The anatomy of job polarisation: Evidence from four decades of French worker flows

Employment polarisation – the decline of middle-skill routine occupations relative to both ends of the skill distribution – is usually documented from repeated cross-sections. However, the time evolution of cross-sections is driven by workers’ flows across occupational categories and in and out of the labour market. Drawing on an administrative panel that follows French workers over 1984–2021, this column shows that polarisation emerged through a collapse in entry into routine jobs from non-employment, not through displacement of incumbent routine workers. The findings redirect policy attention from protecting existing jobs towards the entry margin.

Employment polarisation is one of the most consequential labour market transformations of the past four decades: middle-skill, routine-intensive occupations have contracted while high-skill abstract and, to a lesser extent, low-skill manual jobs have expanded (Autor et al. 2003, Goos et al. 2014). Virtually all the evidence of this comes from repeated cross-sections. Cross-sections record changes in occupational shares, but they are silent about the worker flows behind them, and radically different reallocation processes are consistent with the same aggregate pattern. For instance, the decline of routine employment may reflect the downgrading of incumbent routine workers into manual jobs and non-employment, as in the displacement narrative of the literature (Autor and Dorn, 2013). However, it may also reflect a change in the way workers enter employment: a sorting away from routine work for new workers (Cerina et al. 2021) while incumbents keep moving up the occupational ladder. While observationally equivalent in repeated cross-section analysis, the two reallocation processes involve different workers, imply different distributional consequences, and call for different policies.

Distinguishing between them, however, requires following individual workers over time. A pioneering line of research did this for the US (Cortes 2016, Cortes et al. 2017, 2020), but the available panels combine small samples followed at low frequency with survey-reported occupations, or match workers only over short windows, thus limiting the scope of the analysis. In a recent paper (Cerina et al. 2026), we exploit the French DADS-EDP administrative panel, which follows individual workers across all salaried employment with employer-reported occupations from 1984 to 2021. We classify occupations into manual, routine, and abstract, track workers’ transitions among these categories and in and out of non-employment, and decompose the year-to-year change in each occupation’s employment share into two channels: net job-to-job flows, and net flows between that occupation and non-employment.

1994: When non-employment stopped feeding routine jobs

The French labour market changed regime in the mid-1990s. Before 1994, aggregate occupational shares show no clear trend. After 1994, the routine share declines steadily while abstract and manual employment expands (the employment polarisation pattern). Behind this break lies a striking reversal in a single channel (Figure 1). Before 1994, non-employment is a net feeder of routine work: net flows from non-employment add roughly 0.25 percentage points per year to the routine employment share. After 1994, the channel flips sign and drains roughly 0.1 percentage points per year. Entry rates from non-employment into routine and exit rates from routine into non-employment both fall, but the entry-side collapse dominates: it is the drying-up of hiring into routine jobs from non-employment, not a wave of exits, that flips the net flow. This switch accounts for about four-fifths of the acceleration in the routine decline after 1994.

Within routine occupations, the reversal is concentrated in the cognitive component, represented by clerical, administrative, and supervisory occupations. Net non-employment flows into routine-cognitive work add roughly 0.80 percentage points per year to its share before 1994 and essentially nothing (0.04) afterwards, while the corresponding flows for routine-manual work move in the opposite direction. The timing coincides with the sharp fall in information and communication technology prices, with routine-cognitive tasks being precisely those most directly displaceable by ICT capital.

Figure 1 Net annual worker flows in the pre-polarization (1984–1993) and polarisation (1994–2021) regimes, in percentage points of employment

Note: Blue arrows mark upgrading and entry from non-employment; red arrows mark downgrading or outflows to non-employment.
Source: Cerina et al. (2026).

The displacement that never happened

If routine-biased technological change had pushed incumbent routine workers down the occupational ladder, the panel would show surging flows from routine into manual work and into non-employment. It shows neither. Job-to-job upgrading flows from routine to abstract are large, persistent, and remarkably stable across the whole panel, at 0.52–0.54 percentage points per year, before and after 1994 alike. Flows from routine to manual are small throughout. And the growth of manual employment is driven entirely by the non-employment channel: manual occupations absorb a disproportionate share of entrants from non-employment (+0.25 percentage points per year), while net job-to-job flows actually subtract from the manual share (−0.13 percentage points per year). In the French microdata, the displacement channel is absent.

A non-college story

Splitting the sample by education sharpens the picture (Figure 2). Among non-college workers, the familiar U-shape of polarisation emerges: over 1984–2021 their routine share falls by about 13 percentage points, while manual (+7.7) and abstract (+5.7) both rise. College-educated workers, by contrast, exhibit essentially flat occupational shares during the polarisation era: their shares move by no more than 1.5 percentage points each between 1994 and 2021. This confirms the common view that polarisation is essentially a non-college phenomenon, but for a very different reason than displacement of uneducated middle-skilled workers: non-college workers account for 76% of all net routine-to-abstract flows, in both periods, and within-non-college upgrading alone accounts for over two-fifths of the total rise in the abstract share. Thus, uneducated incumbent workers contribute largely to polarization by moving up the occupational ladder, not down as typically emphasised.

Worker-level characteristics reinforce the picture. Holding age, gender, tenure, and year fixed, the college advantage in moving from routine to abstract is the largest effect we estimate, but it roughly halves after 1994. Since non-college workers vastly outnumber college ones, they still account for most net routine-to-abstract upgrading. Occupational tenure strongly stabilises workers in their current occupation, and women face an additional disadvantage on every upgrading margin. Taken with the stable aggregate upgrading flows, this suggests an optimistic amendment to the routinisation hypothesis, at least for incumbents: routine jobs – in particular routine-cognitive ones – have continued to function as a stepping stone towards abstract work (García-Peñalosa et al. 2023), including for non-college workers. What collapsed is not the ladder, but access to its first rung: the flow of workers entering routine employment from non-employment.

Figure 2 Occupational shares within employment, by education group, 1984–2021

Source: Cerina et al. (2026).

Policy should look at the entry margin

Our entry-margin verdict for France echoes the US evidence (Cortes et al. 2020), suggesting that the mechanism is not driven by the specificities of that economy. Together with our evidence on job-to-job flows, a policy implication follows directly: if polarization operates through who gets hired out of non-employment rather than through who is fired, then policies aimed primarily at protecting incumbent workers from displacement would miss the action and thus, would be ineffective and even counter-productive. Expanding and retooling education and training, supporting re-entry from non-employment, and easing mobility towards expanding occupations are likely to be more effective responses. The welfare burden of this structural change falls on those seeking to enter or re-enter employment, who are disproportionately the less educated, rather than on incumbent routine workers, who continue to upgrade at stable rates. A full welfare assessment requires a quantitative framework in which workers choose occupations over the life cycle and face entry and re-entry margins explicitly – the natural next step this evidence is designed to discipline, and the subject of our ongoing work.

Source : VOXeu