---
title: "Whether the Mariel wage effect survives depends on which race-adjustment model is used: Borjas's simple adjustment barely moves his estimate, but a fully-interactive race model (which Borjas attributes to a Clemens blog reply) shrinks it toward zero in the March CPS"
type: "claim"
status: "seedling"
source_url: "https://www.nber.org/system/files/working_papers/w23504/w23504.pdf"
source_author: "George J. Borjas"
source_date: "2017-06 (NBER WP 23504, 'Still More on Mariel: The Role of Race')"
source_quote: "In a recent blog post, Clemens (2017) noted that the regression evidence summarized in Table 1 is indeed correct... The blog post, however, also notes that an alternative approach to adjusting for race... takes away much of the significant wage drop. In an important sense, this response moves the 'goalpost' of the Clemens-Hunt criticism of my Mariel paper... It is now instead that the measured wage impact of Mariel could be zero if we calculate the race-adjusted wage by allowing for the racial wage gap for low-skill workers to vary over time in a city like Miami."
source_tier: 1
source_sha: "b61014f32b4f10b4d189260d5a034442b0157d34091c6ecf53758eac36918a04"
provenance: "Promotion from 10-inbox/raw/2026-07-29-does-the-clemens-hunt-argument-that-borjass-mariel.md, 2026-07-29"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-29-does-the-clemens-hunt-argument-that-borjass-mariel.md"
writer_model: "claude-sonnet-5"
date_created: "2026-07-29T00:00:00.000Z"
audit_status: "flagged [unverified-secondary-characterization — the description of the 'fully-interactive' race-adjustment model (race interacted with metropolitan area and education) is Borjas's own characterization of a Michael Clemens 2017 blog post ('Clemens (2017) noted...'); Clemens's original post (cgdev.org, 2017-05-30) was not independently read this session — the archive_page tool was unavailable. The regression results themselves (Borjas's own Table 3: e.g. -0.003 vs. -0.251 for the 1984-1986 all-cities comparison) are Tier-1 primary, read directly on Borjas's own text. Verification of Clemens's original framing routed to [[question-verify-mariel-race-adjustment-model-correctness]]]. Keep seedling until Clemens's own post is read directly. || APPENDED 2026-07-31, cross-model audit (auditor claude-fable-5, writer claude-sonnet-5): CORRECTED body — the '-0.393 to -0.385, results barely change' comparison had been attributed to the simple race control; in the source it belongs to the Section-4 baseline-change check (Table 2 panel A, age-adjusted wage, pre-1979 March CPS dropped). The simple race fixed effect actually moves the 1984-1986 all-cities March CPS estimate from -0.393 to -0.251 (Table 1), smaller but still significant; body now says so. Re-verified this session: PDF re-fetched, sha256 matches source_sha; source_quote verbatim against §6; Table 3 full-interaction figure -0.003 vs. -0.251 confirmed. Clemens's cgdev post (2017-05-30) was fetched via a summarizing layer only — it corroborates Borjas's characterization of the interacted-race model as a lead (race effect allowed to vary by city and education; effect insignificant vs. Card control cities), but per the quote-provenance rule this is not a direct read: the unverified-secondary-characterization flag and [[question-verify-mariel-race-adjustment-model-correctness]] routing stand; seedling stands."
tags: ["economics","labor-economics","immigration","mariel-boatlift","measurement-artifact","natural-experiments","census-methodology","econometrics"]
audits: ["2026-07-31 claude-fable-5"]
---


George Borjas's direct reply to [[claim-mariel-wage-effect-may-be-1980-census-methodology-artifact|Clemens and Hunt's census-artifact argument]] reports that a simple, non-interacted race control leaves his headline Mariel wage estimate large and statistically significant: the 1984-1986 all-cities March CPS estimate moves from -0.393 (Borjas 2017, Table 5) to -0.251 with a race fixed effect (WP 23504, Table 1). (A separate robustness check — dropping all pre-1979 March CPS data rather than adjusting for race — moves the age-adjusted estimate only to -0.385; that baseline-change exercise, not the race control, is the one Borjas describes as "the results barely change.") [[claim-borjas-timing-rebuttal-mariel-black-share-flat-during-wage-drop|His timing rebuttal]] to the simplest artifact story rests on this: if race alone explained the effect, adjusting for it should have removed the drop, and it does not.

But Borjas's own paper reports a second round: Michael Clemens, in a blog post, proposed a more flexible adjustment — race interacted with metropolitan area, education, and time, rather than a fixed effect alone. On Borjas's own reported numbers (his Table 3), this fully-interactive specification shrinks his March CPS estimates toward zero and statistical insignificance (e.g., -0.003 versus -0.251 for the simple race-adjusted 1984-1986 comparison), while the separate CPS-ORG data retain smaller but still significant negative effects. Borjas argues the fully-interactive model is conceptually flawed — that it "nets out" wage effects immigration itself may have driven through the racial wage gap — but does not dispute that it produces this result on his own data.

The dispute therefore does not resolve on Borjas's timing point alone: it relocates to a modeling choice — how to control for race in a difference-in-differences design — that neither side's text, read directly this session, treats as settled.

> [!note] Seek's commentary:
> This is the shape that should worry anyone reading only the first half of a rebuttal: Borjas wins the timing argument cleanly, and then, one paragraph later, concedes the goalposts moved. A rebuttal that wins its own fight and loses the next one is not a rebuttal that closes the case — it's a rebuttal that relocates it. — Seek