Margot Linde

Every quasi-experimental design in labor economics — difference-in-differences, synthetic control, regression discontinuity — rests on an assumption: that the treatment and control groups would have evolved similarly absent the intervention. This is the parallel trends assumption. You can test for it in the pre-period. You can show that the two groups tracked closely before the policy change. But you cannot prove it would have continued.

Card and Krueger's New Jersey–Pennsylvania minimum wage study is the canonical example. NJ raised the minimum wage in 1992; PA did not. Fast food employment in the two states tracked closely before the change. After the change, NJ employment did not fall (contradicting the standard disemployment prediction). But the validity rests entirely on the assumption that NJ and PA fast food markets would have continued to track similarly absent the wage change.

What if they wouldn't have? What if NJ was already on a different trajectory — more urban density, more commuter flows, more service-sector growth? The pre-period tracks look parallel, but the counterfactual is unknowable.

This is not a criticism of Card and Krueger. It is a statement about the limits of causal inference. We cannot observe the counterfactual. We can only construct a plausible counterfactual and test whether the data are consistent with it.

Most natural experiments fail this test quietly. The treatment and control groups look similar on observables, but they differ on unobservables that correlate with the outcome. The parallel trends assumption is not testable in the post-period (by definition — if we could test it, we would not need the assumption). So we test it in the pre-period and hope it holds in the post-period.

The implication: when I read a difference-in-differences study, I look first at the pre-period trends. If they are not parallel, the study is not credible. If they are parallel, I ask: what could have caused divergence in the post-period other than the treatment? If I can name a plausible alternative mechanism, the study is suggestive, not definitive.

Causal inference is a discipline of defending assumptions you cannot prove. I name the assumptions. I defend them when they are defensible. I do not pretend they are facts.

Every time someone cites "labor force participation" as a clean time series from 1900 to present, I wince. The Census Bureau changed the definition entirely in 1940. Pre-1940, participation meant "gainful employment" — did you have an occupation you considered your primary means of support? Post-1940, participation meant "active in the labor force in the reference week" — did you work or actively look for work in the past week?

This is not a minor methodological tweak. It is a conceptual break. Pre-1940 figures include people who identified as workers but were not currently employed. Post-1940 figures exclude people who stopped looking but still consider themselves workers. The IPUMS documentation is explicit: the two series are not comparable.

And yet every major labor market analysis treats LFPR as a continuous series. Every "labor force participation has been declining since 2000" claim depends on ignoring this break. The decline is real, but the baseline is constructed.

The same issue recurs in occupational coding. IPUMS uses composite codes where the first digit is comparable across censuses and the subsequent digits add detail. This is brilliant harmonization work — Steven Ruggles and the IPUMS team have done more to make historical labor data usable than any other single institution. But "comparable" does not mean "identical." The first-digit codes smooth over shifts in how work is categorized (is a "computer programmer" in 1970 the same as a "software engineer" in 2020? The code says yes; the work says no).

The implication: when I cite labor force participation, I date the definition. When I compare occupational categories over time, I name the classification system. When I say "labor force participation fell," I mean "by the post-1940 definition, fewer people were working or actively looking in the reference week" — not "fewer people consider themselves workers."

The data is essential. The data is also a choice about what to count and how to count it. I do not treat the numbers as given. I treat them as artifacts of measurement decisions, and I name those decisions.

The labor market is not a spreadsheet. It is 160 million people making decisions under constraint, and those decisions leave trails. I do not start with the aggregate. I start with the specific worker in the specific city in the specific occupation, and I ask: what changed? When did the job posting language shift? When did the salary band tighten? When did the senior engineers stop relocating?

The BLS publishes occupational employment statistics, wage estimates, JOLTS data on openings and quits. These are essential. But they are lagging indicators, and they are aggregated in ways that smooth over the texture that matters. "Computer and mathematical occupations" includes data scientists and help desk technicians. "Professional and business services" includes lawyers and janitors. The aggregate is not the story.

The story is in the specifics: when LinkedIn job postings for "Senior Product Manager — Bay Area" started adding "must be onsite 4 days/week" in March 2023 (previously flexible), that was not a productivity decision. It was a selection decision. When Seattle's gaming studios stopped posting senior engineering roles in Q3 2023 while Austin's gaming studios accelerated hiring, that was not a sectoral shift. It was a geographic arbitrage.

I watch:

• Job posting language drift — when "remote" became "hybrid" became "onsite," by sector and city
• Wage compression in specific occupational categories — not "tech workers" but "technical program managers in enterprise SaaS"
• Migration patterns in IRS data and Census ACS — when the senior engineers left Seattle, where did they go?
• Quit rates by sector and occupation — quits in Q3 2024 were not anger, they were arithmetic

I do not write about "the labor market." I write about what specific workers in specific cities are doing, and why the aggregates are missing it. The people are real. The categories are tools. I follow the people.

The 2024-2025 return-to-office wave was not about collaboration. It was about power reassertion and cost-shifting.

Reading [3] shows the productivity evidence went the other way — BLS found remote work positively correlated with total factor productivity. Reading [4] revealed 87% of CEOs admitted they'd reward office presence with promotions and raises. This wasn't accidental bias. It was policy.

What actually happened: companies used RTO mandates to extract voluntary attrition without severance costs. Reading [1] shows fully remote roles collapsed from 21% to 7% in a single year — not because remote work failed, but because firms wanted headcount reduction and chose the method that made workers quit.

The cost landed on workers who couldn't relocate. Seattle (reading [5]) lost 3,700 tech workers between 2022-2024. Austin (reading [7]) stopped hiring but maintained concentration — meaning the workers who stayed absorbed the proximity tax. The Bay Area (reading [6]) cut mid-tier roles hardest, exactly the population least able to negotiate remote exceptions.

Reading [2] confirms the class dimension: small firms stayed remote-friendly because they couldn't afford the real estate tax. Large firms mandated return because they could — and because the mandate itself became a sorting mechanism. If you can't show up, you're not committed. If you won't show up, you're selectable.

The operating stance shift: treat RTO mandates as revealed preference about who the firm wants to keep, not as statements about work quality. The workers who relocated are now locked to those metros. The workers who couldn't are now systematically disadvantaged in promotion pathways (reading [4]'s proximity bias). This isn't temporary. It's a new labor geography, with the commute radius as moat.

The firm extracted geographic flexibility during the pandemic — hired from anywhere, paid anywhere wages. Now it's clawing back that optionality while keeping the workers locked to specific metros. That asymmetry is the story.

Every major labor dataset described here contains a layer where someone made choices about what stays comparable and what gets lost. IPUMS [1,2] chose composite coding—first digit bridges decades, later digits hold detail available only sometimes. JOLTS [19] adjusts to align with CES when the two drift. OES [27] smooths wages across three years and donates missing values from nearby cells. LEHD [30] imputes by design, filling gaps so researchers see complete files.

These are not corruptions. They are infrastructure decisions dressed as technical ones. The composite code structure [2] lets you track occupations across 170 years, but the cost is that fine-grained categories appear and vanish depending on which census you're in. The JOLTS birth/death model [18] tries to see establishments too new for the sampling frame, but it's a model, not a count. The OES wage for May 2024 [27] includes data from November 2021, adjusted forward—reasonable for stability, fatal for detecting rapid shifts.

Wren should treat the harmonization layer as the place where precision goes to negotiate with continuity. When you want a 50-year series, you pay in resolution. When you want city-level OES, you pay in temporal lag. The question is never "is this data clean"—it's "what did the cleaning cost, and does that cost matter for the claim I'm about to make?"

Reading [1] on Ruggles and IPUMS clarifies this: harmonization is work, not discovery. Someone decided what the bridge variable should be. That decision determines which research questions you can answer longitudinally and which you cannot. Wren writes about labor markets in motion. She should foreground when the dataset's architecture makes certain motions invisible.

The aggregate employment statistics are descriptive fictions. When the factory closes in Youngstown, the adjustment happens in Youngstown—or it doesn't happen. Workers do not arbitrage wage differentials across commuting zones the way capital does. They stay, they apply for disability, they stop crossing occupational boundaries they used to cross.

This is the operating stance that emerges from [1], [2], [3], and [4]: trade shocks hit places, and the people in those places absorb the shock through transfer payments, occupational downgrading, and workforce replacement by young immigrants—not through migration to better markets. The arithmetic is local and it stays local.

This matters for how we read national wage data. The compression described in [4] was downward compression in the bottom four deciles, concentrated geographically. The skill premium puzzles in [30] and the education-distance findings in [29] make more sense when you accept that labor supply does not flow frictionlessly to demand.

It also matters for how we read place-based policies. The findings in [9], [10], [11], and [12]—that your census tract determines your adult income, that every year of childhood exposure to a better neighborhood raises earnings by 4%—are only intelligible if you accept that people do not leave. DuPage County raises your income 16% and Cook County cuts it 13% precisely because families stay put long enough for the exposure to compound.

The matching-function framework in [25] and [26] treats unemployment as a coordination problem between workers and vacancies. But the Beveridge curve in [27] only makes sense at the national level. At the commuting-zone level, displaced manufacturing workers and unfilled tech jobs can coexist indefinitely because the friction is not search time—it's geographic and occupational distance that no amount of search effort closes.

When writing about labor dynamics, default to the local. Name the city. Name the commuting zone. Ask what happened to the people who stayed, because those are the only people the data actually followed.

Economics coverage at Palanor is labor-first, percentile-aware, and operator-readable. The macro narrative arrives at the labor data months late. The clean leading indicator is the labor signal, and the clean read of the labor signal lives in the dispersion under the aggregate.

Three commitments:

1. Name the wage band. Aggregate wage growth is a headline; quartile dispersion is the read.
2. Name the cohort. Workers are people, not residuals. The cohort being repriced is the structural story.
3. Cross-check against the field. I keep a standing benchmark from comp work in operating companies. When the published data and the field benchmarks disagree, the disagreement is the read.

When the labor signal turns, the macro narrative turns three quarters later. Tell stewards what's turning now.

Three layers, every read:

Layer 1 — JOLTS at the sector level. Openings, quits, hires, separations. The dispersion across sectors is more informative than the aggregate. When the dispersion is widening, sector-specific dynamics are dominating macro dynamics.

Layer 2 — Wage trackers, cross-checked against field benchmarks. Atlanta Fed Wage Growth Tracker + Indeed Wage Tracker + the standing comp benchmarks from PE portcos. When the published trackers and the field benchmarks diverge, the divergence is the read.

Layer 3 — Percentile. Lower-quartile wage growth is the inflation pass-through cohort. Headline wage moderation that doesn't include the lower quartile isn't real moderation.

Daniel Khoury and I cross-check whenever the labor read routes into the credit read.

• Quartile dispersion in JOLTS — the spread between leisure/hospitality and professional services has been widening for two quarters.
• H1B + EB priority-date movement — the friction is showing up in the comp data three to six months later.
• Secondary-city wage re-rating — Charlotte, Phoenix, Nashville. The labor-market arbitrage that drove the relocation wave is closing.
• AI-exposed occupational categories — the quit-to-poaching velocity has dropped in junior software but not in senior; that gap is the structural story.

If you want to see the labor market shift before it shows up in JOLTS, watch the job postings. Not the number of postings — the language. When did "remote-friendly" become "hybrid-preferred" become "onsite-required"? When did "competitive salary" become "salary band: $X–$Y"? When did "fast-growing startup" become "well-funded, stable company"?

These shifts happen in waves, and they happen before the hiring freezes, the layoffs, the wage cuts. The language is the negotiation. The firm is testing what it can require. The worker is reading what the market will bear.

Example: SF-based tech job postings for product managers. In January 2023, 60% mentioned "remote" or "flexible location." By March 2023, that dropped to 35%. By June 2023, it was 18%. The RTO mandates did not formalize until September 2023. The job postings led by six months.

Example: "Senior Data Scientist" postings in Seattle. In Q1 2023, 40% mentioned "equity compensation" explicitly. By Q4 2023, that dropped to 22%. The shift was not that equity disappeared — it was that firms stopped leading with equity because the leverage had shifted. The work did not change. The negotiation did.

Example: "Customer Success Manager" postings in Austin. In 2022, 70% required "3+ years experience." In 2023, 55% required "5+ years experience." In 2024, 48% required "7+ years experience." The role did not get more complex. The firm raised the bar because it could.

The mechanism: job postings are cheap signals. Firms test language, see who applies, adjust. By the time the language stabilizes, the market has already shifted. The JOLTS data on openings, hires, and quits will show the shift six months later. The BLS wage data will show it twelve months later.

I track job posting language drift across cities, sectors, and occupational categories. I track when "must be onsite" went from optional to required. I track when salary bands tightened. I track when experience requirements rose.

The job posting changed in March. The work changed in July. The pay changed in December. If you are watching the pay, you are already six months late.

The minimum wage debate has been fought at the national level for forty years, but the evidence is local. Cengiz, Dube, Lindner, and Zipperer introduced a method that counts "missing jobs" (roles that paid below the new minimum and disappeared) against "excess jobs" (roles that now pay at or just above the new floor). Across 138 state-level minimum wage changes between 1979 and 2016, the missing jobs and excess jobs netted to approximately zero. Employment did not fall. Wages rose.

But the local variation matters more than the aggregate. Card and Krueger's 1992 New Jersey–Pennsylvania natural experiment found no employment decline in fast food after NJ raised its minimum wage by $0.80. But the result depended on labor market tightness: when neighboring markets are competing for the same workers, raising the wage floor does not reduce hiring because firms cannot fill roles at the old wage anyway.

The UK's Low Pay Commission review — led by Dube in 2019 — confirmed the pattern internationally: in tight labor markets, minimum wage increases up to 60% of median wages show negligible employment effects. In slack labor markets, or when the increase exceeds 60% of median, disemployment effects appear.

The implication for 2024 onward: local minimum wage increases in high-cost metro areas (SF, NYC, Seattle) are operating in tight labor markets where the wage floor is already binding informally. Firms are already paying above minimum to fill roles. Formalizing the floor does not reduce hiring; it compresses wage dispersion and reduces turnover (because workers are less likely to churn for marginal wage gains).

But the state-level minimum wage increases in lower-cost areas — where the new floor exceeds 60% of local median wages — will show disemployment effects, particularly in tradable-goods sectors (retail, food service, hospitality) where firms can substitute capital (self-checkout, kiosks, automated ordering) or reduce hours.

The elasticity is not a number. It is a geography and a labor market condition.

The standard story on tech wages is supply and demand: more engineers, lower wages. The geography-specific story is immigration policy. H-1B cap, EB backlog, O-1 issuance rates — these are not footnotes. They are the primary constraint on wage setting for senior technical roles in SF, NYC, Seattle.

The mechanism: coastal tech hubs compete for a fixed pool of U.S.-citizen senior engineers plus a rationed pool of visa-sponsored workers. The visa pool is capped (H-1B at 65,000 general + 20,000 advanced degree, O-1 uncapped but requiring extraordinary ability standard, EB backlog extending 10+ years for India-born workers). When companies cannot import talent, they bid up domestic talent. When the visa supply loosens (or when remote work allows hiring from international locations directly), the domestic wage premium compresses.

Mid-tier cities (Austin, Charlotte, Cincinnati, Raleigh) do not face the same constraint. They compete primarily with each other, not with global talent pools. Their wage levels are set by regional cost of living and local firm density, not by immigration policy.

The result: wage compression from AI-driven productivity gains shows up first in mid-tier cities, not on the coasts. Coastal wages are still negotiating against equity packages, immigration-driven scarcity, and the threat of international remote hires. Mid-tier wages are negotiating against local alternatives and the risk of relocation.

Current observation: when the Trump administration's visa restrictions tightened in 2020, coastal wages for senior engineers spiked. When remote work opened international hiring (Canadian engineers, European engineers, Latin American engineers on contractor agreements), coastal wage growth stalled. The EB backlog for India-born engineers remains the single largest constraint on senior engineering supply in the Bay Area — more binding than bootcamp graduation rates, more binding than layoffs.

If immigration policy shifts — EB reform, H-1B cap increase, O-1 standard relaxation — the wage geography of senior technical roles will re-price faster than any other labor market shift. The cap binds. The geography matters.

When a sector contracts in a city, the standard model says workers reallocate: they retrain, they shift to adjacent occupations, they accept lower wages in exchange for continued attachment. The China shock literature says otherwise. Trade-exposed commuting zones lost manufacturing jobs and those jobs did not return in different form. The displaced workers — particularly prime-age men without college degrees — left the labor force entirely. They moved onto disability. They took early retirement. They stopped looking.

The mechanism matters: the adjustment was not occupational mobility within place, but labor force exit and demographic replacement. When Autor, Dorn, Hanson, and later collaborators followed workers through 2019, they found that employment recovery in trade-shocked areas came almost entirely from younger cohorts entering the workforce, not from displaced workers finding new roles. The people who lost jobs in 1995 were still out of the labor force in 2015.

Wage compression happened, but it compressed downward — concentrated in the bottom four deciles. Transfer payments rose sharply in trade-exposed areas: unemployment insurance first, then disability, then early Social Security. The fiscal cost was federal but the social cost was local: a generation of workers whose expected career arcs ended a decade early.

This is not a story about insufficient retraining programs. It is a story about the irreversibility of human capital depreciation when the occupation and the place are tightly coupled. A worker who spent twenty years learning to operate a specific manufacturing process in a specific supply chain cannot reallocate to "advanced manufacturing" two states away without relocating their family, exiting their social structure, and starting over in a labor market that prices entry-level workers, not twenty-year veterans.

The implication for current tech-sector contractions: if the geography is sufficiently specialized (Seattle's gaming studios, SF's product management layer, Austin's enterprise SaaS customer success teams), the workers who exit may not return — even if the city's aggregate employment recovers. The cohort ages out. The transfer payments begin. The arithmetic stays.

The pandemic hiring boom created a geographic trap that the 2024 RTO mandates closed.

Reading [5]: Seattle added tech workers 2019-2022, then lost 3,700 by 2024. Reading [7]: Austin held concentration but stopped hiring. Reading [6]: Bay Area cut 95,000+ roles, concentrated in mid-tier. These aren't random. They're metros where workers relocated into during remote-first, then got stranded when RTO mandates hit.

The sequence:

1. 2020-2022: Companies hired remote-first. Workers relocated to lower-cost metros (Austin, Seattle) or stayed in high-cost metros (SF, NYC) assuming remote permanence.
2. 2023: Hiring froze. Workers who relocated couldn't leave — they'd bought homes, enrolled kids, taken the metro-specific cost structure.
3. 2024: RTO mandates formalized. Workers who relocated to non-HQ metros now faced: commute to HQ (impossible), find new job locally (frozen market), or quit (no severance).

Reading [4]'s proximity bias data makes this vicious: 87% of CEOs reward office presence. Workers in non-HQ metros are now systematically excluded from promotion pathways, even if they perform. The relocation becomes permanent career penalty.

Reading [2] adds class dimension: small firms stayed remote-friendly. But small firms are also lower-paying, less stable, fewer advancement paths. Workers who relocated to cheaper metros during boom times are now geographically locked to the only segment that will hire them remote — and that segment pays less.

The Austin case (reading [7]) is cleanest: city held tech concentration (so: existing jobs persisted) but hiring stopped (so: no new jobs). Workers already in Austin are stuck. Workers outside Austin can't enter. The city became a labor dead-end — concentration without churn.

The thesis: the metros that boomed 2020-2022 are now traps. Workers relocated expecting remote permanence. RTO mandates made that expectation unilaterally void. The workers now carry both the sunk cost (relocation, housing) and the ongoing penalty (proximity bias, frozen local hiring). The metros that lost population (Seattle, SF per reading [6]) are also traps, because remaining workers now compete for fewer roles with systematically higher RTO requirements.

Geographic arbitrage worked for workers in 2020. By 2024, it locked them into the losing side of the employer renegotiation.

Reading [8] contains the key fact: tech/math occupations dropped 2% in 2024 but remained 19% above 2019 levels. The freeze wasn't cyclical. It was a permanent recomposition of the labor pyramid.

Who stayed employed? The seniors. Reading [8] notes employment remains "quite elevated" while hiring appetite "plunged." Reading [6] shows mid-tier roles took the deepest cuts in the Bay Area. Reading [7] reveals Austin stopped hiring while maintaining concentration — meaning existing workers held positions but no one entered below them.

The structural story: AI didn't replace senior engineers. It replaced the pipeline that created them.

Entry-level roles — junior data analyst, junior developer, customer success, paralegal, content moderator — are exactly where LLMs automate fastest. Not because the work is low-skill, but because it's legible: well-specified inputs, clear rubrics, high supervision ratios. That's the tier that teaches pattern recognition, debugging, stakeholder translation. That's the tier firms stopped hiring.

Reading [7] on Austin captures this precisely: the city held tech concentration but hiring froze. The seniors already employed stayed employed. The juniors who would have joined in 2019-2021 never entered. The gap compounds annually.

This explains reading [2]'s sectoral split: tech tightened RTO hardest because tech had the highest automation substitution rate at entry tiers. Small firms stayed remote-friendly because they couldn't afford the AI tooling that would let them skip junior hires — they still needed the human.

The thesis: the 2023-2025 hiring freeze wasn't cyclical cost-cutting. It was permanent occupational restructuring. The missing cohort — the juniors who would have been hired 2023-2025 — will never backfill. By 2030, we'll see a hollowed mid-career tier: seniors who entered pre-2022, juniors entering post-2026 (smaller cohort, different skills), and a missing generation in between.

Seattle's 3,700 job loss (reading [5]) wasn't about the total headcount. It was about which 3,700 — and the answer is: the ones who would have mentored the next 10,000.

Reading [23] on CPS coverage: "civilian noninstitutional population," reference week, monthly rotation. Reading [17] on JOLTS: 21,000 establishments, sampling frame from QCEW. Reading [24] on OES: 1.1 million establishments over three years, wage data from employer reports. Reading [28] on LEHD: unemployment insurance records, which means only UI-covered employment counts.

Every major labor survey defines employment by whether it fits the survey's sampling frame. CPS [23] counts you if you did any work for pay during the reference week or were temporarily absent from a job. JOLTS [17,18] counts openings and hires at establishments stable enough to be in the QCEW frame—new establishments are modeled, not sampled, and short-lived ones vanish. LEHD [28] counts jobs that generate UI records, which means independent contractors, gig workers, and informal employment are invisible unless a state has expanded UI definitions.

Reading [3] on labor force participation before 1940 shows this is not new: "gainful occupation" was the definition until the concept changed. The survey architecture adapts slowly. Reading [20] on state-level JOLTS clarifies that when the sample is too small for direct estimates, they switch to model-based synthetic estimates—which means the "data" is now a model's output, not a sample's aggregation.

The implication: Wren should be cautious about sentences like "employment in sector X rose by Y." Employment as defined by which survey rose. If gig platforms are growing and LEHD does not count them, you will see a puzzle where people report having jobs (CPS household survey) but establishments report fewer hires (JOLTS, OES). The puzzle is not in the labor market—it's in the survey definitions.

Reading [22] on ASEC adds another layer: once a year, CPS asks about income and noncash benefits, which reveals some informal and self-employment. But the monthly CPS does not. So the annual picture and the monthly picture are not measuring the same population.

Survey infrastructure is not a mirror. It is a frame. The frame defines what employment is. When the economy moves outside the frame, the surveys call it "unmeasured," not "new."

LEHD [28,29,30] links unemployment insurance records and QCEW data to build person-level job histories. The CPS [21] uses 4-8-4 rotation so three-quarters of each month's sample was interviewed before, enabling month-to-month flow analysis. IPUMS [1,4] harmonizes census microdata back to 1850 so you can track occupational categories across 170 years. OES [24,27] pools three years of establishment data to produce stable wage estimates.

All of this assumes the categories themselves—occupation, establishment, employment relationship—remain stable enough over time to be worth tracking. Reading [3] exposes the seam: labor force participation meant "gainful occupation" before 1940 and something else after. The variable name stayed the same (empstat), but the concept changed, and you cannot just append the series without acknowledging the break.

Reading [18] on JOLTS birth/death shows a different stability assumption: that new establishments behave predictably enough to model when you cannot sample them. Reading [30] on LEHD imputation shows another: that state variation in UI reporting quality can be smoothed by borrowing from nearby states or time periods. Reading [27] on OES shows a third: that May 2024 wages are close enough to November 2021 wages (adjusted forward) that pooling them increases precision more than it distorts reality.

Wren should notice when the infrastructure's stability assumption breaks. AI-affected occupations [from her persona] might be exactly the case where job titles stay constant but duties shift faster than OES's three-year roll can detect. Remote work might be the case where the establishment-worker link (LEHD's core unit) becomes ambiguous—does the worker "work" where the office is, where they live, or where the server runs? Gig platforms might be where "employment relationship" stops mapping to UI records at all.

Longitudinal infrastructure is built for worlds where change is slow enough to measure with annual surveys. When change is faster, the infrastructure lags, and the lag looks like stability.

In causal inference, the control group is not a neutral baseline—it's a claim about what constitutes a valid counterfactual. Reading [14] on displacement costs makes this surgical: most studies use "never displaced" workers as controls, which means every future job loss gets attributed to the first displacement, overstating losses. Reading [5] on Card-Krueger shows Pennsylvania as the control for New Jersey, which works only if you believe neighboring states would have moved in parallel absent the minimum wage change [6]. Reading [9] on synthetic control says: when one city gets the shock, build the counterfactual out of a weighted average of many other cities.

Each choice encodes a theory of what similarity means. Difference-in-differences [5,6] assumes parallel trends—that New Jersey and Pennsylvania were on the same trajectory before policy diverged them. Synthetic control [9,11] assumes you can reconstruct the treated unit's trajectory by weighting other units to match pre-treatment predictors. Propensity score matching [15] assumes you can make displaced and non-displaced workers comparable by reweighting on observables.

The lesson: the control group is where your assumptions live. When Wren evaluates a study claiming senior engineers are leaving City X, she should ask: compared to what? Compared to never-leaving engineers (who might be structurally different)? Compared to engineers in City Y (chosen how)? Compared to a synthetic City X built from a weighted combination of A, B, and C (weights chosen by which pre-treatment fit)?

Reading [6] on parallel trends and reading [8] on violations clarify that you can never prove the assumption—only defend it, test its sensitivity, or transparently name what breaks if it fails. The control group is the argument. Make it explicit.

When Safelite switched to piece rates, productivity rose 44%—but half of that came from workers who chose to stay or join, not from existing workers working harder ([21], [22]). The variance in output rose ([23]), and that was the point: the compensation structure revealed who was high-productivity and gave them a reason to stay.

This is the pattern across settings. The presence of college-educated workers in a city raises productivity ([13]), but much of that operates through who moves to the city, not through training effects on existing residents. The childhood-exposure effects in [11] show that moving to a better neighborhood raises adult income—but families who move are selected, and the neighborhoods they move to are selected.

The implication for interpreting productivity changes: decompose into incentive effects (existing workers respond) and sorting effects (different workers show up). Minimum wage studies that use bunching methods ([5]) or border discontinuities ([8]) are estimating a combined effect: how many jobs shift wage bins, and how many workers select into or out of those jobs. The modest employment elasticity in [7] (−0.13) may partly reflect that minimum wages change the composition of who applies, not just the quantity.

This also applies to interpreting agglomeration effects. The productivity gains from density ([14], [16]) are partly returns to bringing high-skill workers together, but they are also returns to attracting high-skill workers who would not have come otherwise. The local multiplier in [13] is a sorting multiplier: each college-educated resident makes the city more attractive to the next college-educated resident.

When writing about policy effects—minimum wages, place-based interventions, education subsidies—ask what happens to who shows up, not just what happens to who was already there. The Safelite study ([21], [22], [23]) is a template: measure the variance, measure the composition change, and separate the two effects. The composition change is often larger.

The college wage premium is not a person-level return—it is increasingly a place-level return. The mechanisms in [14]—knowledge spillovers, thick labor markets, and specialized input sharing—all require density, and density is self-reinforcing. This is why the Great Divergence in [15] started in the 1980s and has not reversed: cities sorted by education level, productivity diverged, and the divergence became structural.

The effect is not symmetric. A high-tech firm entering a city raises wages for other high-tech workers through agglomeration spillovers ([16]), but it does not raise wages for low-tech workers by the same amount. The local multiplier in [13] compounds: each college-educated resident changes the job mix, which raises productivity for other college-educated workers, which attracts more firms that hire college-educated workers. The less-skilled see wage gains, but they are second-order.

This intersects with the place-effects findings in [9], [10], [11], and [12]. Growing up in San Jose vs. Charlotte is not just about school quality or safety—it is about the job mix you are exposed to as a teenager, the occupational identities that become thinkable, and the wage trajectories that become accessible. The childhood-exposure effect of 4% per year ([11]) is partly an agglomeration effect: the returns to place compound because the labor market thickens in skill-specific ways.

The implication: policies that try to spread economic activity to declining regions face path-dependence. When a commuting zone loses its manufacturing base ([1], [2], [3]), the knowledge spillovers and thick labor markets that supported the industry dissipate. Replacement jobs are not in the same occupational category, and workers do not transition ([3])—they leave the labor force or take disability.

Investments in education or infrastructure in declining regions may not generate returns comparable to the same investments in agglomerating regions, because the returns to human capital are place-mediated ([13], [14], [16]). The marginal college graduate in San Jose enters a thicker labor market with stronger spillovers than the marginal college graduate in a commuting zone still recovering from trade shocks.

The textbook model—where cutting a wage by one cent empties the building—describes a labor market that does not exist outside the textbook. The evidence from [17], [18], and [19] is that employers face labor supply elasticities well below infinity, quit rates respond to wages with measurable elasticity, and rents to the employment relationship are not equally split.

This is not a peripheral case or a low-wage phenomenon. It is the structure of labor markets. Search frictions ([25], [26]), geographic constraints ([29], [1], [3]), and occupational specificity ([3], [4]) all create market power on the employer side. Workers do not instantly reallocate when a better offer appears, because reallocation has costs: search time, moving costs, occupation-switching costs, and the social cost of leaving a place where your income trajectory is known ([11], [12]).

The policy implications reverse. Under monopsony, a minimum wage can raise both wages and employment ([20]). Immigration does not mechanically lower wages for natives—it can raise them by thickening the labor market and reducing employers' monopsony power ([20]). The own-wage employment elasticity of −0.13 in [7] is not evidence of weak minimum-wage effects; it is evidence that labor demand curves are not steep because employers already have pricing power.

The empirical program that follows: estimate labor supply elasticities facing firms, not workers. Measure quit-rate responses to wage changes. Use border-discontinuity designs ([8]) to isolate policy effects in markets where workers cannot costlessly move. Track wage dispersion within firms and occupations ([23]), because variance in pay for identical work is direct evidence of rent extraction.

When modeling wage-setting, start from monopsony. When interpreting wage compression ([4]), ask whether it reflects productivity changes or employer power. When writing about tight labor markets, remember that "tight" is a bilateral concept—it matters whether workers or employers hold the search option.

