Information and Bargaining through Agents: Experimental Evidence from Mexico’s Labour Courts

Abstract

Well-functioning courts are essential for the health of both financial and real economies. Courts function poorly in most lower-income countries, but the root causes of poor performance are not well understood. We use field experiments with ongoing cases to analyse sources of dysfunction in Mexico’s largest labour court. We provide parties with personalized predictions for case outcomes and show that this information nearly doubles settlement rates and reduces average case duration. The experiment generates the first experimental evidence in live court cases that reducing information asymmetries results in a decrease in delay, an outcome predicted by many theories of bargaining. We also find that the information treatment is effective only when the plaintiff is present to receive it directly, suggesting agency issues between plaintiffs and their private lawyers. For most workers, the treatment appears to improve welfare, as measured by discounted payouts and ability to pay bills.

1 INTRODUCTION

Well-functioning courts underpin markets and constrain private power in developed economies, but courts function poorly in most developing countries. Outcomes are unpredictable, parties are misinformed, and inefficient processes lead to slow decisions and large case backlogs (Djankov et al., 2003). In addition to raising concerns for justice, poorly functioning courts undermine both the financial (Ponticelli and Alencar, 2016) and real (Boehm and Oberfield, 2020; Chemin, 2020) economies. The effects of poorly functioning courts are increasingly well documented, most convincingly through studies analysing changes following the creation of new institutions providing judicial resolution for specific activities (Visaria, 2009; Lichand and Soares, 2014). However, as Boehm and Oberfield (2020) note, even “Newly created courts tend to… accumulate backlogs over time” (p. 2009). Thus, while the existing literature provides evidence on the importance of well-functioning legal institutions, it is much less informative on how to improve the performance of existing institutions, or how to ensure the continued high performance of new institutions. This is due in part to a lack of rigorous empirical work aimed at understanding the micro-analytics of court proceedings, and a particular paucity of randomized experiments in courts (Greiner and Matthews, 2016).

To illuminate the causes of dysfunction in courts, we conduct a randomized experiment with active cases in the Mexico City Labor Court (MCLC). In the experimental treatment, we provide parties in a random subset of labour disputes with case-specific predictions of the amount collected by the plaintiff, the duration of the case, and the probability the case ends by settlement, judgment, or by being dropped. We generate the predictions using parameters from machine-learning models run on data from 5,000 concluded cases and using variables taken from the initial case filing: daily wage, hours worked per week, tenure at the firm, whether or not the worker was registered with Social Security, and other characteristics of the position.

The MCLC is responsible for enforcing labour law for private employers located in Mexico City. Over 95% of its cases involve workers who claim to have been involuntarily separated from their jobs by employers who then failed to transfer to the workers severance payments as required under the labour law. Dismissed workers are entitled to severance payments of a minimum of 90 days’ wages and, depending on circumstances, substantially more. Workers are entitled to the payments following almost any involuntary separation.¹ The MCLC receives more than 35,000 filings each year and had a case backlog of more than 100,000 cases at the time of our experiment. The large backlog stems from long case durations and settlement rates that are low in comparison to those of similar courts in other countries. While severance payments are a fundamental right of workers in Mexico, a national survey of workers we describe later in the article indicates that half of dismissed workers do not receive their legally entitled severance payment. However, only 15% of workers whose employers fail to pay severance choose to pursue a claim in the court. Among the remaining 85%, more than three-quarters said the reason for not pursuing a claim was that the court is either “useless” or “too expensive.”

Legal dysfunction has many potential sources. Our experiment holds fixed any corruption and misalignment of incentives for judges (e.g. Boehm and Oberfield, 2020; Chemin, 2020) and focuses instead on other potential causes of low settlement rates. We show that workers are overly optimistic and misinformed. They wildly overestimate their chances of winning their case and are not able to state basic elements of the law and even what is included in their case filing. Yildiz (2011) reviews the literature showing that overoptimism may lead to delays in bargaining. Kennan and Wilson (1993) show instead that when there is asymmetric information quarrelling parties use delay to communicate information, a behaviour that may lead to failure to reach a settlement beneficial to both parties. While we won’t distinguish between these theories, a prescription common to both is to provide objective information, which is what our information treatment was designed to do.² As most of these models predict, our treatment increases settlement rates and so decreases case duration and court congestion, while leaving plaintiffs at least as well off. However, the information has an impact only when we are able to provide it directly to plaintiffs, suggesting a third element to the story: that agency issues in the market for legal services are also a cause of low settlement rates.

We begin by using the historical case files to document a set of stylized facts about the functioning of the court. These facts show evidence that the court functions poorly and few workers use it even when their employer fails to pay severance required by law. Among filed cases, settlement rates are low, parties are overconfident and plaintiffs, particularly those represented by private lawyers, are misinformed.

Working with the court, we conduct an experiment in three phases, which differ with respect to the point at which we intervene in the process. In Phase 1, we intervene in ongoing cases at any point in the life of the lawsuit. Our first result is that the information treatment increases settlement on the day of treatment by 75%. Settlements are more frequent in cases that are early in the process, suggesting that earlier interventions to be more effective. Given this, in Phase 2, we intervene in the first hearing of each case. We find a treatment effect that is very similar in magnitude, an indication that the intervention replicates. This increase in settlements is consistent with the treatment decreasing informational asymmetries and/or overconfidence.

This first result connecting information to a breakdown in bargaining is, to our knowledge, the first experimental evidence showing that providing information decreases delay in live court cases with bargaining under asymmetric information. This lends empirical support to a large theoretical literature. However, the experiment produces an important second result: in both the first and second phases, the treatment is effective only when the employee is present at the hearing and receives the information directly. When the information is provided only to the plaintiff’s lawyer, there is no effect on settlement either on the day or subsequently.³ Moreover, the interaction of employee presence and the information treatment happens exclusively when the employee is represented by a private lawyer, precisely the cases where the stylized facts and analytical arguments we provide below suggest incentives could generate agency problems.⁴

Does the treatment simply shift to the present settlements that would have happened anyway in the future? We were able to follow cases using administrative data for 42 months after treatment. We find that the treatment effect and the relevance of the employee receiving the treatment directly remain unchanged over this time. The persistent treatment effect rules out the possibility that the lawyer did not have time to consult worker. In fact, an additional 38% of the cases are settled in both the control and treatment groups after the day of our intervention. The importance of the employee’s presence 42 months after treatment suggests instead that lawyers do not convey the information to their clients.

Importantly, data on case outcomes from Phases 1 and 2 show that the increase in settlements improves the outcomes for the typical plaintiff. These patterns suggest that case trajectories are affected by lawyer-plaintiff agency issues. Given this, in Phase 3, we seek to intervene before workers contract with a lawyer and file a case. We work with a sample of recently dismissed workers who come to the court seeking information about their rights and how to file a case. We find that the treatment providing predicted case outcomes is again effective in increasing settlement and that the information decreases overconfidence.⁵

The results of the experiment suggest that information asymmetries between parties to the case and between the worker and her lawyer are underlying causes of malfunctioning courts. A majority of plaintiffs earn below the median wage and have modest levels of schooling; more than 80% are using the court for the first time. Our surveys indicate that 38% of plaintiffs found their lawyer on the sidewalk outside the court, where many lower-quality lawyers find clients. We show analytically that the incentives of the lawyers, who have information advantages, do not always align with those of their clients, even though private lawyers in the MCLC cases almost always receive a share of the award collected by their plaintiffs. Differences in discount rates and the ability to diversify risk results in workers and their lawyers having different preferences over settlement options. Lawyers also charge an initial fee that is generous relative to the time required to file a suit, giving them incentives to file cases even when they have little prospect of winning.

By 2020, almost 90% of both treatment and control cases in the first two phases were resolved. This allows us to examine how treatment affects the welfare of plaintiffs. Relative to the control group, those in the treatment group are 6.7% more likely to have settled, 7.3% less likely to have lost a judgment, and 0.4% less likely to have won a judgment. The additional settlements come mainly from cases with modest recovery amounts. When we compare the net present value of the amount collected by plaintiffs in the treatment and control groups, we find that the treatments improved the outcomes for the typical plaintiff. Though making a precise statement about plaintiff welfare is difficult for reasons we discuss in Section 7, surveys in Phase 3 conducted 2 months after treatment show that treated workers are 7% less likely to report not having enough food to eat or being unable to pay for basic services; reported happiness is insignificantly higher with treatment.

One consistent interpretation of the stylized facts and experimental results together is that there is little settlement in MCLC because those workers who file cases are overconfident of winning, and because they operate in an environment where their (private) lawyers have more information than they do but have no incentives to debias them. On the contrary, the lawyers appear to not share important information (the calculator) that encourages them to settle and get paid—more on average—today, instead of pursuing a long trial at the end of which they often collect nothing. We find that providing statistical information allows workers to update their expectations, decreasing overconfidence and leading to more settlement and higher payoffs on average, consistent with many models of overconfidence and asymmetric information.

Literature. Experiments in courts are uncommon. Greiner and Matthews (2016) review randomized evaluations in the U.S. legal system, finding only 50 studies conducted between 1963 and 2015. Most of these examine mediation and alternative dispute resolution, though a handful evaluate programs that affect the use of lawyers. There are likewise few legal experiments in low-income countries.⁶ As a result, there is little credible evidence on the source of delays and the effect of differences in rules and organizational structure. Indeed, given the scarcity of diagnostic information about courts in developing countries, the stylized facts we derive from administrative records contribute to the literature on courts as institutions for development.

The article also contributes to the literature on experts and moral hazard. The most closely related paper in this literature is Anagol et al. (2017), who conduct an audit experiment in the Indian life insurance market. They show that agents use their informational advantage to induce clients to make decisions that are favourable to the agents’ interest. There is much more extensive observational evidence that agents take advantage of superior information in diverse settings: Schneider (2012) in Canadian auto repair shops; Emons (1997) among doctors in Switzerland; and Levitt and Syverson (2008) among real estate agents in the US.⁷

A third related literature is that of bargaining in the field. Courts are a disciplining device for a bargaining game between the parties to the case. Bargaining may break down if parties have asymmetric information (Myerson and Satterthwaite, 1983) or are overly optimistic (Yildiz, 2011).⁸ Our data indicate that both misinformation and excessive optimism are characteristics in the MCLC cases, and our intervention is the first, to our knowledge, to randomly alter the information of parties in live court cases. Most of the relevant bargaining literature constructs a game between two parties, but court cases typically involve four parties: the plaintiff, the defendant and the lawyers representing either side. This distinction is relevant if agency is important. Theoretically, the addition of lawyers may lead to either more- or less-efficient outcomes. Gilson and Mnookin (1994) model court cases as prisoners’ dilemmas in which the parties play once and the lawyers play repeatedly. As such, the lawyers may cooperate when the parties would not, generating more-efficient outcomes. Ashenfelter and Dahl (2012) examine data from arbitration cases involving emergency services unions and municipalities in New Jersey. In a context where parties sometimes represent themselves and sometimes are represented by lawyers, they show that lawyers provide positive benefits to the party they represent.

Finally, we contribute to the literature on the effects of information provision and decision making. Information has been shown to improve decision making in wide range of other contexts: in the functioning of private markets (Andrabi et al., 2017; Michèle et al., 2019), in schooling decisions (Jensen, 2010; Dizon-Ross, 2019), in political institutions (Reinikka and Svensson, 2011; Chong et al., 2015), and in credit markets (Seira et al., 2017). Our results focus on courts and suggest that it is important that the information be conveyed directly to the party affected by the decisions.

