The Wage Ladder Inside a Collective Agreement
To understand the Italian evidence on bargained pay floors, we first need to identify the object that changes. This chapter focuses on the agreement itself: who sets its minima, how several job grades fit into one wage ladder, and how the researchers connect that institutional schedule to workers' observed pay. The payoff is practical. Once the contract, the worker record, and the outcome are kept distinct, a later estimate can be read as a response to a defined update rather than as an unexplained “union effect.”
A contract contains several floors
In the Italian setting studied by Lorenzo Cappellari and Bernardo Fanfani, national industry-level collective agreements are negotiated by representative trade unions and employers' associations. The agreement is not merely a statement that one lowest-paid worker must receive a particular amount. It generally contains contractual minima for different occupations, skill levels, seniority positions, and job classifications. Each minimum is a floor for a particular grade, while the collection of grades forms a contractual wage ladder.
That structure creates several layers of exposure. A worker on a low grade may be close to the lowest floor. Another worker in the same industry may be assigned to a more skilled or senior grade with a higher contractual minimum. If the agreement updates its pay schedule, both workers can be affected even though the nominal amount or percentage change need not be identical. The update is negotiated at the agreement level, but its application is distributed across the ladder through job classification.
The distinction from a statutory minimum is therefore structural rather than rhetorical. A statutory minimum usually gives the labor market a common legal lower bound. A collective agreement can revise a sequence of lower bounds that already reaches into the middle of the wage distribution. The agreement's minima are also fixed components of base pay within the relevant grade. A worker earning above that grade's floor can still be exposed because the scheduled base component changes, even if the worker is not receiving a new legal entitlement for being below a threshold.
The scale of the ladder is substantial. In the study sample, actual wages are on average about 18% above the median contractual minimum. The gap between the average maximum and minimum contractual floors inside agreements is close to 60%. Those figures should not be read as a promise that every worker is paid according to a perfectly observed grid. They show that the grid spans a meaningful internal wage structure. It is capable of moving pay at several points, not just at the bottom edge.
The ladder is not a collection of unrelated numbers. Floors for different job titles within an agreement tend to move together. When the researchers calculate correlations in periods in which the median floor rises, the relationships among growth at the minimum, median, and maximum positions are roughly 0.84 to 0.90. This synchronized movement is what makes the median floor useful as a summary of an agreement-wide update. It captures the direction and timing of change while leaving room for grade-specific differences.
Those differences matter. Annualized nominal floor growth in the study is about 1.8% at the bottom of the contractual-floor distribution, 2.0% at the median, and 2.2% at the top. The higher growth at the top is real evidence of a modestly uneven schedule, but it is too small by itself to explain the much larger contrast in the estimated wage response between the lower and upper percentiles. The contract's architecture opens the door to a broad response; it does not, by itself, settle who benefits most.
There is also variation across agreements. The study follows about 160 national industry-level agreements, and their updates do not occur on one common timetable. In the trade and metal-manufacturing agreements, for example, real floors rise over the study period but do so with different timing and different growth across job titles. Trade floors temporarily lose ground around 2010–2012 before partly catching up, while the distance between higher and lower floors increases in both agreements. The agreement-level timing gives the researchers useful comparison points, while the cross-agreement variation prevents the schedule from being mistaken for a single national policy.
In Italy, representative agreements are automatically extended to workers and firms that are not members of the signatory organizations. That feature makes contractual coverage broader than union membership. It also changes the interpretation of the subject: the question is not whether a worker chooses to join a union and then receives a premium. The question is what happens when the contractual schedule governing the worker's job is updated. This is why membership rates, although relevant institutional context, cannot substitute for information about the agreement and its floors.
The diagram follows the main interface in the study. Bargaining actors determine the agreement; the agreement supplies a ladder of grade-specific floors; the worker's job determines which part of the ladder is relevant; and administrative records reveal several outcomes. No single box can stand in for the others. The agreement is not the worker's wage, the worker's wage is not monthly earnings, and monthly earnings are not employment.
Turning an agreement into an observable treatment
The researchers build the empirical link from two sources. The first is the Italian Social Security Institute's administrative archive, which contains monthly information on earnings, days worked, and worker and job characteristics for the universe of private non-agricultural employees. The second is a hand-collected archive of contractual minima. Employers report the national collective agreement that applies to each employment spell, allowing the two sources to be merged. The matched archive assigns a time-specific bargained minimum to approximately 75% of private-sector employees during 2006–2017.
The study then makes choices that align measurement with the institution. It retains May and November observations and constructs a semiannual panel. These months tend to have fewer vacation-related interruptions, and observed pay is usually closer to ordinary base pay. One-off payments and contractual bonuses are more common in other months, so using May and November helps the measured wage stay closer to the component regulated by the contractual minima. The researchers use a 25% random sample of workers selected by month of birth to make the very large matched panel manageable while preserving representativeness.