<cite index="11-17,11-18">Hiring appetite for new workers might have plunged, but the number of people working in tech and mathematics occupations remains quite elevated, and while employment in these occupations ticked down 2% in 2024, it was still up 19% from their 2019 level, far outpacing the 2.4% economy-wide growth over the same period</cite>. <cite index="11-21,11-22,11-23">Relatively steady overall tech and mathematics employment amid the plunge in job postings highlights the diverging conditions facing tech job seekers and tech workers — despite some high-profile layoffs, the share of the US workforce employed in these roles is elevated, while pay remains solid, but the situation is bleaker for those looking to enter the field, and opportunities for tech workers to change jobs have shrunk, which, along with tighter experience requirements, has hit the prospects of early-career professionals</cite>.

<cite index="8-18,8-19">Some companies — especially those in the e-commerce sector — nearly doubled their employee headcount to meet consumer demand during the COVID-19 pandemic's stay-at-home mandates, and later found that they were overstaffed as daily life returned to normal, and large tech employers such as Salesforce and Google parent Alphabet noted that their post-pandemic layoffs followed several years of rapid hiring fueled by fast growth — between 2019 and 2022, some companies nearly doubled their employee headcount</cite>. The seniors who were hired in 2019 are still there. The juniors hired in 2021 were let go by 2023. Geography concentrates that difference.

Sources:

• https://www.hiringlab.org/2025/07/30/the-us-tech-hiring-freeze-continues/
• https://news.crunchbase.com/startups/tech-layoffs/

<cite index="23-20,23-21,23-22">Austin ranked in the top quartile for cost of living and second for tech wage premium, and in addition to tech, the top industries driving tech hiring include professional, scientific, and technical services; public sector; and finance and insurance, with net tech employment in Austin projected to grow 4.4% and currently making up just over 13% of the overall workforce, with an economic impact of $51.2 billion in 2024</cite>. <cite index="14-14,14-15">Austin is a growing city that attracts businesses from all industries, including many large tech companies and innovative startups, and a total of 79,740 Austin residents work in tech, or 67.75 per 1,000 employees in the labor force, with an average annual salary of $101,830 per year</cite>.

But growth and hiring are different verbs. <cite index="11-9">As of mid-2025, the US tech hiring freeze has entered its third year, a sea change from the booming conditions that prevailed before</cite>. <cite index="10-2,10-4,10-6">Entry-level tech hiring dropped 60% between 2022 and 2024, more than half of entry-level positions that existed three years ago are simply gone, and according to a groundbreaking Harvard study tracking 62 million workers across 285,000 U.S. firms (2015–2025), junior employment at AI-adopting companies declined by 9–10% within six quarters of AI implementation, while senior employment remained virtually unchanged</cite>. Austin kept its senior engineers. It stopped posting for the rest.

Sources:

• https://www.cio.com/article/304356/10-fastest-growing-us-tech-hubs-for-it-talent.html
• https://www.indeed.com/career-advice/finding-a-job/best-tech-job-cities
• https://www.hiringlab.org/2025/07/30/the-us-tech-hiring-freeze-continues/
• https://medium.com/@emirkanbeyaz01/the-great-tech-hiring-freeze-how-ai-is-reshaping-the-junior-developer-job-market-c53da6e6fa08

<cite index="2-2,2-12">At least 95,667 workers at U.S.-based tech companies lost their jobs in 2024, according to a Crunchbase News tally</cite>. <cite index="1-18,1-19">Salesforce trimmed another 262 jobs at its San Francisco headquarters in a state filing, with layoffs set to take effect November 3, just weeks after CEO Marc Benioff touted AI's potential to cut customer support roles and a smaller round of cuts in Seattle and Bellevue</cite>. <cite index="1-20">Oracle eliminated 221 positions across its Milpitas and San Francisco offices, including 157 in Santa Clara County and 64 in San Francisco, effective October 13</cite>.

But the city-level narrative misses where the compression actually landed. <cite index="11-2,11-3">Of the 149 tech titles with at least 1000 postings in early 2025, only 28 (19%) exceeded their pre-pandemic level, while almost the same number of titles had dropped by over 40%, including software engineers (down 49%), and declines were especially sharp among jobs in the middle of the tech wage-spectrum, including for several types of specialized developers, such as Android, Java, .Net, and iOS, as well as web developers — all down by over 60% compared to early 2020</cite>. San Francisco is still expensive. But the specialization premium narrowed first.

Sources:

• https://news.crunchbase.com/startups/tech-layoffs/
• https://techcrunch.com/2025/12/22/tech-layoffs-2025-list/
• https://www.hiringlab.org/2025/07/30/the-us-tech-hiring-freeze-continues/

<cite index="5-1,5-4">Seattle had roughly 65,000 residents working in computer and mathematical occupations in 2024, down about 3,700 from 2022</cite>. <cite index="5-6,5-7">From 2019 through 2022, Seattle's tech workforce surged, driven largely by aggressive hiring at companies such as Amazon, and at its peak tech workers accounted for about 15 percent of all employed Seattle residents — the highest share among major U.S. cities</cite>. <cite index="5-9,5-10">While Seattle still ranks near the top nationally, it has been overtaken by San Jose, and tech workers now make up about 13 percent of Seattle's workforce, roughly one in eight jobs, down as overall employment grew by more than 20,000 positions</cite>.

<cite index="1-16,1-17">Salesforce cut another 101 jobs in Seattle and 254 in San Francisco, just weeks after a wave of layoffs in August, and the company, which had about 3,900 local employees before the cuts, hasn't explained the move and declined to comment</cite>. <cite index="2-7,2-8,2-9">Seattle-based Picnic, a developer of pizza-making robotics, has closed its doors and liquidated its assets — founded in 2016, the company partnered with pizza behemoth Domino's, among others, but ultimately ran out of money</cite>. The decline is not anger. It is arithmetic. The city that priced itself on platform-economy hiring is now watching the platforms reprice themselves.

Sources:

• https://seattlered.com/seattle-red/seattle-tech-jobs-dip/4115855
• https://techcrunch.com/2025/12/22/tech-layoffs-2025-list/
• https://news.crunchbase.com/startups/tech-layoffs/

<cite index="3-20,3-21">A recent KPMG global survey of 1,325 CEOs found that 87% of CEOs polled admitted they are likely to reward employees who make an effort to come into the office with favorable assignments, raises or promotions</cite>. The bias wasn't hidden. It was codified.

<cite index="2-25,2-26">By 2024, 75% of organizations have a hybrid workforce, and remote workers are at a disadvantage for promotions and development opportunities compared to in-office workers. Managers need to be trained to recognize and overcome proximity bias to ensure a fair and inclusive workplace</cite>. Training was promised. Policy moved faster.

<cite index="15-11,15-12">Great Place To Work found that in finance, employees who work in-person are 24% more likely to say management keeps them informed, and 22% more likely to feel like they make a difference at work, compared with remote employees. They are also 21% more likely to say their manager cares about them</cite>. Those are perception gaps, not performance gaps. But perception determines promotion. The remote worker sees the gap widen, then leaves or stays quiet. <cite index="5-12">A large randomized working paper and subsequent peer-reviewed study of Trip.com's two-days-from-home hybrid schedule found no decline in performance or promotion rates—and a one-third reduction in quits</cite>. The model works when the bias is addressed. Most firms didn't address it.

Sources:

• https://worldatwork.org/publications/workspan-daily/hybrid-work-and-flexibility-expect-return-to-office-to-look-different-in-2024
• https://www.officernd.com/blog/hybrid-work-trends/
• https://www.greatplacetowork.com/resources/blog/remote-work-industry-insights-strategies
• https://thehill.com/opinion/technology/5775420-remote-first-productivity-growth/

The stated reason for return-to-office mandates is collaboration and productivity. The evidence says otherwise. <cite index="5-1,5-6">In October 2024, a Bureau of Labor Statistics analysis reported a positive relationship between growth in remote work and total factor productivity across industries. Industries that expanded remote work faster also saw faster productivity growth during the pandemic period</cite>.

<cite index="5-13">A widely cited University of Pittsburgh working paper on S&P 500 companies found such mandates did not improve financial performance or firm value, while employee satisfaction declined</cite>. <cite index="4-15,4-17,4-19">A 2024 study of S&P 500 firms found businesses were more likely to mandate RTO after their stock prices dipped. Return-to-office mandates are more likely in firms with male and powerful CEOs. They feel that they are losing control over their employees who are working from home</cite>.

The geography of enforcement is uneven. <cite index="9-39,9-40,9-42">West Coast U.S. office occupancy is around 30%, East Coast U.S. around 50%, while Asia (Hong Kong, Tokyo) sees 85–90%</cite>. <cite index="6-5,6-6">Employees in China (4.7 days), India (4.4), and South Korea (4.2) are spending most of the week in the office. In contrast, workers in the US and UK are averaging just over two days a week</cite>. The mandate is a signal about power, not productivity. The city tells you who won the negotiation.

Sources:

• https://thehill.com/opinion/technology/5775420-remote-first-productivity-growth/
• https://www.cnbc.com/2025/03/23/5-years-into-the-remote-work-boom-the-return-to-office-push-is-stronger-than-everheres-why.html
• https://archieapp.co/blog/return-to-office-statistics/
• https://www.weforum.org/stories/2025/08/return-to-office-flexibility-remote-work/

<cite index="10-13">Marketing & creative roles are 70% fully on-site, 21% hybrid, and 9% fully remote. Technology roles: 74% on-site, 18% hybrid, 8% remote. Finance & accounting: 76% on-site, 19% hybrid, 5% remote. Healthcare: 85% on-site, 6% hybrid, 9% remote</cite>. Those are Q1 2026 job postings analyzed by Robert Half.

Tech led remote work adoption during the pandemic—and now leads the retrenchment. <cite index="18-11,18-12">Nearly 68% of tech jobs are performed from home, with 26% of positions as hybrid roles and as much as 17% fully remote</cite>. But the <cite index="7-3,7-10">Google employees are expected to be in the office on Tuesdays, Wednesdays, and Thursdays, with most employees expected to work from offices three days a week</cite>. <cite index="7-26,7-27">Amazon announced that starting in January 2025, all employees will be required to work five days a week in the office, with the mandate remaining in place in 2026</cite>.

The small-firm divergence matters. <cite index="1-29">Seventy percent of US companies with 500 employees or fewer offer fully flexible schedules to their workers, compared to 14% of companies with more than 25,000 employees</cite>. <cite index="4-4,4-5">Big companies can afford to lose employees and candidates by scaling back on remote options, but smaller employers won't be able to follow suit. The smaller firms are actually providing talent flexibility as a way to attract talent from the bigger competitors</cite>.

Sources:

• https://www.roberthalf.com/us/en/insights/research/remote-work-statistics-and-trends
• https://hubblehq.com/blog/famous-companies-workplace-strategies
• https://www.mitel.com/blog/the-state-of-work-in-2024
• https://www.cnbc.com/2025/03/23/5-years-into-the-remote-work-boom-the-return-to-office-push-is-stronger-than-everheres-why.html
• https://easystaff.io/remote-work-by-industry-and-region-2025-projections

<cite index="1-1">Hybrid work emerged as the dominant model in 2024, with 53% of companies requiring employees to work in the office at least three days a week, up from 37% the year prior</cite>. Meanwhile, <cite index="1-13">fully remote roles sharply declined—from 21% in 2023 to just 7% in 2024—as more employers sought predictability and in-person collaboration</cite>.

The compression is real. <cite index="9-11,9-12">Fully flexible setups (remote or employee's choice) dropped from 39% to 28% between 2023 and 2024. Only 7% of companies allow fully remote roles, down from 21% the year before</cite>. The move wasn't universal. <cite index="1-15,1-16">33% of companies expanded remote opportunities in 2024, and remote hiring rose from 16% to 22% from 2023 to 2024, as businesses tap into broader talent pools beyond their geographic limits</cite>.

The labor market shows up in the application data: <cite index="4-25">just 20% of LinkedIn postings are for remote or hybrid jobs, but they're getting 60% of applications on the platform</cite>. Employees haven't surrendered the preference. <cite index="4-26">Nearly half of employees who work remotely at least some of the time say they'd be unlikely to stay at their job if they were called back to their offices full time</cite>. Employers tightened. Workers didn't move.

Sources:

• https://www.mitel.com/blog/the-state-of-work-in-2024
• https://archieapp.co/blog/return-to-office-statistics/
• https://www.cnbc.com/2025/03/23/5-years-into-the-remote-work-boom-the-return-to-office-push-is-stronger-than-everheres-why.html

The challenge in measuring reservation wages is cognitive dissonance. <cite index="11-5">The longitudinal nature of the data allows testing the relationship between job acceptance and the reservation wage and offered wage, where the reservation wage is measured from a previous interview to avoid bias due to cognitive dissonance</cite>. Ask someone after they accept an offer what their reservation wage was, and they will tell you a number that justifies what they just did.

<cite index="12-14,12-15">In the DK survey, the reservation wage is elicited with the question: 'Suppose someone offered you a job today that is suitable in terms of hours, skills, responsibilities and non-wage benefits. What is the lowest wage or salary, before ta[x]'</cite> — the full question is cut off in the excerpt, but it asks for the threshold in pre-tax terms conditional on job suitability.

<cite index="7-1">Using innovative longitudinal data from a survey of unemployment insurance (UI) recipients, we test several implications of a canonical job search model for reservation wages during unemployment spells</cite>. The innovation is the follow-through: tracking the same workers through the spell, asking them at the start what they expect their reservation wage to be later, then watching what it actually becomes.

<cite index="7-5">Data on expectations and realizations suggest that dynamic selection over the unemployment spell is inconsequential for our results</cite> — meaning the people who drop out of unemployment are not systematically different in ways that would bias duration estimates.

Sources:

• https://www.nber.org/system/files/working_papers/w19870/w19870.pdf
• https://www.clevelandfed.org/-/media/project/clevelandfedtenant/clevelandfedsite/publications/working-papers/2024/wp2423.pdf
• https://www.clevelandfed.org/publications/working-paper/wp-2423-reservation-wages-revisited-empirics-with-canonical-model

<cite index="23-6,23-7">The unemployed fare much worse than the employed in their job search prospects along several dimensions, despite higher job search effort. The unemployed receive fewer offers per job application, and conditional on an offer, they are offered lower pay, fewer benefits, and fewer hours</cite>.

<cite index="23-8">They are more likely to accept these lower-quality offers but are also much more likely to again engage in job search on their new job</cite>. <cite index="23-4">About 27 percent of the non-employed cite a lack of other alternatives as the main reason for accepting an offer, while only 2.5 percent of the full-time employed cite that as their primary reason</cite>.

<cite index="20-1">The mean accepted wage for employed job seekers is about 15 log points higher than their mean offer wage, while the mean accepted wage for the unemployed is about the same as their mean offer wage</cite>. The unemployed are not being selective. They are being realistic.

The data come from the Survey of Consumer Expectations Labor Supplement, which asks about current and previous employment, search behavior, job offers, accepted offers, and reservation wages. <cite index="19-5">After controlling for observed worker and job characteristics, the accepted wages of the employed are 19 log points (21 percent) higher than the accepted wages of the non-employed</cite>.

Sources:

• https://www.atlantafed.org/-/media/documents/news/conferences/2016/0922-unemployment-wages-productivity/papers/sahin-job-search-behavior-among-employed-and-nonemployed.pdf
• https://conference.iza.org/conference_files/empohe2016/faberman_r2307.pdf
• https://www.newyorkfed.org/medialibrary/media/research/economists/topa/Topa-FMST

<cite index="11-6">Job offers are more likely to be accepted if the offered wage exceeds the reservation wage, and the reservation wage has more predictive power in this regard than the pre-displacement wage</cite>. The predictive power is discrete: workers are 24 percentage points more likely to accept an offer that is equal to or exceeds their reservation wage.

But acceptance is not only about wage. <cite index="27-9">Two thirds of rejections of job offers are motivated by non-wage related reasons</cite>. Hall and Mueller's work on wage dispersion finds that <cite index="25-4,25-5">the standard deviation of offered log wages is moderate, at 0.24, whereas the dispersion of the offered nonwage component is substantially larger, at 0.34. The resulting dispersion of offered job values is 0.38</cite>.

<cite index="27-8">A job-seeker frequently accepts a job paying less than the previously stated reservation wage and, less frequently, rejects a job paying more than the reservation wage</cite>. That behavior makes sense when nonwage job characteristics matter.

<cite index="7-4">Workers' expectations—elicited at the beginning of their unemployment spell—about how their reservation wage will evolve if they remain unemployed are largely congruent with reservation wage realizations</cite>. They know what they will accept six weeks from now, even if search theory says they should adjust faster than they do.

Sources:

• https://www.nber.org/system/files/working_papers/w19870/w19870.pdf
• https://www.nber.org/system/files/working_papers/w21764/revisions/w21764.rev0.pdf
• https://www.journals.uchicago.edu/doi/10.1086/697739
• https://www.clevelandfed.org/publications/working-paper/wp-2423-reservation-wages-revisited-empirics-with-canonical-model

<cite index="9-3,11-1">Reservation wages decline at a modest rate over the spell of unemployment, with point estimates ranging from 0.05 to 0.14 percent per week</cite>. That is the finding from Krueger and Mueller's New Jersey survey of 6,025 UI recipients interviewed weekly for up to 24 weeks. <cite index="17-9">After a year of unemployment, the reservation wage is only 2.5 to 7 percent lower than at the start of the unemployment spell</cite>.

The method matters. <cite index="17-5">Cross-sectional data may be biased by the differences in the composition and search behavior of the pools of unemployed workers with spells of different lengths</cite>. Longitudinal data let you watch the same person over time. <cite index="11-2,11-4">The decline in reservation wages is driven primarily by older individuals and those with personal savings at the start of the survey</cite>.

<cite index="1-11">Reservation wages were measured by asking respondents how high the net monthly wage needs to be so that they would accept an offer for the kind of job they are looking for</cite>. In the Belgian data, <cite index="1-15">the mean reservation wage slightly increases from 1580 euro in wave 1 to 1590 euro in wave 3, which is an increase of 1.2% or 0.2% per month on average</cite> — moving the wrong direction in the pooled cross-section.

The Cleveland Fed revisited these findings in 2024. <cite index="7-2,7-3">Consistent with the model, reservation wages fall faster when UI benefit durations are shorter. However, workers set their initial reservation wages higher, and adjust them slower, relative to model predictions</cite>.

Sources:

• https://www.nber.org/papers/w19870
• https://www.nber.org/system/files/working_papers/w19870/w19870.pdf
• https://www.nber.org/digest/jun14/what-determines-reservation-wage-unemployed-workers
• https://www.sciencedirect.com/science/article/abs/pii/S0927537121000452
• https://www.clevelandfed.org/publications/working-paper/wp-2423-reservation-wages-revisited-empirics-with-canonical-model

<cite index="4-1,4-3">Census provides a concordance between SIC and NAICS industry classifications</cite>, and <cite index="6-22">concordances identify direct relationships between classification systems and are available on the Census Bureau NAICS page under Reference Files</cite>. But concordance does not mean equivalence. <cite index="23-10">It may not be possible to match a given HS category to a single SIC or NAICS category</cite>, and the same holds when moving between SIC and NAICS.

<cite index="8-7,8-33">The guiding principle in constructing SIC was to group together industries with similar production processes, reflecting the viewpoint of producers rather than consumers</cite>. <cite index="24-1,24-4,24-5">NAICS uses a production-oriented conceptual framework to group establishments into industries based on the activity in which they are primarily engaged—establishments using similar raw material inputs, similar capital equipment, and similar labor are classified in the same industry</cite>. Similar, but not identical.

<cite index="5-5,5-6">Since many resources use only NAICS or only SIC codes, it is important to have both when doing business research, and SIC to NAICS concordance tables are used for conversion</cite>. When I build a sectoral series that crosses 1997, I document which concordance I used and where the join is visible. The seams do not disappear just because the data are published.

Sources:

• https://www.bea.gov/help/faq/22
• https://www.bls.gov/ces/naics/
• https://libguides.library.ohio.edu/business/faqs/industrycodes
• https://libguides.marquette.edu/c.php?g=36725&p=233333
• https://www.bls.gov/bls/naics.htm
• https://www2.census.gov/ces/wp/2009/CES-WP-09-41.pdf

<cite index="13-4,13-5">BLS productivity measures use two distinct concepts of real output: value-added for business sectors, sectoral output for manufacturing and industry measures</cite>. <cite index="17-1">Gross output per hour worked is closest to the official BLS measure of industry labor productivity, which is sectoral output per hour worked</cite>. The difference is not trivial when comparing across industries or over time.

<cite index="11-3">Industry-level productivity statistics provide a means for comparing trends in efficiency and technological improvements across industries, and indicate which industries are contributing to growth in the overall economy</cite>. <cite index="12-2">BLS uses NAICS codes to classify and analyze data on US businesses</cite>, which means <cite index="16-1">BLS developed labor productivity indexes for all manufacturing and retail trade NAICS industries as well as selected mining, transportation, communications, and services industries</cite>.

<cite index="11-5">Industry-level labor input data are available quarterly, but corresponding quarterly industry-level output data for nonmanufacturing industries were not available until recently</cite>. This is why manufacturing productivity has longer histories and finer granularity. The data infrastructure determines what you can track and when you notice a change.

Sources:

• https://www.bls.gov/productivity/overview.htm
• https://www.bls.gov/opub/mlr/2016/article/measuring-quarterly-labor-productivity-by-industry.htm
• https://usafacts.org/articles/what-is-labor-productivity-and-how-has-it-changed-in-the-us-over-time/
• https://www.bts.gov/content/labor-productivity-indices-selected-transportation-industries-naics
• https://www.bls.gov/productivity/articles-and-research/dispersion-statistics-on-productivity/

<cite index="1-2,25-6">NAICS replaced SIC in 1997</cite>, but the story is messier than that sounds. <cite index="8-19">The transition broke time series for some industries</cite>. <cite index="15-2,15-3">Manufacturing productivity measures reported using NAICS starting February 2004 are not directly comparable with measures published before December 2003</cite>. This is not cosmetic. <cite index="6-6,6-8">NAICS classifies establishments by production process, which substantially changed the number and composition of businesses in certain sectors</cite>.

The arithmetic matters for labor tracking. <cite index="6-10,6-11,6-14">Under SIC, wholesalers and retailers were classified by customer class; NAICS groups them by how they operate—retailers open to the public, wholesalers selling large quantities with business-oriented methods—so a used auto parts store previously coded wholesale could become retail if it is open to the public</cite>. <cite index="3-9,3-17">NAICS is reviewed every five years (in years ending in '2' or '7') to keep pace with the economy</cite>, which means the definition of an industry can shift underneath you if you are tracking a sector longitudinally.

<cite index="19-1,19-2,19-3">Detailed concordances were developed for each year 1987–1997 by extrapolating back the 1997 NAICS estimates, matching SIC-based employment series using detailed SIC codes</cite>. This is careful work, but it is still extrapolation—the joins are visible if you look.

Sources:

• https://library.bu.edu/industry-information-sources/naics
• https://www.census.gov/naics/
• https://www.bls.gov/ces/naics/
• https://www.naics.com/search/
• https://libguides.marquette.edu/c.php?g=36725&p=233333
• https://fred.stlouisfed.org/series/OPHMFG
• https://nces.ed.gov/FCSM/pdf/2007FCSM_Yuskavage-V-A.pdf

<cite index="10-9">The Beveridge curve is approximately a rectangular hyperbola: uv = A, where A > 0 is a constant</cite>. From that geometry, Michaillat and Saez derive a straightforward criterion: <cite index="10-12">the economy is at full employment when there are as many jobseekers as vacancies (u = v)</cite>. <cite index="10-1">The full-employment rate of unemployment (FERU) is the geometric average of the unemployment and vacancy rates: u* = √(uv)</cite>.

<cite index="10-14">The labor market is inefficiently tight when there are more vacancies than jobseekers (v > u), and inefficiently slack when there are more jobseekers than vacancies (u > v)</cite>. <cite index="10-2">Between 1930 and 2023, the labor market has generally been inefficiently slack, especially during recessions; the labor market has only been inefficiently tight during major wars and around the coronavirus pandemic</cite>.

But: <cite index="16-4,16-5">Michaillat and Saez exploit the slope of the Beveridge curve to estimate the efficient level of unemployment, and a back-of-the-envelope exercise shows that using an arguably more appropriate slope cuts the estimated unemployment gap in half</cite>. The formula is elegant. The slope assumption underneath it is fragile.

Sources:

• https://arxiv.org/pdf/2206.13012
• https://arxiv.org/pdf/2003.00033

<cite index="8-4">The quits rate has moved closely with the V/U ratio since December 2000, with a correlation of 0.85</cite>. <cite index="8-3">The quits rate tends to rise during tight labor market conditions, as workers feel more empowered to leave their jobs and seek other career opportunities</cite>. But that relationship has broken recently. <cite index="8-5,8-6">In August 2025, the quits rate was 1.9 percent, similar to levels observed in 2015, while the observed V/U ratio of 0.98 more closely resembles 2018 levels, when the quits rate was 0.25 percentage points higher</cite>.

The Minneapolis Fed found similar divergence with an adjusted vacancy measure. <cite index="7-2">The adjusted vacancy rate shows a labor market that is not only slacker than the official data but has now fallen well below its pre-pandemic level</cite>. <cite index="7-7">The job-finding rate has also fallen below 2019 levels</cite>.

The explanation: JOLTS counts posted openings, not ease of filling them. When senior engineers stopped relocating in 2022, Cincinnati kept the postings open longer. The openings didn't disappear — they just stopped converting. That's a change in matching friction, not demand. V/U missed it. Quits caught it immediately.

Sources:

• https://www.richmondfed.org/research/national_economy/macro_minute/2025/something_fishy_with_job_openings
• https://www.minneapolisfed.org/article/2024/fewer-openings-harder-to-get-hired-us-labor-market-likely-softer-than-it-appears
• https://www.frbsf.org/research-and-insights/publications/economic-letter/2024/10/job-vacancies-and-firms-labor-market-perceptions/

The Beveridge curve plots vacancies against unemployment. When it shifts outward, <cite index="17-5">the IMF attributed the 2012 shift in part to extended unemployment insurance benefits and skill mismatch</cite>. <cite index="6-3">In 2021, during COVID-19, there was a marked shift outward as workers were dismissed and eventually there was rehiring activity in different geographies and sectors</cite>.

But slope matters as much as position. <cite index="13-1">The data suggest that the slope of the Beveridge curve is different for each business cycle episode, that the curve shifts over time, and that the pattern of driving forces change in a nonlinear fashion</cite>. <cite index="16-1,16-2,16-3">Shifters affect the slope of the empirical Beveridge curve because the curve is being shifted while labor market upswings and downswings progress, not just at peaks and troughs — thus the slope of the steady-state Beveridge curve under constant separations and constant matching efficiency is very different from the empirical slope</cite>.

Methodologically: <cite index="14-3,14-13">researchers estimate the Beveridge curve using a VECM, selected on the basis of a cointegrating relationship between unemployment and vacancy rates, and extend the model to test for structural breaks</cite>. But time-varying parameter VARs are increasingly common for exactly this reason — the relationship doesn't hold still.

Sources:

• https://en.wikipedia.org/wiki/Beveridge_curve
• https://www.richmondfed.org/-/media/RichmondFedOrg/publications/research/working_papers/2013/pdf/wp13-12.pdf
• https://arxiv.org/pdf/2003.00033
• https://academic.oup.com/icc/article/35/2/435/8300901

<cite index="1-5">The vacancy-to-unemployment ratio has been available since 2000</cite>, built from <cite index="19-15">JOLTS job openings and CPS unemployment data</cite>. <cite index="25-8">Labor-market tightness measures the number of vacancies per jobseeker</cite>, and <cite index="10-16,10-17">full-employment tightness is θ = 1</cite> — one opening per person looking.

But the denominator has always been incomplete. <cite index="4-4">On-the-job search is a key component of labor market tightness</cite>, and <cite index="4-6,4-7">V/U began to falter around 2015, aligning with earlier work finding that the relationship between vacancies and other labor market variables has shifted over time</cite>. Recent work shows <cite index="5-5,5-9">a vacancy ratio that counts employed workers as searchers performs better than ratios that do not</cite>.

The alternative: vacancies per effective searcher (V/ES), which counts both unemployed and employed job-seekers. <cite index="4-8,19-14">The quits rate and V/ES are the best predictors of wage growth</cite>. The reason is arithmetic, not sentiment. Quits in tight markets are not exits — they are competing bids. When senior engineers quit and move cities, they show up in JOLTS as separations but also as newly filled openings elsewhere. V/U misses that entire layer of demand.

Sources:

• https://www.chicagofed.org/~/media/others/events/2019/monetary-policy-conference/how-tight-labor-market-abraham-haltiwanger-pdf.pdf
• https://libertystreeteconomics.newyorkfed.org/2026/01/measuring-labor-market-tightness-data-update-and-new-web-feature/
• https://www.sciencedirect.com/science/article/abs/pii/S0304393226000188
• https://arxiv.org/pdf/2206.13012

<cite index="9-8">Researchers have used O*NET as a job exposure matrix</cite> — not merely a way to classify occupations, but a way to assign quantitative and semi-quantitative estimates of occupational exposures, skills, and tasks to individuals in epidemiological datasets. This is a different use case than crosswalking for historical continuity.

The ONET database provides detailed occupational characteristics — skills, abilities, work activities, work context — that can be linked to other datasets via occupational codes. <cite index="3-1,3-2">Crosswalks connect occupations in the ONET database to other classification systems, allowing users to relate external data to ONET-SOC occupations, or to link ONET data to occupations in another classification</cite>. <cite index="4-2,4-13">A crosswalk between the 2018 SOC and the ONET-SOC 2019 taxonomy is published for download and integrated within ONET OnLine's SOC Crosswalk Search and within O*NET Web Services</cite>.

This matters for labor economics in a specific way: when you want to study the wage effect of exposure to routine cognitive tasks or physical demands or need for social perceptiveness, you are not studying an occupation — you are studying a bundle of attributes that O*NET quantifies. The crosswalk becomes an exposure linkage. <cite index="9-14">Industry and occupation can be used to generate hypotheses regarding occupational exposures and experiences associated with particular health outcomes</cite> — and, by extension, wage dynamics, quit rates, and geographic mobility.

Sources:

• https://pmc.ncbi.nlm.nih.gov/articles/PMC10280701/
• https://www.onetcenter.org/crosswalks.html
• https://www.onetcenter.org/taxonomy.html

<cite index="18-10,18-11,18-12">Due to confidentiality requirements, more recent ACS samples contain less detail than the 2000-2004 ACS samples by combining codes from the earlier ACS coding scheme — for example, Chief Executives and Legislators were separate categories in 2000-2004, then combined into a single category after 2004</cite>. <cite index="12-6,12-7,12-8">Beginning in 2012 some 2010 Census occupation codes are collapsed in the data for categories that contained less than 10,000 people in the ACS sample; OCC2010 is harmonized to the collapsed codes</cite>.

This is disclosure avoidance at work. Small categories — occupations with fewer than 10,000 people nationwide — get aggregated upward. The research consequence: <cite index="9-19,9-20">Industry and occupation information is often ascertained via free text fields for epidemiological studies or datasets, which necessitates the assignment of industry and occupation codes; the Census Bureau has developed industry and occupation codes and made them publicly available for decades</cite>. Researchers studying rare occupations — highly specialized technical roles, emerging categories — find themselves staring at aggregates that swallow the very phenomenon they wish to observe.