The rest of the article proceeds as follows: We begin by describing the context and the Mexican labour law in Section 2. We then detail the data from both administrative records and surveys of litigants and lawyers in Section 3. Section 4 uses those data to describe a set of stylized facts that motivate our experiment. Section 5 describes the experimental protocol and Section 6 the results. We discuss the welfare implications of the results in Section 7 and conclude in Section 8.

2 THE CONTEXT

Government-run unemployment insurance is the most common job-displacement insurance system in higher-income countries, but severance payment programs like the one Mexico uses are the norm in Africa, Asia, and Latin America (Gerard and Naritomi, 2019). Mexican workers dismissed from their job for almost any reason are entitled to 90 days’ wages and often substantially more, with payments made directly by firms to workers. Severance payment provisions in the private sector are governed by federal labour law in Mexico, with adjudication of disputes in most industries assigned to state-level labour courts.⁹ We work with the Mexico City Labor Court, the state court serving Mexico City. Each year, the MCLC receives more than 35,000 new cases and concludes fewer than 30,000 cases. Its portfolio of 100,000 active cases is therefore not only large but growing.

Over 95% of cases at the MCLC are filed by dismissed workers claiming they have not received severance payments they are owed by their employers. The sector of employment determines to which one of 20 “subcourts” a case is assigned. In the first phase of the project, we worked with Subcourt 7, which deals mainly with firms in the retail automotive and transport services industries. In Phase 2, we expanded to four additional subcourts specializing in industries such as private education, security, restaurants, retail banking, large department stores, and medical services. In Phase 3, we worked with dismissed workers approaching the court who might ultimately file a case assigned to any of the 20 subcourts.

On paper, Mexico’s labour law is straightforward and favourable to workers. Dismissed workers are entitled to minimum severance payments of 90 days’ wages and, depending on circumstances, substantially more. Payments must be made even if the dismissal is due to loss of business by the firm. However, while the law itself is generous, the framework for enforcing the law disadvantages workers for at least three reasons. First, workers and firms are often informal (Kumler et al., 2020). Wages are paid in part or entirely in cash, and there is no formal written contract. In these circumstances, dismissed workers may find it difficult to prove wage levels or even the existence of the labour relationship itself. Second, at the time of hiring, firms often take actions that undermine the worker’s subsequent claim of unfair dismissal. For example, firms may require the worker to sign an undated letter of resignation as a condition of hiring. Third, while payments through unemployment insurance systems are often administered by the government, payments in Mexico are made directly from the firm to the worker. Workers who win a judgment often face challenges in recovering payments. Firm assets are typically unregistered and hence, the court cannot place liens on assets at the time of its ruling. When firms refuse to pay voluntarily, workers must pursue seizure of assets and firms have many ways of avoiding asset seizure. For example, they may declare bankruptcy, transfer assets to another entity, or pay bribes (Leo et al., 2014). Using the threat of avoiding payment, firms often negotiate much lower payments even after the court rules against them. Among the 5,000 historical cases that we code, 203 cases ended in a judge’s decision in favour of the plaintiff. The plaintiff was able to collect the full amount awarded in only 41 of these (20%) cases and failed to collect anything in 109 (54%) of them.

We describe the Mexican Labor Law and the court procedures in more detail in Supplementary Appendix A. Plaintiffs may be represented by either private or public lawyers. Private lawyers often charge a fixed fee of as much as 2,000 MXN (USD 100) to file a case, and then receive 30% of any amount recovered by the worker. Lawyers from the Public Attorney’s office are paid a fixed wage and do not charge their clients for services. Employers most often respond to a suit in one of three ways: denying the existence of a labour relationship, offering reinstatement, or claiming the worker resigned voluntarily and producing a letter of resignation signed by the employee.¹⁰ High levels of informality make it more difficult for plaintiffs to prove employment relationships and easier for employers to hide assets. These strategies decrease the likelihood that workers win a judgment or collect the compensation the court awards them.

The parties reach a settlement in around 55% of the cases. Cases that settle usually do so within a year of filing. In the absence of settlement, cases typically involve at least six to eight hearings spread over several years. The hearings are conducted by administrative assistants to the judge, but the judge of the subcourt makes all rulings, based on the written record prepared by the assistants. Judgments within three years of filing are rare.

3 DATA

We use both administrative records and survey data. We describe these data briefly here, and in more detail in Supplementary Appendix A. Through an agreement with the court, we had access to the experimental case files and all of the historical case files from the court. The case files register all the legally relevant information in the lawsuit. Given the scarcity of evidence on the functioning of courts, we view the construction of this data itself as a contribution of the article.

We conducted the first phase of the experiment in Subcourt 7 between March and May, 2016 and the second phase in Subcourts 2, 7, 9, 11, and 16 between October 2016 and February 2017. In Phases 1 and 2, we intervened in case hearings where both parties had been notified and were therefore obligated to attend the hearing. We conducted Phase 3 between May 2017 and August 2018 with a sample of dismissed workers who had not yet filed a case (see Figure 1 or a schematic of court cases). In all three phases, we carried out very brief surveys of the plaintiffs and lawyers when they were present, though in Phase 2 the survey was very limited for logistical reasons.

Figure 1.

Diagram of case origination in each phase Notes: This figure is a stylized representation of the process that case files follow in Mexico City’s Labor Court.

3.1 Administrative data

Historical cases: We began by digitizing the historical case file data with the goal of building predictive models of case outcomes, as we describe below (MCLC, 2011–2020). Given the duration of the average lawsuit, we focus on cases filed in 2011, the earliest year for which the court had digital (pdf) records of all initial case filings. For Phase 1, we digitized 2,158 lawsuits filed in 2011 or 2012, assigned to Subcourt 7, and concluded by December 2015. Only 55 of those cases were concluded by a decision of the judge. In order to increase the sample of cases concluded by the judge’s decision, we reached back to lawsuits filed in Subcourt 7 in 2009 and 2010, identifying 241 additional case files concluded by a judge’s decision. Together with the 2011 and 2012 cases, we use these to calibrate the likelihood of winning and amount collected at trial.

For the second phase of the experiment, we used data from 1,000 concluded cases in each of the five participating subcourts. We again selected cases filed in 2011 and concluded by December 2015. We used all such cases from Subcourt 7, and a random sample of approximately 1,000 cases in each of Subcourts 2, 9, 11, and 16. Thus, the calculator for Phase 2 was calibrated with historical data covering 5005 cases, all filed in 2011 and concluded by December 2015.¹¹

Because we often intervened in the first hearing, the predictive model used only information included in the initial filing. We capture the amount claimed by the plaintiff, the date of the lawsuit, whether the lawyer is public or private, the worker’s gender, age, daily wage, tenure at the firm, weekly hours worked, and industry. The variables are defined in Table A.1 in Supplementary Appendix A. The basic formula for severance payment in the law is in large part a function of the wage, tenure, and hours worked.

We also record the outcome of the suit, the end date, and the amount recovered by the worker. Cases end in one of five ways: being dropped by the plaintiff, expiring due to lack of activity, a judge’s ruling with no collection, a judge’s ruling with a positive collection, or settlement between the parties. The majority of the cases with positive recoveries end in settlement. Settlements are essentially always recorded at the court, in order to assure that the plaintiff does not continue to pursue the case. For cases ending in judgments in favour of the plaintiff, the amount the plaintiff recovers is often different from the amount awarded by the judge for three reasons: first, the law provides that if the judgment is not enforced immediately, additional lost wages may be added to the award; second, the worker may agree to settle for a lower amount to avoid the high costs of enforcing payment; and third, as noted above, the worker may be unable to collect the judgment found by the court.

In addition to providing the raw material for the prediction calculator, the historical data allow us to construct a set of stylized facts about the functioning of the court. We discuss what the data show with regard to trial length, frequency of settlement, amount collected, the fraction of plaintiffs collecting awards, and so forth, in the next section.

Administrative data for ongoing cases: We code the initial case file data from all of the on-going lawsuits involved in Phases 1 and 2 of the experiment (MCLC, 2011–2020). We use these data, combined with the predictive model developed with the historical data, to predict the outcome of the lawsuit. We also use the administrative records to determine who attended the hearing on the day of the experiment, whether the lawsuit ended on that day through a settlement, and the amount of money recorded for the settlement. We then repeat the data collection in 2020 around 42 months after the start of each phase of the experiment.¹² For Phase 3, administrative records from late 2019 show which workers sued or settled. As we noted, settlements are generally registered in court files even for cases that settle out of court, because this is the only way the firm can ensure that the employee does not continue to pursue the case.

3.2 Survey data

We conducted surveys with parties involved in the experiment in all three phases. In Phases 1 and 2, we administered short surveys at the hearing on the day of the experiment. We interviewed the defendant’s lawyer and either the plaintiff, if she was present at the hearing, or her attorney if she was not.¹³ The survey was conducted before parties were aware of their treatment status. We asked about knowledge of the case file and the relevant law, expected case outcomes, and, where the plaintiff was present, demographic characteristics of the plaintiff.¹⁴

In Phase 3, we sample dismissed workers approaching the court in search of information. We provided information according to the protocol described below, and conducted a survey to collect demographic data and information about the worker’s employment necessary to use the calculator to make predictions on the worker’s own case outcomes. We also conducted telephone follow-up surveys 2 weeks and 2 months after the initial contact. The follow-up surveys recorded actions taken after our initial interaction, and elicited updated beliefs about case outcomes. For example, we asked workers if they had talked to a lawyer, and if so, how they had found the lawyer(s). We also asked whether they had filed a suit, settled, or decided not to pursue any claim, and collected a measure of life satisfaction and difficulty in paying bills. We conducted at least one of the two telephone surveys with 89% of the sample. Survey questions and response rates from all three phases are discussed in more detail in Supplementary Appendix A.

Table 2 shows summary statistics of the cases for each of the three phases of the experiment. We use the administrative data for the first two phases and the survey data from Phase 3. Table A.2 in Supplementary Appendix A summarizes data from the plaintiff surveys from Phase 1. The sample for these surveys is determined by the plaintiffs who attend the hearings. Plaintiffs with public lawyers were more likely to attend the hearing: 29% of workers present at Phase 1 and 2 hearings had a public lawyer, while only 10% of the case files in the experiment had a public lawyer. These plaintiffs have modest levels of schooling, with 58% of having no more than lower secondary schooling. Only 7.6% were currently employed, and for those not currently working who were searching for a job, the average reported likelihood of finding a job in the next 3 months was 58%. Of those workers represented by a private lawyer who attended the hearing, most (nearly 82%) said their agreement called for paying their lawyer a fraction of any award (30%, on average).

3.3 Construction of the calculator

In the experiment, we provide personalized predictions on important case outcomes to a subset of plaintiffs and defendants. We developed simple, parsimonious, predictive models using the historical case records. We considered several machine learning models, including boosting, random forest, and regularization methods (e.g. ridge), along with OLS and logit. The construction of the calculator is described in detail in Supplementary Appendix A, but we summarize the main points here. As we noted, the calculator in Phase 1 was developed using 2,158 cases from Subcourt 7 and the calculator used in Phases 2 and 3 used 5,005 cases from the five participating subcourts. The information we provided to the parties also changes somewhat in each phase, as we describe below.

