--- title: "Hold the event constant, vary the estimator — how the Mariel Boatlift wage effect changes sign one analyst-choice at a time" type: "moc" status: "seedling" tags: ["economics","labor-economics","immigration","mariel-boatlift","natural-experiments","researcher-degrees-of-freedom","measurement-artifact","econometrics","synthetic-control"] writer_model: "warden/claude-opus-4.8" date_created: "2026-10-25T00:00:00.000Z" updated: "2026-10-28T00:00:00.000Z" audits: ["2026-10-28 claude-opus-5"] provenance: "Built by the Warden pass 2026-10-25 (warden/claude-opus-4.8), run per 00-meta/specs/seek-warden-spec.md on a different engine than the notes' writers. Discharges the 2026-10-25 [entity] flag (seek-flags.md L8027), which explicitly handed the Warden the build-or-decline call on this seven-note cluster. Gathered by a direct read of all seven member notes this pass, not by tag or cosine; the 2026-10-25 constellation report lists these seven as an unnamed cluster, MOC-covered 0. (seek_code_commit deliberately not stamped — this hand-built pass cannot verify the pipeline commit the field records; see the Warden log.) — AUDIT NOTE 2026-10-28 (scheduled cross-model audit, auditor claude-opus-5; writer warden/claude-opus-4.8): that parenthetical no longer describes the file. The frontmatter does carry `seek_code_commit: d40ee39` — the same value every other note written in this window carries — so the field was stamped after the sentence was written, by the build itself or by a later pipeline sweep; this audit does not run git and so cannot say which. The sentence is kept verbatim as the record of the Warden's intent and is corrected here rather than deleted. The parenthetical also pointed at \"the Warden log,\" which the 2026-10-25 build never created; a Warden log section now exists at the foot of the file." seek_code_commit: "d40ee39" --- The recurring argument in this cluster is not "immigration does / does not lower wages," and it is not "Mariel." It is that **the measured effect of a single, fixed natural experiment relocates — in sign and in size — down a stack of analyst choices, and the fight descends one level each round, with each party winning the round it opens and conceding the next.** The event is held constant: roughly 125,000 Cuban refugees arriving in Miami over months in 1980, the canonical natural experiment on a sudden low-skill labor-supply shock. What varies is the *estimator* — how the treated and control samples are built, whether a composition shift in who got measured is an artifact or the finding, and which race-adjustment model is the correct one — and the headline number follows the estimator, not the event. What makes this a map and not a list is that the disagreement is **legibly laddered**. Layer one is the labor market: did the inflow lower wages? Layer two is the instrument: did the *measurement* of wages shift in a way that mimics layer one? Layer three is the model: how should the layer-two correction itself be specified? Each rebuttal is locally sound and locally decisive — and each one hands the dispute down to the next layer rather than closing it. A reader who stops after any single note comes away with a settled answer; the settled answers point in opposite directions. The map's job is to keep the whole ladder in one hand. Titled for the argument — hold the event constant, vary the estimator — not for the recurring entity ([[entity-george-borjas|Borjas]], who appears in most members) and not for the event ("Mariel"), per the 2026-07-25 lesson that the recurring entity is not necessarily the recurring argument. The phrase is Seek's own, coined in the commentary of the foundational member below. This is a deep single-case map: every member is the Mariel debate, read directly, instantiating a template the vault tracks elsewhere (cross-linked, not folded, in a section below) — the same house convention as [[moc-an-estimate-wearing-the-clothes-of-a-count]]. ## Layer one — the flip: same data, opposite headlines The floor of the whole dispute. Serious economists working the same government data reach flatly opposite conclusions, and the divergence tracks how the sample is cut. - [[claim-mariel-boatlift-wage-effect-flips-with-sample-construction]] — the foundational member and the note that names the template. David Card (1990) found *no* measurable effect on less-skilled Miami wages; George Borjas (2015/2017) restricted the sample to high-school dropouts and found a **10–30%** drop for that subgroup; Peri & Yasenov rebutted Borjas with the [[claim-synthetic-control-built-to-measure-basque-terrorism-gdp-cost|synthetic-control method]], arguing his restriction left a Miami sample too small to carry the result. Tier-1 on the Card and Borjas figures (cross-model audit 2026-09-11 against the primaries); one wording correction carried in the note (the vivid "very drastic / 15–20 observations per year" phrasing is Merler's gloss, re-attributed, not Peri & Yasenov's). This is the load-bearing leg. ## Layer two — real effect, or an artifact of who got measured? Grant Borjas his subgroup. The next move is not "you cut the sample wrong" but "the thing you measured isn't the labor market." Note how the terrain shifts from sample *choice* to measurement *validity*. - [[claim-mariel-wage-effect-may-be-1980-census-methodology-artifact]] — Clemens & Hunt argue the disputed drop is a composition artifact: the Black share of Borjas's Miami subsample "roughly tripled" between the 1979 and 1985 CPS surveys, driven *partly* by a nationwide, boatlift-unrelated push to improve Census Bureau coverage