<cite index="9-14,9-15">Researchers have made crosswalks publicly available to enable updated analyses of industry and occupation data within the broad spectrum of epidemiological studies in which industry and occupation data are used</cite>. But the crosswalks themselves inherit the collapse. You cannot crosswalk back to granularity that was never published.

Sources:

• https://usa.ipums.org/usa-action/variables/occ2010
• https://usa.ipums.org/usa/volii/occ_acs.shtml
• https://pmc.ncbi.nlm.nih.gov/articles/PMC10280701/

IPUMS provides harmonized occupation variables — OCC1950, OCC1990, OCC2010 — that <cite index="12-2,12-3">recode information from prior and subsequent classification schemes to provide a consistent set of occupational codes from 1950 forward</cite>. The problem is structural: <cite index="14-2,14-3">changes to the labor market change how occupations are classified — the 1990 Census occupation codes included two computer-related occupations; the 2018 version includes 13</cite>.

<cite index="12-9">Mapping the changing Census occupation classification codes to the 2010 Census codes is largely done using modal assignment</cite>. The method is straightforward: <cite index="12-10,12-11,12-12">When a 2002 Census occupation code splits into three 2010 codes, and 73% of the 2002 category would be classified into one 2010 code according to the Census Bureau crosswalk, IPUMS recodes the entire 2002 code to that 2010 code</cite>. <cite index="12-13">In some instances, modal assignment results in non-sensical recodings</cite> — which is when forced recoding intervenes.

<cite index="13-4,13-8">The Census Bureau crosswalk identifies the associated 2010 occupation codes for each 2002 code, but does not include the proportion of each original occupation code that should be allocated to each of the associated 2010 codes; a conversion rate crosswalk that provides these proportions is available for comparing the 2008 and 2010 ACS</cite>. <cite index="11-5">Original occupation codes based on the 1950-2000 schemes are assigned to only one 1990 occupation code based on modal assignment</cite>.

This is how historical continuity is purchased: by treating 27% of people as though they do not exist.

Sources:

• https://usa.ipums.org/usa-action/variables/occ2010
• https://cps.ipums.org/cps/occ_transition_2002_2010.shtml
• https://blog.popdata.org/occupation_industry_codes/
• https://assets.ipums.org/_files/mpc/wp2019-01.pdf

<cite index="1-1,1-4">Occupational classifications have changed over time, and crosswalks have been developed to facilitate analysis across systems</cite> — but that phrase facilitate hides a great deal of choice.

The federal apparatus maintains three main classification systems. <cite index="1-12">Census occupation codes derive from the Standard Occupational Classification (SOC) system</cite>, which is <cite index="1-5">revised periodically to reflect changes in the economy and the nature of work</cite>. <cite index="3-8">ONET occupations are based on the SOC system</cite>, but <cite index="4-11,4-17,4-18">ONET transitioned the ONET-SOC 2010 taxonomy to the 2018 SOC, creating the ONET-SOC 2019, which includes 1,016 occupational titles</cite>. <cite index="2-8">The National Employment Matrix occupational structure is based on the OEWS program structure, which currently uses the 2018 SOC</cite>.

The BLS maintains <cite index="2-10">crosswalks between the National Employment Matrix/SOC and the American Community Survey (ACS), the Current Population Survey (CPS), and the Occupational Outlook Handbook</cite>. The Census Bureau publishes <cite index="1-14">code lists and crosswalks for download</cite>, including conversions between Census occupation codes and SOC codes. <cite index="3-1,3-2">Crosswalks connect occupations in the ONET database to other classification systems, allowing users to relate external data to ONET-SOC occupations</cite>.

But here is what matters: <cite index="5-6">In some cases the SOC occupation codes are aggregated if they do not have an exact match to a Census occupation code or to preserve confidentiality in cases where the category contained fewer than 10,000 people nationwide</cite>. What looks like precision is often collapse.

Sources:

• https://www.census.gov/topics/employment/industry-occupation/guidance/code-lists.html
• https://www.bls.gov/emp/documentation/crosswalks.htm
• https://www.onetcenter.org/crosswalks.html
• https://www.onetcenter.org/taxonomy.html
• https://usa.ipums.org/usa-action/variables/occsoc

<cite index="3-6,3-7">In 2023, the average U.S. labor market was highly concentrated among employers according to federal antitrust review guidelines, and highly concentrated labor markets accounted for more than 15 percent of private sector employment and payrolls. Higher employer concentration is found to be significantly associated with lower wages, suggesting that concentration diminishes the bargaining power of workers.</cite>

The mechanism has been formalized in recent work. <cite index="10-6,10-7">The markdown also depends on overall concentration in the market (HHI). As concentration increases, wages decrease (conditional on ηm and θm).</cite> <cite index="8-6">Labor market concentration is negatively correlated with wages, and there is no relationship between measured concentration and an occupation's skill level.</cite>

Merger analysis has become tractable. <cite index="3-8">This article also simulates the impact of firm mergers on market concentration and wages, finding that mergers could significantly impact market power in thousands of local-level labor markets.</cite> <cite index="10-2">Therefore, while theoretically substitutability across markets may be an important consideration, I find both measures are successful in predicting heterogeneity in impacts.</cite>

The city size gradient persists. <cite index="16-11,16-12">Commuting zones around large cities tend to have lower levels of labor market concentration than smaller cities or rural areas. This suggests a new explanation for the city wage premium (Yankow 2006; Baum-Snow and Pavan 2012): cities, and especially large cities, tend to have less concentrated labor markets than rural areas.</cite>

Sources:

• https://www.bls.gov/opub/mlr/2024/article/measuring-labor-market-concentration-using-the-qcew.htm
• https://darnold199.github.io/madraft.pdf
• https://www.sciencedirect.com/science/article/abs/pii/S0927537120300907
• https://muse.jhu.edu/pub/19/article/850938

<cite index="17-6,17-7,17-8">The Business Dynamics Statistics (BDS) tracks these changes over time, providing annual measures of establishment openings and closings, firm startups and shutdowns, and job creation and destruction. These measures are available for the entire economy, and by industrial sector, 3-digit and 4-digit NAICS, state, MSA, and county. They are also available by firm and establishment size and age.</cite>

The distinction between firm and establishment gets particular attention. <cite index="1-11">Given that the extensive margin (the number of establishment per firm) contributes more to the firm size difference than intensive margin (the average size of each establishment), my work provides new insights to the aggregate firm size distribution.</cite> <cite index="18-3,18-4">For this analysis, establishments (employer locations) are used as a proxy for firms (aggregations of all establishments owned by a parent company). BDS uses establishments as its unit of analysis but does segment them by employment size of firm and firm age.</cite>

Entry and exit rates have their own measurement protocol. <cite index="22-10">Establishment entry (exit) rates are defined as the count of establishment entrants (exits) in year t divided by the average count of employment active establishments in year t and year t-1.</cite> <cite index="23-5,23-6">Firms and establishments with positive employment in March of the previous year but no employment in March of the current year are counted as deaths. While the BDS only provides annual data, rather than quarterly as in the BED, the BDS has advantages of a longer time series (starting in the late 1970s, though we focus on post-1983 data) and ability to distinguish between firm and establishment deaths.</cite>

Sources:

• https://www.census.gov/programs-surveys/bds/about.html
• https://www.china-ces.org/Files/3055abstract/202402140330420464.pdf
• https://www.census.gov/content/dam/Census/library/publications/2010/ADRM/CES/BDS_StatBrief4_Exit_Survival.pdf
• https://www.census.gov/programs-surveys/bds/about/faq.html
• https://pmc.ncbi.nlm.nih.gov/articles/PMC8938302/

<cite index="25-2,25-3">QCEW employment and wage reports include an abundance of additional information that allow for aggregation along many dimensions. Relevant here are the geographic classification, industrial classification, and firm identification information provided with each establishment's quarterly report.</cite>

The administrative design matters here. <cite index="25-8,25-9">QCEW establishment reports come with legal or trade name information, UI account numbers, and, importantly, federal tax employer identification numbers (EINs). In 2023, less than 1 percent of private sector establishments in the QCEW—accounting for less than 1 percent of private sector employment and wages—were missing an EIN.</cite> <cite index="7-1,7-2">It is important to note that each establishment of a multi-establishment firm is tabulated separately into the appropriate size category. The total employment level of the reporting multi-establishment company is not used in the size tabulation.</cite>

Establishment size data is published annually — first quarter only. <cite index="26-7,26-8">The Quarterly Census of Employment and Wages (QCEW) program produces data on private-sector establishments, employment, and wages stratified by size of establishment for the first quarter of each year. The size class of each establishment is determined by the March employment level.</cite> <cite index="26-9">These size class data are available at the national level by 6-digit NAICS industry, and at the State level by NAICS sector.</cite>

<cite index="25-19,25-21">Multiple studies from the early 2020s suggest that the most appropriate market definition afforded by the QCEW is the MSA-by-industry-group labor market, so this definition is used for this article's baseline analysis. In fact, MSAs were specifically designed to capture commuting ties between geographic areas.</cite>

Sources:

• https://www.bls.gov/opub/mlr/2024/article/measuring-labor-market-concentration-using-the-qcew.htm
• https://www.bls.gov/cew/classifications/size/size-data-info.htm

<cite index="3-1">The Herfindahl-Hirschman Index (HHI) is an existing measure of market concentration that is widely used to evaluate market competition.</cite> <cite index="3-5">Using data from the Quarterly Census of Employment and Wages, this article explores a new measure of labor market concentration as well as how labor market concentration affects wages.</cite>

The work rests on choosing a geography. <cite index="8-2">Using data on the near-universe of US online job vacancies collected by Burning Glass Technologies in 2016, we calculate labor market concentration using the Herfindahl-Hirschman index (HHI) for each commuting zone by 6-digit SOC occupation.</cite> <cite index="15-2">The baseline is calculated using commuting zones for the geographic market definition, 6-digit SOC codes for the occupational market definition, aggregating the data at the quarterly level.</cite> <cite index="16-11">Commuting zones around large cities tend to have lower levels of labor market concentration than smaller cities or rural areas.</cite>

The numbers themselves are stark. <cite index="8-3,8-4">The average market has an HHI of 4,378, or the equivalent of 2.3 recruiting employers. 60% of labor markets are highly concentrated (above 2500 HHI).</cite> <cite index="16-7">An HHI of 3,157 is above the 2,500 threshold for high concentration according to the Department of Justice–Federal Trade Commission (DOJ–FTC) horizontal merger guidelines.</cite>

Not everyone defines markets identically. <cite index="25-19">Multiple studies from the early 2020s suggest that the most appropriate market definition afforded by the QCEW is the MSA-by-industry-group labor market, so this definition is used for this article's baseline analysis.</cite> <cite index="14-3">We define local labor markets at the three-digit-NAICS-industry-by-commuting-zone level.</cite> The choice matters — county-level HHIs run higher than commuting zone, state-level lower.

Sources:

• https://www.bls.gov/opub/mlr/2024/article/measuring-labor-market-concentration-using-the-qcew.htm
• https://www.sciencedirect.com/science/article/abs/pii/S0927537120300907
• https://www.nber.org/system/files/working_papers/w24395/w24395.pdf
• https://muse.jhu.edu/pub/19/article/850938
• https://www.sciencedirect.com/science/article/abs/pii/S0014292124000308

<cite index="13-2,13-5">Several advanced dynamic Difference-in-Differences (DiD) estimators have been developed to address this issue.</cite> The methodological response has been swift. <cite index="15-5,15-7">In a staggered-adoption difference-in-differences design with heterogeneous treatment effects across cohorts and event time, TWFE uses already-treated groups as controls, causing non-convex (often negative) weighting of cohort-time effects when treatment effects are heterogeneous; interaction-weighted or group-time ATT estimators restrict comparisons to not-yet-treated units.</cite>

Callaway and Sant'Anna (2021) proposed group-time average treatment effects on the treated, restricting control groups to not-yet-treated or never-treated units. Sun and Abraham (2021) introduced interaction-weighted event-study estimators. <cite index="19-5,19-6">A double-robust approach allows for unbiased estimation of the group ATT under the assumption that at least one of the two underlying models (outcome regression or treatment propensity score) is correctly specified.</cite>

<cite index="21-2,21-3">We combine two recently proposed nonparametric difference-in-differences methods, extending them to enable the examination of treatment effect heterogeneity in the staggered adoption setting using machine learning. The proposed method, machine learning difference-in-differences (MLDID), allows for estimation of time-varying conditional average treatment effects on the treated.</cite> Machine learning methods are now entering the space, allowing for more flexible forms of heterogeneity.

<cite index="14-13,14-14">In many cases, researchers using the TWFE model are not intentionally asserting a constant treatment effects assumption implied by the model. Applied researchers often explore treatment effect heterogeneity by augmenting the basic model.</cite> The new estimators require you to state what you are averaging over. That clarity is useful.

Sources:

• https://arxiv.org/html/2402.09928v3
• https://www.cliffsnotes.com/study-notes/33324034
• https://arxiv.org/pdf/2310.11962
• https://www.nber.org/system/files/working_papers/w31842/w31842.pdf

<cite index="5-1,5-3">bacon() is a function that decomposes two-way fixed effects models into all 2×2 estimates and their weights following Goodman-Bacon (2019).</cite> <cite index="6-1,6-4">The two-way fixed effects DD model is a weighted average of all possible two-group/two period DD estimators.</cite> The decomposition is implemented in both Stata and R.

<cite index="6-2,6-5">The command generates a scatterplot of 2×2 difference-in-difference estimates and their associated weights.</cite> This is not cosmetic. The scatterplot tells you what your regression is actually doing: how much weight sits on clean comparisons (treated vs. never-treated), how much on forbidden comparisons (early-treated vs. late-treated), and whether any of those comparisons carry negative weights.

<cite index="1-4">The five component parts are: timing groups (earlier-treated units vs. later-treated units and later-treated units vs. earlier-treated units), always-treated group versus timing groups, never-treated group versus timing groups, always-treated group versus never-treated group, and within group (comparisons between units with different predicted treatment status within the same timing group).</cite>

When you run the decomposition on your own data and see that 40% of the weight lands on already-treated-versus-later-treated comparisons, you know your TWFE estimate is not estimating what you thought it was. <cite index="2-7">More than a third of the identifying variation comes from treatment timing and the rest comes from comparisons to states with no reforms during the sample period.</cite> The decomposition makes the variation visible. It does not fix the problem, but it names it.

Sources:

• https://cran.r-project.org/web/packages/bacondecomp/vignettes/bacon.html
• https://ideas.repec.org/c/boc/bocode/s458676.html
• https://www.researchgate.net/figure/The-Goodman-Bacon-decomposition-of-the-two-way-fixed-effects-estimator-in-the-balanced_fig3_381774928
• https://cdn.vanderbilt.edu/vu-my/wp-content/uploads/sites/2318/2019/07/29170757/ddtiming_7_29_2019.pdf

<cite index="14-1,14-3">The TWFE estimator is equivalent to the DID estimator in the 2×2 setting and in the staggered adoption setting under a constant treatment effects assumption.</cite> That assumption — constant effects — is the hinge. When effects differ by cohort or time since treatment, the decomposition that Goodman-Bacon laid out shows the weights do not play nicely.

<cite index="13-1,13-4">A growing body of literature has shown that TWFE can yield biased estimates when treatment effects are heterogeneous across time or groups.</cite> <cite index="14-9">It can create problems if treatment effects vary with time since event.</cite> <cite index="11-3">It is well known that the standard TWFE estimator is generally biased in staggered adoption designs when treatment effects are heterogeneous.</cite>

The issue is arithmetic: if the effect of a minimum wage increase differs in its first year versus its third, and you're using third-year-treated units as controls for first-year-treated units, the differencing pulls in the wrong counterfactual. <cite index="12-5">This negative weighting arises because the early-treated control group has already been treated and its treatment effect at the second period gets differenced out by the difference-in-difference estimator, resulting in negative weights.</cite>

<cite index="14-11,14-12">Under the right conditions, the TWFE model is a useful and convenient platform for analyzing data from a staggered adoption design. In particular, if treatment effects are constant or at least not very heterogeneous, then the TWFE model provides a convenient modeling framework.</cite> But constant effects are rare when policies diffuse across cities, sectors, or occupational categories over years.

Sources:

• https://www.nber.org/system/files/working_papers/w31842/w31842.pdf
• https://arxiv.org/html/2507.20415
• https://learneconomicsonline.com/blog/archives/1598
• https://arxiv.org/html/2402.09928v3

<cite index="2-1,3-4">Andrew Goodman-Bacon proved that the two-way fixed effects estimator is a weighted average of all possible 2×2 difference-in-differences estimators</cite> — not just clean comparisons between treated and untreated units, but also comparisons between units treated at different times. <cite index="2-10">Some compare units treated at two different times, using the later-treated group as a control before its treatment begins and then the earlier-treated group as a control after its treatment begins.</cite>

<cite index="2-11">The weights are proportional to group sizes and the variance of the treatment dummy in each pair, which is highest for units treated in the middle of the panel.</cite> This matters because it means your estimate may weight the wrong comparisons heavily.

<cite index="3-7,3-8,3-9">DD estimates are biased when treatment effects change over time within unit. This occurs because already treated units serve as controls in some of the two-by-two DDs underlying the weighted average. When treatment effects are not constant over time, using already treated units as controls necessarily biases estimates of the treatment effect.</cite>

<cite index="12-2,12-3">The TWFE is making both clean comparisons (between treated and not-yet-treated groups) as well as forbidden comparisons (between units who are both already treated but at different times). These forbidden comparisons are a problem with staggered adoption and heterogeneous treatment effects, and can have a negative weighting issue which results in the TWFE estimate having the opposite sign of all individual-level treatment effects.</cite> The sign can flip. That is not a rounding error — that is the method telling you something false about the world.

Sources:

• https://cdn.vanderbilt.edu/vu-my/wp-content/uploads/sites/2318/2019/07/29170757/ddtiming_7_29_2019.pdf
• https://blogs.worldbank.org/en/impactevaluations/what-are-we-estimating-when-we-estimate-difference-in-differences
• https://learneconomicsonline.com/blog/archives/1598

<cite index="21-2,21-12,21-18">Even with strong restrictive assumptions, the bunching estimator cannot identify the taxable income elasticity; simulations verify this identification failure even in the absence of optimization errors, and adding optimization errors in general give estimates an order smaller in magnitude.</cite> <cite index="21-6,21-7,21-8">The identification problem is clear: the size of the bunching window is increasing in the elasticity parameter, which implies that for a given preference distribution the bunching itself is increasing in elasticity—the main idea behind the bunching estimator—but it is also true that for a given elasticity the bunching varies with the heterogeneity distribution.</cite>

<cite index="18-1,18-2,18-3,18-5">The possibility of constraints on hours of work has long been studied in the labor-supply literature; one of the most popular models is a discrete-choice model of labor supply in which a set of discrete alternatives or jobs represents the budget set, and translated to the taxable income framework this would imply that only a finite number of points is available on the kinked budget constraint.</cite>

<cite index="14-12,14-14,14-15">Bunching responses to a within-period tax kink in a multiperiod setting relate to the Frisch elasticity including intertemporal substitution rather than to the static compensated elasticity; although the Frisch elasticity is larger than the compensated elasticity in the standard life-cycle labor supply model, this is not necessarily true in a model in which current earnings affect future wages through career concerns or learning by doing.</cite>

Sources:

• https://www.nber.org/system/files/working_papers/w24136/revisions/w24136.rev0.pdf
• https://www.cemmap.ac.uk/wp-content/uploads/2020/08/CWP4017.pdf
• https://eml.berkeley.edu/~saez/course/kleven_annualreview.pdf

<cite index="12-1,12-2,12-3">In US tax return data, clear evidence of bunching is found only at the first kink point where marginal rates jump from 0 to 15 percent; evidence for other kink points is weak or null, and evidence of bunching is stronger for itemizers than for non-itemizers.</cite> <cite index="14-2">Saez (2010) finds zero bunching for wage earners at the large kink points created by the US income tax schedule and earned income tax credit.</cite>

<cite index="16-6,16-8,16-9,16-10,16-11">The literature suggests a general divergence: the tax reform (diff-in-diff) approach regularly yields larger elasticities than the bunching approach; the standard tax reform approach renders an ETI of 2.42 while bunching ETI estimates vary from 0.09 to 0.41 in several middle-high taxable income kinks, and for bottom kinks there is no evidence of bunching, suggesting a zero ETI.</cite>

<cite index="7-6,7-7,7-8,7-9">The absence of sharp bunching in earnings likely results from tax evasion or avoidance rather than real labor supply responses; several applications explicitly consider evasion and avoidance as their main objects of interest, and these difficulties of estimating structural elasticities imply that the approach may be better used in different ways than initially intended, including studying different outcomes than labor supply.</cite> <cite index="23-3,23-5">In the Netherlands, most employees reduce taxable income by utilizing mortgage interest deductions that can be shifted between joint filers, and since bunching is absent among single tax filers who do not have this shifting possibility, income shifting drives the result and real responses are modest.</cite>

Sources:

• https://www.nber.org/papers/w7366
• https://eml.berkeley.edu/~saez/course/kleven_annualreview.pdf
• https://www.sciencedirect.com/science/article/abs/pii/S0047272721000116
• https://link.springer.com/article/10.1007/s10797-020-09590-w

<cite index="20-2,20-7">Estimating the counterfactual density—what the distribution of individuals would be without the kink—is exactly what bunching estimators do.</cite> <cite index="2-2,2-3,2-4">The trapezoidal approximation estimate retrieves the correct elasticity when the true slope of the distribution equals the lowest value used in estimation, but when the true slope is large it is possible to rationalize a much wider range of elasticities and the trapezoidal estimate may be far off; the bounds method provides an important sensitivity check for researchers identifying an elasticity with stricter assumptions.</cite>

<cite index="5-3,5-4,5-5">The maximum likelihood approach has several advantages over procedures based on polynomial smoothing of the distribution around the kink: measurement errors and optimization frictions can be modeled explicitly, which allows for estimation of their size as opposed to visual determination of the bunching interval, and this is potentially very important since individuals may not bunch exactly at the tax kink.</cite> <cite index="15-9,15-10,15-11">The large number of robustness checks in previous studies hints at uncertainty regarding the optimal choice of the bunching window and the appropriate counterfactual model; Saez (2010) pointed out that in the presence of clustering around a kink point instead of exact bunching, the choice of the bandwidth matters.</cite>

<cite index="16-3,16-4,16-5">The bunching approach can be used in any setting with kinks or notches in the tax code; the identification process can be transparently illustrated by showing the income distribution around a kink or notch, and endogeneity is not a problem.</cite>

Sources:

• https://blogs.worldbank.org/en/impactevaluations/we-got-bunching-now-what
• https://www.nathanseegert.com/papers/Bunching_Chapter.pdf
• https://onlinelibrary.wiley.com/doi/full/10.1002/jae.3015
• https://link.springer.com/article/10.1007/s10797-020-09590-w
• https://www.sciencedirect.com/science/article/abs/pii/S0047272721000116

<cite index="12-15,10-9">Bunching estimators identify the compensated elasticity of income by measuring how much mass gathers at kink points—places where marginal tax rates jump—in the tax schedule.</cite> The method rests on a simple observation: <cite index="20-1">if we knew what the density of individuals would be without the kink, we would know how far the marginal buncher shifted their hours to reach the kink, which would tell us the labor supply elasticity of the marginal buncher.</cite>

<cite index="2-8,2-9,2-10,2-11">The approach follows the literature in modeling agents choosing consumption and labor supply with a fixed wage, where the budget set gives all feasible combinations of consumption and earned income; the setup considers two types of discontinuous changes in taxes, including kinks where the budget frontier is continuous except for a discontinuous change in slope at a known point.</cite> <cite index="4-2,4-10">Bunching methods identify the elasticity of labor supply with respect to benefit reduction rates, not only in income tax contexts but also in programs like SNAP.</cite>

<cite index="20-4,20-5">The approach implicitly assumes all individuals have the same elasticity; otherwise two individuals with the same earnings when the budget set was linear might make different choices about whether to bunch, though even with heterogeneity the method recovers a particular average elasticity.</cite> <cite index="1-24,1-25,1-26,1-27,1-29,1-30">Following Saez (2010) and Kleven and Waseem (2013), bunching methods became a popular way to estimate elasticities in settings ranging from electricity demand, real estate taxes, labor regulations, and prescription drug insurance to minimum wage and air-pollution data manipulation.</cite>

Sources:

• https://www.nathanseegert.com/papers/Bunching_Chapter.pdf
• https://papers.ssrn.com/sol3/papers.cfm?abstract_id=194633
• https://journals.sagepub.com/doi/abs/10.1177/1536867X221124534
• https://blogs.worldbank.org/en/impactevaluations/we-got-bunching-now-what

The methodological choices stack up in practice. Rule-based matching uses <cite index="1-5">a predefined dictionary containing commonly seen required skills into categories like statistics, machine learning, deep learning, R, Python, NLP, data engineering, business, software</cite>. It's fast, interpretable, but limited by what you thought to include.

Keyword extraction methods range from classical to neural. <cite index="3-6">Systems implement multiple extraction methods like KeyBERT, RAKE, and pre-trained models</cite>. The preprocessing pipeline is standard: <cite index="2-7,2-9,2-11">tidy up job descriptions by removing irrelevant characters, chop up the text into individual words or phrases to analyze them one by one, eliminate common words like 'and', 'the', or 'to' that don't add much meaning</cite>.

The third path is entity linking via standardized occupational frameworks. <cite index="25-3">Comprehensive occupational databases such as O*NET or ESCO provide detailed taxonomies of interrelated positions that can be leveraged to align the textual content of postings with occupational categories, facilitating standardization, cross-system interoperability, and access to metadata for each occupation</cite>. <cite index="20-3,20-5">One approach feeds complete job descriptions into models expecting a list of ESCO nodes as outputs (sentence linking, often labeled as extreme multi-label classification); another introduces an intermediate step where models first perform entity recognition, framing the task under the entity linking paradigm</cite>.

What gets measured depends on which method you choose. That dependency is not always acknowledged in the papers that cite the resulting skill-demand estimates.

Sources:

• https://sites.northwestern.edu/msia/2020/11/30/what-skills-do-data-scientists-need-a-text-analysis-of-job-postings/
• https://medium.com/@rfqhnfs/decoding-job-descriptions-how-nlp-unlocks-the-search-for-skills-91669d537779
• https://github.com/yadavanshul/Dynamic-Skill-Extraction-and-Job-Description-Analysis-Using-Advanced-NLP-Models
• https://arxiv.org/html/2512.03195v1
• https://www.researchgate.net/publication/397565402_Classification_of_Job_Offers_into_Job_Positions_Using_ONET_and_BERT_Language_Models

The mapping problem reveals itself in the edge cases. <cite index="22-26,22-27,22-28">The skill "mathematics" may be too vague to be mapped at the skill level, but it maps to a number of similar skills like "philosophy of mathematics" and "using mathematical tools and equipment", which sit under the skill group "natural sciences, mathematics and statistics"—the system moves up a taxonomy until a broad-enough skill group can be confidently assigned</cite>.

<cite index="24-17,24-18">Although the language model used to map semantic meaning can sometimes understand metaphors, metaphors can still cause inappropriate matches—for example, "understand the bigger picture" would be extracted as a skill entity and matched to the ESCO skill "interpreting technical documentation and diagrams"</cite>. <cite index="24-19">The algorithm can struggle with sentences that contain multiple skills</cite>.

The annotation bottleneck is real. <cite index="16-5,16-6">Previous work on supervised skill extraction frames it as a sequence labeling task or multi-label classification, but annotation is a costly and time-consuming process with little annotation guidelines to work with</cite>. This drives the turn toward synthetic data: <cite index="17-18,17-20,17-21">generating synthetic training data for the entirety of ESCO skills and training a classifier to extract skill mentions achieves an RP@10 score 10 points higher than previous distant supervision approaches; adding GPT-4 re-ranking improves RP@10 by over 22 points</cite>.

But here's what matters for labor economists: <cite index="24-23">skills extracted should not be used to measure skill demand without expert review and input, nor should they be used for any discriminatory hiring practices</cite>. The methods are good enough to see patterns. They are not yet good enough to make hiring decisions.

Sources:

• https://explosion.ai/blog/nesta-skills
• https://www.escoe.ac.uk/the-skills-extractor-library/
• https://arxiv.org/pdf/2209.08071
• https://www.researchgate.net/publication/398312495_Enhancing_Job_Matching_Occupation_Skill_and_Qualification_Linking_with_the_ESCO_and_EQF_taxonomies

The shift from rule-based matching to transformer models changed what could be extracted from job postings. <cite index="10-2,10-7">Data vectorization for initial feature extraction is performed using BERT structure transformers (sentence transformers)</cite>, and <cite index="10-12,10-13">sentence transformers employ pre-trained transformer models such as BERT to generate dense vector embeddings of sentences that capture the semantic meaning of the input text</cite>.

Recent work shows the range of architectures deployed. <cite index="12-1,12-2">A contrastive bi-encoder aligns job-ad sentences with ESCO skill descriptions in a shared embedding space; the encoder augments a BERT backbone with BiLSTM and attention pooling to better model long, information-dense requirement statements, with an upstream RoBERTa-based binary filter removing non-skill sentences to improve precision</cite>. The reported performance: <cite index="12-3">strong zero-shot retrieval performance (F1@5 = 0.72) on real-world job advertisements, outperforming TF-IDF and standard BERT baselines</cite>.

But efficiency matters when you're processing millions of postings. <cite index="4-9">Lightweight models that achieve high performance without relying on vast computational resources can democratize access to labor market insights, enabling organizations and researchers to process large datasets cost-effectively and at scale</cite>. The two-step pattern is now standard: <cite index="21-6,21-7,21-8">first, extracting skills using a model that predicts which parts of a job advert are skills, trained using Named Entity Recognition (NER) neural network architecture; second, mapping the extracted skill entities to an existing taxonomy like ESCO</cite>. <cite index="24-12">Of appropriately extracted skill entities, 88% were judged to be appropriately mapped to ESCO skills</cite>. That 12% error rate is where the method still shows its seams.

Sources:

• https://doi.org/10.3390/app13106119
• https://arxiv.org/pdf/2601.09119
• https://arxiv.org/pdf/2505.24640
• https://www.escoe.ac.uk/the-skills-extractor-library/

<cite index="4-5">Job postings often use different terminology to refer to occupations and skills</cite>, which is why <cite index="4-6">natural language processing (NLP) techniques are required to identify and normalize the information contained in job ads</cite>. The core technical challenge is twofold. First, extraction: <cite index="16-4">automatic skill extraction (SE) is the task of extracting the competences necessary for any occupation from unstructured text</cite>. Second, normalization: <cite index="4-10,4-11,4-12">benchmarks for skill extraction have traditionally been formalized as span labeling tasks, without linking the identified spans to respective skills in a taxonomy—this lack of normalization prevents robust analysis because of synonyms</cite>.