Our goal was to provide predictions on the expected amount collected by the plaintiff, the duration of the case, and the probability the case ends by being dropped, expiring, judgment with zero recovery, judgment with positive recovery, or settlement. The main explanatory variables were all taken from the initial case filing: daily wage, hours worked per week, tenure at the firm, gender of the plaintiff, type of lawyer, whether or not the worker was registered with Social Security, if s/he was employed in a position of high trust (an “at will” worker), the specific claims in the case (reinstatement, overtime, back pay, vacation pay, end-of-year bonus, statutory profit sharing, severance pay) and the industry of the firm. We used 70% of the data to fit the models and the remaining 30% for testing. For each outcome, we used cross-validation to choose the model and variables with the best fit on the testing sample, measured by the correlation between predicted and actual values. Tables A.5 and A.6 in Supplementary Appendix A present goodness of fit measures for all the models and highlight those we selected. For the settlement amount, the out-of-sample correlation between predicted and actual outcomes is 0.61 in Phase 1 and 0.69 in Phase 2/3, while for the discrete outcomes on the way the case ends, accuracy rates are between 0.61 and 0.93.¹⁵

The models allow us to produce individualized predictions that we shared with parties present at the hearing in cases randomized into the calculator treatment. Figure 5 displays an English translation of the template we used in Phase 1. The template shows the minimum legal entitlement based on the law if the plaintiff were to win on the issue of unfair dismissal and the probability the case ends in each of the five possible endings. For each of these five endings, we showed the average amount recovered by the plaintiff. We also show the expected payout across all endings and the percentage of cases that were still unresolved after three years. In Phase 1, we provided the same information sheet to both sides of the case. For Phase 2, we adjusted the format, first to simplify the information so that it could be explained to parties more quickly, and second to address concerns raised by court officials. In particular, conciliators working for the court suggested that we provide the expected settlement amount, conditional on characteristics, and then provide each side with data indicating the contingency they faced if they did not settle. We developed separate templates for the plaintiffs and defendants, translations of which are shown in Supplementary Figure C.1. For the worker, the no-settlement contingency was the percentage of cases where workers collected nothing, and for firms it was the average amount collected by plaintiffs that won judgments. For the firms, we also showed the recovery amount implied by the law. In addition to using the calculator as a treatment in the experiment, we use it to build a proxy of average overconfidence, as we describe below.

There are several potential sources of bias in the predictions based on the historical data. One is that our sample is composed of cases that were concluded when the models were estimated in early 2016, and 29% of cases filed in 2011 and 2012 were still ongoing at that time. If concluded and ongoing cases have different potential outcomes, then although our predictions are unbiased for the concluded cases, they may be biased for a random sample of ongoing cases. Note that if cases end in settlement, they almost always do so within the first 24 months after filing. Since the historical data used in the calculator models cover more than 24 months after filing, very few of the 29% of historical cases that were unresolved are likely to end in settlement. Therefore, the projected average payment for cases ending in settlement—the most important variable in the calculator information—is not affected by this censoring issue.

For cases ending in other outcomes—being dropped, expiring, or ending in judgment, the censoring is a larger concern. This potential bias was communicated to the parties when the calculator information was provided. We perform two exercises to estimate how large any bias might be. First, we compare characteristics of the experimental cases with those of the concluded historical cases used in the models. In Supplementary Figure A.4, we show that the two sets of cases are similar on observables.¹⁶ Second, we compare the characteristics of completed and continuing lawsuits within the historical data. To do this we drew a random sample of 956 cases filed in 2011 that were not finished by 2015. Supplementary Figure A.5 shows that these 956 unresolved cases are very similar to the completed cases used to develop the models.

A second issue is that even if our predictions are unbiased on average, they are not unbiased for a specific case. Parties may have information unobservable to us about the strength of their case. Again, we made clear to the parties that the predictions were based on average outcomes, and outcomes of individual cases will vary depending on the circumstances of the case.

Finally, the calculator predictions will build in any biases contained in previous court decisions. For example, if workers collect less because they are unable to prove wage payments made in cash or because firms avoid making payments ordered by the court, then the calculator will implicitly assume these conditions will continue to apply in the future. Although reforming the institutions so that payments more faithfully reflect the law should be a goal of judicial reforms, we believe the calculator faithfully reflected conditions that the parties in our experiment faced. The situation might be seen as analogous to providing parents and children with accurate information on returns to public schooling. The information allows them to make better schooling decisions given the actual quality of education but may not directly lead to improvements in the quality of schooling. In our case, the goal of the experiment is to uncover the sources of inefficiencies in the courts to allow reformers to focus on the most critical issues while also providing information with which plaintiffs can make more informed decisions given the way the court actually functions.

4 OUTCOMES AND EXPECTATIONS: STYLIZED FACTS

We use the administrative and survey data from Phase 1 to document a set of stylized facts about the court. These serve as a motivation for the experiment we implement but also provide some insight on the functioning of the court. We note whether the source of data for each stylized fact is the historical administrative data or survey data.

Fact 1: Labour courts are seldom used (Survey data)Half of dismissed workers do not receive severance, but only 13% of those pursue a claim in labour courts.

In late 2021, we conducted a survey of formal workers registered with the Mexican Social Security Administration (IMSS). The survey was conducted through a link sent in an email from IMSS to workers registered in August 2021. Among other questions, we asked workers if they had been dismissed by any employer during the 3 years prior to the survey, and if so, whether they had receive the severance payment required by law. Among the 27,934 respondents saying they had been dismissed, 48% said their employers had not made the required severance payments. Yet only 13% of those not compensated pursued a case through labour court. Among the 87% choosing not to pursue their case in the court, the most common reason for not pursuing the case was that the court “is too expensive” (49% of respondents) or “useless” (29%).

Fact 2. Plaintiffs receive little (Historical data):The amount collected is only 20% of the amount claimed on average, and 50% of what the law mandates.

Figure 2 uses the sample of concluded cases to show the amounts claimed and recovered for the four main outcomes: settlement, drop, judgment, and expiry. Both the historical data and the Phase 2 case files suggest that around 55% of cases end by settlement, 20% are dropped or expire and 25% end with a judge’s decision.¹⁷ For each outcome, the first bar shows the average amount of money claimed by the plaintiff, the second bar the estimated minimum compensation by law;¹⁸ the third bar the amount of money collected, on average; and the final bar the average amount collected conditional on collecting a positive amount. The amount collected is zero when the lawsuit is dropped, the time expires, and when the plaintiff either loses or wins but is unable to collect anything from the defendant. In cases ending with a judgment the worker recovers a positive amount only 24% of the time. For either settlements or judgments, the amount received is a small percentage of the amount claimed.

Figure 2.

Differences in claims and compensation by case file outcome—historical data Notes: The figure shows the average amount asked for in the filing, the minimum legal compensation based on the case characteristics, and the amount actually received at the end of the process (overall and conditional on recovering a positive amount). Data are displayed in thousands of pesos by type of case ending, using the 5,005 historical case files. The amounts are discounted at the rate of 50%/year (3.43 per month). Cases end in any one of four ways: settlement; being dropped by the plaintiff; court ruling; or expiring from lack of activity. The share of completed cases in the historical case files by type of ending is shown at the bottom. Note that these do not match the share of all cases by type of ending because these data exclude cases that were unresolved at the end of 2015. Workers recover nothing when cases are dropped or expire. We define settlements as an agreement followed by payment, so settlements always imply a positive recovery for the worker. Workers recover a positive amount only 24% of the time when the case goes to judgment. These data do not distinguish between the judge ruling against the worker and the judge ruling in favour but the worker being unable to collect anything from the firm.

In the 24% of court judgments where the worker recovers a positive payment, she receives on average 170% of the minimum legal compensation for her case and 37.5% of her claim. Figure 2 shows that plaintiffs recover, on average, less than the minimum compensation according to the law and only 8% of their claim in a court judgment. Supplementary Figure C.3 shows the CDF of realized recoveries from our 5,005 historical case files. Plaintiffs with private lawyers receive negative (discounted) payments in around 40% of cases. This is largely the result of the cases where the plaintiff recovers nothing but pays some upfront cost, including perhaps the case filing fee. However, around 7% of the settlements are also for amounts that imply a negative net present value for the plaintiff.

Fact 3. Long suit duration (Historical data):30% of trials started in 2011 had not finished by December 2015. Even conditional on reaching a settlement, the average duration is almost 1 year.

Figure 3 shows the distribution of case length by type of case ending. Even conditional on being concluded in December 2015, cases ending in judgment take 2.4 years on average. Given that many of the 30% of cases filed in 2011 and still open in 2016 are likely to end in a judgment, the unconditional average is much higher. Settlements occur, on average, 10 months after filing, but settlement rates are low by international comparison. Only around 55% of MCLC cases are settled. By way of comparison firing disputes are settled after filing in 79% of the cases in Australia, in 80% of the cases in the U.S., and in 90% of the cases in Sweden (Ebisui et al., 2016).

Figure 3.

Time duration Notes: The figure uses the historical data (5,005 case files) to plot the cumulative distribution of the duration of the case in months, by type of ending, for the 70% of cases concluded by the end of 2015.

The long delays and low settlement rates help to explain the large backlog of cases in the court. Delay has direct costs in the form of court staff time, lawyer fees, and the opportunity cost of litigants’ time. But, delay also results in a collective welfare loss if, as is likely the case, awards result in payments from the party with a lower discount rate (the firm) to the party with the higher discount rate (the plaintiff). These delays represent pure efficiency losses.

Fact 4. Inflated expectations (Survey data):The subjective probabilities of winning for plaintiffs and defendants (in the same case) sum to 1.47,¹⁹indicating aggregate overconfidence. There is average overconfidence relative to the calculator’s prediction as well.

Excessive optimism of the parties may result in there being no settlement that is acceptable to both parties, even in cases where settlement would be possible with more realistic expectations.²⁰ We asked parties present at the hearing the likelihood they would win the case and, conditional on the plaintiff winning, what amount would be paid. In Phases 1 and 2, the average expected probability of winning reported by plaintiffs is 0.79 and 0.80, respectively, while for firm lawyers it is 0.68 in Phase 1 and 0.40 in Phase 2. These probabilities sum to 1.47 and 1.20 in the two phases, respectively. Data from the Phase 3 surveys of workers approaching the court suggest that the overoptimism at least initially comes from workers themselves: prior to beginning the process, workers’ stated probability of winning was 89%.

By comparison, the probability of the worker winning predicted by our calculator in the same cases is 41% in Phase 1 and 33% in Phase 2. In Phases 1 and 2, there are also large differences in the expected amount of the award conditional on winning. Both the worker and her lawyer estimate average amounts more than twice those of defendants. We can build a proxy of overconfidence as the difference between the subjective expectation and the calculator prediction. Figure C.4 in Supplementary Appendix C plots the distribution of overconfidence for different parties for peso amounts conditional on winning and probabilities of winning, and Table C.4 in Supplementary Appendix C shows that, in Phase 1, plaintiffs and plaintiff lawyers are equally overconfident both with regard to the probability of winning and the expected amount recovered.

Fact 5. Misinformation (Survey data):Only one-third of plaintiffs understand their main legal entitlement. Only half know what they are asking for in their own suit.

The main legal entitlement for unfair dismissal is 90 days severance pay, a right so fundamental that it is enshrined in the Mexican Constitution and taught in elementary schools. Less than 30% of plaintiffs know the number of days covered by this entitlement. Even more strikingly, the plaintiffs often do not know what they are asking for in their own suit. In the survey, we asked plaintiffs to: “… mark the items you are asking for in your suit among the following…”, listing: Constitutional payment, reinstatement, overtime, holiday bonus, Sunday bonus, and insurance. We assess accuracy by comparing the responses to the case file. Figure 4 shows the proportion of time the plaintiffs responded correctly to questions regarding severance pay, reinstatement, and overtime claims. The figure shows responses of plaintiffs represented by private and public lawyers separately. We see that between 35 and 50% of respondents represented by private lawyers answered each element correctly. Plaintiffs represented by public lawyers are significantly more knowledgeable about the content of their cases, with between 80 and 100% answering the questions correctly. Knowledge of both the law and the case increases in the level of education.

Figure 4.

