From a Shrinking Middle to a Measurable Flow
This chapter makes the French evidence on occupational polarization legible as a flow problem. A changing routine share is an outcome; to understand it, we need to know how many workers move directly between occupations, how many leave employment, and how many enter or return to employment in each destination. The value of the accounting frame is not that it supplies a theory by itself. It tells us which empirical channel must change for a particular interpretation of polarization to be credible.
The distinction is essential because a repeated cross-section observes employment stocks at separate dates but not the route between them. The French study instead follows workers through a longitudinal matched employer–employee panel. It can therefore connect a change in the number of people observed in manual, routine, or abstract work to the transitions that produced that change, while also stating where its observation window is incomplete.
A panel that follows occupations through time
The study uses the DADS–EDP administrative panel, which links annual employer–employee records with a demographic panel. After sample restrictions, it covers roughly 1.9 million workers observed between 1984 and 2021. The analysis focuses on people aged 25–64, a prime-working-age window that avoids most apprentices and most workers past statutory retirement age. Each worker can be located in an occupation in one year and in the same or another state in the next.
The occupations are grouped into three broad, task-based categories. Manual work includes low-skill service and manual occupations. Routine work includes clerical, supervisory, and production occupations. Abstract work includes managerial, professional, and technical occupations. The categories are not simply wage bins, although the average wage order is stable in the data: abstract work pays more than routine work on average, and routine more than manual. That distinction matters because the analysis is about the tasks and transitions associated with occupational groups, not a claim that every worker in a category has the same wage or skill level.
A fourth state is needed: out of employment. It allows the analysis to record a worker who was employed in one year but not observed in salaried employment in the next, and a worker who was not observed in one year but appears in an occupation in the next. The paper is careful about the label. DADS–EDP does not cover every form of self-employment, so the state can include both genuine non-employment and some people whose work is outside the administrative coverage. It is more accurate to think of it as non-observed salaried employment than as a perfectly measured unemployment state.
The panel also has technical discontinuities that matter for an honest reading. The DADS series lacks 1990, so the 1989-to-1991 interval is treated as a two-year gap and its flows are converted to annual equivalents. The panel expanded sharply in 2002 and again in 2006, which can create artificial jumps in one-year flow counts; the corresponding transition pairs are excluded from the baseline analysis. These choices do not make the data useless. They make the conditions under which the estimates are comparable visible.
Four states, three kinds of movement
Let an occupation's employment stock be the number of workers observed in that occupation at a given date. Between two dates, that stock changes through two broad channels.
The first is direct job-to-job movement. A worker can move from manual to routine, routine to abstract, abstract to manual, or between any other pair of employed occupational groups. For a destination occupation, the relevant quantity is the net flow in: moves into it from the other employed groups minus moves out of it to those groups. A positive net flow means that direct occupational switching adds workers to the destination relative to the reverse movements.
The second is movement between employment and the out-of-employment state. Workers can leave a destination occupation for non-employment, or enter it from non-employment. The net effect is entries minus exits. This channel includes re-entry after a period away from observed employment, not only a young person's first job. It is precisely the channel that a simple comparison of occupational stocks cannot see.
The stock identity is therefore straightforward in words:
Change in an occupation's employment stock equals its net direct job-to-job inflow plus its entries from non-employment minus its exits to non-employment.
The study observes direct job-to-job moves and exits in the matched panel. Entries are harder to observe because a worker's status before their first appearance in the panel is not always known. The baseline analysis recovers entries from the identity: once the change in the occupation's stock, its exits, and its direct flows are known, the implied entry flow is the amount needed to make the accounts balance. This is not a guess disguised as an observation. It is a transparent residual construction, and the paper compares it with a fully panel-based alternative in robustness checks. But it remains a reason to avoid describing every estimated entry as a directly recorded spell.
There is a second adjustment when the study moves from employment counts to employment shares. Suppose total employment is growing. An occupation can receive many new workers and still lose share if it receives fewer than the amount needed to maintain its previous proportion of employment. To handle this, the study compares the occupation's net flow with non-employment against a share-neutral benchmark: the flow it would receive if it simply kept its current share while aggregate employment changed. The difference is the excess non-employment flow.
A positive excess means that the occupation is being replenished from non-employment more than its existing share would predict. A negative excess means that the occupation is receiving too few entrants, or losing too many workers to non-employment, relative to that benchmark. The measure is useful because it separates an occupation's own entry and exit dynamics from a general expansion or contraction of employment.
The benchmark also clarifies why a headcount and a share can tell different stories. Imagine that total employment grows and that routine work receives more entrants than it did in the previous year. If total employment grows even faster, routine can still occupy a smaller fraction of all jobs. Conversely, a routine group can lose workers and yet preserve its share if the whole employed population is shrinking at a similar rate. The excess measure asks whether the non-employment channel is doing more or less than the occupation's existing position in the employment distribution would require. It is therefore a comparison against a counterfactual of share stability, not a claim that every worker would otherwise have taken a routine job.