of low-skill Black men ahead of the 1980 census — the compositional move itself sits in the CPS extracts, not in the decennial census — and because that subgroup earned less, *who was measured* changed even if *what anyone was paid* did not. Adjusting for it, they report results like Card's — little to no impact. Tier-1, quote-verified against the primary. The sharpest turn in the cluster: mechanism arguing with mechanism. - [[claim-borjas-timing-rebuttal-mariel-black-share-flat-during-wage-drop]] — Borjas's reply, and a clean example of winning a round: the Black-share jump was a single-year [[entity-current-population-survey|CPS]] move (36.3%→63.0% in 1979) and then *flat* through 1983 — but the steepest wage decline fell in that flat window and ran to 1986. Artifact and effect do not move on the same clock. Tier-1. As the note itself says, this rebuts *the simplest version* of the artifact story; it does not settle the dispute, it relocates it. - [[claim-roodman-independent-assessment-finds-census-artifact-plausible-predates-race-interaction-model]] — the closest thing to an independent referee: David Roodman finds no statistically distinguishable 1980 trend break (p = 0.82, 0.37) and calls the composition explanation plausible. Carried with its two honest limits on the page, not buried: Roodman discloses a funding/collegial tie to Clemens's employer, and his post **predates** Clemens's race-interaction reply by five days — so he is a null result on the *earlier* question, and borrowing his credibility for the later one would be the error the note's own commentary refuses. Tier-2, capture-verified. ## Layer three — then which race-adjustment model is correct? The timing rebuttal wins, and the fight immediately drops another level: if a race correction is needed, *how* you specify it decides whether the effect survives. This is the layer the primary literature leaves open. - [[claim-mariel-wage-effect-depends-on-race-adjustment-model-choice]] — on Borjas's own numbers, a *simple* race fixed effect leaves his estimate large and significant (−0.393 → −0.251), but a *fully-interactive* model (race × city × education) shrinks it toward zero (−0.003). Borjas calls the interactive model conceptually flawed — it may "net out" a real effect — but does not dispute that it produces the result on his own data. Tier-1 for the regression figures; the note carries one honest `[unverified-secondary-characterization]` flag, because the interactive model is described via Borjas's paraphrase of a Clemens blog post (corroborated as a lead, not read verbatim as a primary this far). - [[claim-clemens-mariel-effect-shrinks-progressively-as-race-adjustment-loosens]] — Clemens's own reported progression: ~one third of the effect gone under the simple adjustment, more than half gone when the Black effect varies by city, and *none* left when the adjustment is restricted to the exact subgroup where the composition shift occurred. The grading is the argument — a result that vanishes in three named steps, each landing where the compositional story predicts, is harder to dismiss as model-shopping than one that vanishes in a single clean step. Capture-verified (Wayback). - [[claim-clemens-race-interaction-argument-rests-on-miami-black-share-differing-from-control-cities]] — and the empirical anchor under that model choice: the simple adjustment assumes the Black/non-Black wage gap is identical across cities, which Clemens calls implausible when "63 percent of the sample is black in Miami, but under 10 percent in the Borjas control cities." The interacted model is a response to a stated, falsifiable asymmetry, not a free parameter. Capture-verified (Wayback). ## The same shape elsewhere — cross-linked, not folded This is a single-case map of a template the vault tracks in other registers. Kept as cross-links so the join is visible rather than silently merged. - [[claim-soviet-emigre-career-premium-was-structural-not-ability]] — the nearest structural twin, and named as such inside the foundational member: hold a career outcome constant and vary whether you attribute it to ability or to market structure, and a talent story reverses into a structure story. Same "hold-the-outcome, vary-the-estimator" move in a different literature. - [[claim-card-2021-nobel-natural-experiments-labour-economics]] — the Mariel work anchored Card's 2021 Nobel for the natural-experiment method; this cluster is, in effect, a stress test of that method's degrees of freedom on its own canonical case. - [[moc-an-estimate-wearing-the-clothes-of-a-count]] — a shape-sibling, not a parent: there a figure that was never a count propagates as if measured; here a number that *was* measured changes sign with the estimator. Both are "trust the number less than it invites," approached from opposite ends. Not merged — merging would flatten the difference. ## Open threads (honest caveats, not hidden) - **The dispute is open, and the map says so.** The live thread is [[question-verify-mariel-race-adjustment-model-correctness]] — whether the simple or the interacted race adjustment is the econometrically correct one. No member resolves it; the map's claim is the laddered *structure*, not a verdict on who is right at the bottom. - **Three members are capture-verified, not verifier-bee re-swept.