Knowledge about law and their own claims in lawsuit Notes: Data are from baseline survey of Phase 1. The figure shows averages of correct answers for several questions, grouped into: knowledge of the law (panel 1) and knowledge of the content of their own lawsuit (panels 2 to 4). The question for panel (1) is “In the case of unjustified dismissal, law gives you a constitutional indemnification: Do you know how many days of salary this represents?”. Panels (2) to (4) correspond to the question: “Mark the benefits that you claimed in this suit: 1. Constitutional indemnification, 2. Medical insurance, 3. Reinstatement, 4. Overtime, 5. Premium for working Saturdays, 6. Aguinaldo (bonus), 7. Don’t know”. The figure indicates whether the respondent correctly answered items 2, 3, and 4.

Fact 6. Private lawyers file higher claims, but do not recover more (Historical Data):Controlling for observables, private lawyers ask for 86% more than public lawyers, but win no more. After paying lawyer fees, the average plaintiff therefore recovers much less with a private lawyer compared with a public lawyer.

The filing fees private lawyers often charge upfront exceed the small marginal cost of filing a standard case. Filing fees give private lawyers an incentive to inflate claims in order to convince workers to file a suit. With regard to case outcomes we find that, conditioning on five basic variables coded from the initial filing,²¹ private lawyers ask for 86% more, on average. However, the ratio of the amount clients recover to the amount demanded is 0.06 lower for private lawyers. These combine to produce average recoveries that are very slightly (0.1%) and insignificantly lower for private lawyers. We show this analysis in Supplementary Table C.1.

While the amount recovered is the same for public and private lawyers, plaintiffs receive all of the recovered amount with a public lawyer and (because of the 30% fees charged by private lawyers) only about 70% of the recovered amount with a private lawyer. Hence, plaintiffs with public lawyers receive much larger payouts than plaintiffs with private lawyers, conditioning on characteristics. Of course, these data are only descriptive, and we make no attempt to adjust for the endogenous selection of lawyers beyond the five control variables described above.

5 EXPERIMENTAL INTERVENTION

The stylized facts presented above show an environment in which workers are uninformed about their legal entitlements and their own lawsuit, and parties to the case are overconfident on average. Both ingredients feature very prominently in theoretical models of bargaining delay/impasse. Our experiment is designed to address a fundamental question: Given these conditions, does the provision of personalized statistical predictions increase settlement rates?

5.1 The treatment

In each phase of the experiment, we compare the effects of providing statistical predictions of case outcomes against a control group. In Phases 1 and 2, hearings were randomly assigned to either the treatment arm or the control group. In the Phase 3, individuals were assigned to treatment or control as they approached the court, with treatment randomized at the day level. We describe the treatments here, and also a describe a placebo treatment that was implemented during Phase 2, designed to test whether experimenter effects were driving outcomes.

The calculator: Subjects in the treatment arm received a personalized prediction of their case’s expected outcomes based on the statistical model described above and the covariates of their own case. The predictions were presented in a single sheet of paper like the one shown in Figure 5, which was used for Phase 1.²² We extracted the data needed to customize the calculator predictions from the initial filing (Phases 1 and 2) or from a survey (Phase 3). The data were typed into a user interface. This was done in the presence of the parties in Phase 1, but for logistical reasons, away from the parties in Phases 2 and 3. The predictions were then printed and given to all of the parties present at the hearings in Phases 1 and 2, and to the worker in Phase 3. A highly trained enumerator working for the research team spent about 5 minutes explaining to the parties the meaning of the numbers. The enumerators explained that these were only statistical approximations and that they were based on concluded cases from historical records. Enumerators gave no additional legal advice. Where the treatment was administered at a hearing, after explaining the calculator information, the enumerators asked the parties if they wanted to delay the start of their hearing for a few minutes to negotiate with the assistance of a court conciliator.

Figure 5.

Calculator Treatment Format (example)—Phase 1 Notes: The figure shows an example of the calculator use in Phase 1 treatments. The top half described the entitlement by law if the judge rules in favour of the plaintiff, based on data in the case filing. The second half shows what fraction of cases end which way, the average duration and amount for each ending, and the expected value ex ante saliently in red in the bottom box. The worker and firm name are removed from the example shown here. Parties were told that this information comes from a statistical exercise based on completed historical cases, and that it gives average prediction based on variables of the initial lawsuit described in the calculator treatment.

Placebo: We were concerned that simply making parties aware of the court’s conciliation services, or having research assistants present and carrying out surveys, might change the behavior of the parties. With this in mind, we implemented a “placebo” treatment in Subcourt 7 during Phase 2 in which we provided a leaflet (see Supplementary Appendix C Figure C.9) describing the role of conciliators in the court process. The leaflet was provided in format similar to the calculator information, but rather than quantitative predictions it simply said: “Do you know that you could resolve this conflict today? Conciliation is fast, free, confidential, and impartial. Subcourt 7 has conciliators. Ask for help!”. If a party receiving the placebo treatment asked to talk with the conciliators, our enumerators showed them where they were situated.

5.2 Implementation

The first phase of the experiment started in Subcourt 7 on 2 March 2016 and continued daily for 12 weeks. The “subcourt” is not a single courtroom, but rather a room with a waiting area and eight counters conducting simultaneous hearings. Subcourt 7 manages about 55 hearings per day. Each night the court gave us a list of hearings scheduled for the following day, along with their notification status. We worked with the subset of hearings for which both parties were notified and therefore required to be present. Among the 20 case files meeting this criterion on a typical day, we excluded hearings scheduled to start at the court’s opening hour of 9 a.m. because the court did not want a delay the start time of the first hearings to cause cascading delays through the day. Note that cases are assigned to hearing times randomly, so our agreement not to consider 9 a.m. hearings does not compromise the validity of the experiment. On a typical day, this reduced our sample by around 1.5 cases to roughly 18.5 cases per day. In the first phase, the sample cases were at different stages of the process—that is, not all were new suits.²³

After receiving the list of cases for the following day, we randomized the eligible cases to the treatment and control group in equal proportions.²⁴ Control cases followed business as usual, except for the surveys we administered. Each morning we set up a survey table and a calculator module in the waiting area just outside the hearings counters. The hearings were displayed on a screen and parties were called up by the subcourt judge’s assistants. Except for the 9 a.m. hearing slot, the start time of hearings is typically delayed, and we carried out surveys and treatments during parties’ waiting time.

Table 1 shows details of the treatments. We began by administering the baseline survey. The survey was conducted blind to the experimental assignment for both the parties and our enumerators. We were able to isolate the survey area from the calculator treatment area, and so avoid contamination, because our sample was only about 18 cases per day. All the parties present were asked to complete the survey, but compliance was optional; in about 70% of the hearings, at least one party completed the survey.²⁵ Treatment status was revealed after the baseline survey, and parties were channelled to their assigned experimental condition and given the appropriate treatment protocol described above.

Table 1.

Description of treatment arms

	Types of cases	Court	# cases	# treated cases	Treatment	Measurements	Dates
Phase 1	All hearings	Subcourt 7	636	315	Point estimates on the probability, time, and monetary returns of case outcomes.	Baseline survey and admin data 42-month follow-up	02 March 2016 to 27 May 2016
Phase 2	1st hearings	5 subcourts	1,097	768	Point estimate on settlement amount, conditional on settling. Point estimate of not receiving any compensation, conditional on court ruling.	Baseline survey and admin data 42-month follow-up	14 October 2016 to 10 February 2017
Phase 3	Dismissed workersbefore finding a lawyer	All workers approaching the information booth.	1,922	922	95% confidence intervals regarding expected returns and duration of the lawsuit.	Baseline, 2-week and 2-month follow-ups, admin data for 30 months	15 May 2017 to 17 August 2018

Notes: This table presents the main characteristics of the three treatment arms.

Table 2.

Summary statistics

Variable	HD	HD Subcourt 7	Phase 1	Phase 2	Phase 3
Panel A: Plaintiff outcomes
Won by plaintiff	0.65	0.69
	(0.48)	(0.46)
Amount won (000)	23.92	20.36
	(56.30)	(47.49)
Total asked (000)	343.1	301.6	598.3	644.1
	(655.2)	(551.2)	(1,438.5)	(3,861.0)
Conciliation	0.63	0.67	0.22	0.2
	(0.48)	(0.47)	(0.41)	(0.40)
Losing court ruling	0.07	0.01
	(0.25)	(0.10)
Winning court ruling	0.02	0.02
	(0.14)	(0.12)
Duration (years)	1.02	0.98
	(0.96)	(0.71)
Panel B: Case file basic variables
Public Lawyer	0.10	0.15	0.11	0.06
	(0.30)	(0.36)	(0.32)	(0.24)
Female	0.48	0.35	0.42	0.46	0.45
	(0.50)	(0.48)	(0.49)	(0.50)	(0.50)
At will worker	0.07	0.06	0.18	0.08
	(0.25)	(0.23)	(0.39)	(0.27)
Tenure (years)	4.17	3.72	4.71	4.33	3.71
	(4.99)	(4.72)	(6)	(5.74)	(4.74)
Daily wage	470	455	740	605	323
	(1,100.8)	(656.2)	(1,361.0)	(1,007.8)	(635.0)
Weekly hours	57.33	57.36	57.39	56.15	52.86
	(15.47)	(15.57)	(16.87)	(13.38)	(15.53)
Observations	5,005	857	599	1,087	1,922

Notes: All monetary units in MXN. Columns 1 and 2 show data from the historical cases filed in 2011 and completed by December 2015. Column 1 uses the complete sample of historical cases for 5 subcourts used in the Phase 2 experiment, and Column 2 shows only the 2011 cases from Subcourt 7 used in Phase 1. The remaining columns show data from the experimental case files, for comparison to the historical files. Column 3 shows the Subcourt 7 case files that were subject of the Phase 1 experiment and Column 4 shows the files from the five subcourts for the Phase 2. Column 5 shows data from the surveys of workers in Phase 3 for variables that are available from the survey. Panel A shows outcomes. We define a case as being won by the worker if either there is a settlement or the plaintiff wins a judgment and makes a positive recovery. The amount won is the average recovery by the plaintiff including includes zeros. Note that this may not coincide with what the judge ordered. Similarly, the fraction of plaintiffs winning a court ruling is limited to those with a positive recovery. Panel B shows some of the main characteristics of the plaintiff’s case from the initial filing. The variables from “At will worker” through “Weekly hours” are essential for calculating the amount of money that the worker is owed under the law for unfair dismissal.

The implementation of the experiment differed slightly in Phase 2. First, randomization was at the case level in Phase 1 and at the day level in Phase 2. This change was made for logistical reasons, given that during the Phase 2 we were working with a larger number of the subcourts. The second is we intervened in cases at all stages of the process during Phase 1 but focused on cases holding their first hearing in Phase 2. All first hearings are held on Fridays. Otherwise, the protocol was not materially changed from Phase 1.

We randomize across Fridays in each of the five subcourts during the experimental window. To save time, we shortened the survey and we pre-filled and pre-printed the calculator. The subcourts did not agree to allow us to delay the hearings, so if after receiving the calculator the parties wanted to negotiate with one another, they themselves had to request a delay in the hearing to sit with the court conciliator.

We implemented the placebo treatment in Subcourt 7 during the Phase 2 experimental window, using ongoing cases with hearings Monday through Thursday. We randomized the placebo at the bi-weekly level with cases during 2 weeks treated and cases in two adjacent weeks serving as a control group without any intervention. For both groups, we coded the variables in the case file and recorded whether there was a settlement on the day of the hearing.