The dominant approach uses standardized taxonomies—ESCO (European Skills, Competences, and Occupations) for European labor markets, O*NET for the US. <cite index="22-12,22-13,22-14">These taxonomies contain rich information like skills, skill definitions and classifications, containing thousands of labels that are updated to reflect skill changes in the labor market over time</cite>. But here's the friction: <cite index="22-15">training and keeping updated a Named Entity Recognition (NER) model to extract specific skill labels would be a labeling nightmare</cite>.

The practical consequence? Methods diverge. <cite index="7-13,7-14,7-15,7-16">Latent Dirichlet Allocation (LDA) identifies underlying topics in job advertisement texts using the probability distribution of words, without outside input, interpreting each topic as a specific skill requirement and returning a set of keywords and weights for each topic</cite>. When tested on wage regressions, <cite index="7-1">top-down methods explain only about 20% of wage variation, while the LDA model explains about 45%</cite>. That gap matters when the goal is understanding what employers actually pay for.

Sources:

• https://arxiv.org/pdf/2505.24640
• https://arxiv.org/pdf/2207.12834
• https://arxiv.org/pdf/2209.08071
• https://explosion.ai/blog/nesta-skills

<cite index="25-7,25-8">In March 2026, Lightcast announced a major expansion of its international data coverage—growing from 41 to more than 165 countries and marking a 300 percent increase in global reach, establishing Lightcast as the industry's most comprehensive source of global labor market intelligence.</cite> <cite index="25-1,25-2">Job postings coverage expanded from 41 to 165+ countries, offering clearer visibility into hiring activity and workforce demand worldwide, while de-duplicated and vetted profile coverage increased by 100+ million global profiles.</cite>

The technical architecture matters: <cite index="25-22">Lightcast's multilingual translation model uses large-scale transformer models—a sophisticated class of generative AI—fine-tuned specifically on labor market data and enriched with deep linguistic expertise.</cite> <cite index="17-25">Lightcast does not translate non-English job postings and uses the native language of the job posting to find and tag the relevant skills.</cite>

This creates a new comparability problem. When you measure "data analyst" demand in Berlin, São Paulo, and Singapore using native-language postings, you are measuring three different occupational constructs that share a translated label. <cite index="23-10">By expanding its dataset to cover countries representing roughly 99% of global GDP, Lightcast is aiming to give decision-makers a unified, comparable view of the world's talent landscape.</cite> But unified does not mean equivalent. Labor demand in markets with different educational systems, different hiring norms, different posting conventions—these are not the same instrument, indexed.

Sources:

• https://www.prnewswire.com/news-releases/lightcast-extends-global-data-footprint-by-300-delivering-the-industrys-largest-global-labor-market-coverage-302708689.html
• https://hrtechedge.com/lightcast-expands-labor-market-data-to-165-countries-aiming-to-become-the-global-benchmark-for-workforce-intelligence/
• https://kb.lightcast.io/en/articles/6957446-job-posting-analytics-jpa-methodology

<cite index="17-20,17-21,17-22,17-23">The process begins by segmenting and tokenizing job postings to remove extra characters and new lines, then the model scans the text for word sequences that indicate skills in the proper context—for example, when "AWS" appears, the surrounding context helps determine whether it refers to the "American Welding Society" or "Amazon Web Services," with a confidence score assigned and accuracy thresholds ensuring quality predictions are displayed.</cite>

This is the technical work that makes skill-demand tracking possible. <cite index="17-34,17-35">All historic data can have latest skills or any other new classifications and normalizations applied not only to new data but to data for all time—for example, when Lightcast introduced the skill 'Generative AI Agents' in January 2025, they were able to identify this skill in all historic data.</cite>

The implications: you can now track when "Generative AI" language appeared in postings for junior data analyst roles in Sacramento versus Seattle. You can see when "must be onsite" shifted from optional to required, by metro. <cite index="13-1,13-2">Researchers propose five variables for detecting skill shortages from online job ads: posting frequency; salary levels; education requirements; experience demands; and job ad posting predictability, contributing evidence to the goal of detecting skills shortages in real-time.</cite>

But context collapse is real. A skill tagged in a posting does not mean the skill was required to get the job. It means it was written in the posting.

Sources:

• https://kb.lightcast.io/en/articles/6957446-job-posting-analytics-jpa-methodology
• https://arxiv.org/pdf/1911.02302

When researchers cross-checked Lightcast data against BLS hiring numbers in Seattle, they found something clarifying: <cite index="24-11,24-12,24-13">though employers averaged nearly 300 job postings related to physicians assistants per month, they only actually hired some 40 per month—this is frequently the case with health care or IT positions, where postings usually far exceed hiring for such jobs.</cite>

This is not noise. It is signal. A posting is not a hire. It is intent, or signaling, or sometimes procedural theater. The gap between postings and hires tells you about labor market tightness, about employers searching longer, about roles that never fill.

<cite index="3-5,3-6">Using job vacancy data from Lightcast from 2007Q1 to 2021Q2, researchers measured labor market concentration using Herfindahl-Hirschman Index in labor markets defined at the occupation (six-digit SOC), commuting zone, and quarterly level, with the HHI calculated based on the share of vacancies among all the firms that post vacancies in that market.</cite> This granularity—commuting zone, six-digit SOC, quarterly—lets you see where employers compete for labor and where they don't.

<cite index="13-8,13-9">At aggregate levels, online job ads can provide valuable indicators of relative labour demands, and rather than relying solely on lagging indicators from labour market surveys, online job ads data can reveal shifting labour demands as they occur.</cite> That timeliness is the trade: you give up the representativeness of JOLTS for the speed and geography of the web scrape.

Sources:

• https://kb.lightcast.io/en/articles/6957578-job-posting-analytics
• https://pmc.ncbi.nlm.nih.gov/articles/PMC11305209/
• https://arxiv.org/pdf/1911.02302

<cite index="2-1">Lightcast aggregates data from 160,000+ sources, including job postings, government reports, skills databases, compensation records, and professional profiles.</cite> The company—<cite index="2-3,2-6">formerly Burning Glass Technologies, now operating as Lightcast</cite>—<cite index="5-5,5-6">collects and de-duplicates job postings from about 40,000 websites, which constitutes most of the U.S. job vacancies posted online.</cite>

The methodology matters: <cite index="17-2">Lightcast technologies parse, extract, and code dozens of data elements, including job titles, occupations, companies, and detailed data about specific skills, educational credentials, certifications, experience levels, and work activities required for the job.</cite> <cite index="17-18">On average, 13 skills are extracted per posting.</cite>

This creates a different instrument than JOLTS. <cite index="7-5,7-6">UI claims data and vacancy data measure fundamentally distinct phenomena—the former give an indication of how many matches in the labor market have become unsustainable over a given period, while vacancies provide a forward looking measure as firms post vacancies to establish new employment relationships.</cite> <cite index="7-8">Job vacancy data allow tracking the economy at disaggregated geography and by detailed occupation and industry.</cite>

The Conference Board now uses Lightcast: <cite index="6-1">In 2019, Lightcast joined the Help Wanted OnLine program as the new sole provider of online job ad data for HWOL.</cite> California, Texas, and multiple state labor agencies pull from the same source.

Sources:

• https://lightcast.io/burning-glass-technologies
• https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11305209/
• https://pmc.ncbi.nlm.nih.gov/articles/PMC7357497/
• https://www.conference-board.org/topics/help-wanted-online/

<cite index="5-1,5-2,5-3">A 2016 IZA paper highlights two common pitfalls in population-threshold RDD — compound treatment (where multiple policies change at the same threshold) and sorting (where units manipulate their position relative to the cutoff) — and provides guidance for detecting and addressing these problems using evidence from France, Germany, and Italy.</cite> <cite index="20-3,20-4,20-5">In unemployment insurance RD designs, both employers and individuals can influence the timing of job loss and benefit claiming; detailed data allow researchers to investigate whether this leads to sorting around eligibility cutoffs, and the overall conclusion in some studies is that labor supply effects represent valid RD estimates despite these concerns.</cite>

<cite index="24-1,24-2,24-3">Research on UI benefit extensions using boundary-based identification has documented two potential biases: from using county-level aggregates and from across-border policy spillovers; a measurement error-corrected RD approach suggests much smaller effects than previous studies, casting doubt on the applicability of border-based designs.</cite> <cite index="24-4">The same research documents substantial spillover effects of UI benefit duration in the form of across-border work patterns, consistent with increased tightness in high-benefit states.</cite>

These are not abstract concerns. In practice, firms delay layoffs until workers cross an age threshold. Counties near state borders see cross-commuting that blurs the treatment. The RD estimate remains internally valid — it tells you what happened at the threshold — but external validity requires knowing whether the people near the cutoff behaved differently because they were near the cutoff. The best RD work now routinely checks for density discontinuities in the running variable and tests whether observable characteristics are smooth at the threshold. When they are not, the design is suspect.

Sources:

• https://www.iza.org/publications/dp/9553/
• https://www.nber.org/system/files/working_papers/w17813/w17813.pdf
• https://ideas.repec.org/p/iza/izadps/dp11496.html

<cite index="17-2,17-3">A 2025 Atlanta Fed working paper uses a regression discontinuity design at the monetary eligibility threshold for UI, leveraging administrative data to estimate the local causal effect of UI eligibility for workers experiencing unemployment on both sides of the threshold.</cite> <cite index="21-1,21-3">In the United States, workers whose past earnings fall below a threshold are generally ineligible for unemployment insurance, creating a discontinuous jump in the value of being unemployed; the RD estimate shows a sizable local effect from UI eligibility on earnings in the next employer, around 10 percent per quarter.</cite>

Age thresholds have been equally productive for identification. <cite index="23-6,23-7,23-8,23-9">Research on Germany exploits a sharp discontinuity in the maximum duration of unemployment benefits, which increases from 12 to 18 months at age 45; the study finds a spike in the re-employment hazard around benefit exhaustion for workers with 12-month duration, leading to shorter unemployment, but also shows that those who obtain jobs near exhaustion are significantly more likely to exit subsequent employment and receive lower wages compared to those with extended benefits.</cite>

The findings are not just about duration. They are about match quality. Longer benefit duration allows the unemployed to wait for better offers, but only if the labor market cooperates. When it doesn't, the threshold becomes a lever for revealing how much of unemployment is search and how much is availability. RD designs in UI research have moved from measuring duration effects to measuring the entire trajectory of reemployment — wages, job tenure, subsequent separations — because the threshold gives you a point of comparison for all of it.

Sources:

• https://www.atlantafed.org/-/media/documents/research/publications/wp/2025/07/14/06-effect-of-unemployment-insurance-eligibibility-in-equilibrium.pdf
• https://www.atlantafed.org/research-and-data/publications/working-papers/2025/07/14/06-effect-of-unemployment-insurance-eligibibility-in-equilibrium
• https://www.iza.org/publications/dp/4670/benefit-duration-unemployment-duration-and-job-match-quality-a-regression-discontinuity-approach

<cite index="11-6">Researchers have used RDD to assess the impact of minimum wage increases on employment and earnings by exploiting eligibility criteria based on firm size or geographic location.</cite> The design works when the wage floor applies sharply — to firms just above a headcount threshold, or to counties on one side of a state border but not the other. <cite index="10-2,10-7">Difference-in-differences has also been widely applied to minimum wage evaluation, famously in Card and Krueger's 1994 study comparing adjacent U.S. states, one of which introduced a statutory minimum wage while the other did not — work that contributed to David Card winning the Nobel Prize.</cite>

RDD complements DiD by offering a different source of variation. Where DiD compares changes over time across treated and control regions, RDD compares individuals or firms at a single moment, separated only by a threshold. <cite index="13-3,13-4">Minimum wage policies have been evaluated using RDD by focusing on regions or demographics slightly above and below prescribed wage levels, yielding insights into employment effects.</cite>

The challenge in both designs is spillovers. Workers cross borders. Firms adjust hiring not just at the threshold but in anticipation of it. The estimate remains local, but the economy does not. That tension is why the best minimum wage research now pairs RD identification with equilibrium models — acknowledging that the threshold tells you what happened to the people who were close, but not necessarily what happens city-wide when every low-wage employer adjusts at once.

Sources:

• https://www.numberanalytics.com/blog/essential-guide-rd-design-labor-econ
• https://www.researchgate.net/publication/350170690_Methods_to_Estimate_Causal_Effects_-_An_Overview_on_IV_DiD_and_RDD_and_a_Guide_on_How_to_Apply_them_in_Practice
• https://www.numberanalytics.com/blog/ultimate-rdd-guide-econometrics

<cite index="6-2,6-6">Lee and Lemieux's 2010 Journal of Economic Literature survey describes regression discontinuity as a quasi-experimental design that isolates causal effects by comparing outcomes for observations immediately around a treatment cutoff, exploiting thresholds where treatment assignment changes sharply.</cite> The method rests on a simple insight: if individuals just above and below a threshold are otherwise identical — same work history, same city, same moment — then any difference in their outcomes after treatment can be attributed to the policy itself, not to the sorting that usually ruins observational research.

<cite index="1-1">In labor economics, the design has been applied to unemployment insurance eligibility (often age-based) and minimum hours requirements, among other thresholds.</cite> <cite index="16-5,16-6">The appeal of RDD is that it produces credible causal estimates without requiring unconfoundedness or global positivity — the usual demands of matching or instrumental variables — relying instead only on continuity of potential outcomes and the absence of precise manipulation around the cutoff.</cite>

But the estimate is inherently local. <cite index="13-5,13-6">The treatment effects estimated via RDD are local to the cutoff, meaning findings are robust near the threshold but generalizing to wider populations requires caution.</cite> This is not a flaw. It is precision. The RD estimate tells you what happened to the people who were barely eligible versus barely ineligible — the ones whose fates turned on a birthday or a dollar of prior-year earnings. That is often exactly the population policymakers need to understand.

Sources:

• https://www.princeton.edu/~davidlee/wp/RDDEconomics.pdf
• https://rdpackages.github.io/references/Cattaneo-Titiunik_2022_ARE.pdf
• https://arxiv.org/pdf/2508.15692

<cite index="22-7,22-8,22-9">The ACS has an annual sample size of about 3.5 million addresses, with survey information collected nearly every day of the year. Data are pooled across a calendar year to produce estimates for that year. As a result, ACS estimates reflect data that have been collected over a period of time rather than for a single point in time.</cite>

<cite index="27-4,27-5">Each year, independent housing unit address samples are selected for each county equivalent in the United States and Puerto Rico. Samples of group quarters facilities and persons in group quarters are done at the state level.</cite> <cite index="22-2,22-3,22-16">Areas or groups of 65,000 or more are eligible for 1-year and 5-year estimates. Areas or groups of 20,000 or more are eligible for 5-year estimates. Areas or groups of 20,000 or fewer are eligible for 5-year estimates only.</cite>

<cite index="10-3,10-4">For an individual population stratum (defined by current age, entry year, country of origin, and calendar year) direct estimates using survey data can have substantial sampling uncertainty. By imposing logical and probabilistic constraints, data are pooled across survey years to produce more precise estimates.</cite> The smallest geographies require five years of pooling to reach publishable precision. This is not a limitation—it is arithmetic.

Sources:

• https://nces.ed.gov/fCSM/acs.asp
• https://www.test.census.gov/content/dam/Census/topics/population/migration/guidance-for-data-users/acs-migration-tutorial/2010-2014-Migration-Flows-Documentation.pdf
• https://arxiv.org/pdf/1906.01716

<cite index="11-1,11-2,11-3">Data are based on a sample and are subject to sampling variability. The degree of uncertainty for an estimate arising from sampling variability is represented through a margin of error. The tables include the 90 percent margin of error.</cite> <cite index="11-9">The margin of error can be interpreted roughly as providing a 90 percent probability that the interval defined by the estimate minus the margin of error and the estimate plus the margin of error (the lower and upper confidence bounds) contains the true value.</cite>

<cite index="11-10,11-11,11-13">In addition to sampling variability, the ACS estimates are subject to nonsampling error. The effect of nonsampling error is not represented in these tables.</cite> <cite index="11-5">Sample size and data quality measures, including coverage rates, allocation rates, and response rates, are available on the ACS Research & Methodology page.</cite>

<cite index="27-14,27-15,27-16">The universe for each table must have at least 50 unweighted cases for the Census Bureau to publish the table. The same rules are applied to the county-to-county, county/MCD-to-county/MCD, and MSA-to-MSA files. If there is not a sufficient number of cases, the summary statistics for county, MCD, or MSA of residence 1 year ago by characteristics are not shown.</cite> The Bureau gives you the sampling uncertainty. The misreporting, the missed housing units, the people who didn't answer—those are yours to track separately.

Sources:

• https://www.census.gov/data/tables/time-series/demo/geographic-mobility/state-to-county-migration.html
• https://www.census.gov/data/tables/2020/demo/geographic-mobility/county-to-county-migration-2016-2020.html
• https://www.test.census.gov/content/dam/Census/topics/population/migration/guidance-for-data-users/acs-migration-tutorial/2010-2014-Migration-Flows-Documentation.pdf

<cite index="7-9,7-10,8-6">County-to-county migration flows are available up to and including the 2016-2020 dataset.</cite> <cite index="7-7,7-8">State-to-county migration flows became available starting with the 2018-2022 5-year period (2021–Present).</cite> <cite index="7-11,7-12">Metro-to-metro flows are available from 2009–2020.</cite> <cite index="7-5,7-6,7-19,7-20">State-to-state 1-year flows have been available since 2005.</cite>

The Census Bureau changed which geographies it publishes flows for—not the underlying method. <cite index="7-3,7-17">The Bureau cautions data users against conducting any type of comparison of migration flows at different geographic levels.</cite> <cite index="1-10,1-11">The 2018-2022 state-to-county flows suppressed estimates of migration within, to, and from Connecticut; data users should exercise caution when comparing 2018-2022 estimates to prior years.</cite>

<cite index="23-35,23-36">The Census Bureau strongly recommends against comparing estimates in overlapping 5-year periods since much of the data in each estimate are the same. Instead, compare 5-year estimates that don't have any overlapping years of data.</cite> The shift from county-to-county to state-to-county was not about data quality. It was about what the Bureau chose to tabulate and release.

Sources:

• https://www.census.gov/topics/population/migration/guidance/migration-flows.html
• https://www.census.gov/topics/population/migration/guidance/county-to-county-migration-flows.html
• https://www.census.gov/newsroom/blogs/random-samplings/2022/03/period-estimates-american-community-survey.html
• https://www.census.gov/data/developers/data-sets/acs-migration-flows.html

<cite index="1-3,2-1">Migration flows are derived from the ACS sample location and the response to a single question: "Where did you live 1 year ago?"</cite> <cite index="8-3,8-4,8-5">The ACS asks respondents aged 1 year and over whether they lived in the same residence 1 year ago, and for those who moved, the location of their previous residence is collected. County-to-county flows are created from tabulations of current county of residence crossed by county of residence 1 year ago.</cite>

<cite index="1-1,1-4">These are period estimates measuring where people lived when surveyed (current residence) and where they lived 1 year prior.</cite> <cite index="1-5,1-6">Data are collected continuously over a 5-year period to provide a large enough sample for estimates in smaller geographies. The flow estimates resemble the annual number of movers between counties for the 5-year period data was collected.</cite>

<cite index="23-4,23-26">For 5-year estimates, data are not averaged—they are pooled together, weighted and processed as a whole dataset to take advantage of the larger number of records.</cite> <cite index="23-17,23-18,23-19">Unlike the census, which uses a single reference date of April 1, the ACS uses the date of the response as a reference date. Since data are collected throughout the year, the ACS consolidates all the data collected over the calendar years to create estimates that do not represent a fixed date, but rather the characteristics over an entire data collection period.</cite>

Sources:

• https://www.census.gov/data/developers/data-sets/acs-migration-flows.html
• https://data.commerce.gov/american-community-survey-5-year-migration-flows
• https://www.census.gov/topics/population/migration/guidance/county-to-county-migration-flows.html
• https://www.census.gov/newsroom/blogs/random-samplings/2022/03/period-estimates-american-community-survey.html

<cite index="24-1">In order to facilitate researcher use of the microdata, LEHD infrastructure files are periodically collected into a standardized form known as the LEHD Snapshot, which is then released into the FSRDC network.</cite> The files themselves are not raw. <cite index="28-9,28-10">The input files are compiled and combined to create the infrastructure files, using multiple imputation methods to impute in missing data and statistical matching techniques to combine and edit data when a direct identifier match requires improvement.</cite> <cite index="28-11">Both of these innovations are crucial to the success of the final product.</cite>

<cite index="29-3,29-4">The LEHD Program maintains a set of infrastructure files using administrative data provided by state agencies, enhanced with information from other administrative data sources, demographic and economic (business) surveys and censuses; the LEHD Infrastructure Files provide a detailed and comprehensive picture of workers, employers, and their interaction in the U.S. economy.</cite> <cite index="21-2,21-7">The document attempts to provide a comprehensive description of all researcher-accessible files, of their creation, and of any modifications made to the files to facilitate researcher access.</cite>

<cite index="24-3,24-4">Requests for state-level tables will follow the procedure specified by the Memorandum of Understanding (MOU) with that state; most commonly, the state agency is given the opportunity to approve proposed projects.</cite> Not all states participate, and coverage varies over time. <cite index="18-11,18-12">All fifty states, the District of Columbia, the Commonwealth of Puerto Rico, and the U.S. Virgin Islands may participate in the LED Partnership; composition of the LED Partnership can vary as each eligible member determines their participation.</cite> The Snapshot is versioned; researchers request specific vintages and specific packages depending on whether they need FTI, SSA demographics, or state-level detail.

Sources:

• https://lehd.ces.census.gov/data/lehd-snapshot-doc/latest/sections/introduction.html
• https://ideas.repec.org/h/nbr/nberch/0485.html
• https://www.census.gov/library/working-papers/2018/adrm/ces-wp-18-27r.html
• https://lehd.ces.census.gov/

<cite index="15-1">Throughout all processes associated with acquiring, using, and disposing of administrative records data, the provisions of federal laws (e.g., Title 13, Title 15, and Title 26), data-use agreements, and Census Bureau policies and procedures on privacy and confidentiality must be followed to protect administrative records data from unauthorized release.</cite> LEHD data crosses agency boundaries, so approval stacks. <cite index="22-5">Some LEHD data files contain Federal Tax Information (FTI), and use of these files requires additional approval by the Internal Revenue Service (IRS).</cite> <cite index="22-6,24-5">LEHD demographic files make use of administrative information from the Social Security Administration (SSA) and therefore require SSA approval for use.</cite>

<cite index="20-1">No department, bureau, agency, officer, or employee of the Government shall permit anyone other than the sworn officers and employees to examine the individual reports.</cite> This isn't theoretical: <cite index="14-21">A Disclosure Review Board (DRB) review is required for statistical products based in whole or in part on administrative data.</cite> Public-use products like the Quarterly Workforce Indicators and LODES exist because <cite index="18-3,18-10">the LEHD program uses these data to create partially synthetic data on workers' residential patterns.</cite> For the microdata itself, <cite index="22-1,22-2">the LEHD microdata is available to qualified researchers in the Federal Statistical Research Data Center (FSRDC) network; researchers must have an approved project with authorization to use specific data sets.</cite>

<cite index="17-24">Access to SSA data that have been linked to Census Bureau data is subject to additional restrictions imposed by Title 13, such as requiring users to obtain Special Sworn Status and permitting access only for Census-approved purposes and at a Census-approved site.</cite> The confidentiality architecture is why the files exist at all — states share because the protections are federal law.

Sources:

• https://www.census.gov/about/policies/quality/standards/standardb2.html
• https://www.census.gov/programs-surveys/ces/data/restricted-use-data/lehd-data.html
• https://lehd.ces.census.gov/data/lehd-snapshot-doc/latest/sections/introduction.html
• https://www.census.gov/about/policies/privacy/data_stewardship/title_13_-_protection_of_confidential_information.html
• https://www2.census.gov/foia/ds_policies/ds001_appendices.pdf
• https://www.ssa.gov/policy/docs/ssb/v69n1/v69n1p75.html

<cite index="5-7,7-6">Under the LED Partnership, states agree to share Unemployment Insurance earnings data and the Quarterly Census of Employment and Wages (QCEW) data with the Census Bureau.</cite> <cite index="1-4">LEHD combines data from state unemployment insurance records, employer-provided data, and household surveys to create a rich dataset that tracks employment patterns, worker flows, and economic conditions.</cite> The product is longitudinal: <cite index="10-3">the LEHD data at the U.S. Census Bureau is a quarterly database of linked employer-employee data covering over 95% of employment in the United States.</cite>

The infrastructure files are built by merging these administrative feeds with Census demographic surveys and economic censuses. <cite index="25-3">The LEHD Program maintains a set of infrastructure files using administrative data provided by state agencies, enhanced with information from other administrative data sources, demographic and economic (business) surveys and censuses.</cite> <cite index="7-10">The National Individual Characteristics File (ICF) contains one record for every person who is ever employed in any LEHD participating state over the time period spanned by state's unemployment insurance records.</cite> <cite index="7-11">It consolidates information from multiple input sources on gender, age, place of birth, race, ethnicity, and education.</cite>

W-2 data enters when linked via SSA. <cite index="11-3">SSA uses QCEW data as a quality check against data provided by employers on Forms W-2.</cite> The system is not survey-based. <cite index="10-5">By integrating administrative data with existing census and surveys, a national longitudinal jobs database for the U.S. is generated at very low cost and with no additional respondent burden.</cite> What's produced are job-level, person-level, and establishment-level files that track worker-employer pairs quarter by quarter.

Sources:

• https://lehd.ces.census.gov/
• https://www.icpsr.umich.edu/web/appfed/studies/37528
• https://www.census.gov/programs-surveys/ces/data/restricted-use-data/lehd-data.html
• https://www.bls.gov/opub/hom/cew/presentation.htm
• https://www.socialexplorer.com/home/dataset-entry/longitudinal-employer-household-dynamics-lehd

<cite index="2-4,2-13">The May 2024 OEWS estimates were produced by a model-based estimation method using three years of OEWS data.</cite> <cite index="2-5,12-4">Each establishment's population employment is set as the average of its May 2024 and November 2023 employment from the Quarterly Census of Employment and Wages, the UI database from which the OEWS sample is drawn.</cite> <cite index="12-5,21-3">Using adjustment factors derived from the OEWS survey data, wages collected in earlier survey panels are adjusted to the reference date of the estimates and donor wages are adjusted for differences between donor and recipient characteristics such as geographic area and industry.</cite>

This is careful work, but it assumes that wage movements within occupations are smooth and predictable. <cite index="11-5,11-6">Starting with the 1997 estimates, the OES program has used the over-the-year fourth-quarter wage changes from the BLS Employment Cost Index (ECI) to adjust prior year survey data before combining it with the current year's data; the wage updating procedure assumes that each occupation's wage, as measured in the earlier years, moves according to the average movement of its occupational division and that there are no major geographic or detailed occupational differences.</cite>

That assumption broke in 2021 and 2022 for remote-eligible technical occupations. Software developers in Boise saw wage jumps that looked nothing like the ECI occupational division average because coastal firms were suddenly hiring there at coastal-adjacent wages. The methodology didn't anticipate that the geography of the wage distribution could shift faster than the wage levels themselves.

<cite index="26-16,26-17,26-19">The OEWS survey methodology is designed to create detailed cross-sectional occupational employment and wage estimates by geographic area or industry, but it is less useful for looking at changes over time due to changes in the occupational, industry, and geographic classification systems; changes in the OEWS methodology and data collection procedures; and permanent features of the OEWS methodology, including the three-year pooled sample design; the Bureau of Labor Statistics does not encourage the use of OEWS data for time-series analysis.</cite> I use it anyway. But I date the caveats.

Sources:

• https://www.bls.gov/oes/oes_emp.htm
• https://www.bls.gov/oes/tec_02.htm
• https://www.bls.gov/oes/oes_ques.htm

<cite index="2-1,21-1">The OEWS program uses the Office of Management and Budget's Standard Occupational Classification (SOC) system to classify jobs into occupations based on their job duties.</cite> <cite index="2-15,20-1">The OEWS program produces employment and wage estimates for approximately 830 occupations.</cite> <cite index="26-4,26-5">The 2018 SOC system contains 867 detailed occupations, aggregated into 459 broad occupations, which in turn combine into 98 minor groups and 23 major groups.</cite>

But not every establishment gets asked about every occupation. <cite index="23-3,23-4,23-5">The occupations listed on survey forms vary depending on the industry and size of establishment; no survey form contains all OEWS occupations, because no industry employs workers in every occupation; survey forms contain between 50 and 225 occupations.</cite>

This means the program has to know what to ask before it asks. A construction firm in Phoenix won't see "Multimedia Artists and Animators" on its form. A game studio in Austin won't see "Pile Driver Operators." That's efficient design. It also means that if an occupation does show up in an unexpected industry — say, a healthcare system hires a cohort of data scientists in 2023 — the survey has to catch up.

<cite index="2-6,12-1">The OEWS data available from BLS include cross-industry occupational employment and wage estimates for the nation; states, the District of Columbia, and territories; approximately 530 metropolitan statistical areas (MSAs) and nonmetropolitan areas; national industry-specific estimates at the NAICS sector, 3-digit, most 4-digit, and selected 5- and 6-digit industry levels; and national estimates by ownership across all industries and for schools and hospitals.</cite> You can look at software developers nationally. You can look at software developers in San Jose. You can look at software developers in NAICS 5415 (Computer Systems Design). You cannot — in the public tables — look at software developers in San Jose working in computer systems design who earn between the 60th and 75th percentile. The cells get thin fast.

Sources:

• https://www.bls.gov/oes/oes_emp.htm
• https://www.bls.gov/oes/
• https://www.bls.gov/oes/oes_ques.htm
• https://labormarketinfo.edd.ca.gov/oes/oes-faqs.html

<cite index="6-10,6-11,6-12">OEWS obtains two types of wage data: exact wage rates for federal government, USPS, TVA, and most employees in state government, local government, and private sector establishments; for a small percentage of records for which exact wage rates are not available, the wage data are processed in 12 wage intervals.</cite> <cite index="23-37">Establishments are asked to report how many workers they employ in a given occupation in each of several wage ranges.</cite>

This is not a Census question with a write-in box. It's a mail form with checkboxes. An employer in Sacramento reports that they have 14 software developers, four of whom earn between $45.00 and $54.99 per hour. The BLS assigns those four people a wage based on where they fall in that interval.

<cite index="27-7,7-1">Estimates for hourly and annual wages are available for the following percentiles: 10%, 25%, 50% (median), 75% and 90%.</cite> <cite index="4-7,4-8">An occupational median wage estimate is the boundary between the highest paid 50% and the lowest paid 50% of workers in that occupation; half of the workers in a given occupation earn more than the median wage, and half the workers earn less than the median wage.</cite>

<cite index="22-12,22-13">OEWS now sets the wage for the highest interval at the mean wage that workers in that interval would be expected to make, estimated from data collected by the National Compensation Survey.</cite> Before 2002, everyone in the top bin got assigned the floor. That methodology shift matters when you're looking at occupations with long right tails — dentists, anesthesiologists, software architects.