The resulting analysis sample contains more than 52 million worker-semester observations and more than 5 million individuals. It includes actual wages, monthly earnings, days worked, job characteristics, and the applicable agreement. The employment analysis later aggregates observations into cells defined by agreement, two-digit sector, and local labor market. The individual-level wage analysis and the cell-level employment analysis answer related but different questions, so their estimates should not be blended as though they came from one measurement unit.
The main wage-rate outcome is the log of full-time-equivalent daily wages. In plain terms, it is intended to track the price of a standardized day of work more closely than a monthly paycheck does. Monthly earnings are a different outcome: they reflect the wage rate and the amount of labor supplied during the month. If days worked fall, monthly earnings can lag behind a stable daily rate. If workers who remain employed work more days, monthly earnings can rise even when the daily wage response is modest. This distinction is essential for reading a distributional result.
There is one measurement complication. The records identify the applicable agreement, but they do not identify the precise job title within the agreement's wage grid for every worker. The researchers therefore use the minimum associated with the median job title as a proxy for the overall dynamics of the contractual floors. The high correlation of floor changes across job titles supports this choice as a measure of timing and direction. It does not imply that the median title is every worker's title or that each grade receives an identical adjustment.
This is an important interface between the written agreement and the observed worker. The bargaining parties decide an agreement-level schedule, but the schedule reaches an individual through the job classification recorded for that employment spell. The administrative data then show what was paid, how many days were worked, and whether the person remained in the employment records. A change at the first interface is therefore not identical to a raise observed at the last one. The study's proxy is designed to identify the common timing of the first change, while the distributional estimates ask how the final outcomes moved across workers.
The treatment variable is nominal, whereas some of the paper's descriptive trends are expressed in real terms after accounting for consumer prices. That distinction is useful rather than contradictory. Nominal growth identifies whether a contract's scheduled amount was updated between two adjacent periods; real growth describes what happened to the purchasing power of the schedule over a longer interval. A nominal update can be large enough to be observed in the contract while still producing a smaller real improvement if prices are rising. The causal comparison is about the response to the contractual event, not a complete estimate of the change in workers' living standards.
The same discipline applies to the outcomes. A worker-level record can show a change in daily FTE pay among people observed in both periods. A cell-level count can show that the number of workers or total FTE days changed in a contract, sector, and local labor market. The first speaks more directly to pass-through among observed workers; the second captures a quantity response that may alter who remains visible. Keeping these units separate is what allows the study to ask both whether the price of labor changed and whether the amount of labor available changed.
The treatment comparison uses that proxy. The researchers create rolling panels of two consecutive periods covering 2006–2017. An agreement is treated in a panel when its median nominal minimum rises by at least 1% from the first period to the second. Agreements with no nominal growth provide the comparison group. Agreements with a small increase below 1% are left out rather than being forced into either category. Among treated agreements, the average median-floor increase is 2.19%; the corresponding average change is 1.92% at the lowest floor and 2.33% at the highest.
This is a comparison of adjacent changes, not a before-and-after portrait of one national wage policy. It asks whether workers covered by agreements that experienced a clear update changed differently from workers covered by agreements whose floors were flat over the same two-period window. Individual-by-event-panel fixed effects absorb stable differences among workers in each comparison. Contract-by-panel effects absorb stable differences among agreements, and time-by-panel effects absorb shocks common to the comparison window. The details serve one interpretive purpose: the estimate is built from within-worker and within-comparison changes rather than from a simple wage gap between union and nonunion workers.
The design also tests for a differential movement before the actual update. In the placebo exercise, future-treated workers and controls are compared over the preceding period when the relevant minima were unchanged. A small and statistically insignificant placebo is consistent with the idea that the later divergence is not simply a pre-existing trend. It does not prove that every negotiation date is random, especially because contract renewals can be anticipated. It is a check on whether the selected comparison shows evidence of a different trajectory before treatment.
The resulting estimand is deliberately narrow. Under the design's assumptions, it is a short-run reduced-form response to a change in bargained contractual minima within a centralized and representative Italian bargaining system. It is not the effect of collective bargaining relative to individual wage setting, the effect of belonging to a union, or the effect of replacing a statutory minimum. It also does not imply that the same percentage increase reaches every grade or every worker.
This narrowness is a strength, not an embarrassment. A broad question such as “Do unions raise wages?” bundles together membership, coverage, bargaining power, contract design, enforcement, and employment selection. The Italian study holds the institutional setting in view and asks a more answerable question about a schedule of contractual floors. Once the schedule and the comparison are defined, the next issue is no longer what a pay floor means in the abstract. It is where the observed wage rate moves when that schedule rises.