5.3 Integrity of the experiment

Table A.3 in Supplementary Appendix A shows treatment and survey compliance rates for the first two phases of the experiment. We define compliance as the parties being present and willing to receive the treatment. The table shows compliance for each party and at the case level. At least one party received the treatment in 80% of cases in Phase 1 and 87% in Phase 2. We estimate the intention to treat (ITT) in all reported results. Table A.4 in Supplementary Appendix A shows that the variables are well balanced across the experimental groups in both phases: only 6 out of 33 tests are significant, 3 at the 5% level and further 4 at the 10% level. We have no reason to believe the imbalance is driven by anything other than chance. Indeed, as discussed in the notes on Supplementary Table A.4, the imbalance in several variables is driven by five extreme observations that all were allocated to control, and none of our results are changed when these five observations are excluded.

6 RESULTS

The historical data and survey responses show that plaintiffs are overconfident and uninformed about the law and even their own case. Settlement rates are low and case durations are long. Our intervention aims to understand if there is a connection between these two sets of facts: If we increase information and reduce overconfidence, do a rates increase?

6.1 Effects on settlement

Given that treatment is randomized, we estimate the causal effect of treatment by estimating the following equation by OLS:

$$y_{it}={\alpha }_t+{\beta }_t{T}_i+\gamma X_i+{ϵ}_{it}$$

(1)

where $y_{it}$ is the outcome (e.g. settlement) of case i at time t, the constant ${\alpha }_t$ estimates the mean for the control group and $T_i$ indicates assignment of the case to the calculator treatment arm. Thus, ${\beta }_t$ estimates the ITT effect at a given point in time t. We estimate separate regressions for each t, with t indicating the day of the hearing or 42 months after treatment (the latter measured in January 2020 for Phase 1 and July 2020 for Phase 2), or 2 months after treatment in Phase 3. $X_i$ is a vector of controls including subcourt and year-of-filing fixed effect.²⁶ Finally, since the effect may differ according to which parties received the treatment, we also interact the two treatment arms with an indicator for whether the employee was present (EP) when we delivered the treatment, while controlling for EP itself. In Phase 2, we add subcourt fixed effects.

The first six columns of Table 3 focus the short-term outcome of same-day settlement. The dependent variable is a dummy for whether there was a settlement on the day of the intervention. The first two columns of Table 3 use data from Phase 1 of the experiment; Columns 3 and 4 use data from the second phase of the experiment; and Columns 5 and 6 combine data from the first two phases. Column 1 shows that 6% of the control cases in Phase 1 settle on the day of the hearing, while the settlement rate of the treatment group is approximately 5 percentage points higher. The treatment effect is significant at the 5% level. A similar significance level is suggested by randomization inference, reported at the bottom of the table.

Table 3.

Effect of treatment on settlement

	Months after treatment
	Same day settlement							Long run
	Phase 1		Phase 2		Phase 1/2			Phase 3
	OLS						CF OLS	OLS
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
Calculator	0.050**	0.026	0.040*	0.0059	0.047***	0.016	0.0023	−0.012	0.064**
	(0.023)	(0.019)	(0.021)	(0.023)	(0.015)	(0.016)	(0.016)	(0.025)	(0.027)
Emp present (EP)		0.14***		0.15**		0.15***	0.70***	0.051
		(0.051)		(0.060)		(0.040)	(0.21)	(0.052)
Calculator#EP		0.13		0.14*		0.14***	0.12**	0.11*
		(0.081)		(0.072)		(0.052)	(0.051)	(0.064)
Residual							−0.31***
							(0.11)
Control group mean	0.063	0.063	0.16	0.16	0.11	0.11	0.11	0.41	0.39
Interaction control group mean		0.19		0.27		0.23	0.23	0.46
Observations	631	631	1,092	1,092	1,723	1,723	1,723	1,723	1,920
$R^2$	0.140	0.217	0.056	0.117	0.073	0.139	0.144	0.122	0.007
$H_0:Calc+Calc\mathrm{\#}EP=0$		0.052		0.030		0.0020	0.011	0.091
Calculator p-value RI	0.034	0.20	0.15	0.83	0.0080	0.34	0.91	0.69	0.056
Interaction p-value RI		0.14		0.064		0.013	0.026	0.077

Notes: This table estimates the main treatment effects for both experimental phases. Columns 1 through 7 measure settlement on the same day as the treatment, with the sample in Columns 1 and 2 limited to Phase 1, the sample in Columns 3 and 4 limited to Phase 2, and the sample in Columns 5 and 6 combines data from the two phases. Column 7 uses treatment-day settlements but adds a control for potential endogeneity of the presence of the employee, as described in the text and shown in Table C.5. Column 8 shows the effect of treatment on settlement around 42 months after the treatment, using the combined sample from the first two phases. Column 9 shows the effect on settlement 2 months after treatment in Phase 3, using data from the follow-up phone surveys. Note that in Phase 1, treatment is at the case file level, while in Phase 2 it is at the subcourt-day level. Standard errors are clustered at the level of treatment. “Calculator” is a dummy indicating that the case file or day was randomly assigned to receive the calculator. “Emp present” is an indicator for the employee being present on the treatment day, and “Calculator#EP” is the interaction between the employee present and treatment dummies. Columns including Phase 2 observations include subcourt dummies. ^*Significant at 10%; ^**significant at 5%; ^***significant at 1%.

Column 3 shows that in the second phase of the project, just over 16% of the cases settled on the day of the hearing. Recall that the second phase was conducted with cases holding their first hearing, and the higher settlement rate likely reflects this fact.²⁷ However, the effect of the calculator treatment is similar in magnitude to that in the first phase: settlement rates on the day increase by 4 percentage points in the treatment group compared with control, an effect that is just significant at the 0.10 level. Thus, our main result is that statistical information facilitates bargaining. To the best of our knowledge, this is the first experimental evidence from live court cases showing that decreasing information asymmetries helps to resolve bargaining impasses.

Columns 2 and 4 show our second main result: the treatment effect occurs only when the employee is present. In these regressions, we interact treatment with a variable indicating the plaintiff herself was present, while also including a variable indicating that the plaintiff was present. First, note that in both Phase 1 (Column 2) and Phase 2 (Column 4), settlement on the day is much more likely when the employee is present. In the control group, 19% of the Phase 1 cases and 27% of the Phase 2 cases are settled on the day of the intervention when the employee is present. But treatment increases settlement rates by 15.3 percentage points in Phase 1 (0.026 + 0.127) and 14.3 percentage points in Phase 2 when the employee is present. The joint effect of the treatment and the treatment/employee present variables are significant at just below the 5% level in Phase 1, and at the 0.05 level in Phase 2.²⁸ The calculator treatment in either phase when the employee is not present is close to zero and highly insignificant particularly in the second phase (Row 1 in Columns 2 and 4).²⁹ The effect of the treatment when the employee is present increases settlement rates by enough to significantly close the gap with those of developed countries referenced above.

The Phase 2 results provide a replication within the experiment, and the similarity of results in Phases 1 and 2 is reassuring. Combining the samples increases statistical power. We do that in Columns 5 and 6 using the specification from Columns 1 and 2, respectively. Not surprisingly, we find very similar treatment effects, with the treatment—employee present interaction effect now itself significant at the 1% level (or the 2% level using RI), and the effect of the calculator when the employee is not present remaining very close to zero.

The regressions in the first six columns measure the effect of treatment on immediate settlement and suggest that lawyers do not act on the calculator information in the absence of their client. But might they share the information with their client after the hearing, producing a delayed effect on settlement? We use the court’s administrative records to track cases over time. Column 8 shows the effect of treatment 42 months after treatment in December 2019/January 2020 for Phase 1 and July 2020 for Phase 2. The 42-month window allows for several additional hearings in the case and, indeed, is after almost 90% of cases are resolved.

Our third main result is that the effect of treatment does not change materially 42 months after the intervention, even though the number of cases settled in both the control and treatment groups increases substantially. Focusing first on cases where the employee was not present at the hearing, comparing Column 6 with Column 8, we see that in the control group, the settlement rate increases from 11% to 41%. Meanwhile, the effect of the calculator when the employee was not present (Row 1) remains a fairly precisely estimated zero (and, indeed, is slightly negative). Where the employee was present to receive the treatment, the treatment effect also remains almost unchanged over time. The 14 percentage point effect on the day of the hearing drops (insignificantly) to 11 percentage points after 42 months; the effect of the calculator when the employee was present remains statistically significant at the 7% level.

Treatment is random conditional on the presence of the employee at the hearing, and hence the results in Column 8 are internally valid. However, inferring agency from this result requires that we assume treatment would have a similar effect among those employees not present at the hearing. We might be concerned with this interpretation if plaintiffs are present when there is potential for the case to be settled, and not present when there is little potential for settlement. However, the long-run follow-up data suggest that the plaintiff’s presence on the day is not determinant of settlement in the control group. First, among cases in the control group where the employee was present on the day, the effect of the employee’s presence dissipates over time; 42 months after treatment, the effect in the control group is much smaller and not statistically significant (Row 2, Column 8 of Table 3).

Second, among the control group cases where the employee was not present on the day, an additional 30% of the cases settled over the following 42 months. Taken together, these results imply that neither the presence nor absence of the plaintiff on the day of the intervention determined settlement in the longer run among the control group cases. On the other hand, settlement in the treatment group was affected by the presence of the employee, both on the day of treatment and in the longer run.

To address residual concerns with the endogeneity of the plaintiff’s presence, in Column 7, we use a control function approach (Wooldridge, 2015), using settlements on the day of treatment as the outcome.³⁰ Hearing times are assigned to cases randomly, and we find that a dummy variable indicating the two early (9:00 or 9:30) and two late (12:00 or 12:30) hearing times is highly significant in predicting employee presence.³¹ The results indicate that the control function variable itself is significant, but the control function has little effect on the magnitude or significance of the interaction between employee presence and treatment.

In Phase 3 of the experiment, we provide the calculator to all of the dismissed workers in the treatment group. Recall that our sample for Phase 3 is dismissed workers approaching the court seeking information. Column 9 shows the treatment on settlement in this sample prior to filing a case. Note that almost two in five (39%) of the control group in Phase 3 settle their case by 2 months after they come to the court. The calculator nevertheless significantly increases settlement: an additional 6.4% of workers assigned to receive the calculator treatment settle before filing. This represents 10.4% of the 61% of workers who would not have settled without treatment, an effect size only slightly smaller than the effect when the plaintiff was present in Phases 1 and 2.³²

Collectively, the results from Phases 1 and 2 suggest that the lawyers do not share the calculator information with clients. The historical and survey data give us reason to suspect that agency issues might be particularly relevant in cases with private lawyers. In Table 4, we separate plaintiffs according to whether they are represented by a private or public lawyer. We repeat the regressions in Columns 6 and 8 of Table 3 for each type of lawyer. The results show that the calculator has an effect only when the plaintiff is represented by a private lawyer.³³ In cases where the plaintiff is represented by a private lawyer, the calculator treatment increases settlement by 19 percentage points when the plaintiff is present, and not at all when the plaintiff is not present. Both of these outcomes change only slightly 42 months later. Moreover, for the control-group cases represented by private lawyers, the effect of employee presence drops from 14% to 5% over the 42 months, suggesting again that the presence of the employee on the day of the experiment is not in itself determinant of longer-run outcomes in the case. Meanwhile, the treatment has no significant effect on settlement in the much smaller sample of plaintiffs represented by public lawyers. That agency underlies this pattern is also suggested by the data on Figure 4, which shows that plaintiffs using private lawyers are significantly less informed about the contents of their case than are plaintiffs using public lawyers.

Table 4.