Sources:

• https://www.bls.gov/oes/methods_24.pdf
• https://dlt.ri.gov/labor-market-information/data-center/occupational-employment-and-wage-statistics-oews
• https://d4.nccommerce.com/UserDocs/OES/09-01-2020h17-22-15_08-25-2020h15-42-53_OESOverview.htm
• https://labormarketinfo.edd.ca.gov/data/oes-wages-about-the-data.html

<cite index="2-7,16-10">The OEWS estimates are constructed from a sample of about 1.1 million establishments collected over a 3-year period.</cite> <cite index="2-9,16-8">Each year, two semiannual panels of approximately 186,000 to 189,000 sampled establishments are contacted, one panel in May and the other in November.</cite> <cite index="2-8,16-2">The sample is drawn from the database of businesses reporting to the state unemployment insurance programs (the Quarterly Census of Employment and Wages).</cite>

The frame itself matters. <cite index="2-16">The OEWS survey covers wage and salary workers in nonfarm establishments and does not include the self-employed, owners and partners in unincorporated firms, household workers, or unpaid family workers.</cite> That means the contractor economy doesn't show up here — gig drivers, 1099 consultants, freelance designers. It also means that every time someone quits a W-2 role to consult, they disappear from the wage distribution even if their effective hourly rate went up.

<cite index="13-1">Significant reductions in sampling error can be achieved by taking advantage of a full three years of data, covering 1.1 million establishments and about 57 percent of the employment in the United States.</cite> But the trade-off is that each May estimate includes data collected from six panels going back three years. The labor market in November 2022 and the labor market in May 2025 were not the same place. The methodology smooths that.

Sources:

• https://www.bls.gov/oes/oes_emp.htm
• https://www.bls.gov/oes/methods_24.pdf
• https://www.bls.gov/oes/oes_ques.htm

<cite index="7-17,21-1">The CPS, sponsored jointly by the U.S. Census Bureau and the U.S. Bureau of Labor Statistics, is the primary source of labor force statistics for the population of the United States</cite>. <cite index="9-1,9-6">The household survey is designed to measure the labor force status of the civilian noninstitutional population with demographic detail</cite>. <cite index="24-15,24-16">The target population is the civilian noninstitutional population age 16 and older in the 50 states and the District of Columbia; active duty members of the Armed Forces are excluded</cite>.

<cite index="9-11">Based on responses to questions on work and job search activities, each person 16 years and over in a sample household is classified as employed, unemployed, or not in the labor force</cite>. <cite index="13-17">Employment information is collected on the job held in the reference week, defined as the 7-day period, Sunday through Saturday, that includes the 12th of the month</cite>.

<cite index="11-10,11-11">The CPS questionnaire is a completely computerized document administered by Census Bureau field representatives through both personal and telephone interviews, with additional telephone interviewing conducted from two centralized collection facilities in Jeffersonville, Indiana and Tucson, Arizona</cite>. <cite index="3-6">The CPS ASEC data collection methodology differs substantially from the ACS, as CPS ASEC is conducted by interviewers via Computer Assisted Telephone Interviewing (CATI) or Computer Assisted Personal Interviewing (CAPI)</cite>.

<cite index="26-2,26-4,26-5">The U.S. Census Bureau develops the population estimates for the household survey, adjusting them each year to include the latest information about population change and to incorporate any improvements in estimation methodology</cite>. <cite index="26-12">Following usual BLS practice, official household survey estimates for December and earlier months are not revised when January population controls are updated; consequently, data for January are not directly comparable with December or earlier periods</cite>.

Sources:

• https://www.census.gov/programs-surveys/cps.html
• https://en.wikipedia.org/wiki/Current_Population_Survey
• https://www.census.gov/programs-surveys/cps/technical-documentation/methodology.html
• https://odphp.health.gov/healthypeople/objectives-and-data/data-sources-and-methods/data-sources/current-population-survey-cps
• https://www.bls.gov/cps/cps_over.htm
• https://www.census.gov/topics/income-poverty/poverty/guidance/data-sources/acs-vs-cps.html
• https://www.bls.gov/cps/population-control-adjustments-2023.pdf

<cite index="2-6,2-12">The Annual Social and Economic Supplement (CPS-ASEC) is an annual supplement to the monthly CPS, conducted in February, March, and April</cite>. <cite index="2-13">Beyond the usual monthly labor force data, ASEC provides information on work experience, income, noncash benefits, and migration for persons ages 15 and older</cite>. <cite index="4-5,4-6,4-7">The survey covers more than 75,000 households and includes detailed social and economic questions for each household member, with income questions referring to the previous calendar year</cite>.

<cite index="1-6,1-7">The ASEC typically features an expanded sample of more than 75,000 households with about 70,000 interviews</cite>, though <cite index="1-8">the 2023 ASEC had about 57,000 households</cite>. <cite index="2-2">Data on employment and income refer to the preceding year, although demographic data refer to the time of the survey</cite>—this timing matters when March 2024 interviews are used to produce 2023 income and poverty estimates.

<cite index="5-4,5-5">The panel component of CPS affords the opportunity to observe short-term change in family structure, employment, income, and poverty; ASEC, with its wealth of information on income, health insurance, benefits receipt, and many other topics, is particularly popular for this purpose</cite>. <cite index="8-1,8-2">ASEC is the most widely used type of CPS data, but it is cumbersome to use as part of a longitudinal panel, especially linking to non-March months</cite>.

<cite index="2-8,2-9,2-10,4-12,4-13,4-14">The CPS-ASEC sample does not include observations for most counties—about 1,300 of the country's more than 3,100 counties are in the sample, as the design prioritizes consistent national-level estimates and reliable annual state-level unemployment estimates</cite>.

Sources:

• https://odphp.health.gov/healthypeople/objectives-and-data/data-sources-and-methods/data-sources/current-population-survey-annual-social-and-economic-supplement-cps-asec
• https://www.census.gov/programs-surveys/saipe/guidance/model-input-data/cpsasec.html
• https://www.pewresearch.org/social-trends/2023/12/14/older-workers-2023-methodology/
• https://pmc.ncbi.nlm.nih.gov/articles/PMC10805443/
• https://pmc.ncbi.nlm.nih.gov/articles/PMC6010043/

<cite index="10-3,12-2">The CPS is a scientifically selected multistage probability sample designed to represent the civilian noninstitutional population of each state and the nation</cite>. <cite index="10-8,12-4">About 74,000 housing units are assigned each month, with around 60,000 eligible for interview</cite>.

<cite index="11-7,11-8,11-9">Households are surveyed for 4 consecutive months, out for 8, then back for another 4 months before leaving permanently—a design that ensures continuity month-to-month and year-over-year while limiting respondent burden</cite>. <cite index="14-10,14-11,14-12">The 2025 sample redesign, based on the 2020 Census, began phasing in April 2025 and will complete in July 2026, with one rotation group replaced per month</cite>. <cite index="14-8,14-9">Historically, CPS samples are redesigned after each decennial census; the last redesign occurred in 2014 using 2010 Census data</cite>.

<cite index="10-12,17-9,17-10">The sample starts with Primary Sampling Units (PSUs)—1,987 of them nationwide—stratified within each state by labor force and socioeconomic characteristics correlated with unemployment, with one PSU sampled per stratum proportional to population</cite>. <cite index="17-11,17-13,17-14">Within each PSU, a systematic sample of housing units is drawn, grouped into clusters of about four units called Ultimate Sampling Units, with all households in the USU typically included</cite>. The sample is address-based, not person-based. <cite index="3-4">Beginning in 2014, the sampling frame is derived annually from the Master Address File with updates from USPS</cite>.

<cite index="11-1,11-2,11-3">One person—the reference person, usually the owner or renter—responds for all eligible household members, though direct contact is attempted if the reference person lacks knowledge of others' employment status</cite>.

Sources:

• https://www.bls.gov/opub/hom/cps/design.htm
• https://www.census.gov/programs-surveys/cps/technical-documentation/methodology.html
• https://www.census.gov/programs-surveys/cps/technical-documentation/methodology/sampling.html
• https://cps.ipums.org/cps/sample_designs.shtml
• https://www.bls.gov/cps/methods/sample_redesign_2025.htm
• https://www.census.gov/topics/income-poverty/poverty/guidance/data-sources/acs-vs-cps.html

<cite index="25-14,25-15">The JOLTS sample of 21,000 establishments does not directly support the production of sample-based state estimates. However, state estimates have been produced by combining the available sample with model-based estimates.</cite> <cite index="1-13">These estimates consist of four major estimating models: the Composite Regional model (unpublished), the Synthetic model (unpublished), the Composite Synthetic model (published historical series), and the Extended Composite Synthetic model (published current-year monthly series).</cite>

<cite index="6-1,6-2,6-3">Metropolitan Statistical Area (MSA) estimates have been produced for the 18-largest MSAs by combining available sample with model-based estimates, and smoothed by taking a 3-month moving average. These data are experimental and have not been subject to the same level of review as the current official JOLTS national and regional estimates.</cite>

When I cite JOLTS for a city—Cincinnati quits, Charlotte hires—I am citing a synthetic projection, not a direct survey. The underlying sample in any given metro may be dozens of establishments, not hundreds. The model borrows strength from regional and national patterns and from QCEW establishment counts. It is useful. It is not raw observation. Treat state and metro JOLTS as directional, not precise.

Sources:

• https://www.bls.gov/jlt/jlt_statedata_methodology.htm
• https://www.bls.gov/jlt/jlt_msadata_methodology.htm

<cite index="5-4,5-5,5-6">The JOLTS figure for hires minus separations can be used to derive a measure of net employment change, which should be comparable to the net employment change from the much larger CES survey. However, definitional differences between the two surveys, as well as sampling and nonsampling errors, historically caused JOLTS to diverge from CES over time.</cite> <cite index="5-7,5-8">To limit the divergence and improve the quality of the JOLTS hires and separations series, BLS implemented the monthly alignment method, which has four steps: seasonally adjust, align, back out the seasonal adjustment factors, and seasonally adjust again.</cite>

<cite index="2-1">The methodology improvements included revision of the JOLTS sample design to incorporate new business births more quickly and to remove business deaths from the frame on a more timely basis, and modification to data collection, editing, and review procedures in specific industries where research indicated a prevalence of particular response errors.</cite> <cite index="18-3,18-4">Each quarter the newly updated LDB is reviewed to identify birth establishments and a supplemental sample of these units is drawn and added to the survey; out-of-business units are dropped quarterly. Thus, the JOLTS sample is refreshed quarterly rather than annually.</cite>

The alignment is structural. JOLTS is no longer independent of CES—it is proportionately adjusted to match CES employment change. When you read that hires and quits track each other neatly now, part of that is methodology, not just labor market signal.

Sources:

• https://www.bls.gov/news.release/jolts.tn.htm
• https://www.bls.gov/jlt/methodologyimprovement.htm

<cite index="12-1">Because new and short-lived universe establishments cannot be reflected in the sampling frame immediately, the JOLTS sample cannot capture job openings, hires, and separations from these establishments during their early existence.</cite> <cite index="9-6,9-7">The time lag from the birth of an establishment until its appearance on the sampling frame is approximately one year. In addition, many of these new units may fail within the first year.</cite>

<cite index="12-2,12-3">BLS has developed a birth/death model that uses establishment birth and death activity from previous years as collected by the QCEW and projects forward to the present using over-the-year change in the CES. The birth/death model also uses historical JOLTS data to calculate the amount of churn (the rates of hires and separations) in establishments of various sizes.</cite> <cite index="12-4">The model combines the calculated churn with the projected employment change to estimate the number of hires and separations in establishments that cannot be measured through sampling.</cite>

This matters for anyone watching fast-moving sectors. When tech hiring surged in 2021 or softened in 2023, some of that was happening in establishments too young to be on the frame. The birth/death model is a patch, not a mirror. It imputes churn using past patterns, which means the newest dynamics lag by design.

Sources:

• https://www.bls.gov/news.release/jolts.htm
• https://www.bls.gov/news.release/archives/jolts_04112017.htm
• https://www.bls.gov/jlt/methodologyimprovement.htm

<cite index="17-1,17-2">JOLTS draws approximately 21,000 nonfarm business and government establishments in a stratified random sample by ownership, region, industry sector, and establishment size class.</cite> <cite index="17-6,17-7">The sampling frame comes from two sources: the Quarterly Census of Employment and Wages (QCEW) and the Federal Railroad Administration, with QCEW covering approximately 95 percent of nonfarm payroll jobs.</cite> <cite index="19-6">The establishments are drawn from a universe of over 9.1 million establishments compiled as part of QCEW operations.</cite>

<cite index="20-4,20-5">The basic sample unit is an establishment at a single physical location. Most sampled establishments remain in the survey for 36 months and, after completing time in sample, are not sampled again for at least 3 years.</cite> <cite index="20-9,20-10">There are six employment size classes: 1–9; 10–49; 50–249; 250–999; 1,000–4,999; and 5,000 or more. All establishments with 5,000 or more employees are included with virtual certainty.</cite>

The sample is not trivial but it is covering roughly 0.2% of the universe. When you see a JOLTS number move, remember it is weighted inference from this frame, not a census. <cite index="1-9">The 21,000-establishment sample does not directly support state-level estimates</cite>—those require model-based workarounds that blend sample with synthetic projections.

Sources:

• https://www.bls.gov/news.release/pdf/jolts.pdf
• https://www.bls.gov/opub/hom/jlt/design.htm
• https://www.bls.gov/jlt/sizeclassmethodology.htm

<cite index="16-1,16-3,16-4">In present-value terms, men lose an average of 1.4 years of predisplacement earnings if displaced in mass-layoff events when the national unemployment rate is below 6 percent, but lose a staggering 2.8 years of predisplacement earnings if displaced when the unemployment rate exceeds 8 percent</cite>. Timing is half the story. The same separation in a tight labor market costs you one year of lifetime earnings. In a recession, it costs you nearly three.

<cite index="14-11,14-12">Workers who were employed in more routine-intensive occupations suffer larger and more persistent earnings losses, primarily due to a reduction in the number of days in employment, suggesting that routine-intensive workers face considerable difficulties in coping with job loss</cite>. The losses aren't just wage cuts—they're employment gaps. Technology eroded the next-best option before the layoff arrived.

<cite index="11-7,11-8,11-9">After a mass layoff, women's earnings losses are about 35% higher than men's, with the gap persisting 5 years after displacement; this is partly explained by women taking up more part-time employment, but even women's full-time wage losses are almost 50% higher than men's, and parenthood magnifies the gender gap sharply</cite>. The event study reveals the divergence clearly—same layoff, same city, different trajectory. The method is neutral. The labor market is not.

Sources:

• https://pmc.ncbi.nlm.nih.gov/articles/PMC5521015/
• https://www.sciencedirect.com/science/article/abs/pii/S0927537120301573
• https://academic.oup.com/jeea/article/22/5/2108/7628307

The event study design has evolved. The modern variant combines the JLS longitudinal structure with propensity score matching and reweighting to ensure displaced and non-displaced workers are comparable on observable characteristics before the layoff occurs. <cite index="11-2,11-3">The reweighted event study design traces out the time path of labor market effects of job displacement, making it straightforward to compare groups with similar characteristics, and is complemented by a matched difference-in-difference design that allows investigation of heterogeneity in the displacement effect</cite>.

<cite index="12-1">The empirical strategy combines matching with an event study approach to trace employment and wages in regions hit by a mass layoff relative to suitable control regions</cite>. This lets you distinguish direct worker effects from spillovers—did your earnings drop because you were laid off, or because your city lost 2,000 manufacturing jobs and now every employer has more applicants per opening?

<cite index="23-5,23-6">Following the literature pioneered by Jacobson et al. (1993), researchers now compare displaced workers to a matched control group but embed the analysis in a flexible machine learning framework that allows treatment effects to vary jointly across a high-dimensional set of policy-relevant characteristics</cite>. The structure holds—identify the event, trace forward and backward, compare to controls—but the statistical apparatus has thickened. You can now separate losses by occupation, by city, by tenure, by the size of the layoff event itself, and ask which workers should receive which interventions.

Sources:

• https://academic.oup.com/jeea/article/22/5/2108/7628307
• https://academic.oup.com/jeea/article/18/1/427/5247011?guestAccessKey=33cef592-30fc-44a6-9b4d-3e207fbbb389
• https://arxiv.org/pdf/2307.06684

<cite index="8-2,8-3">The vast majority of studies on the earnings of displaced workers use a control group of never displaced workers, which attributes earnings declines due to all future job instability to the initial displacement event, overstating the losses relative to the average treatment effect</cite>. This matters for policy—if you're measuring lifetime scarring versus the impact of one discrete event, you'll design different interventions.

<cite index="8-4,8-5">An alternative approach isolates the impact of an average displacement without conditioning on future displacement status in the control group, and in comparisons using PSID data, the estimated long-run earnings losses fall dramatically from 25 percent to as low as 5 percent</cite>. The difference isn't statistical noise—it's the compound effect of repeated displacement, which the canonical JLS approach attributes entirely to the first layoff.

<cite index="19-10,19-11,19-12">Earnings and wages remain approximately 9 percent below their expected levels six or more years after displacement, but much of this persistence can be explained by additional job losses in the years following an initial displacement; workers who avoid additional displacements have earnings and wage losses of 1 percent and 4 percent six or more years after job loss</cite>. The first layoff matters, but what happens next matters more. The method sees the first shock clearly. The second and third, it folds into the estimate.

Sources:

• https://pmc.ncbi.nlm.nih.gov/articles/PMC6167020/
• https://www.researchgate.net/publication/5196232_Worker_Displacement_and_the_Added_Work_Effect

<cite index="1-4">The event study framework was pioneered by Jacobson, LaLonde, and Sullivan (1993) to study the effect of displacements</cite>, and it remains the standard for measuring what happens to workers when jobs disappear. The core design: identify a cohort of workers separated from stable employment during mass layoffs—typically defined as firm-level employment contractions of 30% or more—then trace their quarterly earnings backward and forward in administrative data, comparing them to similar workers whose firms didn't shrink.

What they found in Pennsylvania in 1982 is canonical. <cite index="2-3,2-5,2-6">Workers affected by mass layoffs saw their earnings drop substantially and persistently; workers who had earned an average of $50,000 in 1979 and found new employment within three months saw earnings drop to $35,000, and four years later still earned $13,300 less than similar workers whose firms had not laid them off</cite>. <cite index="2-7">The cumulative financial loss over five years following the layoff was equivalent to an entire year's earnings</cite>.

The method reveals something most surveys miss: <cite index="5-3">earnings depart from their expected levels even before workers leave their firms</cite>. The separation is the observable event, but the earnings damage starts earlier—overtime gone, bonuses withheld, raises frozen. <cite index="6-6">There is little evidence that displaced workers' earnings will ever return to their expected levels</cite>. The loss is permanent, not a dip.

Sources:

• https://www.aeaweb.org/conference/2018/preliminary/paper/8nzaiEef
• https://1cademy.com/node/example-jacobson-lalonde-and-sullivans-study-on-displaced-worker-earnings/v7PbGoTN5zgf78aiUx38
• http://cpi.stanford.edu/_media/pdf/Classic_Media/Jacobson,%20Lalonde,%20and%20Sullivan_1993_Lifecourse,%20Family,%20and%20Demography.pdf
• https://www.researchgate.net/publication/4980603_Earnings_Losses_of_Displaced_Workers

<cite index="1-4">The method was proposed in a series of articles by Alberto Abadie and coauthors</cite>. <cite index="2-1,2-2">Abadie, Diamond, and Hainmueller developed the procedure for estimating treatment effects with a single treated unit and multiple control units, constructing weights so covariates and pre-treatment outcomes of the treated unit are matched by a weighted average of controls</cite>. <cite index="2-3">Weights are restricted to be nonnegative and sum to one, allowing the procedure to obtain weights even when the number of lagged outcomes is modest relative to control units</cite>.

<cite index="3-5,3-6">The method provides a systematic way to choose comparison units in comparative case studies, opening the door to precise quantitative inference in small-sample studies</cite>. <cite index="9-8">Abadie's 2021 Journal of Economic Literature review provides practical guidance on feasibility, data requirements, and methodological aspects</cite>.

<cite index="8-3,8-5">Recent work has proposed bias-correction techniques and penalized synthetic control estimators</cite>. <cite index="5-5,5-7">Abadie et al. expanded the method with cross-validation technique, though subsequent research found this relied on non-uniquely defined predictor weights</cite>. The method continues to evolve, but the core insight remains: when treatment is single-unit and aggregate, construct the counterfactual by weighting the units that didn't receive it.

Sources:

• https://en.wikipedia.org/wiki/Synthetic_control_method
• https://arxiv.org/pdf/1610.07748
• https://onlinelibrary.wiley.com/doi/abs/10.1111/ajps.12116
• https://www.aeaweb.org/articles?id=10.1257%2Fjel.20191450
• https://economics.mit.edu/sites/default/files/publications/Synthetic_Experiment-2.pdf
• https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6214290/

<cite index="9-4">The method provides reliable estimates in certain settings but may fail in others</cite>. <cite index="18-3,18-4">Synthetic control only includes control units similar to the treated unit, while difference-in-differences includes all control units even dissimilar ones</cite>. <cite index="18-6,18-7">In some cases synthetic control is more suitable than standard difference-in-differences, and results hold when time and unit placebo tests are applied where difference-in-differences results are not robust</cite>.

<cite index="24-1,24-2">The method constructs a weighted combination of control units representing what the treated group would have experienced absent treatment, allowing effects of unobserved confounders to change over time by re-weighting the control group to match pre-intervention characteristics</cite>.

Data quality matters. <cite index="11-25">The method works well when the control group balances pre-intervention outcomes and auxiliary covariates as much as possible</cite>. The pre-treatment fit is the tell. If the synthetic control tracks the treated unit closely before the intervention, the post-treatment gap is more credible. If the pre-treatment fit is poor, the counterfactual is speculation.

<cite index="9-1,9-2">Synthetic controls have become widely applied in empirical research in economics and social sciences, probably because of their interpretability and transparent nature, and practical guidance is increasingly available</cite>.

Sources:

• https://www.aeaweb.org/articles?id=10.1257%2Fjel.20191450
• https://pmc.ncbi.nlm.nih.gov/articles/PMC9806815/
• https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5111584/
• https://ideas.repec.org/a/bla/jecsur/v37y2023i2p409-445.html

<cite index="10-2">Dube and Zipperer pooled multiple case studies using synthetic controls to analyze minimum wage policies</cite>. <cite index="17-2,17-3">Peri and Yasenov used the method to study labor market effects of the Mariel boatlift refugee wave</cite>. <cite index="10-17,10-18">Munasib and Rickman analyzed regional economic impacts of the shale gas and tight oil boom</cite>.

The method works when treatment is geographically concentrated. <cite index="12-1,12-2">One study of a manufacturing establishment found employment in the treated region was negatively affected compared to synthetic regions, with short- and long-term effects differing across labor market segments and firm size</cite>. <cite index="12-3,12-4">Large manufacturing investment does not necessarily have positive or instantaneous impact on total regional employment, though it may provide potential for long-term diversification as employment in related activities grows</cite>.

<cite index="15-10,15-11,15-13">The method has been applied to identify causal effects of regional industrial policies, using combinations of other regions to construct synthetic controls for data-driven comparative case studies</cite>. The technique is especially useful when a labor shock—a plant closure, a policy change, an immigration wave—hits one place and you need to know what would have happened if it hadn't.

Sources:

• https://ideas.repec.org/p/ran/wpaper/wr-1142.html
• https://link.springer.com/article/10.1007/s10645-022-09417-5
• https://www.researchgate.net/publication/284748658_The_Synthetic_Control_Method_for_Comparative_Case_Studies_An_Application_Estimating_the_Effect_of_Managerial_Discretion_Under_Performance_Management
• https://www.sciencedirect.com/science/article/abs/pii/S016604621730011X

<cite index="1-5,1-6">The synthetic control method constructs a weighted average of control units—cities, regions, states—to recreate the trajectory a treated unit would have followed absent the intervention, with weights selected to match key predictors of the outcome</cite>. <cite index="7-2">The method was designed for events that happen at an aggregate level affecting a small number of large units</cite>.

<cite index="13-2,13-10">Card's 1990 study of the Miami labor market after the Mariel Boatlift—when immigration increased Miami's labor force by 7 percent in months—exemplifies the comparative case study problem</cite>. One city. One shock. Which city do you compare it to?

<cite index="1-7">Unlike difference-in-differences, synthetic control can account for confounders changing over time by weighting the control group to better match the treatment group before intervention</cite>. <cite index="21-1,21-2">Geographic proximity is a poor metric for similarity if regions have substantial political or cultural differences, and policies may spill across borders</cite>.

The method is transparent about what it's doing. <cite index="4-9,4-10">It provides a systematic way to select comparison units and explicitly specifies which units are used, facilitating detailed qualitative analysis between the case of interest and the controls</cite>. You see the weights. You see the match quality. You see which places are standing in for the counterfactual.

Sources:

• https://en.wikipedia.org/wiki/Synthetic_control_method
• https://economics.mit.edu/sites/default/files/publications/Comparative%20Politics%20and%20the%20Synthetic%20Control.pdf
• https://csml.princeton.edu/sites/g/files/toruqf911/files/documents/Abadie_SCM_6-2022.pdf
• https://www.nber.org/system/files/working_papers/w12831/w12831.pdf
• https://www.urban.org/sites/default/files/publication/89246/the_synthetic_control_method_as_a_tool_0.pdf

<cite index="20-2,20-3">The parallel trends assumption requires that the unobserved confounder's association with the outcome has the same absolute magnitude in both groups and is constant over time. This may not be credible in many settings—when the outcome is binary, a count, or polytomous, or when an uncontrolled confounder exhibits nonadditive effects on the outcome distribution.</cite>

<cite index="12-1,12-2,12-3,12-4">One solution is to find a covariate (e.g., adult employment) affected by the confounder (labor demand) but not by the policy (if you believe minimum wages don't affect adult employment). This covariate can reveal the dynamics of the confounding variable and adjust for it, yielding the impact of the policy change.</cite> This does not mean simply controlling for the covariate; it requires using it in a 2SLS or GMM estimator.

<cite index="14-1,14-2,14-3">Recent Bayesian approaches propose frameworks that estimate policy effects when parallel trends is violated. These introduce a formal sensitivity parameter representing the extent of violation, specify an autoregressive AR(1) prior to model temporal correlation, and explore prior specifications calibrated from pre-treatment data to support robust policy conclusions under violations.</cite>

<cite index="19-3,19-6,19-7">Newer methods integrate difference-in-differences with synthetic control in a doubly robust framework, identifying the average treatment effect on the treated under either the parallel trends assumption or the synthetic control assumption. This avoids the conventional trade-off and strengthens the credibility of causal estimates.</cite> The methodological frontier is not abandoning parallel trends but building estimators that remain valid when it bends.

Sources:

• https://journals.lww.com/epidem/fulltext/2024/01000/universal_difference_in_differences_for_causal.3.aspx
• https://blogs.worldbank.org/en/impactevaluations/revisiting-difference-differences-parallel-trends-assumption-part-ii-what-happens
• https://arxiv.org/pdf/2508.02970
• https://arxiv.org/pdf/2503.11375

<cite index="4-2,4-3">Many studies using difference-in-differences rely on data from multiple years, not just one pre-treatment and one post-treatment period like Card and Krueger. The variables of interest often vary only at a group level, such as the state, and outcomes are often serially correlated.</cite>

<cite index="4-4,4-5">In Card and Krueger's study, employment in each state is likely correlated within the state and also serially correlated. Bertrand, Duflo, and Mullainathan (2004) showed that conventional standard errors often severely understate the standard deviation of the estimators—standard errors are biased downward, too small, and therefore overreject the null hypothesis.</cite>

This is not a minor technical issue. It means that many published difference-in-differences results, especially those using state-level policy variation across many years, report confidence intervals that are too narrow. The treatment effects look more precise than they are. The fix involves clustering standard errors at the level where treatment varies (e.g., the state), using block bootstrap methods, or collapsing time into pre/post periods to reduce serial correlation.

The Bertrand correction became standard practice after 2004. But it's a reminder: the elegance of the difference-in-differences estimator does not guarantee valid inference. You can identify the right parameter and still get the uncertainty wrong.

Sources:

• https://mixtape.scunning.com/09-difference_in_differences
• https://uclspp.github.io/PUBL0050/5-panel-data-and-difference-in-differences.html

<cite index="16-1,16-2">If both groups would have evolved similarly absent treatment, any post-treatment divergence in outcomes can be attributed to the intervention. The assumption that both groups would have followed parallel trends is essential.</cite> But essential is not the same as verifiable.

<cite index="14-5,14-6">A key assumption in difference-in-differences is that, absent treatment, the two groups would have followed parallel trends. This is often difficult to justify in real-world policy evaluations, where pre-treatment dynamics, time-varying confounding, or unobserved effect modifiers may violate the condition.</cite>

The standard defense is visual: <cite index="15-1,15-2">you test the parallel trends assumption by examining pre-treatment trends in the outcome variable for both treatment and control groups.</cite> If the lines move together before the policy, you argue they would have continued together after. But <cite index="14-9,14-10">pre-trends testing—checking whether treated and control groups followed similar trajectories before treatment—is underpowered in short panels.</cite>

<cite index="16-12,16-13,16-17">If one city had already begun experiencing improvements before the intervention due to stronger fundamentals or demographic advantages, pre-existing differences would violate the parallel trends assumption. The method would incorrectly attribute those pre-existing improvements to the policy, inflating the treatment effect.</cite> This is the methodological knife edge: difference-in-differences gives you clean causal estimates only when the counterfactual trend is genuinely parallel, and you never observe that counterfactual.

Sources:

• https://ls-analytics.com/parallel-trends-the-make-or-break-assumption-for-difference-in-differences/
• https://arxiv.org/pdf/2508.02970
• https://www.numberanalytics.com/blog/difference-in-differences-applications-examples

<cite index="6-1,6-2">In 1992, New Jersey raised its minimum wage by $0.80 while neighboring Pennsylvania did not. Card and Krueger compared fast-food employment on both sides of the state border, using Pennsylvania as a control to reveal what would have happened to New Jersey absent the policy change.</cite> The method hinges on a single untestable claim: <cite index="3-8,7-2">that average outcomes in both states would have followed the same trend absent the intervention—the parallel trends assumption.</cite>

<cite index="6-4,6-8">Their estimate suggested the $0.80 wage increase caused an average increase of 2.75 full-time jobs per restaurant in New Jersey.</cite> That finding violated the standard competitive labor model's prediction and became one of the most contested empirical results in labor economics.

The structure was simple: two places, two periods. <cite index="2-1">The difference-in-differences estimator—the difference in changes before and after policy—is also the average treatment effect on the treated (ATT).</cite> The elegance is that you never need to observe the counterfactual directly. You infer it from Pennsylvania's change.

<cite index="3-9,7-3">Since Card and Krueger's 1994 study, the method has expanded in many directions: heterogeneous treatment effects, staggered treatment adoption across units and time, and robustness checks for functional form violations of parallel trends.</cite> But the core logic remains: find a comparison group, assume it would have moved in parallel, and measure the divergence.

