticity and for the possibility that the outcomes in cases that have been grouped into the same proceeding may be correlated. We use the notation of *** to denote significance at the 0.01 level. Similarly ** denotes significance at the 0.05 level and * denotes significance at the 0.10 level. All models include dummy variables for the year when the employee separated from the firm and the quarter when the employee separated from the firm. Industry dummies are included in the models for tribunal 15. Each observation is given the weight of the inverse of its ex-ante probability of being included in the sample. See text for details. have 0.01 the the of excluded err the include 0.11 0.36 0.27 0.34 0.40 excluded err 0.14 0.56 0.34 0.48 0.41 that std std at separated predicts weight cases *** *** *** 0.15 882 cases ** 0.08 409 cases models settlement settlement in the coef -0.47 -1.46 -0.73 0.34 -0.22 coef -0.29 -0.81 0.12 0.62 0.20 All employee perfectly significance details Dropped Dropped level. the given is outcomes for the denote 0.10 when variable text to the that at See *** 15: err 6: quarter observation err 0.10 0.32 0.27 0.35 0.39 0.14 0.57 0.35 0.46 0.41 of dummya the std std Each dummy. termination trial *** *** ** 0.12 1,076 trial * 0.06 547 possibility and cases, of Tribunal Tribunal notation significance 15. that coef 0.42 1.02 0.56 -0.48 0.07 coef 0.25 0.62 the -0.19 -0.63 -0.15 firm the some to mode for the In of and use tribunal included err case 0.09 0.32 0.26 0.25 0.32 included err denotes* case 0.19 0.41 0.39 0.44 0.41 We from for std std and sample. models cases ** 0.05 1,076 cases 0.04 487 models the corresponding dropped dropped level in Logit coef 0.01 0.65 0.26 0.04 0.38 coef -0.07 0.53 0.27 0.02 -0.10 separated correlated. the 8: in Dropped Dropped heteroscedasticity be 0.05 err err the included Table 0.08 0.27 0.20 0.22 0.28 0.12 0.41 0.30 0.33 0.34 for may at employee observations std std the *** *** ** 0.08 1,076 * 0.05 547 the included being settlement settlement are of drop coef -0.26 -1.06 -0.52 0.22 -0.29 coef -0.11 -0.73 -0.10 0.31 0.14 allowing when to proceeding significance year same dummies filing) filing) the probability calculated at at the denotes for are ** necessaryit into Industry ex-ante claim years obs claim years obs its workers 15 2 of workers 15 2 of errors making worker R worker R variables firm. of grouped Similarly the ln(worker's multiple female at-will tenure Pseudo Number ln(worker's multiple female at-will tenure Pseudo Number Standard been level. dummy from inverse outcome Table 9: Effects of worker's assertion, our calculation of the worker's real claim, and mode of termination on final payment (Dependent Variable: Log of final payment) Dropped cases included Tribunal 15: Tribunal 6: Tobit OLS Tobit OLS coef std err coef std err coef std err coef std err ln(worker's claim) -0.10 0.40 -0.08 0.30 -0.42 0.58 -0.35 0.46 ln(imputed claim) 0.40 0.51 0.33 0.38 0.81 0.60 0.73 0.49 trial -4.97 *** 1.12 -3.27 *** 0.70 -2.83** 1.15 -2.06 ** 0.83 female -0.41 0.61 -0.35 0.45 -1.02 0.82 -0.77 0.62 at-will worker 0.50 0.58 0.43 0.44 0.51 0.77 0.38 0.61 tenure 15 years 1.18 0.87 0.96 0.67 1.25 * 0.76 1.09 * 0.62 R2 0.14 0.12 Number of obs 1,076 1,076 547 547 Censored obs 288 122 Dropped cases excluded Tribunal 15: Tribunal 6: Tobit OLS Tobit OLS coef std err coef std err coef std err coef std err ln(worker's claim) -0.20 0.22 -0.18 0.20 -0.18 0.21 -0.17 0.20 ln(imputed claim) 0.75 ** 0.32 0.69 ** 0.28 0.53 0.34 0.53 0.33 trial -5.98 *** 0.73 -5.27 *** 0.60 -4.23*** 0.90 -3.83 *** 0.78 female 0.36 0.28 0.32 0.25 -0.36 0.32 -0.32 0.30 at-will worker 0.59 ** 0.29 0.57 ** 0.27 0.32 0.33 0.30 0.31 tenure 15 years 2.14 *** 0.41 2.01 *** 0.37 1.30 *** 0.50 1.24 *** 0.47 R2 0.54 0.36 Number of obs 882 882 457 457 Censored obs 94 32 Notes: The dependent variable is the log of the amount awarded to the employee in December 1998 pesos. In cases in which the amount awarded was zero, we set the log of the award to zero. Additionally, these cases are treated as censored observations in the Tobit model. Standard errors are calculated allowing for heteroscedasticity and for the possibility that the outcomes in cases that have been grouped into the same proceeding may be correlated. We use the notation of *** to denote significance at the 0.01 level. Similarly ** denotes significance at the 0.05 level and * denotes significance at the 0.10 