Treatment effects conditional on type of lawyer

	Private		Public
	Same day	Long run	Same day	Long run
	Phase 1/2
	(1)	(2)	(3)	(4)
Calculator	0.020	−0.001	0.012	0.010
	(0.017)	(0.029)	(0.059)	(0.132)
Emp present (EP)	0.143***	0.051	0.249**	0.261*
	(0.042)	(0.060)	(0.115)	(0.136)
Calculator#EP	0.194***	0.159**	−0.054	−0.035
	(0.058)	(0.074)	(0.137)	(0.182)
Control group mean	0.115	0.421	0.163	0.429
Control group interaction mean	0.229	0.446	0.269	0.538
Observations	1,546	1,546	131	131
$R^2$	0.144	0.086	0.084	0.077
Calculator p-value	0.000142	0.0253	0.734	0.845
RI p-value	0.197	0.349	0.685	0.981
RI interaction p-value	0.003	0.021	0.676	0.873

Notes: The table reproduces the regressions in Columns 6 and 8 of Table 3. Columns 1 and 2 use the sample of cases in which the plaintiff uses a private lawyer and the Columns 3 and 4 the sample of cases where the plaintiff uses a public lawyer. We are unable to determine whether the lawyer is public or private in 60 of the cases. The dependent variable is a dummy indicating the case was settled by the time indicated on the column heading. All regressions include sub-court dummies. A Wald test from a fully flexible specification on the full sample indicates that the p-value for the difference between calculator × EP in Columns 1 and 3 (private versus public lawyer in the short run) is 0.084, while the p-value for the difference in Columns 2 and 4 (private versus public lawyer in the long run) is 0.310. ^*Significant at 10%; ^**significant at 5%; ^***significant at 1%.

One concern is that the parties may believe the calculator information is provided by experts, and so simply agree to settle for the amount presented to them. Figure C.5 in Supplementary Appendix C shows the ratio of the agreed settlement to the calculator predicted settlement for all cases ending in settlement. We find that only 29% of settlements in the treatment group are within 25% of the calculator prediction. This is only slightly higher than the 24% of control-group settlements that fall within this band. These outcomes suggest that the calculator served only as an initial guide for the bargaining between the parties. Moreover, Figure 6 shows the density and CDF of the predicted case value among the settled cases in the treatment is shifted left relative to control group, indicating settlements induced by the treatment come disproportionately in cases with lower predicted case values. Further, Figure 7 shows conversely that, among the unresolved cases, the calculator predicts much larger recoveries for the treatment cases than for the control cases, conditional on winning a court judgment.

Figure 6.

Treatment and settlement of cases (a) PDF, (b) CDF Notes: This figure shows the distribution of amounts predicted by the calculator for treatment and control cases ending in settlement. Panel (a) shows the distribution functions, while panel (b) shows the cumulative distribution function. The dotted line at the bottom of Panel B shows the net difference in the two distributions at each percentile. The two distributions are not statistically different from one another at conventional levels at any single point or in total.

Figure 7.

Unresolved cases: calculator prediction conditional on court win Notes: This figure presents the distribution of recovery amounts predicted by the calculator, conditional on winning a judgment, for treatment and control cases that were unresolved at the time of last data access. Data are truncated at the 99th percentile to compress the scale. Note that the differences between predicted outcomes in treatment and control reflect in part the fact that more of the unresolved cases in the treatment group are from Phase 2, and hence higher-value cases. Figure C.7 in Appendix C shows the distributions by phase. The patterns are similar though somewhat more muted.

We read the collective results as indicating that plaintiff-lawyer agency issues are important in this context. Private lawyers appear not to transmit evidence to plaintiffs who are not present to receive the information directly. Three alternatives to the agency explanation are possible. First, lawyers may have attempted to explain the calculator results to their clients but were unable to convey the meaning of the data provided. Given the simplicity with which the data are presented, we find this unlikely. Second, lawyers may have simply forgotten, though we view this as itself a version of misaligned incentives and moral hazard. Third, the worker trusts the calculator more than her lawyer. Even if the failure to pass on the information to their clients simply reflects the difficulty perfectly aligned lawyers have in explaining the calculator, the lack of a treatment effect when the plaintiff is not present indicates that the plaintiff does not fully trust her lawyer to make decisions on her behalf.³⁴

A final result is that the placebo has no effect on settlement.³⁵ The placebo makes parties aware of the availability of the court conciliation process, but provides no predictive information on their own case. We interpret the lack of any effect of the placebo treatment as evidence that the content of the calculator information matters.

6.2 Case outcomes

What is the counterfactual outcome for the cases induced to settle by treatment? We examine this first in Table 6 by looking at the pattern of case outcomes in the control and treatment groups for the first two phases of the experiment. We accessed administrative records for the Phase 1 cases in December 2019 and January 2020 and the Phase 2 cases in July 2020. By those dates, only 12% of the cases remained unresolved. The outcomes on Table 6 suggest that, in aggregate, the treatment shifts cases from court rulings without collection to settlements. Compared with the control group, settlement rates are 6.5 percentage points higher and court judgments without collection 7.3% lower in the treatment group. For the purpose of Table 6, we have classified plaintiffs as winning if the judge rules in their favour, regardless of whether or not they are able to collect the award from the defendants. As we have noted, collection of the award is far from automatic. Indeed, as of the last date we accessed the records, there were 50 cases where plaintiffs had won but not yet collected anything from the defendants. Moreover, in 17 cases where the plaintiffs had won judgment and collected a positive amount, they recovered only 52% of the judgment, on average.

6.3 Effects on overconfidence

The calculator treatment provides information on likely outcomes of the case. One channel through which the treatment may be effective is by reducing excessive optimism of the parties. Ideally, we would measure beliefs both before and after treatment in both the treatment and control group. We faced operational challenges in constructing this measure in all three phases of the experiment. In Phase 1, the baseline survey data indicate initial overconfidence, but compliance rates with the follow-up survey conducted after the hearing were low, as parties were anxious to leave immediately after the hearing. Nevertheless, the data available from Phase 1 indicate that the treatment lowered the expectations of overconfident plaintiffs.³⁶ The data from the Phase 3 provide stronger evidence that the treatment tempered optimism, albeit somewhat modestly. As we noted above, in the Phase 3 baseline survey indicates the average worker believed they had an 89% chance of winning their case. After presenting them with the calculator information, we elicited expectations a second time. Table 5 shows the difference between the initial and subsequent expectations among the treated sample. After reviewing the calculator information, respondents decreased their probability of winning by 6 percentage points, with 30% reporting a lower probability of winning, and 26% reducing the amount they expect to recover. However, on average they remain optimistic relative to the historical outcomes.³⁷

Table 5.

Immediate expectation updating

	Probability		Amount
	Dummy (lowered)	Continuous	Dummy (lowered)	Continuous
	(1)	(2)	(3)	(4)
Calculator	0.30***	−0.06***	0.26***	1,993
	(0.02)	(0.01)	(0.03)	(3,287)
Observations	1,251	1,251	573	573
$R^2$	0.19	0.06	0.16	0
RI p-value	<0.001	<0.001	<0.001	0.646

Notes: The table uses the Calculator treatment to estimate its effect on immediate expectation updating. Column (1) regresses an indicator for “decreased expectations” (a variable a dummy = 1 if the the subjective belief of the probability of winning measured in the immediate follow up survey is lower than in the baseline survey) on treatment arms dummies and basic variable controls. On the other hand, Column (2) regresses the amount by which expectations changed with the same regressors. Column (3) is the same as Column (1) except that the dummy refers to amount to be won conditional on winning instead of probability of winning, while Column (4) mimics Column (2) referring to the quantitative difference of the amount the subject thought she might win. ^*Significant at 10%; ^**significant at 5%; ^***significant at 1%.

Table 6.

Case outcomes by treatment arm

	Control (%)	Calculator (%)	Difference
Resolved with recovery:
Settled	41.18	47.73	6.55
Court ruling with payment	6.19	5.76	−0.44
Resolved with no recovery:
Court ruling without payment	26.93	19.59	−7.34
Expired/Dropped	13.62	13.74	0.12
Unresolved:
Continues	12.07	13.18	1.11
N	646	1,077	1,723

Notes: The table shows the status of the experimental cases in Phases 1 and 2 as of January 2020 (for Phase 1) or July 2020 (for Phase 2). These outcomes are taken from the administrative records of the court. “Court ruling with payment” indicates the plaintiff was awarded a judgment by the court. However, 22.23% of those awards were uncollected as of January/July 2020 and it is likely that some will never be collected.

6.4 Effects on case duration

We should expect that the increase in the number of settlements on the day of the treatment will result in shorter average case duration. Table C.8 in Appendix C reports the results of regressing the duration of cases with a private lawyer against the treatment, employee present and the interaction of the two. We define duration as the number of days from the date of the case filing to the final resolution of the case. For cases still unresolved on the date the records were last checked, we use the date we last checked administrative records as the final date. We run both an OLS and a Cox proportional hazard model. Using either specification, we find that the treatment has no effect on case duration when the plaintiff is not present, but reduces duration by almost 4 months when the plaintiff is present, significant at the 10% level. Table C.8 also reports a hazard regression on the sample excluding the cases that were settled. In that subsample, we find a small and insignificant effect on duration, suggesting that the treatment affects duration only through the increase in the settlement rate.

7 IS THE INCREASED SETTLEMENT RATE BENEFICIAL?

The increased settlement rate generated by treatment helps the court meet its goal of reducing the case backlog. There is evidence from other contexts that reducing backlogs leads to more efficient outcomes (e.g. financial contracting in Brazil, Ponticelli and Alencar, 2016). But are the settlements induced by treatment directly beneficial to the parties who receive the treatment? A necessary condition for the increase in settlements to be beneficial is that the untreated settlement rate is inefficiently low. In this section, we first discuss a simple framework that shows why settlement rates might be inefficiently low. We then analyse the data on the actual outcomes of cases in Phases 1 and 2, and surveys of well-being from Phase 3. Because almost 90% of cases in Phases 1 and 2 were resolved by 2020, we are able to measure outcomes for plaintiffs directly, with few cases requiring projected outcomes.

The discussion highlights reasons that the plaintiff’s incentives may be misaligned with those of her lawyer, underlining analytical reasons that settlement rates may be inefficiently low. While contingency contracts help to align incentives, they do not eliminate agency issues, particularly with regard to settlement decisions. The empirical analysis shows that the plaintiffs are made better off, on average, by the increase in settlement rates, a finding that is robust to different assumptions on outcomes for the continuing case files.

7.1 Risk, discounting, and settlement incentives

Treatment increases the probability of settlement, but only when the employee is present. We also find that plaintiffs and the parties collectively are excessively optimistic about their chances of winning their case. While there are four parties involved in the case, for this discussion we focus on cleavages between the plaintiff and her lawyer, which appear to play a central role in the results shown on Tables 3 and 4.

The bargaining literature defines the “settlement range” as the range of outcomes that both parties prefer to continued bargaining. The settlement range will be empty if bargaining is costless, and the parties to the dispute are risk-neutral, have identical discount rates and have identical expectations over outcomes. However, if, for example, future outcomes are uncertain and parties are risk-averse, then both parties may prefer to settle for the (certain) expected outcome to remove the risk associated with the uncertain future outcome. Myerson and Satterthwaite (1983) shows that, when parties have asymmetric information, bargaining can break down even when there is a non-empty settlement range—that is, even when both parties prefer settlement over continued bargaining. This classical explanation for sub-optimal settlement rates may be a factor in generating inefficiently low settlement rates at the MCLC, and why providing information to parties that decreases information asymmetry may increase settlement rates. But, the experimental results also suggest that plaintiff-lawyer agency also plays a role. Given the apparent divergence between the plaintiff and her lawyer, we focus on factors that drive wedges in settlement preferences between these two.