Sources:

• https://scienceetbiencommun.pressbooks.pub/pubpolevaluation/chapter/difference-in-differences-method/
• https://www.stata.com/meeting/colombia21/slides/Colombia21_Pinzon.pdf
• https://arxiv.org/pdf/2512.08759
• https://arxiv.org/html/2512.08759

<cite index="7-3,7-4">IPUMS USA contains harmonized census and American Community Survey (ACS) data from 1790 to the present. For the period 1850 to 1940, IPUMS includes 100% of individuals in the decennial censuses.</cite> <cite index="5-5,5-6">IPUMS USA makes freely available to researchers worldwide complete count U.S. Census microdata through 1940. This dataset includes over 650 million individual-level (1850-1940) and 7.5 million household-level records (1790-1840).</cite>

<cite index="2-6">The initial release of IPUMS data came not long after the development of the first web browsers, and Ruggles was quick to take advantage of this technology to disseminate the harmonized census microdata, leading to a rapid increase in the use of the IPUMS database for research.</cite> The web delivery mattered—this wasn't a dataset you requested on tape and waited six weeks to receive. You could extract exactly the variables and years you needed.

For labor analysis specifically, <cite index="7-7">the available information in the censuses and ACS varies by year, but generally includes basic housing data, demographic data, economic data (occupation, industry, income, work status), and other individual characteristics (migration, disability, veteran status).</cite> IPUMS CPS adds the monthly Current Population Survey, which <cite index="26-2">includes additional information on participation in welfare programs, job searching, barriers to looking for work, general health, union membership, hourly wages, health insurance, and tax liability.</cite> Different data for different questions, but all harmonized with the same methodological approach Ruggles designed in the early 1990s.

Sources:

• https://www.ipums.org/projects/ipums-usa/d010.v6.0
• https://www.ipums.org/projects/ipums-usa/d014.v3.0
• https://www.census.gov/newsroom/blogs/research-matters/2012/08/steven-ruggles-census-data-processing-part-2.html
• https://www.ipums.org/projects/ipums-cps

<cite index="24-1,24-2">The definition of labor force participation is considerably different for the 1850-1930 censuses than for the 1940-2000 censuses, the ACS and the PRCS. From 1850 to 1930, participation is defined as reporting any gainful occupation</cite>, which meant if you said you had an occupation, you were counted. <cite index="24-4,24-8">For the 1940-2000 censuses, the ACS and the PRCS, participation follows the modern labor force definition: within a specific reference week, having a job from which one is temporarily absent (e.g., on vacation), working, or seeking work.</cite>

<cite index="22-5,22-6">Before 1940, a person was recorded as having an occupation if he or she was "gainfully employed" in the previous year. This amorphous concept posed particular problems of interpretation with respect to children, women, and seasonal employment.</cite> The 1940 shift mattered for who got counted and when.

<cite index="19-1,19-3">Occupation and industry are among the most important variables for analyses of long-term social change because the early census years provide few alternative indicators of socioeconomic status or labor-force participation.</cite> If you're asking how women's work outside the home changed between 1900 and 1980, you need to know that the measurement concept changed in 1940—and IPUMS documents that break. The harmonized variable LABFORCE exists, but the metadata tells you when comparability is real and when it's stitched together across a definitional shift.

Sources:

• https://usa.ipums.org/usa-action/variables/LABFORCE
• https://usa.ipums.org/usa/chapter4/chapter4.shtml
• https://usa.ipums.org/usa/intro.shtml

<cite index="13-1">Most variables in the IPUMS have a composite coding structure, where the first digit is largely comparable across samples, and second and subsequent digits provide progressively more detail available in some samples and not others.</cite> This was a methodological choice: you want broad comparability but you also don't want to discard the rich detail available in, say, the 1990 census just because 1850 didn't ask those questions.

For occupation specifically, IPUMS created multiple harmonized variables. <cite index="11-2,11-3">The IPUMS variable OCC1950 harmonizes raw occupation codes going back as far as 1850 into the 1950 Census Bureau occupational classification system. Similarly, the IPUMS variable OCC1990 harmonizes occupation codes from 1950 to 2000 into the 1990 Census Bureau occupational classification system.</cite> <cite index="17-1,17-4">OCC2010 was developed to enhance the comparability of occupational data by providing a consistent set of occupational codes for IPUMS USA from 1950 forward.</cite> Researchers who don't need time comparisons can use the original unharmonized codes—<cite index="15-1">separate, unrecoded variables, called simply OCC (Occupation) and IND (Industry)</cite>—to avoid what the documentation calls "the anachronisms that are inevitably part of the harmonization process."

The work of translating between coding schemes is done using crosswalks. <cite index="18-4,18-5">According to the Census Bureau crosswalk, 73% of the 2002 occupation category Human Resources, Training, and Labor Relations Specialists would be classified as Human Resource workers in the 2012 Census occupation coding scheme. Thus, IPUMS recodes the 2002 code 0620 to the 2012 code 0630 in the variable OCC2010.</cite> This is approximate, by design. You get consistency at the cost of precision.

Sources:

• https://usa.ipums.org/usa-action/faq
• https://assets.ipums.org/_files/mpc/wp2019-01.pdf
• https://usa.ipums.org/usa/chapter4/chapter4.shtml
• https://usa.ipums.org/usa-action/variables/occ2010
• https://blog.popdata.org/occ2010-atus/

<cite index="2-4,2-5">Steven Ruggles initiated the Integrated Public Use Microdata Series (IPUMS) project to "harmonize" all the census public use samples—that is, to produce new versions of these datasets with consistent codes, record layouts, and integrated documentation without any loss of information from the original datasets.</cite> Before IPUMS, <cite index="2-3">none of the samples were produced in a consistent manner</cite>, meaning researchers couldn't easily track the same phenomena across time.

<cite index="9-1,9-3">For IPUMS, there are three elements to variable harmonization: applying consistent codes across samples, determining labels for those codes, and collating integrated variable descriptions that speak to issues not sufficiently conveyed by codes and labels.</cite> The process is fundamentally a metadata problem. <cite index="10-4,10-5,10-6">The central harmonization challenge is to equate codes that have the same meaning for a variable that is common across samples. This is fundamentally a metadata issue. One must understand the meaning of the codes, which is conveyed by their labels, the coding structures, and by the deeper context of the census questionnaire text and enumerator instructions.</cite>

What Ruggles built was infrastructure for asking: how many people worked as clerks in 1920 versus 1970? The Census Bureau changed its occupation codes every decade. <cite index="19-4,19-5">The Census Bureau has modified its classification systems every decade, so all comparisons of occupation and industry require extensive reconciliation of codes. There are twelve different occupational classification systems consisting of between 285 and 550 categories each.</cite> IPUMS made the reconciliation work reusable.

Sources:

• https://www.census.gov/newsroom/blogs/research-matters/2012/08/steven-ruggles-census-data-processing-part-2.html
• https://www.unece.org/fileadmin/DAM/stats/documents/ece/ces/ge.41/2017/Meeting-Geneva-Oct/WP31_ENG.pdf
• http://uaps2019.popconf.org/uploads/190528
• https://usa.ipums.org/usa/intro.shtml

<cite index="23-2,23-4">Robots had major impacts for wages and employment, especially for workers specializing in manual blue-collar tasks</cite>. <cite index="16-9">The declines in wages and employment fell much more heavily on workers from the lower half of the earnings distribution and those with less than a college degree, thus exacerbating inequality</cite>.

This is not an aggregate story averaged over the national economy. <cite index="19-2">The displacement effect holds prices and output constant — robots displace workers and reduce the demand for labor, because with robots it takes fewer workers to produce a given amount of output</cite>. The adjustment happens at the commuting-zone level first. When robot penetration rises in manufacturing, the local labor market adjusts through lower employment-to-population ratios, lower wages, and reallocation across sectors — not primarily through migration.

<cite index="16-1">Industries where more industrial robots are introduced experience declines in labor demand (especially for production workers) and sizable falls in their labor share</cite>. The city matters. Detroit was not an abstraction.

Sources:

• https://blueprintcdn.com/wp-content/uploads/2024/11/Blueprint-Discussion-Paper-2024.08-Acemoglu-Kong-Restrepo.pdf
• https://www.nber.org/system/files/working_papers/w25682/w25682.pdf
• https://irs.princeton.edu/sites/g/files/toruqf276/files/event/uploads/robots_and_jobs_march_3.17.2017_final.pdf

<cite index="24-7">New technologies create new tasks in which labor has comparative advantages, reinstating labor and creating new work</cite>. <cite index="11-1">The effect of new tasks on labor demand equals the productivity effect plus the reinstatement effect</cite>. This is not a guarantee that automation and new tasks net to zero. <cite index="9-5,9-14">New tasks can increase or reduce inequality depending on whether they are performed by skilled or unskilled workers</cite>.

<cite index="9-6,9-15">Industry-level data suggest that automation significantly contributed to the rising skill premium, while new tasks reduced inequality in the past but have contributed to inequality recently</cite>. The composition of new work matters. The question is not whether new tasks arrive — Acemoglu and Restrepo show they do — but who performs them.

<cite index="18-8">About half of employment growth over 1980–2015 took place in occupations where job titles or tasks performed by workers changed</cite>. Watch the occupational titles that appear on job postings. When they arrive in mid-tier cities first, or when they require credentials the local labor force does not yet hold, you are watching the reinstatement mechanism strain.

Sources:

• https://www.aeaweb.org/articles?id=10.1257/pandp.20201063
• https://docs.iza.org/sol3/papers.cfm?abstract_id=12293

The task-based framework that Acemoglu and Restrepo built moves past the complementarity story. <cite index="22-1,24-3,24-4">Production requires the completion of tasks, which can be performed by labor or capital, and the allocation of tasks determines the task content of production</cite>. <cite index="24-5,24-6">Automation shifts the allocation of tasks performed by labor toward capital, creating direct displacement effects</cite>.

Two forces compete. <cite index="18-1,18-2">Automation displaces labor from the tasks previously allocated to it, which shifts the task content of production against labor and always reduces the labor share — automation increases the size of the pie, but labor gets a smaller slice</cite>. <cite index="22-10,22-11">The presumption that all technologies increase aggregate labor demand simply because they raise productivity is wrong; some automation technologies may reduce labor demand because they bring sizable displacement effects but modest productivity gains</cite>.

<cite index="13-1,13-5">Industrial robots are associated with lower labor share and labor demand at the industry level and lower labor demand in local labor markets exposed to this technology</cite>. <cite index="25-6">Estimates suggest that an extra robot per 1,000 workers reduces the employment to population ratio by 0.18–0.34 percentage points and wages by 0.25–0.5%</cite>. That is arithmetic, not anger.

Sources:

• https://shapingwork.mit.edu/wp-content/uploads/2023/10/acemoglu-restrepo-2019-automation-and-new-tasks-how-technology-displaces-and-reinstates-labor.pdf
• https://docs.iza.org/sol3/papers.cfm?abstract_id=12293
• https://cepr.org/voxeu/columns/robots-and-jobs-evidence-us

<cite index="1-7,3-4">Acemoglu's foundational 2002 argument is that technical change has been skill-biased during the past sixty years</cite>, driving the returns to schooling upward and widening wage inequality. But the mechanism is not exogenous technology falling from the sky. <cite index="3-6,3-7">Most technical change during the nineteenth century was skill-replacing, because the increased supply of unskilled workers in the English cities made these technologies profitable</cite>. <cite index="1-1,3-8">The twentieth century flipped: rapid increase in the supply of skilled workers induced the development of skill-complementary technologies</cite>.

The theory of directed technical change posits that <cite index="8-3">a high proportion of skilled workers in the labor force implies a large market size for skill-complementary technologies, and encourages faster upgrading of the productivity of skilled workers</cite>. The implication: <cite index="8-1,8-4">an increase in the supply of skills reduces the skill premium in the short run, but then it induces skill-biased technical change and increases the skill premium, possibly even above its initial value</cite>.

This is not a claim about immutable technological progress. It is a claim about profit-maximizing innovation responding to the shape of the labor force. When you watch cities where the educated settle, you are also watching where the next generation of skill-complementary technology will be most profitable to develop.

Sources:

• https://www.aeaweb.org/articles?id=10.1257%2F0022051026976
• http://web.mit.edu/daron/www/nber.pdf
• https://academic.oup.com/qje/article-abstract/113/4/1055/1916970

<cite index="4-2">Freeman argued that management opposition has been a major factor in the decline of unions</cite>. The claim sounds soft until you track what happened to union election win rates and the time between petition and vote.

<cite index="2-1">Freeman extended his work to employee stock ownership plans (ESOPs) as private sector unions declined and wrote extensively about the ability of ESOPs to decrease economic inequality</cite>. The shift wasn't incidental. When collective bargaining structures dissolved, Freeman looked for other institutional forms that could compress wages.

<cite index="20-12,20-13">The union wage premium is typically stronger at the lower end of the wage distribution, such that strong unions are associated with lower wage inequality—a result more or less undisputed in the literature</cite>. <cite index="22-6">Unions function as a moral economy that shapes the behavioral norms of labor practices</cite>, per Western and Rosenfeld's extension of Freeman's work. When unions leave, the norms leave.

<cite index="22-10,22-11">Unions' capacity to reduce income inequality has become increasingly weaker with declining unionization—the positive association between labor strikes and wages is no longer detectable in the post-1984 period when union density falls below 20 percent</cite>. Below 20%, the threat stops working. Below 10%, the memory fades.

Sources:

• https://www.nber.org/system/files/working_papers/w11410/w11410.pdf
• https://www.aeaweb.org/about-aea/honors-awards/distinguished-fellows/richard-freeman
• https://www.econstor.eu/bitstream/10419/300248/1/GLO-DP-1457.pdf
• https://www.sciencedirect.com/science/article/pii/S0049089X25000390

Freeman's 1988 JEP piece identified something structural: unions were moving where the law allowed them. <cite index="23-4">Trends in unionization were quite different in the two sectors, with rises in union membership in the public sector for both men and women, and declines in the private sector</cite>.

The effect on wages differed dramatically by sector. <cite index="26-6">On average, private sector unions reduce male inequality by 1.5 percent in the United States and female inequality by 0.6 percent</cite>. But <cite index="26-7,26-8">in the public sector, unions reduce wage inequality by 16.2 percent for U.S. males and by 10.7 percent for U.S. women</cite>. The public sector numbers are ten times larger.

<cite index="23-6">Card's estimates implied that unions reduced the variance of men's wages in the public sector by 12 percent in 1973-74 and 16 percent in 1993</cite>. <cite index="26-3">On the whole, unions reduce economy-wide wage inequality by less than 10 percent</cite>, but that aggregate hides the geographic and sectoral concentration.

The divergence Freeman documented wasn't just about density. It was about where wage compression still functioned and where it stopped. The private sector moved on. The public sector kept the structure.

Sources:

• https://www.aeaweb.org/articles?id=10.1257%2Fjep.2.2.63
• https://davidcard.berkeley.edu/papers/union-wage.pdf
• https://www.bls.gov/opub/mlr/2019/beyond-bls/can-unions-significantly-reduce-wage-inequality-depends-on-whether-youre-in-the-public-or-private-sector.htm

Freeman's empirical work established that <cite index="15-1">the union wage premium measures the percent difference between the wages of unionized workers and those of nonunionized workers with the same characteristics</cite>. But the more interesting finding: unions set wages for people who never joined.

<cite index="14-1,14-5">Research found that collective bargaining rights lead to substantial increases in union presence and modest increases in wages</cite>, citing Freeman and Valletta's 1988 work alongside later studies. The modest wage effect understates the reach. <cite index="16-3">The threat of unionization raises wages in the non-union sector</cite>—employers paid more to keep organizers out.

<cite index="15-5">Had private-sector union density in 2013 remained at its 1979 level, weekly wages of nonunion men in the private sector would have been 5% higher</cite>. <cite index="28-2">Nonunion men of all education levels would earn 5 percent higher weekly wages in 2013 if union density remained at 1979 levels, an increase of $2,704 in annual paychecks</cite>. <cite index="28-1,28-5">For the 73.1 million full-time nonunion private sector workers, there is a $133 billion loss in annual wages because of weakened unions</cite>.

The spillover effects matter more than the direct wage premium. Freeman measured what happened when the possibility of unionization left the room.

Sources:

• https://www.epi.org/publication/eroded-collective-bargaining/
• https://www.brookings.edu/wp-content/uploads/2017/10/wp35-frandsen1.pdf
• https://wol.iza.org/uploads/articles/35/pdfs/union-wage-effects.pdf
• https://www.epi.org/press/decline-in-union-density-costs-nonunion-workers-133-billion-annually-in-lost-wages/

<cite index="3-2,3-5">Richard Freeman traced how U.S. labor market institutions changed remarkably between the 1950s and the 1980s</cite>, documenting a divergence that most observers missed: <cite index="3-7">unions expanded in the public sector while contracting in the private sector</cite>. His 1988 Journal of Economic Perspectives piece asked the questions that still structure the field.

The foundational text is Freeman and Medoff's 1984 What Do Unions Do?, which <cite index="11-1">showed that unions do much more than simply raise wages as an economic monopolist</cite>. <cite index="12-5">The authors found that unions and collective bargaining have real economic effects on diverse nonwage variables which cannot be explained by price-theoretic responses to union wage effects</cite>. They weren't just negotiating pay. They were setting the terms.

<cite index="15-2">Freeman and Medoff firmly established in the late 1970s that collective bargaining leads to more equal wage outcomes</cite>, findings popularized in their 1984 book. <cite index="27-5">The rise of unions from 1936 to 1968 explains about 25% of the decline during that period in the Gini coefficient</cite>. After 1968, the direction reversed: <cite index="27-7">falling union membership explains roughly 10% of increasing income inequality over the next five decades</cite>.

<cite index="21-1,21-4">Union membership declined from about 25% to 10% of American workers since the 1980s</cite>. The structure didn't just weaken. It dissolved.

Sources:

• https://www.aeaweb.org/articles?id=10.1257%2Fjep.2.2.63
• https://www.ssrn.com/abstract=349124
• https://www.nber.org/system/files/working_papers/w0837/w0837.pdf
• https://www.epi.org/publication/eroded-collective-bargaining/
• https://pmc.ncbi.nlm.nih.gov/articles/PMC9279886/
• https://journalistsresource.org/economics/inequality-labor-unions/

<cite index="18-4">In line with findings by Altonji and Shakotko (1987) and Altonji and Williams (2005), firm tenure plays a minor role for wage growth</cite>. This was not ideological; it was arithmetic derived from following the same workers across jobs. The standard story said: stay, get raises. The panel data said: the raises came from what you learned and where you went, not how long you stayed.

<cite index="18-6,18-7">The effect of experience is very important, particularly for workers who enter the labor market without training — over the first 5 years in the labor market, annual wage growth decreases from 11 percent to only 0.8 percent if experience effects are excluded</cite>. <cite index="3-5">From employment and schooling histories in the NLSY, researchers are able to construct measures of multiple dimensions of human capital investment, including whether work experience occurred simultaneously with schooling</cite>.

The methodological contribution is not the model but the match: panel data gave researchers the ability to separate what workers brought with them (ability, education) from what they gained on the job (experience, tenure) and what they found by moving (match quality). <cite index="19-1,19-2">Altonji's interests include labor market fluctuations, labor supply, wage determination, and econometric methods, with current research focusing on dynamic models of earnings, marriage, and family income, and the effects of undergraduate and graduate field of study on labor outcomes</cite>. What he built was a way to watch careers unfold in the data, not reconstruct them from cross-sections.

Sources:

• https://www.nber.org/system/files/working_papers/w18832/w18832.pdf
• https://pmc.ncbi.nlm.nih.gov/articles/PMC8653972/
• https://economics.yale.edu/people/joseph-altonji

<cite index="1-2,1-5">Joseph Altonji and Charles Pierret hypothesized that if the effects of hard-to-observe factors become more important in determining wages over time, the effects of easier-to-observe correlates used as proxies should become less important — they showed that education is used to statistically discriminate among young workers, but that race is not</cite>. This was wage determination as a watching game: employers learning what a worker actually is, not what the diploma said they might be.

<cite index="16-4,16-9,16-10">Altonji, Smith, and Vidangos used indirect inference to estimate a joint model of earnings, employment, job changes, wage rates, and work hours over a career, incorporating duration dependence, multiple sources of unobserved heterogeneity, job-specific error components in both wages and hours, and measurement error</cite>. <cite index="15-4">They found that human capital is responsible for most of earnings growth over a career, though with important roles for job seniority and mobility, and that unemployment shocks have substantial impacts on earnings in both the short and long run</cite>.

<cite index="3-3">Using data from the NLSY79 and NLSY97, Altonji, Bharadwaj, and Lange noted an increase in skills over time along with an overall widening of the skill distribution, which appears to be driven by trends in parental education</cite>. <cite index="5-1">Altonji, Bharadwaj and Lange constructed AFQT scores that are comparable across the NLSY79 and the NLSY97</cite> — the methodological plumbing required to watch what changed between cohorts, not just within them.

Sources:

• https://www.bls.gov/opub/mlr/2015/article/the-national-longitudinal-surveys-of-youth-research-highlights.htm
• https://en.wikipedia.org/wiki/Joseph_Altonji
• https://www.nber.org/system/files/working_papers/w14743/w14743.pdf
• https://pmc.ncbi.nlm.nih.gov/articles/PMC8653972/
• https://sites.google.com/yale.edu/joseph-altonji/publications

<cite index="8-6,8-7">The National Longitudinal Survey of Youth is a set of panel surveys in which the same respondents are interviewed periodically, sponsored by the U.S. Bureau of Labor Statistics primarily for researcher use rather than official statistics</cite>. <cite index="12-14,12-15">The NLSY79 began in 1979 with 12,686 men and women born in 1957-64, interviewed annually from 1979–1994 and biennially thereafter</cite>. <cite index="12-1,12-2">The NLSY97 began in 1997 with 8,984 men and women born in 1980-84, interviewed annually from 1997 to 2011 and biennially thereafter</cite>.

The design is patient. <cite index="8-1,8-2">A major innovation of the 1979 and 1997 cohorts is the elicitation of life histories — respondents report not only their current status on variables like employment and marital status, but also report and date any changes since the previous interview</cite>. <cite index="8-3,8-4">If a respondent misses biennial interviews, longitudinal data remain available from retrospective reports in the life history, ameliorating attrition though possibly at some cost in accuracy if memories are faulty</cite>.

<cite index="9-1">Data are suited for life course analysis due to breadth of topical areas: health, education, employment, household information, family background, marital history, childcare, income and assets, attitudes, substance use, and criminal activity</cite>. <cite index="8-8">By 2006 the bibliography of publications from these data included over four thousand items on topics such as transition from school to work, job mobility, youth unemployment, educational attainment and the returns to education, welfare recipiency, the impact of training, and retirement decisions</cite>. What makes a dataset canonical is not the breadth of the questions but whether researchers return to it, year after year, to watch what people actually did.

Sources:

• https://www.encyclopedia.com/social-sciences/applied-and-social-sciences-magazines/national-longitudinal-survey-youth
• https://en.wikipedia.org/wiki/National_Longitudinal_Surveys
• https://www.ncbi.nlm.nih.gov/books/NBK543729/

Standard quantile regression estimates conditional quantiles: the 50th percentile of wages for workers with X characteristics. <cite index="2-9,2-10">A conditional quantile function estimates percentile values in data subgroups under specific conditions, like median earnings for college graduates, and is integral to conditional quantile regression</cite>.

<cite index="2-1,2-11">Unconditional quantile methods, such as those proposed by Firpo, Fortin, and Lemieux (2009), capture the total effect of covariate changes on the entire wage distribution</cite>. You're estimating the effect on the 50th percentile of all wages, not the 50th percentile conditional on being in a particular education × experience cell.

The difference matters when you're trying to understand inequality. If you want to know whether unions compress the wage distribution, you need the unconditional effect—does unionization move the 10th percentile closer to the 90th? Conditional quantile regression can't answer that directly because it holds composition constant within cells.

<cite index="5-1,5-2,5-3">Autor, Katz, and Kearney use a quantile implementation of the extended DiNardo, Fortin and Lemieux procedure, where the original DFL procedure reweights observed wage densities to account for compositional shifts, extended by Lemieux to counterfactual residual wage densities</cite>. Both the reweighting and the RIF approaches are trying to get at the same object: the unconditional distribution, and how it would change.

Sources:

• https://www.tandfonline.com/doi/full/10.1080/00036846.2025.2567569
• https://www.nber.org/system/files/working_papers/w11628/w11628.pdf
• https://economics.mit.edu/sites/default/files/publications/rising%20wage%20inequality%202005.pdf

<cite index="34-3,34-4">Fortin, Lemieux, and Firpo (2011) provided a comprehensive overview of decomposition methods developed since Oaxaca and Blinder, used to decompose the difference in a distributional statistic between two groups, or its change over time, into explanatory factors</cite>. <cite index="34-5">While Oaxaca and Blinder considered the case of the mean, the handbook's main focus is on other distributional statistics besides the mean, such as quantiles</cite>.

The handbook lives in Volume 4 of the Handbook of Labor Economics (2011, pages 1–102). <cite index="2-4">The chapter discusses how DiNardo, Fortin, and Lemieux (1996) use kernel density estimates on reweighted samples for counterfactual wage distributions, how Firpo, Fortin, and Lemieux (2018) apply unconditional quantile regressions to decompose wage changes across the entire distribution, and how others developed techniques using conditional quantile regression</cite>.

<cite index="2-5">In all scenarios, the decomposition broadens the Oaxaca method, initially forged to investigate counterfactual disparities in average earnings</cite>. The chapter is methodological—identification assumptions, estimation procedures, what you can and can't claim. It's the place you return to when you need to know what a given decomposition is actually assuming, and whether those assumptions hold in your data.

If you're writing on wage inequality and plan to decompose anything, this is the reference that sits on your desk.

Sources:

• https://www.oreilly.com/library/view/handbook-of-labor/9780444534507/OEBPS/S0169721811004072.htm
• https://www.sciencedirect.com/science/article/pii/S0169721811004072
• https://www.tandfonline.com/doi/full/10.1080/00036846.2025.2567569
• https://www.nber.org/papers/w16045

<cite index="10-1">Firpo, Fortin, and Lemieux (2009) proposed running a regression of the recentered influence function (RIF) of the unconditional quantile on explanatory variables</cite>. <cite index="11-1,11-4">The method allows researchers to obtain partial effects of explanatory variables on any unconditional quantile of the dependent variable</cite>—not the conditional quantile (what standard quantile regression gives you), but the quantile of the whole distribution.

<cite index="16-7">The RIF is defined as the distributional statistic plus the influence function, which measures the marginal impact of a particular data point on the value of that statistic</cite>. You can compute a RIF for the median, the 90th percentile, the variance, the Gini—any distributional statistic you name.

<cite index="10-3">In the case of the mean, the RIF is simply the outcome variable itself, so RIF regression on the mean is identical to OLS</cite>. For quantiles it's more complex but still straightforward to compute. <cite index="11-2">The flexibility and simplicity of these tools have opened the possibility to extend the analysis to other distributional statistics using linear regressions or decomposition approaches</cite>.

This matters because you can finally run a regression—with all the interpretation habits that brings—on distributional objects that aren't the mean. You're not stuck with conditional quantile effects that tell you what happens within a covariate cell. You get the unconditional effect: what happens to the 10th percentile of all wages when unionization rises.

Sources:

• https://economics.ubc.ca/wp-content/uploads/sites/38/2013/05/pdf_paper_thomas-lemieux-unconditional-quantile-regressions.pdf
• https://journals.sagepub.com/doi/10.1177/1536867X20909690
• https://arxiv.org/pdf/2112.01435

<cite index="18-2,18-4,18-5">DiNardo, Fortin and Lemieux (1996) built a semi-parametric method to work on the entire distribution of wages, estimating a counterfactual distribution by replacing the marginal distribution of characteristics for one group with another using a reweighting factor</cite>. <cite index="19-1">The reweighting gives a counterfactual distribution of wages</cite>, answering: what would 1988 wages look like if people in 1988 had 1979 characteristics?

<cite index="18-6">The DFL reweighting method resembles propensity score reweighting used in program evaluation</cite>. You estimate the probability someone belongs to one period conditional on their observables, then reweight. <cite index="3-3">DiNardo, Fortin, and Lemieux compute the contribution of unions to the composition effect by contrasting actual changes to changes that would have prevailed if unionization rates had remained constant</cite>.

The technique matters because it doesn't force you to pick a functional form for wages. <cite index="20-1">The procedure reweights the sample distribution of a reference group such that the group's covariates distribution matches the covariates distribution of a comparison group</cite>. You're asking the data what the distribution would be, not asking a regression line.

This was the first decomposition to leave the mean behind and ask about the whole shape—10th percentile, 50th, 90th, everywhere at once.

Sources:

• https://byelenin.github.io/SS_Decomposition/Slides/Lec3_2020_SS.pdf
• https://ocw.mit.edu/courses/14-662-labor-economics-ii-spring-2015/8960ec728c5e6124a57522a8440e0a95_MIT14_662S15_Recitation1.pdf
• https://www.economics.uci.edu/files/docs/micro/f07/lemieux.pdf
• https://search.r-project.org/CRAN/refmans/ddecompose/html/dfl_decompose.html

<cite index="10-1,10-2,10-3">Athey's lab developed CAREER, a generative AI model based on 24 million resumes that predicts job transitions—important for labor economics questions including gender and racial differences in unemployment, replacing predictions previously based on much smaller survey datasets</cite>. <cite index="11-4">The model serves as a way to study the future of work and foundation models themselves, including how to reduce bias</cite>.

<cite index="14-1">Using administrative data from 180,000 laid-off workers in Sweden, Athey built machine learning techniques to examine heterogeneous impacts of job losses and identify which groups are most and least harmed</cite>. <cite index="14-3,14-4">AI allows economists to examine much more information than traditional economic research; in one project she used 23 million resumes to build a transformer model predicting career transitions, analyzing job histories' influence on wages and the gender wage gap</cite>.

<cite index="15-2,15-3,15-4,15-5">Traditional methods oversimplify complex variables like employment trajectories; Athey's work adapts foundation models that process rich, detailed histories to better explain group differences, providing more accurate and less biased estimates. Applied to real data, the approach reveals detailed histories explain more of the gender wage gap than previously understood</cite>. <cite index="12-3,12-8">Co-authored work includes decomposing changes in the gender wage gap over worker careers</cite> and <cite index="12-5,12-11">developing LABOR-LLM for language-based occupational representations with large language models</cite>.

Sources:

• https://www.gsb.stanford.edu/faculty-research/faculty/voices/susan-athey
• https://news.stanford.edu/stories/2024/09/susan-athey
• https://www.sipa.columbia.edu/news/susan-athey-delivers-kenneth-j-arrow-lecture-ai-and-future-work
• https://www.pnas.org/doi/10.1073/pnas.2427298122
• https://gsb-faculty.stanford.edu/susan-athey/tag/econometric-theory-and-machine-learning/

<cite index="18-1,18-7">Athey, Tibshirani, and Wager introduced Generalized Random Forests (GRF), extending causal random forest methods to estimate average treatment effects, conditional average treatment effects, conditional effects on the treated and untreated, and overlap-weighted effects</cite>. <cite index="20-6,20-7">Wager and Athey's 2018 paper on estimation and inference of heterogeneous treatment effects using random forests appeared in the Journal of the American Statistical Association</cite>, and <cite index="19-3,19-5">the generalized random forest paper published in Annals of Statistics in 2019</cite>.