level. All models include dummy variables for the year when the employee separated from the firm and the quarter when the employee separated from the firm. Industry dummies are included in the models for tribunal 15. Each observation is given the weight of the inverse of its ex-ante probability of being included in the sample. See text for details. Table 10: Effects of worker's assertion and our calculation of the worker's real claim on final payment (Dependent Variable: Log of final payment) Dropped cases included Tribunal 15: Tribunal 6: Tobit OLS Tobit OLS coef std err coef std err coef std err coef std err ln(worker's claim) -0.89 * 0.46 -0.65 * 0.33 -0.78 0.55 -0.61 0.43 ln(imputed claim) 1.24 ** 0.60 0.95 ** 0.43 1.17 ** 0.57 1.00 ** 0.47 female -0.75 0.62 -0.57 0.44 -1.01 0.82 -0.77 0.62 at-will worker 0.53 0.60 0.45 0.45 0.71 0.76 0.53 0.60 tenure 15 years 0.96 0.81 0.84 0.63 1.12 0.76 1.02 0.62 R2 0.08 0.09 Number of obs 1,076 1,076 547 547 Censored obs 288 122 Dropped cases excluded Tribunal 15: Tribunal 6: Tobit OLS Tobit OLS coef std err coef std err coef std err coef std err ln(worker's claim) -1.37 *** 0.35 -1.21 *** 0.29 -0.87*** 0.33 -0.81 *** 0.30 ln(imputed claim) 2.00 *** 0.49 1.81 *** 0.41 1.21 *** 0.40 1.16 *** 0.38 female -0.14 0.35 -0.12 0.30 -0.36 0.38 -0.32 0.35 at-will worker 0.62 0.40 0.59 * 0.36 0.72 * 0.39 0.65 * 0.36 tenure 15 years 1.84 *** 0.38 1.75 *** 0.35 1.16 ** 0.50 1.13 ** 0.47 R2 0.20 0.14 Number of obs 882 882 457 457 Censored obs 94 32 Notes: The dependent variable is the log of the amount awarded to the employee in December 1998 pesos. In cases in which the amount awarded was zero, we set the log of the award to zero. Additionally, these cases are treated as censored observations in the Tobit model. Standard errors are calculated allowing for heteroscedasticity and for the possibility that the outcomes in cases that have been grouped into the same proceeding may be correlated. We use the notation of *** to denote significance at the 0.01 level. Similarly ** denotes significance at the 0.05 level and * denotes significance at the 0.10 level. All models include dummy variables for the year when the employee separated from the firm and the quarter when the employee separated from the firm. Industry dummies are included in the models for tribunal 15. Each observation is given the weight of the inverse of its ex-ante probability of being included in the sample. See text for details. n excluded err bee firm. 0.21 0.35 0.29 0.33 0.45 excluded err 0.28 0.57 0.36 0.54 0.44 ex-ante std std the have level. dummy its it cases *** *** ** 0.20 882 cases *** * 0.15 409 settlement settlement that from of 0.01 coef -1.36 -1.28 -0.70 0.08 -0.45 coef -1.31 -0.98 0.13 0.80 making -0.04 cases the include Dropped Dropped in at inverse separated the models of outcome All the termination outcomes significance employee weight of 15: err 6: err the level. 0.20 0.31 0.28 0.34 0.45 0.28 0.60 0.37 0.50 0.42 the the predicts std std that mode trial *** *** ** 0.18 1,076 trial *** 0.11 547 denote 0.10 to given on Tribunal Tribunal the when is coef 1.37 0.95 0.59 -0.26 0.19 coef 1.08 0.71 -0.19 -0.72 0.05 *** at perfectly details claims possibility of quarter for included err the text case 0.19 0.32 0.26 0.26 0.33 included err the variable case 0.33 0.42 0.39 0.41 0.40 std std for notation significance and observation See cases * ** 0.06 1,076 cases 0.04 487 and the firm Each dropped dummya coef -0.32 0.68 0.27 0.12 0.40 dropped coef -0.01 0.53 0.28 -0.05 -0.14 use the 15. dummy. "non-credible" of Dropped Dropped We denotes* cases, that err err and from 0.15 0.27 0.21 0.22 0.28 0.19 0.40 0.31 0.33 0.33 tribunal to effects std std level for some *** *** ** * 0.09 1,076 *** * 0.07 547 heteroscedasticity In correlated. The settlement settlement separated coef 0.03 coef 0.40 0.00 for be 0.05 11: -0.61 -0.99 -0.50 -0.47 -0.64 -0.78 -0.12 the models may sample. at the corresponding the Table allowing employee in claim) claim) in the proceeding significance when included imputed/ imputed/ calculated included observations are are same year the being claim years obs claim years the denotes the obs of drop workers to 15 2 of 15 of errors into ** for worker R workers 2 worker R dummies ln(worker's multiple female at-will tenure Pseudo Number ln(worker's multiple female at-will tenure Pseudo Number Standard grouped Similarly variables Industry probability necessary