This distinction is especially important for interpreting the entry margin. A fall in routine's excess flow can arise because fewer people enter routine, because more routine workers leave for non-employment, or because both happen at once. The accounting frame does not settle that question until the two components are separated. It does, however, prevent an analyst from treating the residual movement as an undifferentiated “technology effect.” The same logic applies to manual and abstract work: a rising share may reflect direct upgrading, a larger inflow from non-employment, or the fact that another occupational group is shrinking faster.
The annual transition window creates another interpretive discipline. If a worker changes jobs twice within one year, the panel may record only the beginning and end states. A direct routine-to-abstract observation could therefore contain an unobserved intermediate job, while a year spent outside observed employment may conceal shorter spells or several employment changes. This does not invalidate the broad flow categories, but it means they describe annual net movement rather than a complete biography of each transition. The study's conclusions concern the balance of these annual channels and their changes across periods.
The diagram's arrows should not be read as a causal model of why a worker chooses a job. They are an accounting map. It tells us how the observed stock change is assembled, not whether technology, education, employer strategy, or worker preferences caused each transition.
What changed around 1994?
The French employment shares show a clear change of regime. From 1984 to 2021, the within-employment routine share fell from 69.0% to 53.0%. The abstract share rose from 13.7% to 26.4%, while the manual share rose from 17.2% to 20.7%. But the averages before and after 1994 look different.
Before 1994, the routine share was nearly flat: it moved from 69.0% in 1984 to 67.4% in 1993. Manual employment declined slightly, from 17.2% to 16.5%, while abstract employment increased from 13.7% to 16.3%. After 1994, routine employment contracted more quickly, manual employment turned upward, and abstract employment accelerated. The average annual routine-share change was about −0.09 percentage points before the break and −0.52 points afterward. Manual changed from −0.20 to +0.12, and abstract from +0.29 to +0.40.
Those figures establish the timing of the polarization pattern, but the denominator matters. The figures above are shares within employment. The study also constructs augmented shares that include the employment rate and the non-employed share. On that broader scale, the routine share falls by roughly six percentage points rather than the sixteen-point decline implied by the within-employed comparison. The difference is not a contradiction. It shows that changes in overall employment and flows through non-employment are part of the story.
The 2008–09 crisis makes the routine decline steeper, especially in the augmented measure, but the study's more durable interpretation is not that the financial crisis created the entire pattern. The non-employment buffer around routine work appears to disappear around 1994 and remains absent afterward; the crisis amplifies the decline. Separating the long-run change from the cyclical amplification is another reason to work with flows rather than attribute every visible movement to the most dramatic downturn in the chart.
Why the accounting frame changes the question
Without the flow decomposition, several explanations remain open. Routine workers might be leaving for abstract jobs, which would reduce routine employment but represent upward movement for the workers involved. They might be moving into manual jobs, which would suggest a different kind of adjustment. They might be exiting employment, or they might be replaced by a new set of entrants who are directed elsewhere. The stock data alone cannot rank these possibilities.
The accounting frame also prevents a common mistake about entry. If routine employment falls, it is tempting to say that routine jobs are disappearing and that the workers who held them must be the people who lose. But the stock can shrink because the occupation stops receiving the same flow of new or returning workers. In that case, a worker already in routine work may continue to upgrade at the previous rate even as the routine share declines. The eventual composition of employment changes because the pipeline into the occupation narrows.
That possibility is not a semantic refinement. It changes the population for whom adjustment policy is designed. A displaced incumbent may need income protection, recognition of prior experience, and a bridge into a new employer. A potential entrant or re-entrant may need a first foothold, updated training, or a route around a credential that used to be less important. The two groups can overlap, but they are not interchangeable, and the data must tell us which flow is doing the work before recommendations become specific.
There are limits to what this framework can establish. The occupation groups are useful analytical categories, not universal natural kinds. Annual observations can bundle several moves within the same year. The out-of-employment state is broader than unemployment in the strict statistical sense. Entries are residuals in the baseline accounting, and the education field is directly observed only from 2002 before being projected backward for some workers. These constraints do not erase the French result; they define its scope.
The central gain is that the next question is now precise. The routine share changes after 1994, and the accounting separates direct occupational switching from the non-employment channel. Which of those channels actually changes? Does routine lose workers because exits rise, because upgrading accelerates, or because fewer people enter and re-enter it? And does the answer remain the same inside the broad routine category? Those are questions about the mechanism behind the shrinking middle, not merely about its appearance.