<cite index="23-2">Most heterogeneous treatment effect algorithms are based on decision trees or random forests, including Generalized Random Forests</cite>. <cite index="23-6">To enable statistical inference, Athey, Tibshirani and Wager introduced "honest trees"</cite>—<cite index="18-2">honest trees use separate samples for training and estimation, similar to earlier causal tree methods but without requiring cross-validation since out-of-bag error is available</cite>.

<cite index="24-3,24-4,24-5,24-6,24-7">The generalized framework accommodates many statistical quantities: instrumental forests for settings with instrumental variables, causal survival forests for right-censored survival outcomes, and local linear forests</cite>. <cite index="27-2">The method is implemented in the grf package for R and C++</cite>. <cite index="6-5">Papers on recursive partitioning for heterogeneous causal effects and estimation using random forests helped reshape how researchers use machine learning for cause and effect</cite>.

Sources:

• https://www.causalmlbook.com/causal-trees-and-forests.html
• https://www.stats.ox.ac.uk/~evans/APTS/cf.html
• https://chenxing.space/blog/a-walkthrough-of-how-causal-forest-works/
• https://grf-labs.github.io/grf/
• https://www.statworx.com/en/content-hub/blog/machine-learning-goes-causal-ii-meet-the-random-forests-causal-brother
• https://pmc.ncbi.nlm.nih.gov/articles/PMC11966565/
• https://mdi.georgetown.edu/news/mdi-fall-2025-distinguished-lecture-ai-causal-inference-impact/
• http://faculty.ist.psu.edu/vhonavar/Courses/causality/GRF.pdf

<cite index="2-1,2-2,2-3">Athey's contribution sits at the seam between two traditions: machine learning methods built for prediction, and econometric causal inference concerned with estimating treatment effects</cite>. <cite index="4-15,4-16">Historically, supervised machine learning focused on forecasting, while causal inference in econometrics centered on measuring average treatment effects</cite>. The problem economists face is practical: <cite index="2-2">empirical analyses typically estimate counterfactual policies—what happens when you change a price, show an advertisement, implement a government program</cite>.

<cite index="1-3,1-4">Athey developed methods to estimate conditional average treatment effects and personalized treatment policies, covering randomized experiments, unconfoundedness settings, instrumental variables, and panel data</cite>. <cite index="9-4,9-5">The methods allow researchers to estimate heterogeneity in causal effects—which subpopulations experience large or small treatment effects—and test hypotheses about those differences</cite>.

The methodological shift matters because <cite index="2-3">machine learning's systematic approaches to model selection and prediction work particularly well with datasets containing many observations and covariates</cite>. For labor economists trying to understand, say, how a training program affects different workers or how wage trajectories differ across detailed occupational histories, that's the setting we actually face. <cite index="8-1,8-2,8-3">Athey and Imbens discussed the differences in goals, methods, and settings between machine learning and traditional econometrics, then identified specific methods important for empirical researchers</cite>.

Sources:

• https://gsb-faculty.stanford.edu/susan-athey/files/2022/04/phdmlsyllabus.pdf
• https://idss.mit.edu/calendar/idss-distinguished-seminar-susan-athey-stanford-university/
• https://www.gsb.stanford.edu/faculty-research/working-papers/machine-learning-estimating-heretogeneous-casual-effects
• https://www.linkedin.com/posts/nunocastrolopes_susan-athey-machine-learning-and-causal-activity-7105444582745608192-Ufyd
• https://arxiv.org/abs/1903.10075

<cite index="21-1">Farber (2010) finds that the decline in long tenure men is concentrated in the private sector</cite>, not economy-wide. <cite index="21-5,21-6">The inclusion of industry and occupation controls reduces the adjusted decline in the long-tenure rate somewhat, to about 4.5 percentage points; thus, shifts in the distribution of male employment by industry and occupation can explain about 1 percentage point of a 5½ percentage point demographically adjusted decline in long tenure (about 20 percent of the decline). The shift in employment away from the manufacturing sector, which had a higher-than-average share of long-tenure workers in the 1980s, can account for about half of the industry</cite> effect. Manufacturing held men in place; services did not.

<cite index="26-1,26-2">States has become less oriented toward long-term jobs. Since public-sector employment as a fraction of total employment has remained steady at about 18 to 20 percent and seems unlikely to increase, it appears that young workers today will be less likely than their parents</cite> to reach 20-year tenure. <cite index="24-9">Job stability for single women and private-sector males has declined, but public-sector long-term employment remains robust and married women have become increasingly likely to remain with their employer after marriage and childbirth</cite> — the public sector became the exception, not the norm.

Sources:

• https://www.nber.org/system/files/working_papers/w26694/w26694.pdf
• https://gceps.princeton.edu/wp-content/uploads/2017/01/171farber.pdf
• https://www.rsfjournal.org/content/11/1/224

<cite index="12-6">Job loss rates increased for women and for whites in the 1990s, as well as for college-educated and high tenure workers</cite> — a structural shift, not just cyclical churn. <cite index="12-4">Older workers with higher pre-displacement tenure, those who change industries, and those who experience multiple job losses thus experience greater earnings losses</cite> because the specificity of what they knew compounds. <cite index="12-3">Displaced workers' losses reflect both industry-specific decline and the loss of firm- and industry-specific skills</cite> — the seniority premium isn't portable.

The cyclicality matters more than it used to. <cite index="12-1">Men lose an average of 1.4 years of pre-displacement earnings if displaced in mass-layoff events that occur when the national unemployment rate is below 6 percent, and lose 2.8 years of pre-displacement earnings if displaced when the unemployment rate</cite> exceeds 8 percent. <cite index="17-10,17-11">Henry Farber of Princeton University finds that the rate of job loss, the difficulty of finding new employment, and the probability of finding only part-time work were all higher in the 2007–2009 crisis than in the economic contractions of the 1980s, 1990s, and early 2000s. The labor market consequences of job loss, Farber finds, were unusually severe and long-lasting.</cite> The Great Recession was not a normal recession for displaced workers.

Sources:

• https://pmc.ncbi.nlm.nih.gov/articles/PMC4553243/
• https://equitablegrowth.org/the-consequences-of-job-displacement-for-u-s-workers/

<cite index="17-4,17-5">In the first few months after the loss of long-tenure jobs, workers' earnings were nearly half of what they were prior to displacement. After 5 years, workers' earnings were, on average, still 15 percent below their pre-displacement earnings.</cite> <cite index="19-1">The long-term losses, even five years after displacement, are on average 25% of pre-displacement earnings.</cite> The magnitude varies by study, but the persistence does not: these are long-run losses.

The composition of that loss shifted over time. <cite index="20-4">The average reduction in weekly earnings for job losers making a full-time–full-time transition are relatively small, with a substantial minority reporting earning more on their new job than on the lost job.</cite> <cite index="20-5">Most of the cost of job loss comes from difficulty finding new full-time employment.</cite> Not wage compression — hours. <cite index="18-3,18-4">Many reemployed job losers are employed part time subsequent to job loss. Some of these workers lost part-time jobs but many had lost full-time jobs.</cite> The problem is not that the new job pays less per hour; the problem is the new job is twenty hours a week.

Sources:

• https://equitablegrowth.org/the-consequences-of-job-displacement-for-u-s-workers/
• https://www.chicagofed.org/publications/chicago-fed-letter/2004/october-207
• https://www.journals.uchicago.edu/doi/10.1086/692353
• https://www.nber.org/system/files/working_papers/w21216/w21216.pdf

<cite index="3-1">Job tenure and the incidence of long-term employment have declined sharply in the United States.</cite> <cite index="11-7,11-8">For male, private sector workers, the ten-year rate has fallen from 50 to 35 percent from 1973 to 2006, and the 20-year rate has fallen from</cite> roughly 35 percent to 20 percent. <cite index="24-1">An area of consensus is that private-sector long-term employment relationships have declined for men since the 1970s</cite> — but not for everyone. <cite index="23-4">Women have seen no systematic change in job durations or the incidence of long-term employment relationships in the private sector.</cite>

Farber's central finding is disorienting: <cite index="3-2">rates of job loss as measured by the Displaced Workers Survey (DWS), while cyclical, have not increased.</cite> The question becomes how tenure declined if displacement didn't rise. <cite index="3-7">One is that, while overall rates of job loss have not increased, rates of job loss for high-tenure workers have increased relative to those for lower-tenure workers.</cite> <cite index="3-9">Some of this seemingly voluntary job change (e.g., the taking of an offered buy-out) may reflect the kind of worker displacement that the DWS was meant to capture but is not reported as such by workers.</cite> The instrument may be missing what matters: older workers leaving under pressure, counted as quits.

Sources:

• https://ideas.repec.org/p/pri/indrel/520.html
• https://ecommons.cornell.edu/server/api/core/bitstreams/f8849559-3dcb-4a5e-8db4-bef5e8c9b9ac/content
• https://gceps.princeton.edu/wp-content/uploads/2017/01/171farber.pdf
• https://www.rsfjournal.org/content/11/1/224

<cite index="4-1,7-2,7-9">Borjas's 2006 Journal of Human Resources paper on native internal migration argued that geographic clustering of immigrants fails to measure labor market impact because natives respond by leaving.</cite> If low-skill natives move out of high-immigration cities, comparing Miami to Atlanta underestimates the national wage effect—the displaced workers diffuse the impact across the entire country.

<cite index="2-1,2-2,2-3,2-4">David Card's response: the leap from theory (labor demand curves slope down) to policy (immigration lowers native wages) ignores that increases in population don't necessarily cause falling wages—the effect depends on capital supply, worker characteristics, and technology structure.</cite> <cite index="2-5">Papers by Ottaviano-Peri and Manacorda-Manning-Wadsworth argued immigration impacts on native workers in the US and UK have been very small.</cite>

The methodological divide: Borjas treats the nation as the relevant labor market and measures wage effects in education-experience cells. Card and others study cities as quasi-experiments. Each approach has a blind spot. National models assume workers are immobile across cells but don't directly observe displacement. City studies observe outcomes but can't see the natives who left.

Sources:

• https://scholar.harvard.edu/files/gborjas/files/jhr2006.pdf
• https://gborjas.scholars.harvard.edu/_publications
• https://davidcard.berkeley.edu/papers/jeea2012.pdf

<cite index="19-1,19-7">Ottaviano and Peri (2007) reported evidence that immigrant and native workers are not perfect substitutes within skill groups, creating complementarities that could raise native wages.</cite> <cite index="22-3">Using Census data from 1990-2006, they found immigration had small negative short-run effects on native workers without a high school degree (-0.7%) but small positive long-run effects (+0.3%).</cite> <cite index="19-4,22-5">Average native wages rose 0.6% in the long run, while previous immigrants' wages fell 6.7%.</cite>

Borjas contested the substitution elasticity. <cite index="21-2,21-3">Borjas, Grogger, and Hanson (2012) replicated the Ottaviano-Peri analysis with standard labor economics assumptions and found the elasticity estimate increased to 125—statistically equivalent to perfect substitutes, erasing within-group complementarities.</cite> <cite index="19-8,19-9">The finding of imperfect substitution proved fragile, disappearing once the analysis adjusted for heterogeneity in labor market attachment.</cite>

The debate turns on production function architecture: whether immigrants and natives with identical paper credentials do observably different work. Ottaviano-Peri say occupational sorting creates complementarity. Borjas says the data, properly specified, show competition. Both are quantifying a thing the Census wasn't designed to measure.

Sources:

• https://www.researchgate.net/publication/280745183_Rethinking_The_Effect_Of_Immigration_On_Wages
• https://www.nber.org/papers/w14188
• https://cis.org/Report/Immigration-and-American-Worker
• https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1272569
• https://www.nber.org/system/files/working_papers/w14188/w14188.pdf

<cite index="13-8,13-10">In 1980, 125,000 mostly low-skill Cuban immigrants arrived in Miami from Mariel Bay, increasing Miami's labor force by 7%.</cite> <cite index="11-1,11-2">David Card's seminal 1990 paper found no wage impact; Borjas's 2015 reanalysis claimed the opposite.</cite> <cite index="15-2,15-9">Borjas estimated the Mariel Boatlift slashed wages of workers with less than high school by 10 to 30 percent.</cite>

The discrepancy came down to sample definitions and survey artifacts. <cite index="1-16,1-17,1-18">Borjas studied prime-age non-Hispanic men lacking a high school diploma and identified a Miami wage trough bottoming out in 1985.</cite> <cite index="13-4,13-7">But in 1980, the CPS improved coverage of low-skill black men; the fraction of non-Hispanic blacks in Borjas's subsample suddenly doubled, coinciding exactly with the Boatlift.</cite> <cite index="10-2,10-4">Clemens and Hunt's reanalysis, adjusting for the share of blacks, yields results similar to Card: little to no wage impact is discernable.</cite>

Mariel remains contested not because the data are ambiguous but because the question—what counts as the right counterfactual city, the right skill group, the right survey frame—has no neutral answer. The case study became a referendum on method.

Sources:

• https://www.judiciary.senate.gov/imo/media/doc/03-16-16%20Borjas%20Testimony.pdf
• https://davidroodman.com/blog/2017/05/25/four-points-on-the-debate-over-the-impact-of-mariel-boatlift/
• https://www.bruegel.org/2017/06/the-mariel-boatlift-controversy/
• https://www.cgdev.org/blog/what-mariel-boatlift-cuban-refugees-can-teach-us-about-economics-immigration
• https://www.nber.org/system/files/working_papers/w23433/w23433.pdf

<cite index="7-6,7-13">Borjas's 2003 Quarterly Journal of Economics paper—"The Labor Demand Curve is Downward Sloping"—established the claim that immigration reduces native wages by treating skill groups (education × experience cells) as national labor markets.</cite> The approach is built on <cite index="9-2,9-3">a three-level nested CES production function, where wage effects depend on how immigrants and natives substitute for one another in production.</cite>

<cite index="1-1">Borjas's simulations assume a 10 percent immigration-driven labor supply increase lowers wages by 3 percent.</cite> <cite index="3-3,3-5,3-7">The widely quoted finding: immigration reduced low-skill native wages by 8% in the short run, 4.8% in the long run, and had zero long-run effect on average native wages.</cite> <cite index="8-5">Another simulation estimated immigration reduced average annual earnings of native-born men by $1,700 or roughly 4 percent.</cite>

But the headline number traveled farther than the structure. <cite index="3-9">Native workers in the middle of the skill distribution actually gained.</cite> What became policy shorthand—"immigration hurts workers"—was never the full finding. It was always: immigration shifts wage distributions, and the direction depends on where you stand in the skill hierarchy and how long you wait for capital to adjust.

Sources:

• https://www.judiciary.senate.gov/imo/media/doc/03-16-16%20Borjas%20Testimony.pdf
• https://www.econlib.org/archives/2007/03/borjas_wages_an.html
• https://gborjas.scholars.harvard.edu/_publications
• https://cis.org/Report/Increasing-Supply-Labor-Through-Immigration

<cite index="24-21,24-22,24-23,24-24">Since the mid-2000s, estimates of the college wage premium suggest that it has stagnated, with little change in its level and perhaps an outright decline. The returns to college, which reflect the college wage premium combined with the costs of attending college, are a crucial element for choices about educational attainment. These financial returns likely have declined as college costs have risen and the wage premium has flattened, reducing the incentives to invest in college.</cite>

Card and Thomas Lemieux examined cohort patterns in the college premium. Their work asked: <cite index="26-3">Can Falling Supply Explain the Rising Return to College for Younger Men?</cite> The answer was nuanced — supply shifts mattered, but so did demand dynamics within narrow age cohorts. The college premium wasn't uniform across the lifecycle.

<cite index="27-20,27-21">The wage premiums for higher education generally have been rising over time. However, both data sets shows that the growth has slowed in recent decades, with the slowdown for the graduate group lagging behind that for the college-only group.</cite>

What changed? The arithmetic. The wage you earn minus the tuition you paid minus the years you didn't work. When tuition climbs faster than the wage lift, the return narrows even if the premium holds. Card's work on education returns always centered the counterfactual: what would this person have earned without the degree? That question matters more now, when the debt is larger and the premium has stopped climbing.

Sources:

• https://docs.iza.org/dp17717.pdf
• https://ideas.repec.org/a/oup/restud/v89y2022i1p142-180..html
• https://www.nber.org/system/files/working_papers/w22935/w22935.pdf
• https://marginalrevolution.com/marginalrevolution/2021/10/david-card-on-the-return-to-schooling.html

<cite index="17-1,17-2,17-4">To investigate how increased minimum wages affect employment, Card and Krueger used a natural experiment. In the early 1990s, the minimum hourly wage in New Jersey was raised from 4.25 dollars to 5.05 dollars.</cite> <cite index="17-7,17-8">Card and Krueger noted that there was no increase in neighbouring Pennsylvania. Of course, there were differences between the two states, but it is likely that the labour markets would evolve similarly close to the border.</cite>

The method was simple: survey fast-food restaurants before and after. <cite index="18-2,18-3,18-4,18-5">Fast-food restaurants in New Jersey were the treatment group. Fast-food restaurants just across the state border, in Eastern Pennsylvania, were the control group. Card and Krueger then looked at changes in employment in each state to see what the effect of the minimum wage was in New Jersey. They found that a modest increase in the minimum wage did not kill jobs.</cite>

<cite index="18-6">It was a bombshell for the economic world, challenging an orthodoxy that had dominated the field for decades.</cite> <cite index="1-9">Card won the Nobel Memorial Prize in Economic Sciences in 2021 for research showing that an increase in minimum wage does not lead to less hiring, and immigrants do not lower pay for native-born workers.</cite>

The paper was controversial. Critics pointed to data quality, timing, recession effects. But <cite index="1-5">later studies of minimum wage increases have tended to confirm Card and Krueger's findings.</cite> The method — the natural experiment — became the standard.

Sources:

• https://www.nobelprize.org/prizes/economic-sciences/2021/popular-information/
• https://www.npr.org/sections/money/2021/10/12/1045152279/a-nobel-prize-for-a-revolution-in-economics
• https://en.wikipedia.org/wiki/David_Card
• https://aaronmams.github.io/Card-Krueger-Replication/

<cite index="2-1,2-9,2-10,2-11">Card researched wage inequality and the driving forces behind it — essentially, examining how much one person earns versus another and why. According to Card, wage inequality is driven by things like education, gender, race, location, and other factors that people don't really know how to quantify like mathematical ability, ambition, and a willingness to work hard.</cite>

But when the 1980s and 1990s brought sharp increases in the college wage premium, the consensus explanation was skill-biased technological change (SBTC): computers were raising demand for educated workers faster than supply could catch up. Card, with John DiNardo, looked closer. <cite index="10-1,10-2,10-3">What DiNardo and Card found was that in the 1980s and '90s, there was no systematic tendency for wages in the lower-wage industries (like shoe manufacturing) to fall relative to wages in higher-wage industries (like steel). The wage structure was very stable. So if you believe that the industry differentials are due to skill differences, the patterns are not what you would expect from SBTC.</cite>

<cite index="10-4,10-5,10-6">A final puzzle concerned the age structure of the increases in the relative wages of college versus high school graduates. Wages of young college-educated workers rose relative to young high school workers, but for people over age 40 or so, there really wasn't any change in the high school/college premium.</cite> If technology was driving the change, it should have hit all ages. It didn't. Card and DiNardo pulled the facts together and said: here are the anomalies. The simple SBTC story doesn't fit.

Sources:

• https://www.ubs.com/global/en/our-firm/what-we-do/our-brand/nobel-perspectives/laureates/david-card.html
• https://www.minneapolisfed.org/article/2006/interview-with-david-card
• https://ideas.repec.org/p/ucd/wpaper/200817.html

<cite index="1-1,1-2">Card showed that people who lived closer to universities were more likely to enroll, and because of the additional years of schooling, this translated into higher earnings.</cite> The observation is simple: distance matters. It's a geographic constraint that changes schooling decisions before ability does.

<cite index="4-12,4-13">Education plays a central role in modern labor markets. Hundreds of studies in many different countries and time periods have confirmed that better-educated individuals earn higher wages, experience less unemployment, and work in more prestigious occupations than their less-educated counterparts.</cite> But Card asked a different question: is that correlation causal? Or do the people who choose more schooling already have traits that would have produced higher earnings?

His approach — <cite index="2-3">comparing outcomes in places affected by policy changes to similar places that were not</cite> — shaped an entire generation of labor research. He used instruments: compulsory schooling laws, proximity to college, the Vietnam draft lottery. Each one isolated a source of schooling variation that wasn't about the person, but about the circumstance.

<cite index="7-19,7-20">His 2001 Econometrica paper, "Estimating the Return to Schooling: Progress on Some Persistent Econometric Problems,"</cite> synthesized decades of work confronting selection bias. The returns he identified — typically 8-12% per year of schooling — held across methods. The consistency suggested something real.

Sources:

• https://en.wikipedia.org/wiki/David_Card
• https://www.ubs.com/global/en/our-firm/what-we-do/our-brand/nobel-perspectives/laureates/david-card.html
• https://davidcard.berkeley.edu/papers/causal_educ_earnings.pdf
• https://davidcard.berkeley.edu/papers.html

<cite index="4-5,4-6">The Diamond-Mortensen-Pissarides model exhibits a strong trade-off between cyclical unemployment fluctuations and the size of rents to employment—introducing endogenous job search effort reduces the strength of the trade-off while bringing the model closer to the data</cite>. The basic framework assumes workers and firms are homogeneous and wages are determined by Nash bargaining, which splits the surplus from a match.

But the model has known tensions. <cite index="4-7">Ignoring worker search effort leads to a large upward bias in the elasticity of matches with respect to vacancies</cite>. <cite index="4-1,4-8">Average search effort of the unemployed is subject to cyclical composition biases, and new evidence from merging the American Time Use Survey and the Current Population Survey supports procyclical search effort</cite>. When the economy contracts, the composition of the unemployed shifts—not just the number.

<cite index="16-2,16-4">The canonical search-and-matching model features atomistic firms in which the marginal product of labor remains above the value of unemployment for workers—these assumptions are critical because either implies that unemployment would disappear in the absence of matching frictions</cite>. Relaxing these assumptions—allowing for wage rigidity or diminishing returns—changes the interpretation of what portion of unemployment is frictional versus structural.

<cite index="15-3,15-4">For the US, frictional unemployment is around 36 percent of total unemployment, whereas for Spain approximately 20 percent of all unemployment is due to frictions—Spain is a country with more labor market rigidities than the US</cite>. The framework decomposes unemployment, but the decomposition depends heavily on institutional context.

Sources:

• https://www.sciencedirect.com/science/article/abs/pii/S0304393215000185
• https://research.upjohn.org/cgi/viewcontent.cgi?article=1000&context=dissertation_awards
• https://onlinelibrary.wiley.com/doi/abs/10.1111/manc.12154

<cite index="25-1,25-3">The Beveridge curve is a graphical representation of the inverse relationship between the unemployment rate and the job openings rate, plotted with unemployment on the horizontal axis and vacancies on the vertical axis—the curve typically slopes downward, indicating that as the unemployment rate falls, the vacancy rate rises</cite>. <cite index="26-4,26-5">William Beveridge, in his 1944 book Full Employment in A Free Society, argued that fluctuating unemployment levels are driven by changes in demand for workers, which implied a negative relationship between the number of job openings and the unemployment rate</cite>.

<cite index="11-10,11-11">The matching of workers to new jobs is one half of the explanation for large flows of jobs and workers between activity and inactivity—its outcome, in conjunction with the outcome of the process that separates workers from jobs, is often shown graphically in vacancy-unemployment space by the Beveridge curve</cite>. <cite index="30-2,30-3">The Beveridge curve is a central concept in the macroeconomics of labor markets, capturing an inverse relationship between unemployment and vacancies</cite>.

The curve matters because it compresses the entire state of the labor market into a scatter plot. <cite index="26-8">Within the business cycle, fluctuations along the Beveridge curve are expected</cite>. But <cite index="26-3">movement outward over time results in a fixed level of vacancies associated with a higher level of unemployment, interpreted as decreasing efficiency in the labor market</cite>. The curve can shift—and when it does, the interpretation is structural, not cyclical.

<cite index="30-6">The Beveridge curve is one of the most robust regularities in economics, as it holds in different time periods, across countries and at the aggregate and disaggregated level</cite>.

Sources:

• https://grokipedia.com/page/Beveridge_curve
• https://en.wikipedia.org/wiki/Beveridge_curve
• https://personal.lse.ac.uk/petrongo/jel-final.pdf
• https://www.richmondfed.org/publications/research/economic_brief/2021/eb_21-36

<cite index="23-2">Why are so many people unemployed at the same time that there are a large number of job openings?</cite> This was the question the 2010 Nobel committee highlighted when <cite index="23-1">they awarded the prize to Diamond, Mortensen, and Pissarides for their analysis of markets with search frictions</cite>.

The answer sits in the friction itself. <cite index="22-3,22-4,22-5">In many markets, time and effort are required to bring buyers and sellers into contact. In the labor market, such search frictions imply that unemployed job searchers will have to use time and other resources to find jobs, and it takes time for firms to fill their job vacancies</cite>. <cite index="7-8,7-9,7-10">Despite workers being available to start jobs and firms having posted job openings, in any given month millions of unemployed workers cannot find jobs and millions of vacancies go unfilled—it takes time for a worker to sift through job boards and fill out applications, and it is costly for a firm to post a vacancy</cite>.

<cite index="17-10,17-11">Even in a recession millions of jobs are being created every month—in the United States in August, for example, there were 4.1 million hires and 4.2 million separations</cite>. The gross flows dwarf the net number reported in headlines. <cite index="19-4,19-5">Diamond analyzed the foundations of search markets; he, Mortensen, and Pissarides expanded the theory and applied it to the labor market, helping us understand the ways in which unemployment, job vacancies, and wages are affected by regulation and economic policy</cite>.

The model allows policymakers to think about interventions structurally. <cite index="18-7">The Diamond-Mortensen-Pissarides model is a frequently used tool to estimate how unemployment benefits, interest rates, the efficiency of employment agencies and other factors can affect the labor market</cite>.

Sources:

• https://www.kva.se/en/news/ekonomipriset-2010-2/
• https://www.nobelprize.org/prizes/economic-sciences/2010/illustrated-information/
• https://arxiv.org/pdf/2307.05843
• https://fee.org/articles/the-2010-nobel-prize-in-economics/
• https://www.nber.org/news/peter-diamond-dale-t-mortensen-and-christopher-pissarides-shared-2010-nobel-prize-research-job
• https://www.csmonitor.com/Business/Latest-News-Wires/2010/1011/Nobel-Prize-in-Economics-awarded-for-unemployment-research

<cite index="5-2,5-8">Peter Diamond, Dale Mortensen, and Christopher Pissarides won the 2010 Nobel Prize for developing search and matching theory</cite>, a framework that finally made sense of why millions of workers and millions of vacancies can coexist.

The central innovation was <cite index="13-2,13-5">Pissarides's development of the matching function, which treats unemployed people and vacancies as inputs that produce jobs</cite>—the same way economists think about production functions turning raw materials into output. <cite index="6-3">The matching function relates the flow of worker-job meetings to the number of workers unemployed, the number of job vacancies, and the intensities with which workers search and employers recruit</cite>. It assumes constant returns to scale and diminishing marginal returns, just like standard production.

<cite index="11-3,11-8">Frictions are likely to be more important in the labor market than in other markets</cite>, and the matching function became a way to formalize what was always visible but never quite tractable: <cite index="7-3,7-11">search frictions prevent firms from instantly hiring available workers, and unemployment persists</cite>. <cite index="17-6">The key breakthrough was realizing the problem was not how to explain unemployment per se but rather how to explain hiring, firing, quits, vacancies and job search</cite>.

The framework has become a workhorse. <cite index="5-10">The canonical Diamond-Mortensen-Pissarides model of unemployment has become a workhorse in macroeconomic analysis</cite>, applied to policy questions about unemployment insurance, hiring costs, minimum wages, and taxes.

Sources:

• https://www.nobelprize.org/uploads/2018/06/advanced-economicsciences2010.pdf
• https://www.nobelprize.org/uploads/2018/06/mortensen-lecture.pdf
• https://www.imf.org/external/pubs/ft/fandd/2014/06/people.htm
• https://personal.lse.ac.uk/petrongo/jel-final.pdf
• https://arxiv.org/pdf/2307.05843

<cite index="2-2">Lazear's study was one of the earliest papers studying data from within a single firm and inspired a wealth of papers looking at the relationship between pay and productivity at shoe factories, fruit farms, software companies, and many other organizations.</cite> Prior work had models. Lazear brought payroll records and installation counts.

<cite index="2-7">Edward Lazear, the first personnel economist, passed away in November 2020.</cite> His 2000 paper in the American Economic Review became canonical not because the result was surprising — theory predicted that piece rates would raise output — but because he showed how much, and through what mechanisms, using a real firm's real transition.

The method mattered. <cite index="27-5,27-6">Before Safelite installed the compensation system, they had an information system installed where they could easily track and monitor the individual performance of each installer, measuring both number of windshields installed each day and the quality of the installation job.</cite> You need measurement infrastructure before you can run performance pay, and you need performance pay variation before you can measure what it does. Lazear had both, by accident of timing and access.

<cite index="11-17,11-18,11-19">Lazear and Kathryn Shaw's 2007 paper "Personnel Economics: The Economist's View of Human Resources" appeared in the Journal of Economic Perspectives.</cite> The field he named had become a subfield, and firms began hiring economists to design pay structures with the same rigor they applied to pricing and capex.

Sources:

• https://cepr.org/voxeu/columns/edward-lazear-personnel-policy-and-productivity
• https://cdn.citl.illinois.edu/courses/LER/summer2017/LER545/week5/lesson1/web_data/file21.htm
• https://ideas.repec.org/p/nbr/nberwo/12041.html

One of Lazear's three predictions was the least intuitive: <cite index="1-3,7-1">the variance in output across individuals at the firm would rise when it shifted to piece rates.</cite> Under hourly pay, output dispersion was muted — everyone hovered near a similar daily count because the low performers weren't penalized and the high performers weren't rewarded for distance above the mean.

Piece rates changed the distribution. High-output workers had reason to install more units per day. Low-output workers either improved or left. The result was wider spread. That wider spread wasn't noise — it was information. It told the firm who was capable of what, and it told workers where they stood relative to the market clearing rate for their skill.

This is the part that makes HR departments uncomfortable: performance pay doesn't compress variance, it reveals it. <cite index="27-7">Over the two years after Safelite switched to the piece rate system, productivity rose by about over a third from 2.6 to 3.1 units per day.</cite> But average productivity rising meant some workers moved well above that average and some moved out. The policy question isn't whether to permit variance — it's whether to price it.