Are there primitive factors that lead to plaintiffs preferring to accept a given settlement offer while their lawyers prefer to pursue the case? We write down a simple framework in Supplementary Appendix B that highlights two factors: the ability to diversify risk and differences in discount rates. The framework formalizes what we view as fairly intuitive points. The plaintiff and her lawyer differ in their ability to diversify risks and, given income differences, likely in their discount rates as well. The framework in Supplementary Appendix B shows why differences in these two dimensions lead to differences in preferences over taking a certain payoff now (i.e. a settlement) or an uncertain payoff later (i.e. a court judgment).

In deciding between an offer of settlement now and an expected outcome from continuing to pursue the case, each party will account for the difference in timing of the two potential payments, and the uncertainty of the payoff from continuing the case. Taking uncertainty first, we note that lawyers typically handle many cases simultaneously. For example, the median number of ongoing cases among lawyers surveyed in Phase 1 was more than 30. Because the outcomes of each of a given lawyer’s cases are largely independent of one another, the lawyers are able to diversify the payoff risk inherent in pursuing any single case. Plaintiffs, on the other hand, are a party to only a single case, and therefore cannot diversify that risk. As both intuition and the formal model show, the difference in the ability to diversify risk leads plaintiffs sometimes to prefer the certainty of settlement when their lawyer would prefer the uncertain outcome inherent in pursuing the case: the settlement range between the firm and the plaintiff’s lawyer is a subset of the settlement range between the firm and the plaintiff.

The parties may also differ in the rate at which they discount the future. Because plaintiffs will have recently lost a job and will generally have lower incomes even when working, we expect them to be less patient than their lawyer. For any expected future payout following a judge’s decision several years later, the higher discount rate would lead the plaintiff to accept a lower settlement today than her lawyer would be willing to accept. The differences in either the ability to diversify risk or the rate of discount drive a wedge between preferences of the plaintiff and her lawyer, with the plaintiff having a larger settlement range than her lawyer.³⁸ Again, our objective here is not to be exhaustive in the modelling, but to show that there are plausible reasons that the incentives of the plaintiff and her lawyer are misaligned, and that the misalignment results in too few settlements from the perspective of the plaintiff.

The divergence in preferences of the plaintiff and her lawyer provides an incentive for the lawyer not to de-bias the expectations of the plaintiffs—for example, by showing the calculator information to the plaintiff. Hence, it provides a rationale for the calculator treatment being effective only when the plaintiff is present to receive the information directly. Finally, the divergence is also sufficient to produce a settlement rate which may be sub-optimally low from the perspective of both the plaintiff and the firm.

7.2 Empirical evidence on the induced settlements and plaintiff welfare

The framework motivates why the plaintiffs might be better off with more settlement, but what do the data say about whether they are made better off by the additional settlements our treatments induce? We examine those data here to provide evidence on the welfare of the plaintiffs induced to settle through the treatments.

Table 6 showed that, on average, the primary effect of treatment was to shift outcomes for the plaintiff from a losing court judgment to a settlement. However, these average outcomes may mask more nuanced underlying patterns. Here, we instead analyse the effect of treatment on the amount awarded to the plaintiffs. Table 7 reports the results of regressing the net present value of the amount awarded against treatment. For cases with a private lawyer, we reflect legal fees by multiplying the recovery by 70% to reflect the plaintiff’s share and then subtracting the 2000 MXN up-front fee. We use the amount awarded for both settlements and court judgments. While settlements are collected in essentially all cases, judgments are not. So for judgments, this will overstate the amount recovered. For the 12% of cases that were continuing at the time of the latest update, we assume either that the cases recover the amount that the calculator predicts when their case goes to judgment³⁹ (Columns 1 through 3), or that they recover nothing (Columns 4 and 5). We discount all payments to the date of filing using the discount rate reported reported by the median plaintiff in the Phase 1 survey, which is 50%/year.⁴⁰ We revisit the assumption on the discount rate below.

Table 7.

Recovery after 42 months, phase 1/2 samples

	Net present vaue of amount awarded
	Calculator imputed			0s imputed
	NPV	IHS NPV	IHS NPV	NPV	IHS NPV
	(1)	(2)	(3)	(4)	(5)
Calculator	−405	−0.45	−0.42	−952	−0.83*
	(650)	(0.49)	(0.49)	(606)	(0.46)
Employee present (EP)	3,023**	−0.18	−0.21	2,940**	0.25
	(1,488)	(0.95)	(0.96)	(1,321)	(0.92)
Calculator#EP	511	2.63**	2.68**	1,313	2.82***
	(1,679)	(1.08)	(1.09)	(1,503)	(1.09)
Control group mean	7123	2.205	2.477	5382	0.193
Control group interaction mean	9803	2.716	2.945	7818	1.149
Observations	1,698	1,698	1,698	1,698	1,698
$R^2$	0	0.12	0.11	0	0.12
Case file Controls	Yes	Yes	Yes	Yes	Yes
Calculator p-value	0.949	0.0285	0.0247	0.811	0.0501
RI p-value	0.539	0.329	0.361	0.102	0.0730
RI interaction p-value	0.768	0.0140	0.0110	0.408	0.00500

Notes: We measure the net present value of the amount awarded to plaintiffs. The dependent variable in Columns 1 and 4 is the NPV, winsorized at the 95th percentile, and in Columns 2, 3, and 5 is the inverse hyperbolic sine of the NPV. For cases not yet resolved, Columns 1 through 3 impute the calculator predicted value for a court judgment, while Columns 4 and 5 impute zeros to unresolved cases. In Columns 1, 2, 4, and 5, we discount payments to the date of filing using a discount rate of 3.43%/months (50%/year). In Column 3, we use a discount rate of 2.22%/month (30%/year). Standard errors are clustered on the day of treatment. All regressions include dummies for the phase, the subcourt, the year the case was filed, the number of plaintiffs in the case and an indicator some information was missing from the case file data. ^*Significant at 10%; ^**significant at 5%; ^***significant at 1%.

The regressions reported in Table 7 interact treatment with the presence of the plaintiff on the day of treatment while controlling for subcourt fixed effects, year of filing, number of plaintiffs in the case, and use of a public lawyer. The outcomes have a very long right-hand tail, and so we should be concerned about the influence of a handful of observations. On the one hand, there is no reason to believe they do not represent real outcomes, though in fact many of the largest awards are not collectable or are reduced through post-trial bargaining. In Columns 1 and 4, we measure the NPV in levels, but winsorize the data at the 95th percentile to reduce the influence of the upper tail of the data. In levels, we find no significant effects of treatment on the NPV of outcomes, whether we use the calculator to impute values for unresolved cases (Column 1) or impute 0’s for those cases (Column 4). The measured effect of the treatment when the plaintiff is not present is negative, and the measured effect when the plaintiff is present is positive. However, neither effect is significant. Column 2 instead uses the inverse hyperbolic sine of the outcomes. The IHS specification yields a pattern similar to the effects of treatment on settlement. We find that treatment leaves plaintiffs better off only when the plaintiff was present to receive the treatment directly. Figure C.6 in Appendix C, shows the distribution of the amount awarded when the plaintiff was present at the hearing, by phase. The distribution of awards in the treated sample has less mass in both tails, the left-hand tail reflecting the results from Table 6, and the right-hand tail perhaps reflecting settlements of cases that would have yielded a higher award if they had proceeded to judgment. Because the IHS specification raises the weight of observations in the left-hand tail and lowers the weight of those in the right-hand tail, the results are more favourable for treatment.

Columns 3 though 5 of Table 7 show that the estimated effects are robust to changes in two key assumptions. First, in Column 3, we discount all payments at an annual rate of 30%—lower than a typical credit card rate in Mexico during this period. Comparing Columns 2 and 3, we see that the discount rate has almost no effect on the estimated treatment effect. Second, although almost 90% of the cases were resolved by 2020, the analysis of the effects of treatment on plaintiff outcomes still depends on assumptions about the remaining unresolved cases. In the first three columns, we assumed that the unresolved cases would all result in plaintiff being awarded the amount predicted by the calculator, conditional on the case proceeding to judgment. In Columns 4 and 5, we instead assume that the plaintiff loses the case and collects nothing. Columns 4 (levels) and 5 (IHS) show that the assumption on recovery from the unresolved cases has little effect on the estimated treatment effects, though the effect of treatment when the plaintiff is not present is now marginally significantly negative in the IHS specification. The general robustness to how we impute values for the unresolved cases is perhaps not surprising given that, as we noted earlier (Figure 6), treatment has less effect on settlement among plaintiffs with the highest-value cases.

The cases of the workers involved in Phase 3 of the experiment are not far enough along to carry out a similar analysis. However, in the survey conducted 2 months after the treatment in Phase 3, we asked workers about well being more directly. In particular, we asked them a standard question about general happiness and two questions about difficulties in paying bills. We examine the effect of treatment on responses to these questions on Table 8.⁴¹ We find no effect on happiness, though the measured effect is positive. However, we do find that the calculator treatment reduced the likelihood they report having had trouble paying bills or not having money for food. These are, of course, short-term impacts, but they are consistent with the higher settlement rates for Phase 3 treatment we observed on Table 3, and also consistent with the importance of cash from settlement while the dismissed workers search for a new job.

Table 8.

Phase 3: effects on welfare

	Happiness	Stopped paying serv.	Lack of money	Works
	(1)	(2)	(3)	(4)
Calculator	0.13	−0.073***	−0.070***	0.020
	(0.12)	(0.024)	(0.025)	(0.025)
Source	2m	2m	2m	2m
Observations	1455	1470	1472	1472
$R^2$	0.0079	0.0075	0.0096	0.021
DepVarMean	7.90	0.60	0.56	0.46

Notes: The 2 month survey for Phase 3 included questions measuring proxies for welfare. We asked (Happiness) On a scale of 1 to 10, where 1 means “not happy at all” and 10 means “totally happy”, in general, how happy do you feel about your life lately? (Stopped paying serv.) In the past 3 months have you had to stop paying for a basic service such as electric power, water, or rent due to lack of money? (Lack of money for food) In the past 3 months, have you lacked money to spend on food one or more days? (Works) Are you currently working?^*significant at 10%; ^**significant at 5%; ^***significant at 1%.

In sum, the data provide evidence that the treatments left the plaintiffs at least no worse off when they were present to receive the treatment directly. They gain both from receiving payment earlier and from receiving payments in cases that would otherwise have been lost. Factoring in risk and the likelihood that the additional settlements come from cases with relatively weak unobservable characteristics, we believe the data are consistent with the typical plaintiff benefiting from the treatment.

8 CONCLUSION

There is evidence that delays in court proceedings are a drag on the economy. The Mexico City Labor Court is emblematic of dysfunctional courts in lower- and middle-income countries. The MCLC suffers from large backlogs and parties to cases face uncertain outcomes. The results of our experiment show that, even in a context where corruption and judges’ incentives may be issues, a straightforward information intervention increases settlement rates both on the day of treatment and in the longer run. These results are consistent with predictions of theories of bargaining under asymmetric information and theories of overconfidence. The induced settlements remove cases from the court docket more quickly, reducing case backlogs. At scale, more efficient courts might either induce workers to file suits more often or, conversely, induce employers to make required severance payments more often. Evidence from other settings shows positive effect on financial (Ponticelli and Alencar, 2016) and other markets (Boehm and Oberfield, 2020; Chemin, 2020) from more rapid resolution of cases.

