What the Bar-Exam Threshold Lets Us Learn
This chapter focuses on the Italian bar-exam threshold as a way to separate a licensing effect from the traits of the people who obtain a license. The useful question is not whether the exam “sorts” candidates in some broad sense, but what can be learned by comparing law graduates just above and just below the first-attempt passing score. By the end, the reader should be able to distinguish a descriptive earnings gap from a local causal estimate and to identify the people to whom that estimate applies.
The gate is imperfect, but it changes access
The Italian legal profession has a useful institutional sequence. Candidates first complete a five-year law degree and then a two-year apprenticeship. They can take the bar exam only after that preparation. The exam has a written stage and an oral stage. The written stage consists of three essay questions, including two legal briefs and a court brief, and a candidate must reach 90 points to proceed. Those who pass the written stage face an oral examination; those who fail must wait for a later attempt.
The pass rule is sharp, but the life consequence is not. About 40 percent of candidates pass the written test, while about 86 percent of written passers pass the oral stage. Candidates can retake the exam without a fixed attempt limit, and roughly 60 percent of people who fail the written stage on their first attempt eventually pass in a later attempt. Passing once therefore does not guarantee that someone will practice law, and failing once does not permanently prevent it.
That imperfection is exactly what makes the setting a fuzzy regression discontinuity. The threshold sharply determines who may continue immediately to the oral stage, but it only changes the probability of eventually holding a license. In the study, passing the written exam on the first attempt raises the probability of ever obtaining a license by about 20 percentage points. The exam is a gate with a hinge, not a wall: it changes access substantially without determining every later choice.
This matters because the comparison the study needs is not between a stereotypical lawyer and a stereotypical non-lawyer. It is between candidates who reached nearly the same point in the same professional route, then landed on opposite sides of a rule that changed their chance of entry. The closer they are to the cutoff, the less plausible it is that the score difference captures a large, pre-existing difference in ability or life circumstances. That is the source of the design's leverage.
The contrast with a simple comparison is important. People who eventually obtain a license are younger on average, perform better in the descriptive education measures available to the researchers, and earn more over the observed period than people who never pass. Those patterns may reflect ability, persistence, family connections, preferences, or other traits that also affect earnings. If we simply subtracted the average earnings of the two groups, we would not know how much of the gap came from legal practice and how much was already present before the license.
How a score cutoff becomes a causal comparison
A regression discontinuity starts with a rule that assigns treatment according to a measured variable. Here, the measured variable is the written score on a candidate's first bar attempt, and the rule is the 90-point threshold. The researchers compare observations close to the cutoff rather than treating the entire population of exam takers as interchangeable.
The logic can be stated in three steps. First, crossing 90 changes access to the oral stage and therefore the chance of eventually becoming licensed. Second, candidates just above and just below 90 are treated as locally comparable, provided they cannot precisely manipulate their score or sort around the threshold in a way related to future earnings. Third, the discontinuity in later earnings at the threshold is scaled by the discontinuity in licensing probability. The result estimates the effect of licensing for the candidates whose legal-career status was changed by barely passing.
It helps to separate the two discontinuities. The first is the change in the probability of treatment: how much more likely is someone just above 90 to ever hold a license than someone just below 90? In this study, the answer is about 20 percentage points. The second is the change in the outcome: how much do later earnings jump at the same score threshold? Because not every person above the cutoff becomes a lawyer and some people below it later do, the earnings jump is not divided by one as if the threshold perfectly assigned careers. It is divided by the licensing jump. That adjustment converts the threshold's reduced-form effect into an estimate of the return to licensing for the people whose status the threshold actually changed.
The comparison is local in two senses. It is local in score: the relevant contrast is among candidates close enough to 90 that their potential outcomes would otherwise be expected to vary smoothly with the score. It is local in treatment response: the estimate concerns people who comply with the threshold-induced opportunity. A candidate far above 90 may have different ability, networks, financial resources, or prospects. A candidate far below it may have a different willingness to retake. The design does not need those people to be identical to the marginal candidates, but it cannot use the marginal result to erase those differences.
The same-degree setting strengthens the interpretation without making it absolute. The candidates have already completed the education and apprenticeship required to sit for the exam, so crossing the cutoff does not suddenly give one group a different law degree. The treatment is access to a regulated occupation. That is important because studies of education or college majors often combine what people learn with the jobs their qualification allows them to enter. Here, the paper is trying to hold the learning route broadly fixed and isolate the value of the occupational door.