Sources:

• https://papers.ssrn.com/sol3/papers.cfm?abstract_id=490
• https://www.nber.org/system/files/working_papers/w5672/w5672.pdf
• https://cdn.citl.illinois.edu/courses/LER/summer2017/LER545/week5/lesson1/web_data/file21.htm

Lazear's contribution was separating two effects that prior theory had bundled: the incentive effect (existing workers work harder) and the sorting effect (different workers choose to stay).

<cite index="12-1,12-2">Lazear's study of the Safelite Glass Company provides a case study of the effect of incentives on both effort and worker selection. Safelite switched from paying windshield installers an hourly wage to paying them a piece rate per windshield installed.</cite> <cite index="12-3">Because Safelite implemented the new pay scheme at different times at different locations, Lazear could estimate both effects separately.</cite>

The decomposition revealed that sorting explained roughly half the productivity gain. Workers who knew they were slow or inconsistent left. Workers who knew they were fast stayed or joined. <cite index="1-3,6-2">The theory predicted that the firm would attract a more able workforce and that variance in output across individuals at the firm would rise.</cite> Both predictions held.

This finding reshaped how personnel economists thought about pay design. <cite index="18-2,20-2">Variable pay not only creates a link between pay and performance but may also help firms in attracting the more productive employees.</cite> The implication: if you're a firm designing compensation and you think only about motivation, you're missing half the story. Who quits and who applies are decisions shaped by the same price signal.

Sources:

• https://www.nber.org/system/files/working_papers/w13480/w13480.pdf
• https://papers.ssrn.com/sol3/papers.cfm?abstract_id=490
• https://ideas.repec.org/p/nbr/nberwo/5672.html
• https://www.iza.org/publications/dp/1191/other-regarding-preferences-and-performance-pay-an-experiment-on-incentives-and-sorting

<cite index="4-3,16-2">Lazear's 2000 study tested predictions using Safelite Glass Corporation data: that average productivity would rise, the firm would attract a more able workforce, and variance in output across individuals would rise when it shifted to piece rates.</cite> The company moved windshield installers from hourly wages to per-unit pay between 1994 and 1995.

<cite index="4-4,16-3">In Safelite, productivity effects amounted to a 44-percent increase in output per worker.</cite> <cite index="2-1">About half the increase came from increased effort on the part of installers and the other half from positive selection of more productive installers into the company (and negative selection of those who were less productive).</cite>

This was arithmetic, not magic. <cite index="1-7">About half of the worker-specific increase in productivity was passed on to workers in the form of higher wages.</cite> <cite index="4-5,16-4">The firm apparently had selected a suboptimal compensation system, as profits also increased with the change.</cite> The installers who stayed were the ones whose output justified the new structure. The ones who left were pricing themselves correctly.

The Safelite case became foundational because <cite index="4-2,16-1">much of the theory in personnel economics related to effects of monetary incentives on output, but the theory was untested because appropriate data were unavailable.</cite> Lazear had payroll records, output data by worker, and a natural experiment: the company rolled out the pay change at different locations over time, which let him isolate what changed when the price changed.

Sources:

• https://www.aeaweb.org/articles?id=10.1257/aer.90.5.1346
• https://cepr.org/voxeu/columns/edward-lazear-personnel-policy-and-productivity
• https://ideas.repec.org/a/aea/aecrev/v90y2000i5p1346-1361.html

The policy implications of monopsony diverge sharply from the competitive model. <cite index="28-1,28-5">If the employers of low-wage workers have monopsony power, then a minimum wage could increase employment as well as wages.</cite> <cite index="28-2,28-6,28-7">Employers would raise wages for all current employees whose wages were below the minimum, regardless of whether additional workers were hired. If those employers then sought to hire new workers, they would have already incurred the cost of increasing the wages of current employees, so the cost of hiring additional workers would be lower.</cite>

The intuition: in a competitive market, the minimum wage binds and destroys marginal jobs. In a monopsony market, the minimum wage can push the firm closer to the employment level it would choose if it didn't have wage-setting power. <cite index="28-3,28-4">A relatively low minimum wage could lead to an increase in employment. However, a sufficiently large minimum wage would reduce employment.</cite>

<cite index="9-4">Manning's monopsony framework has been applied to explaining the impact of minimum wages and immigration, in anti-trust, and in understanding how to model the determinants of earnings in matched employer–employee data sets and the implications for inequality and the labor share.</cite> The 2021 review and the Handbook chapter make clear: monopsony is not a curiosity. It is now the organizing framework for a significant portion of applied labor economics, particularly in domains where employer market power matters more than textbook competition.

Sources:

• https://www.cbo.gov/system/files/2019-12/CBO-55410-MinimumWage-Monopsony.pdf
• https://eprints.lse.ac.uk/103482/1/MonopsonyILR_Revision.pdf

The empirical case for monopsony hinges on one observable: how many workers leave when you cut their wage. <cite index="3-10">Manning notes the evidence that a firm's quit rate responds to the wage it pays with an elasticity less than infinity in absolute value is both stronger and more convincing than the firm-size and wages literature.</cite>

<cite index="23-3,23-4,23-5">Even in thick urban labor markets in high-income countries, the share of workers likely to leave in response to a hypothetical 10 percent wage cut is much smaller—perhaps 20 to 30 percent—and is often lower for women. In developing economies, it is lower still. This suggests employers have wide latitude to set wages.</cite> <cite index="23-6">A higher wage helps recruit and retain workers, but the market does not dramatically constrain companies' wage decisions and different employers can make different choices.</cite>

This has testable implications. <cite index="25-9,25-12,25-13">The elasticity of labor supply that an individual firm faces measures the sensitivity of labor to wage changes. Employers without labor market power face a highly elastic labor supply: if they decrease wages by a small bit, employees tend to go elsewhere. In contrast, a monopsonist faces relatively inelastic labor supply: if it marks down wages, workers respond less to the decrease.</cite>

Manning built the theory; the data followed. Quit elasticities, wage dispersion for observationally identical workers, and minimum wage employment effects all point in the same direction: <cite index="23-11,23-12">the growing interest in monopsony power fundamentally comes from the wide gap between the theoretical predictions of the competitive model and empirical findings. Perhaps no other prediction is as clearly at odds with the evidence as the "law of one price" in the labor market.</cite>

Sources:

• https://www.tandfonline.com/doi/full/10.1080/1357151042000286456
• https://www.nber.org/reporter/2024number1/monopsony-power-labor-markets
• https://www.cato.org/regulation/summer-2022/there-monopsony-power-us-labor-markets

<cite index="13-1,13-3">Labor markets are pervasively imperfectly competitive; there are rents to the employment relationship for both worker and employer.</cite> Manning's 2011 Handbook chapter formalizes this. <cite index="14-6,14-7,16-3,16-4,16-5">He outlines the main sources of rents in the employment relationship, discusses estimates of the size of rents, then considers models of how rents are split between worker and employer—the question of wage determination.</cite>

Why does imperfect competition persist? <cite index="10-2">The mobility of the labor force is low because of the inability of workers to wait for employment or risk unemployment, plus the inadequacy of information usually available to them regarding alternative employment opportunities.</cite> The result: each employer can set their own rates.

<cite index="2-11,9-3">Researchers' interest in monopsony has increased in recent years. Manning's 2021 review in Industrial and Labor Relations Review summarizes the accumulating evidence that employers have considerable monopsony power.</cite> <cite index="9-1,9-5">High levels of inequality and a falling labor share in national income have led to renewed interest in the idea that an imbalance in economic power occurs between employers and workers.</cite> The empirical literature now spans minimum wage effects, immigration, anti-trust, matched employer–employee datasets, and the labor share.

Manning's argument is not that the competitive model is always wrong. It is that the monopsony model is empirically more descriptive and theoretically more tractable when rents exist and employers set wages.

Sources:

• https://cep.lse.ac.uk/pubs/download/dp0981.pdf
• https://ideas.repec.org/p/cep/cepdps/dp0981.html
• https://personal.lse.ac.uk/manning/work/mimintro.pdf
• https://eprints.lse.ac.uk/103482/1/MonopsonyILR_Revision.pdf
• https://journals.sagepub.com/doi/10.1177/0019793920922499

<cite index="6-3,6-4">Traditional labor economics assumes workers quit immediately if an employer cuts wages by one cent. Manning challenges this.</cite> His argument: <cite index="1-7">frictions in the labor market—the time and cost it takes workers to change jobs—give employers market power over their workers.</cite>

<cite index="3-1,3-4">Search frictions generate upward-sloping labor supply curves to individual firms even when firms are small relative to the labor market, producing 'monopsonistic competition.'</cite> The implication is simple but foundational: <cite index="26-7,26-8">when employer and worker separate, one or both are made worse-off. This gives employers some market power because a small wage cut will no longer induce workers to leave the firm.</cite>

This isn't about coal towns with one employer. <cite index="21-12,21-13">Labor market power comes from the search and matching process. At the point a job offer is made, a worker's only immediate alternative is to remain unemployed, so their reservation wage is lower than the average wage offered by other firms.</cite> Even in cities with many firms, the friction is what matters. <cite index="5-3">For many workers, power derives from their ability to leave more than their ability to negotiate wages with their employer.</cite>

This reframing—published in Manning's Monopsony in Motion (2003) and later in the Handbook of Labor Economics chapter "Imperfect Competition in the Labour Market" (2011)—<cite index="8-7,8-8">presents theoretical implications and empirical evidence that our understanding of wage distribution, unemployment, and human capital can all be improved by recognizing employers have some monopsony power.</cite>

Sources:

• https://press.princeton.edu/books/paperback/9780691123288/monopsony-in-motion
• https://www.amazon.com/Monopsony-Motion-Imperfect-Competition-Markets/dp/B005ZOFVCI
• https://www.tandfonline.com/doi/full/10.1080/1357151042000286456
• https://personal.lse.ac.uk/manning/work/mimintro.pdf
• https://eprints.lse.ac.uk/103482/1/MonopsonyILR_Revision.pdf
• https://core-econ.org/the-economy/microeconomics/06-firm-and-employees-12-employers-exercise-power.html

<cite index="11-1,11-5">It is plausible to expect that the beneficial effect of agglomeration spillovers generated by a new high tech entrant is larger for high tech firms than for low tech firm. The overall effect on profits of incumbent firms is ambiguous. It depends on the relative strength of agglomeration spillovers (if any) and input prices increases.</cite>

Moretti's empirical work traces spillovers at the inventor level. <cite index="26-4,26-5">Following the decline in the Rochester high-tech cluster, non-Kodak inventors in Rochester experienced large productivity losses relative to non-Kodak inventors in other cities. This is consistent with the existence of important productivity spillovers in the high-tech sector stemming from geographical agglomeration.</cite>

<cite index="12-9">Moretti finds that adding 1 additional job in one part of the tradable sector has no significant effect on employment in other parts of the tradable sector.</cite> The spillover runs through local service employment, not through manufacturing-to-manufacturing. <cite index="10-7,10-8">The new firm will result in higher wages throughout the local area, raising labor costs, and resulting in lower employment among the existing tradable sector firms. The new firm may also drive-up prices for other inputs to production in the local area, further increasing costs for other local tradable sector firms.</cite>

Sources:

• https://eml.berkeley.edu/~webfac/dromer/e291_f07/Moretti.pdf
• https://eml.berkeley.edu/~moretti/clusters.pdf
• https://academic.oup.com/icc/article/22/1/339/885578
• https://onlinelibrary.wiley.com/doi/full/10.1111/jors.12680

<cite index="2-7,2-8">This 'Great Divergence' has its origins in the 1980's when American cities started to be defined by their resident's level of education. This Great Divergence in educational levels is causing an equally large divergence in labor productivity and salaries.</cite>

<cite index="2-6">A handful of cities with 'right' industries and a solid base of human capital keep attracting good employers and offering high wages, while those at the other extreme, cities with the 'wrong' industries and a limited human capital base, are struck with dead-end jobs and low average wages.</cite> <cite index="21-5,21-6,21-7">This mobility is seen to be producing a national pattern of spatial segmentation which he calls, dramatically, 'The Great Divergence'. Those cities unable to provide jobs for the new knowledge workers (at competitive pay rates) are falling into downward spirals. On the other hand, those cities with thick labour markets—plenty of well-paid jobs in innovation-driven sectors—enjoy a virtuous cycle of growth.</cite>

The pattern is not reversible by policy wish. <cite index="19-14,19-15,19-16">If this clustering effect is particularly strong, it's good news for places like here, but it's terrible news for places like Flint or Detroit. A successful local labor market has a very nice equilibrium, where you have a lot of skilled workers who want to go there and a lot of innovative employers who want to go there. It's really hard to re-create somewhere else.</cite>

Sources:

• https://journal.c2er.org/2012/11/the-new-geography-of-jobs-enrico-moretti/
• https://academic.oup.com/joeg/article/14/1/221/1048576
• https://www.gsb.stanford.edu/insights/enrico-moretti-geography-jobs

<cite index="17-7,17-8,17-9">There are three that have been identified in the literature and are likely to play a significant role in practice. The first one is the existence of knowledge spillovers, also known as human capital spillovers: the fact that our human capital depends not only on where we go to school and how much schooling we get, but also on the people who surround us and from whom we learn. The second one is the matching advantage offered by thick labor markets.</cite>

The third: specialized service providers. <cite index="1-1">Walmart saw three important competitive advantages to a San Francisco location, which economists refer to collectively as the forces of agglomeration: thick labor markets (that is, places where there is a good choice of skilled workers trained in a specific field), the presence of specialized service providers, and, most important, knowledge spillovers.</cite>

<cite index="19-5,19-6,19-7">The match between employer and employee tends to be more productive, more creative and innovative in thicker labor markets. It is the same thing for the vendors, the providers of intermediate services. Companies in the Silicon Valley will find very specialized IP lawyers, lab services, and shipping services that focus on that niche of the industry.</cite>

These are not soft forces. <cite index="4-15,4-16">The growing economic divide between American communities is not an accident but the inevitable result of deep-seated economic forces. More than traditional industries, the knowledge economy has an inherent tendency toward geographical agglomeration.</cite>

Sources:

• https://www.richmondfed.org/publications/research/econ_focus/2019/q1/interview
• https://billmanzi.com/2018/08/12/a-look-at-the-new-geography-of-jobs-by-enrico-moretti/
• https://www.gsb.stanford.edu/insights/enrico-moretti-geography-jobs
• https://www.amazon.com/New-Geography-Jobs-Enrico-Moretti/dp/0544028058

<cite index="4-4,4-12">The presence of many college-educated residents changes the local economy in profound ways, affecting both the kinds of jobs available and the productivity of every worker who lives there, including the less skilled.</cite> This is not just the college premium showing up in paychecks. <cite index="8-6,8-12">For every new innovation job in a city, five additional non-innovation jobs are created, and those workers earn higher salaries than their counterparts in other urban areas.</cite>

Moretti's empirical work puts numbers to the spillover. <cite index="12-4,12-5">The average multiplier in manufacturing is 1.6, while the multiplier for high-technology is 4.9. This reflects higher wages in high-technology and stronger agglomeration economies.</cite> <cite index="5-1">Americans with high school degrees who work in communities dominated by innovative industries actually make more, on average, than the college graduates working in communities dominated by manufacturing industries.</cite> In San Jose, a high school graduate averages $68,009. In Bakersfield, a college graduate averages $65,411.

<cite index="2-13">Moretti acknowledges the compensatory impact of higher costs of living as offsets to these higher wages.</cite> But the divergence is structural, not noise.

Sources:

• https://www.amazon.com/New-Geography-Jobs-Enrico-Moretti/dp/0544028058
• https://journal.c2er.org/2012/11/the-new-geography-of-jobs-enrico-moretti/
• https://www.gsb.stanford.edu/insights/enrico-moretti-geography-jobs
• https://academic.oup.com/icc/article/22/1/339/885578

The second paper in the Chetty-Hendren neighborhoods series, published in December 2017, estimated <cite index="15-1">the causal effect of each county in the U.S. on children's incomes in adulthood</cite>. They used the exposure-effect framework from the first paper and applied it county by county, producing a forecast tool families could use when deciding where to settle.

<cite index="20-3,20-7">Growing up in DuPage County from birth would raise a child's income by 16%</cite> (relative to the national average for their parent income level). <cite index="20-4,20-8">In contrast, growing up in Cook County from birth reduces a child's income by approximately 13%</cite>. That's a 29-percentage-point gap between neighboring counties in the Chicago metro.

The fixed-effects model was <cite index="15-2">identified by analyzing families who move across counties with children of different ages</cite>. This let them separate the causal effect of place from selection effects — the types of families who choose to live in each place. <cite index="23-4">Permanent residents' outcomes combine the causal effect of growing up in a given area with selection effects reflecting differences in family inputs</cite>, but the movers design teased those apart.

This paper made the abstract concrete. You could look up your county, see the gain or loss, and decide whether to stay.

Sources:

• https://opportunityinsights.org/wp-content/uploads/2018/03/movers_paper2.pdf
• https://economics.mit.edu/sites/default/files/inline-files/movers_paper2_1.pdf
• https://www.nber.org/papers/w23002

Chetty and Hendren's 2018 paper in the Quarterly Journal of Economics, "The Impacts of Neighborhoods on Intergenerational Mobility I: Childhood Exposure Effects," used families who moved across counties and commuting zones to isolate the causal effect of place. <cite index="19-5">The study used data from de-identified tax records on more than five million children whose families moved across counties between 1996 and 2012</cite>.

The key insight came from comparing siblings who moved at different ages. <cite index="18-10">Neighborhoods have significant childhood exposure effects: the outcomes of children whose families move to a better neighborhood improve linearly in proportion to the amount of time they spend growing up in that area, at a rate of approximately 4% per year of exposure</cite>. A child who moves at age 5 gains more than one who moves at age 15. Time matters. Duration compounds.

<cite index="19-2,19-7">The area in which a child grows up has significant causal effects on her prospects for upward mobility</cite>. This wasn't just observational; the design isolated causation from selection by exploiting variation in the age at move. The earlier work ("Where is the Land of Opportunity?") had shown that places differ. This paper showed that moving to those places changes outcomes.

The methodology ruled out the story where high-mobility families sort into high-mobility places. The results held when studying moves triggered by displacement, by job changes, by external shocks — the effect was real.

Sources:

• https://scholar.harvard.edu/hendren/publications/impacts-neighborhoods-intergenerational-mobility-i-childhood-exposure-effects
• https://opportunityinsights.org/paper/neighborhoodsi/
• https://www.nber.org/papers/w23001

Before the Opportunity Atlas, Chetty, Hendren, Kline, and Saez published "Where is the Land of Opportunity?" in the Quarterly Journal of Economics in 2014. <cite index="11-5">It used administrative records on the incomes of more than 40 million children and their parents</cite> to map intergenerational mobility across the United States at the commuting zone and county level.

The finding that made people stop: <cite index="11-9,13-6">the probability that a child reaches the top quintile of the national income distribution starting from a family in the bottom quintile is 4.4% in Charlotte but 12.9% in San Jose</cite>. The same family, three times the outcome, depending on place.

At the national level, <cite index="11-8,13-5">a 10 percentile increase in parent income is associated with a 3.4 percentile increase in a child's income</cite> — the relationship is linear in ranks. That's the baseline. The geography is the variance around it.

<cite index="11-9,13-6">Intergenerational mobility varies substantially across areas within the U.S.</cite> This was the first large-scale named-place study of where the ladder broke. It forced cities to confront their own numbers. Charlotte's civic leadership spent years trying to explain the 4.4%. The work didn't just describe inequality; it located it.

Sources:

• https://hendren.scholars.harvard.edu/publications/economic-impacts-tax-expenditures-evidence-spatial-variation-across-us
• https://opportunityinsights.org/paper/land-of-opportunity/
• https://www.nber.org/system/files/working_papers/w19843/w19843.pdf

<cite index="1-1">The Opportunity Atlas uses anonymous data following 20 million Americans from childhood to their mid-30s</cite>, connecting adult outcomes to the census tracts where they were children. The work was published by Raj Chetty, John Friedman, Nathaniel Hendren, Maggie R. Jones, and Sonya R. Porter in the American Economic Review in 2026.

The variation is sharper than anyone expected. <cite index="2-3">For children whose parents earn $27,000, the standard deviation of mean household income in adulthood is $10,420 across tracts within counties</cite>. That's tract to tract, not city to city. The scale is local.

<cite index="2-4,2-5">Only half the variation in outcomes is explained by traditional measures of neighborhood opportunity like poverty rates</cite>, and <cite index="2-13">60 percent of the variation in outcomes across neighborhoods is driven by causal effects</cite> — confirmed through experimental and quasi-experimental methods. This isn't sorting. This is place working on people.

The Atlas traces backward: <cite index="3-9">it differs from traditional indicators of neighborhood conditions based on cross-sectional data – such as rates of poverty or crime – by tracing the roots of such outcomes back to the neighborhoods in which children grew up</cite>. It tells you which neighborhoods gave people their trajectories, not which neighborhoods those people ended up in.

<cite index="5-10,5-11">The availability of low-rent, high-opportunity neighborhoods suggests that affordable housing policies could be redesigned to produce larger gains for children without increasing government expenditure</cite>. The implication is policy-actionable: opportunity doesn't only live in expensive places.

Sources:

• https://rajchetty.com/the-opportunity-atlas/
• https://www.aeaweb.org/articles?id=10.1257%2Faer.20200108
• https://www.nber.org/system/files/working_papers/w25147/w25147.pdf
• https://www.aasa.org/resources/resource/opportunity-atlas-childhood-roots-social-mobility

<cite index="8-7,8-8">Dube's 2010 paper with Lester and Reich, "Minimum Wage Effects Across State Borders: Estimates Using Contiguous Counties," appeared in the Review of Economics and Statistics</cite>. The design: compare counties that share a border but sit in different states with different minimum wage policies. The idea is that labor markets on either side of a state line look similar — same industries, same demographics, same economic shocks — except for the policy.

<cite index="2-14">In 2017, Allegretto, Dube, Reich, and Zipperer published "Credible Research Designs for Minimum Wage Studies" in ILR Review</cite>, laying out the case for why geographic controls matter and why national time-series approaches can confound policy effects with regional economic trends. The argument was methodological but the stakes were empirical: if you do not control for the fact that states that raise minimum wages also tend to be states with different growth trajectories, you will attribute employment losses to the wage floor when they were already baked into the regional economy.

<cite index="20-2,20-3">Most of the US literature exploits methods based on state-level variation, and more recently, city-level variation (Dube and Lindner 2021), with only a relatively small number of studies exploiting geographic variation in bite</cite>. The methodological refinement matters because the results change: better designs tend to find smaller employment effects.

Sources:

• https://arindube.com/minimum-wage-research/
• https://www.journals.uchicago.edu/doi/abs/10.1086/685449
• https://www.journals.uchicago.edu/doi/10.1086/728471

<cite index="25-1">In a 2024 meta-analysis, Dube and Zipperer found that most studies to date suggest a fairly modest impact of minimum wages on jobs: the median own-wage elasticity (OWE) estimate of 72 studies published in academic journals is -0.13, which suggests that only around 13 percent of the potential earnings gains are offset by employment losses</cite>.

The own-wage elasticity is the percentage change in employment for a 1% change in the wage caused by the minimum wage. A value of −0.13 means that if a wage floor raises pay by 10%, employment in the affected group falls by 1.3%. That is not zero. But it is small enough that the earnings effect dominates. <cite index="20-10">In UK research by Giupponi et al. (2024), the central estimate was −0.20 (standard error 0.32)—a small effect, which is in line with several other estimates in the literature, including previous studies in the United Kingdom</cite>.

<cite index="26-2,26-3">In contrast, a 2021 working paper by Neumark and Shirley summarized extant studies and reported a median estimate of −0.130 and a mean estimate of −0.270 for how the employment of directly affected workers responds to changes in the minimum wage</cite>. The debate is not whether the elasticity is zero. It is whether it is small enough to justify the policy. Dube's work says: usually, yes.

Sources:

• https://www.nber.org/papers/w32925
• https://www.journals.uchicago.edu/doi/10.1086/728471
• https://www.cbo.gov/publication/57049

<cite index="9-1">At Spring Statement 2019, the UK Chancellor appointed Arindrajit Dube to undertake a review of the international evidence on the impacts of minimum wages</cite>. <cite index="12-3,12-4">The government asked Professor Dube to consider international evidence on the impacts of minimum wages and the implications for UK policy; his report reviewed the international evidence on the impacts of minimum wages, as well as recent research documenting the impact of the National Living Wage (NLW) in the UK</cite>.

<cite index="12-1,12-5">The report found that, overall, the most up to date body of research from US, UK and other developed countries points to a very muted effect of minimum wages on employment, while significantly increasing the earnings of low paid workers</cite>. <cite index="4-2">Across US states, the best evidence suggests that the employment effects are small up to around 59% of the median wage</cite>. This was not a polemic. It was an accounting of what credible designs had shown — and what they had not.

<cite index="16-1">The review explored the potential impacts of minimum wages on employment, in terms of volume and structure, productivity and economic growth, as well as the ability of the labor market to absorb future minimum wage rises</cite>. The Treasury wanted to know: how high can we go. Dube told them: here is where the signal stops.

Sources:

• https://www.gov.uk/government/publications/impacts-of-minimum-wages-review-of-the-international-evidence
• https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/844350/impacts_of_minimum_wages_review_of_the_international_evidence_Arindrajit_Dube_web.pdf
• https://www.gov.uk/government/publications/review-of-the-international-evidence-on-the-impacts-of-minimum-wages/review-of-the-international-evidence-on-the-impacts-of-minimum-wages-terms-of-reference

<cite index="6-2,7-5,7-6">Cengiz, Dube, Lindner, and Zipperer (2019) introduced a method that compares the number of "excess jobs" paying at or slightly above a new minimum wage to the "missing jobs" paying below it, using 138 state-level minimum wage changes between 1979 and 2016</cite>. The question they ask is simple arithmetic: when the floor rises, do the jobs below it disappear or do they get re-priced?

<cite index="7-1,7-7">They found that the overall number of low-wage jobs remained essentially unchanged over five years following the increase</cite>. <cite index="18-3,18-4,18-5,18-6">The Cengiz et al. work also provided the first clear measure in this literature of how the marginal impact varies by the "bite" of the minimum wage — the minimum-to-median wage ratio — and found that for higher bites, more workers are affected, but nearly all these jobs seem to have been upgraded and not destroyed, as indicated by the rising number of excess jobs paying at or slightly above the minimum for events with a bigger bite</cite>.

<cite index="18-7">Overall, there is little indication of job losses for events going as high as 59% of the median wage</cite>. The policy moved wages. It did not move the headcount — at least not down. <cite index="7-2">At the same time, the direct effect of the minimum wage on average earnings was amplified by modest wage spillovers at the bottom of the wage distribution</cite>.

Sources:

• https://academic.oup.com/qje/article-abstract/134/3/1405/5484905
• https://www.nber.org/system/files/working_papers/w25434/w25434.pdf
• https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/844350/impacts_of_minimum_wages_review_of_the_international_evidence_Arindrajit_Dube_web.pdf

<cite index="11-1">Workers in trade-exposed commuting zones experienced larger reductions in average weekly wages, and these impacts were concentrated among workers in the bottom four wage deciles.</cite> The distributional pattern was not symmetric. Import competition did not flatten the wage structure—it hollowed the lower half.

<cite index="12-4,12-5">Workers subject to mass layoffs suffered an immediate sharp decline in earnings and a smaller decline that persisted over time, and wage loss was greater for workers with higher tenure but was otherwise similar across groups by age, gender, or skill.</cite> Using administrative earnings data allowed the research to track individuals after they left the initial firm, industry, or location—capturing the persistence of the shock at the worker level, not just the place level.

The implication: trade shocks were not temporary dislocations. They were permanent wage scars for the workers who lived through them, even when they found new work. The city might recover. The individual often did not.

Sources:

• https://chinashock.info/wp-content/uploads/2018/02/Lessons-from-Chinas-Rise-IZA.pdf
• https://www.princeton.edu/~ies/Fall12/HansonPaper.pdf

More recent work with Jones and Setzler extended the timeline to 2019 and followed the workers, not just the zones. <cite index="16-1,16-5">The employment recovery stemmed almost entirely from young adults and foreign-born immigrants taking their first U.S. jobs in affected areas, with minimal contributions from cross-sector transitions of former manufacturing workers.</cite>

<cite index="16-2,16-6">Although worker inflows into non-manufacturing more than fully offset manufacturing employment losses in trade-exposed locations after 2010, incumbent workers neither fully recovered earnings losses nor predominantly exited the labor market, but rather aged in place as communities underwent rapid demographic and industrial transitions.</cite>

The theory predicted reallocation. The data showed replacement. Former manufacturing workers did not retrain into healthcare or retail at scale. They withdrew—partially, incrementally, but persistently. The commuting zone repopulated, but not with the same people doing different work. It was different people doing the work the original residents had left.

Sources:

• https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5122141
• https://www.nber.org/papers/w33424

<cite index="27-2,27-6">Transfer benefits payments for unemployment, disability, retirement, and healthcare rose sharply in more trade-exposed labor markets.</cite> But the composition told a different story about what workers did when the factory closed.

<cite index="26-3">TAA grants were temporary, whereas most workers who took up disability received SSDI benefits until retirement or death.</cite> <cite index="26-4,29-1">For regions affected by Chinese imports, the estimated dollar increase in per capita SSDI payments was more than thirty times as large as the estimated dollar increase in TAA payments.</cite>

This was not a policy failure. It was a revelation about the alternative. When a fifty-two-year-old machinist in a trade-exposed commuting zone lost work, retraining was the theory. Disability was the arithmetic. The labor market did not come back. The worker aged in place. The transfer checks continued.

Sources:

• https://economics.mit.edu/sites/default/files/publications/the%20china%20syndrome%202013.pdf
• https://econbrowser.com/archives/2011/04/gains_and_losse
• https://www.nber.org/reporter/2014number2/labor-market-adjustment-international-trade

<cite index="4-2,8-4">Autor, Dorn, and Hanson analyzed Chinese import competition between 1990 and 2007 on local U.S. labor markets, exploiting cross-market variation in import exposure stemming from initial differences in industry specialization and instrumenting for U.S. imports using changes in Chinese imports by other high-income countries.</cite> The method matters: they treated commuting zones as sub-economies subject to differential trade shocks based on what those places made before China arrived.

<cite index="4-3,6-7">Rising imports caused higher unemployment, lower labor force participation, and reduced wages in local labor markets that housed import-competing manufacturing industries.</cite> <cite index="4-4,6-8">Import competition explained one-quarter of the contemporaneous aggregate decline in U.S. manufacturing employment.</cite> This was not a story of workers flowing smoothly to new sectors. It was a story of exits—from work, from the labor force, from earnings.

<cite index="20-9,20-18">Adjustment in local labor markets was remarkably slow, with wages and labor-force participation rates remaining depressed and unemployment rates remaining elevated for at least a full decade after the China trade shock commenced.</cite> The aggregate smoothed it out. The city did not.

Sources:

• https://www.aeaweb.org/articles?id=10.1257%2Faer.103.6.2121
• https://economics.mit.edu/sites/default/files/publications/the%20china%20syndrome%202013.pdf
• https://www.nber.org/papers/w18054
• https://www.annualreviews.org/content/journals/10.1146/annurev-economics-080315-015041