There are two additional reasons to believe that speeding up the resolution of cases improves overall welfare. First, the worker-dismissal cases filed in the MCLC involve a payment from one party (the firm) to other parties (the plaintiff and her lawyer). These payments themselves are zero sum. In this setting, discount rates of firms are almost certainly lower than those of plaintiffs; in those circumstances, earlier payment will improve the collective welfare of the parties. Second, reaching and enforcing agreements consume resources of the court; faster settlement reduces these administrative costs. Intervening before cases are filed would appear to maximize these administrative benefits.

Importantly, the experiment also allows us to illuminate on the underlying causes of the delays. First, parties to the case, and plaintiffs, in particular, show excessive levels of optimism and limited knowledge of the law. Second, lawyers appear to take advantage of less informed plaintiffs, extending cases when workers would prefer settlement. While there are undoubtedly other issues, particularly those related to enforcement of court rulings, our results suggest the need to better understand the functioning of the market for lawyers in this context.

More broadly, the results of the experiment suggest the need to merge the insights of the bargaining literature with those from the literature on expert agents. The literature on bargaining and settlements has focused on the relationship between the plaintiff and defendant, and de-emphasized the importance of agency issues between a party and her lawyer. Given the importance of the employee being present to receive the information directly, we should view the bargaining game as one that involves more than two parties.

Finally, the experiment provides a window on the functioning of the court as an institution. In conducting the experiments, we worked closely with the court, and provided officials there with evidence about effective and easily scalable policies; thus, the research has contributed to the policy dialogue on general policies at the court. Indeed, as part of a major constitutional reform of labour law, the court proposed that federal labour law include both statistical information customized to the case and compulsory pre-filing conciliation hearings—proposals that were grounded in the evidence from the experiment.⁴²

Acknowledgments

We would like to thank Sebastian Garcia, Andrea Fernandez, Sergio Lopez-Araiza, Enrique Miranda, Isaac Meza, Diana Roman, and Monica Zamudio for superhuman research assistance, Vince Crawford, Jonas Hjort and Bentley MacLeod for helpful discussions, and seminar and conference participants at Cambridge, Columbia, Toulouse, ITAM, Monash, NYU—Abu Dhabi, Oxford, Tinbergen, MIT/Harvard, George Mason, Maryland the Latin American Workshop on Law and Economics, and the American Law and Economics Association for comments. All errors are ours. We acknowledge financial support from the Government Partnership Initiative of the Abdul-Latif Jameel Poverty Action Lab (JPAL) at the Massachusetts Institute of Technology, the Economic Development and Institutions program funded by the UK government through UK Aid, and the Asociacion Mexicana de Cultura. We also acknowledge crucial institutional, operational, and human resource support from the Mexico City Labor Court and its president, Darlene Rojas Olvera. The research reported in the paper was carried out with the approval of the ITAM Institutional Review Board. Phase 3 of the experiment is registered as AEARCTR-0002339.

Supplementary Data

Supplementary data are available at Review of Economic Studies online.

Data Availability Statement

The data and code underlying this research is available on Zenodo at: https://dx.doi.org/10.5281/zenodo.10074874.

Footnotes

1

The exception is firing for malfeasance or repeated failure to report to work on the part of the worker.

2

Kennan and Wilson (1993) argue that it may not be important to distinguish between these two causes as it is likely that overconfidence arises from parties having different information.

3

Lawyers representing plaintiffs always have a power of attorney and therefore do not need permission of the worker to settle.

4

Both the choice of lawyer and the presence of the employee at the hearing may be endogenous to outcomes. However, the treatment is orthogonal to either. We discuss this issue in more detail below.

5

The Phase 3 experiment outlined in AEA trial registry AEARCTR-0002339 includes additional treatment arms that are the subject of ongoing analysis. In this article, we use only the arm of the broader Phase 3 experiment that overlaps with the narrower Phases 1 and 2 experiments.

6

Two notable exceptions are Aberra and Chemin (2019), who provide randomly selected plaintiffs access to lawyers in cases involving land disputes in Kenya, showing that legal assistance leads to greater investment in land held by the plaintiffs; and Sandefur and Siddiqi (2015), who provide access to paralegals that lower the cost of accessing the formal legal system in Senegal. They find that access to the formal legal system is particularly valuable for those disadvantaged in the customary legal system—for example, women who are suing men. Kondylis and Stein (2018) use an abrupt change in procedural rules rolled out across six civil courts in Senegal to identify the effect of the reforms, showing that the new regulation resulted in in faster pre-trial phases of cases.

7

Hubbard (1998, 2000) suggest that reputation is effective in controlling agency in the automobile emissions testing market in California. However, in labour courts, the development of reputations among lawyers is limited both because plaintiffs typically use the court only once, and because the plaintiffs seldom connect with one another.

8

Overoptimism may also reflect self-serving bias (Babcock and Loewenstein, 1997).

9

Disputes in a few “strategic” industries named in the Mexican constitution—oil and gas, pharmaceuticals and auto manufacturing, for example—are handled by a federal-level labour court.

10

In a large range of low to mid-level jobs, entering employees are commonly obliged to sign a letter of resignation (or a “blank letter”) in advance. After firing, the firm adds a date to the letter.

11

For Phase 1, the calculator used the full set 2,158 cases filed in 2011 and 2012 in Subcourt 7. We also include the 2012 cases from Subcourt 7 in the descriptive data we present in the article.

12

Our last data collection was just after the onset of the Covid-19 pandemic. The MCLC closed for long stretches during the pandemic, disrupting normal proceedings for the small share of cases ongoing at its start.

13

Lawyers for the plaintiff and defendant were almost always present, but the plaintiff was present only 18% of the time, and the defendant only 1.4% of the time. In the interest of time, we surveyed only one individual from each side of the case. At least one party completed the baseline survey in 71% of the cases. Survey compliance rates are detailed in Table A.3 in Supplementary Appendix A.

14

We initially attempted to administer post-hearing follow-up surveys, but compliance rates were very low, as parties were anxious to leave immediately after the hearing.

15

We strove for the highest accuracy possible and we believe the goodness of fit is reasonably high. Note however that the results of the experiments are not dependent on this magnitude. Supplementary Figures A.1, A.2, and A.3 show that industry dummies, hours worked, tenure, and wage often belong to the set of important predictors.

16

An exception is that the experimental cases have a higher rate of claiming reinstatement. This is likely because cases demanding reinstatement typically have longer duration, so that they are less likely to be found in a database of concluded lawsuits.

17

The percentages show on Figure 2 reflect the outcomes of the cases that were settled by 2015. As we note later, 30% of the cases filed in 2011 or 2012 were unresolved at the end of 2015. These unresolved cases are very likely to end either in judgment or by being dropped/expired. The percentages in the text account for this censoring.

18

This includes severance pay of 90 days at the stated wage, one year of end-of-year bonus and vacation pay, and a tenure bonus mandated for unfair dismissal of up to twice the minimum wage for 12 days per year worked. We refer to this as the minimum since it does not include salary lost since firing and other claims.

19

This is the measure of overconfidence used by Yildiz (2011) to explain delay or conciliation in a theoretical bargaining model.

20

Yildiz (2011) shows that optimism does not necessarily translate in bargaining delays in a dynamic model. However, excessive optimism can lead to an empty contracting zone so that, in the absence of learning, settlement does not occur even when it be efficient in the absence of optimism.

21

The variables are: gender, at will worker, tenure, daily wage, weekly hours

22

The format and content change somewhat in Phases 2 and 3. Examples of those information sheets are shown in Supplementary Appendix C Figures C.1 and C.2.

23

The experiment in Phase 1 also included a second treatment arm, in which parties were referred to the court conciliator. We focus here exclusively on the calculator treatments, leaving the conciliator treatment to future work.

24

Thirty-four of the 705 Phase 1 cases had more than one hearing during the experimental window. We were not able to determine that a case was coming into the experimental sample for a second time, and hence the case was again randomized into treatment or control. For the analysis, we delete the data from the second occurrence of any case and define the treatment status as the assignment the first time we interacted with the case.

25

In Phase 1, those completing the survey were told that they would be asked to complete a follow-up survey after their hearing and were informed they would receive a prize if they did. However, compliance with the post-hearing survey was much lower, as parties did not want to stay after their hearing ended. We do not use these post-hearing survey data for any of the main analysis, though it is included in some of the additional analysis shown in the Supplementary Appendix.

26

Our results are robust to inclusion of the case characteristics shown on Panel B of Table 2, but we do not include these controls in the reported results.

27

Indeed, regressing a dummy variable indicating settlement on treatment using the combined Phase 1 and Phase 2 data show that there is no difference in settlement rates in the two phases once we control for the age of the case at the time of treatment.

28

The significance levels using RI are slightly weaker, falling below the 0.10 level in Phase 1 and to the 6% level in Phase 2.

29

Table C.2 in Supplementary Appendix C shows balance in case characteristics within subsample (across treatment and control) and across subsamples with the employee present or not. Employees are more likely to attend in cases with public lawyers and when they had a long tenure at the firm, and less likely to attend when they worked longer hours at the firm. Supplementary Table C.3 shows that controlling for the imbalances across subsamples in both levels and interacted with treatment has essentially no effect on the main effect for the calculator.

30

Wooldridge (2015) shows that the control function approach is equivalent to instrumental variables when all specifications are linear but has advantages when the first stage is non-linear (e.g. a probit model) and the second stage includes interaction terms. Both of these hold in our case.

31

The first stage regression is shown on Supplementary Table C.5. We might worry about the exclusion restriction, namely that the time of the hearing affects settlement for reasons other than the plaintiff’s presence. IV restrictions are not verifiable, but we note that in the control group, the time-of-hearing dummies do not significantly predict settlement when the employee is not present (⁠$p=0.2$⁠), nor are they correlated with other measures of the hearing (e.g. length of the deposition).

32

Recall that all employees are present in Phase 3.

33

As with the employee being present, the choice of lawyer is endogeneous, but the treatment is orthogonal to the type of lawyer. Table C.2 in Appendix C shows balance in case characteristics within subsample (across cases with private and public lawyers) and across subsamples with private and public lawyers. As with employee present, Table C.3 shows that controlling for the imbalances across subsamples in both levels and interacted with treatment has essentially no effect on the main effect for the calculator.

34

The possibility that, when the calculator is explained to both the plaintiff and her lawyer in person, the lawyer does not understand the calculator while the plaintiff does seems highly implausible given that the lawyers have both more education and more experience in labour cases.

35

Regression results available from authors on request.

36

Results available from authors on request.

37

Note that the regressions on Table 5 use only the treated sample from Phase 3. We initially asked the control group for updated expectations at the end of the baseline survey, but the most common response was agitation that we were asking them the same question again almost immediately, with no reason for their answer to change. As a result, we stopped asking for expectations a second time from the control group. However, the small sample from the initial control group surveys suggest that no change in expectations is the appropriate counterfactual for the calculator treatment group.

38

This could be offset by differences in the opportunity cost of time of the lawyer. However, lawyers typically manage many cases in the same court building, reducing the opportunity cost of time spent on a given case.

39

That is, the recovery conditional on winning the case times the probability they win the case, conditional on the case ending by a court ruling.

40

Figure C.8 in Supplementary Appendix C shows the discount rate data elicited from surveys, along with comparable data from the Mexican Family Life Survey for 2012 (Rubalcava and Teruel, 2013b, 2013a). Microcredit interest rates in Mexico are closer to 100% per year.

41

The table shows all of the outcomes included in the survey except the time spent at the MCLC. There is a small positive and marginally significant effect of treatment on time spent at the court during the 2 months between treatment and the follow-up survey, though because of the effect of treatment on settlement, we expect this effect will reverse over time.

42

These proposals were passed by the Federal Senate in mid-2019. See text at: https://goo.gl/9AZ6H7.

Author notes

The editor in charge of this paper was Jerome Adda.