The threshold also does not randomize every feature of working life after the exam. It changes the chance of legal practice, and legal practice changes the mixture of self-employment, clients, hours, flexibility, and exposure to demand. Those are consequences of entering the occupation, not necessarily confounders to be “controlled away.” If the license raises earnings because it gives access to a different and riskier work environment, that occupational environment is part of the return being estimated. The design identifies the effect of changing access; it does not define the return as the value of a piece of paper detached from practice.
This distinction protects the study from two opposite mistakes. The first is to call the result merely correlational because candidates choose whether to retake or practice. The threshold supplies an external push that changes the probability of licensing before those later choices. The second is to call the result a pure effect of exam passage independent of all later behavior. The instrument changes a career path, and the later behavior is part of how that path produces earnings. The credible claim sits between those extremes: near the cutoff, the exam-induced change in licensing status has a measurable effect on the subsequent occupational trajectory.
The research records make that sequence observable. The authors digitized five complete Turin exam sessions from 1996 through 2000 and used earlier records to identify first-time takers. Of 2,174 first-time candidates in the main cohorts, 1,972 were successfully linked to Italian Social Security records. Those records cover earnings from private and public employment, self-employment, and professional work, allowing the researchers to follow the candidates for 23 years after the first attempt.
The restriction to first-time candidates is not a cosmetic sample choice. Once candidates can retake an exam, the people who return for a later attempt may be selected by their earlier score, persistence, or decision to remain in the queue. Using later scores could therefore make the composition of candidates change at the very threshold whose effect is being measured. The first attempt gives every candidate a common starting point after the apprenticeship and before later retake choices have reshaped the pool.
The authors also test whether the threshold comparison is credible. Earnings do not show a pre-exam discontinuity: candidates who will later be pushed toward legal practice by barely passing do not already earn more before the exam. Predetermined characteristics are balanced around the cutoff, including age, gender, place of birth, education measures where available, and an indicator of family connection to a lawyer. That last measure is useful because family influence could otherwise make a passing score look like a random opportunity when it is actually an advantage concentrated among connected candidates.
The score distribution contains a noticeable concentration at exactly 90, which could raise a concern about manipulation. The paper argues that the written exam was anonymous and that essay graders commonly rounded scores to multiples of five, making a mechanical pile-up at the threshold more plausible than selective favoritism. The balance patterns around the cutoff are consistent with that interpretation. The authors also report that estimates remain stable when they exclude candidates at 90 and people just below the threshold, a “donut” check designed to reduce the influence of the mass point.
These checks do not turn the design into a universal law of exams. They support the narrower claim that, in this setting, candidates near the first-attempt cutoff are sufficiently comparable for the threshold to serve as an instrument for eventual licensing. The design is also robust to alternative bandwidths, polynomial specifications, and the inclusion of individual controls. Each check addresses a different concern: whether the comparison window is arbitrary, whether the score relationship is being forced into a particular curve, and whether observed characteristics are doing the work that the threshold is supposed to do.
No single check carries the whole argument. A balanced age or gender variable does not prove that every unobserved characteristic is balanced. The absence of a pre-exam earnings gap does not prove that no later selection occurs. And a stable estimate after excluding scores at 90 does not prove that the examiners never exercised discretion. These checks matter because they make specific threats less plausible together: selective sorting would tend to leave traces in predetermined characteristics, pre-existing career differences would tend to appear before the threshold could affect practice, and a result driven only by the score pile-up should change when that pile-up is removed.
The data linkage also clarifies what the outcome means. The researchers observe earnings year by year, rather than a one-time salary immediately after the exam, and they identify whether a person ever obtains the license, not only whether they pass on the first attempt. That is necessary because legal practice may begin well after the written test and because a first-time failure can be followed by a later success. The design uses the first attempt to preserve a common starting point, then follows the consequences of the different access probabilities over the long run.
The final interpretive step is the most important. The estimate is a local average treatment effect, often abbreviated LATE: it describes the effect for “compliers,” the candidates who become lawyers because they barely passed the first attempt but would not have done so had they barely failed. The paper estimates that this group is about 20 percent of the sample. A complier is not the same as an average lawyer, an average law graduate, or even every person who passes the exam.
That boundary changes how the result should be used. The threshold design gives the study a credible answer to a difficult causal question: what happens when marginal access to legal practice changes while prior education is held broadly constant? It does not answer what would happen to a highly prepared candidate far above the cutoff, to a person considering a different license, or to the entire legal profession if the rule were changed. Nor does it say that the license itself is the only difference in a person's future; the license changes entry into an occupation, and the occupation brings its own work, market, and risk conditions.
With that distinction in place, the bar-exam study can support more than the observation that lawyers earn more. It can support a carefully bounded estimate of what marginal access to legal practice does to later earnings. The next question is how that effect unfolds over a career—and whether the average is spread across most practitioners or pulled upward by a small group at the top.