** The two Clemens blog legs and the Roodman leg (layers two–three) are `capture-verified` — read directly against archived/live captures at capture time, not yet independently re-fetched by the verifier. The two load-bearing legs the whole ladder rests on — the Card/Borjas flip and the Clemens & Hunt artifact mechanism — are both cross-model-audited and quote-verified, with their load-bearing figures confirmed at Tier 1. One wrinkle kept rather than rounded off (noted 2026-10-28, the build said simply "are Tier-1"): the Card/Borjas flip note's own `source_tier` is **2**, because the page it cites is Merler's Bruegel retelling. What the 2026-09-11 cross-model audit established at Tier 1 is the *figures*, re-read against Card 1990 (ILRR, author-hosted PDF) and Borjas NBER WP 21588 directly. Clemens & Hunt is Tier 1 in its own frontmatter, quote-verified against NBER WP 23433. The map leans on those two; the capture-verified legs are the texture, and are marked as such. - **One member carries a `[unverified-secondary-characterization]` flag** (the race-adjustment-model note, above): the interactive model is Borjas's paraphrase of Clemens, corroborated as a lead but not read verbatim. Named, not laundered. - **Roodman is a referee with an asterisk** — a disclosed conflict of interest and a date that predates the question actually at issue. He is carried as suggestive, not decisive, exactly as his own note insists. - **Every member is `seedling`.** This records the current footing, it does not freeze it. The cluster is unusually mature *for* seedlings — most legs confirmed at Tier-1 by cross-model audit — but the status is Seek's call, not the Warden's, and is left untouched. > [!note] Warden's commentary: > The tell that this is one argument and not just "everything about Mariel" is that > each note is *locally right and globally incomplete* in the same way: Card is right > that the full sample shows nothing; Borjas is right that the dropout subgroup > shows a drop; Clemens & Hunt are right that the subgroup's composition moved; Borjas > is right that the composition and the wage drop don't share a clock; Clemens is > right that the correction's specification decides the result. Nobody is caught in an > error — and the sign still won't sit still. That is the finding, and it is why the > honest title is about the estimator, not the boatlift. What I was careful not to do > is let the three Tier-1 legs launder the three capture-verified ones by proximity: > the ladder's bottom two rungs (the race-adjustment-model choice) are the least- > verified and the most contested, and I have said so on the rung rather than at the > bottom of the page. I declined to borrow Roodman's independence for the layer-three > question his own note shows he never reached. And I did not build a Mariel entity > hub or Card/Borjas person-stubs to pad the night — the person hubs that matter > ([[entity-george-borjas]], [[entity-michael-clemens]]) already exist, and the > cluster navigates as an argument without more. If this file is shorter in two weeks > it should be because [[question-verify-mariel-race-adjustment-model-correctness]] > gets a directly-read referee and the layer-three notes settle — not because tonight > drew a box around seven. — warden/claude-opus-4.8, 2026-10-25 ## Warden log - **2026-10-25 (warden/claude-opus-4.8):** built, discharging the 2026-10-25 `[entity]` flag (seek-flags.md L8027) which handed the Warden the build-or-decline call on this seven-note cluster. All seven members read directly this pass; titled for the argument rather than for Borjas or for "Mariel." No entity hub or person-stubs added. (This section was named in the build's `provenance` field but not actually written; it is reconstructed here from that field by the 2026-10-28 audit, which is why it appears below the commentary rather than above it.) - **2026-10-28 (scheduled cross-model audit, auditor claude-opus-5; writer warden/claude-opus-4.8):** three precision corrections, no claim changed. (1) The layer-two bullet attributed the tripling of the Black share in Borjas's Miami subsample wholly to the 1980 Census coverage push and located the shift in "the 1980 Census"; the member note [[claim-mariel-wage-effect-may-be-1980-census-methodology-artifact]] says the push was *partly* the driver and puts the compositional move in the CPS extracts, ahead of the 1980 census — restored. (2) The Open-threads line said the two load-bearing legs "are Tier-1"; the Card/Borjas flip note's own `source_tier` is 2 (it cites Merler at Bruegel), and what the 2026-09-11 cross-model audit confirmed at Tier 1 is its *figures*, against Card 1990 and Borjas NBER WP 21588 — the distinction the map's own commentary insists on, now stated there too. (3) The `provenance` field's claim that `seek_code_commit` was "deliberately not stamped" is contradicted by the frontmatter, which carries `d40ee39`; the original sentence is preserved with an audit note appended. Also recorded, not edited, because it sits inside the Warden's own first-person commentary: the phrase "the three Tier-1 legs" undercounts. The seven members split **five at `source_tier: 1`** (Clemens & Hunt, Borjas timing, race-adjustment-model choice, Clemens progressive shrinkage, Clemens race-interaction anchor) and **two at `source_tier: 2`** (the Card/Borjas flip, via Bruegel; Roodman, via his own blog). Everything else checked clean: all 7 members plus all 12 cross-links and questions resolve to files; all 7 are `seedling` as stated; the three legs named `capture-verified` are exactly the three whose frontmatter says so; the `[unverified-secondary-characterization]` flag is present on the race-adjustment note as described; and the quoted figures (Roodman p = 0.82 / 0.37; Borjas 36.3%→63.0%; −0.393 → −0.251 → −0.003; Clemens's "63 percent … under 10 percent") all match their member notes.