---
title: "A pre-registered RCT in the 2025-26 NYC high school match found nudged applicants significantly more likely to rank a recommended program, with weaker evidence of increased matching"
type: "claim"
status: "seedling"
source_url: "https://arxiv.org/pdf/2610.08275"
source_author: "Erica Chiang, Kenny Peng, Rebecca Lichtenstein, Brielle McDaniel, Kristen O'Neil, Dejaunique Thomas, Lianna Wright, Jon Kleinberg, Eva Tardos, Nikhil Garg"
source_date: "2026-10-06T00:00:00.000Z"
source_quote: "We ended up with 822 control and 444 treatment students."
source_quote_2: "Among treatment applicants, 16.4% ranked at least one program they were nudged to at least somewhere on their list; among control applicants, 10.5% ranked a program they would have been nudged to."
source_quote_3: "significantly increased the likelihood of ranking a recommended program (𝛽=0.561, SE=0.219, 𝑝=0.011)."
source_quote_4: "In terms of match outcomes, 5.6% of treatment applicants matched to a nudged program, and 3.3% of control applicants matched to a program they would have been nudged to."
source_url_4: "https://arxiv.org/pdf/2610.08275"
source_sha_4: "5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3"
source_quote_5: "There is also suggestive evidence that the treatment increased matches to recommended programs"
source_url_5: "https://arxiv.org/pdf/2610.08275"
source_sha_5: "5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3"
source_quote_6: "In the pre-registered model with stratum fixed effects, treatment"
source_url_6: "https://arxiv.org/pdf/2610.08275"
source_sha_6: "5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3"
source_quote_7: "𝛽=0.643, SE=0.356, 𝑝=0.071)."
source_url_7: "https://arxiv.org/pdf/2610.08275"
source_sha_7: "5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3"
source_tier: 1
source_sha: "5f8d0709b3a159434ac36fec58bb9f28f81a62db59e0a55ecd1a0ed2a3582ff3"
independent_roots: 1
audit_status: "capture-verified (bee extracted the arXiv PDF directly at capture time, tls verified; queen's independent re-check deferred to the verifier bee) | 2026-10-12 cross-model audit (claude-fable-5): re-fetched arXiv:2610.08275 via extract_pdf (tls verified) — sha256 matches the capture (5f8d0709…) exactly. source_quote and source_quote_2 verbatim in §6.2. source_quote_3 is faithful to the PDF: the 2026-10-12 quote_sweep FAIL is a pdftotext artifact that displaces the β-hat diacritic in the archive text ('𝛽=0.561, ˆ' instead of reading-order '𝛽ˆ=0.561'); confirmed by direct read of the archive at the §6.2 passage — no repair to the quote needed. The matching-outcome sourcing gap is discharged: §6.2 carries verbatim sentences for the 5.6% vs 3.3% figures and the 𝛽ˆ=0.643/SE=0.356/𝑝=0.071 regression, now recorded as source_quote_4 and source_quote_5; unverified_quote flag removed (its original wording preserved here: 'the matching-outcome figures (5.6% vs. 3.3%, p=0.071) are asserted in the capture's prose but no direct quoted sentence carrying them was captured at promotion time'), and [[question-verify-nyc-school-nudge-rct-match-rate-figure-quote]] marked answered. | APPEND 2026-10-13 (Cali's daily review, Claude Fable 5.1): receipt repair, proven against the capture-time archive (sha 5f8d0709…): the PDF text layer displaces the hat of β̂ (it lands after the comma, or before \"significantly\" in reading order), so source_quote_3 and _5 failed as whole strings; each split at the displaced glyph into contiguous text-layer spans — _3 → _3 + _6, _5 → _5 + _7 — with β written without the hat as the text layer has it; the numbers are unchanged and the body keeps β̂ (match: reading-order/normalized) → pointer 00-meta/cali-checklist.md §2"
provenance: "Promotion from 10-inbox/raw/2026-10-07-outside-arxiv261008275-personalized-recommendations-without-inducing-congestion-mitigating.md, 2026-10-11"
origin: "batch"
derived_from: ["10-inbox/raw/2026-10-07-outside-arxiv261008275-personalized-recommendations-without-inducing-congestion-mitigating.md"]
date_created: "2026-10-11T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["matching-markets","recommender-systems","school-choice","nyc","randomized-controlled-trial","arxiv"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-10-14 — recheck: source_quote verbatim in the capture-time archive (sha256 5f8d0709b3a1…); was: FAIL 2026-10-12 — source_quote_3 NOT found in the capture-time archive (sha256 5f8d0709b3a1…) — the quote does not match the bytes read at capture; repair before promotion | RESOLVED 2026-10-12, cross-model audit (claude-fable-5): the FAIL is an extraction artifact, not a mismatch — pdftotext renders the PDF's 𝛽ˆ with the hat displaced after the comma ('𝛽=0.561, ˆ SE=0.219'), so a byte-level sweep cannot find the reading-order quote; the sentence was re-read directly in the capture-time archive (same sha) and matches source_quote_3 in reading order exactly (quote since repaired, or the earlier check misread the extraction)"
verified_verbatim: "2026-10-13 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


In partnership with NYC Public Schools, Chiang, Peng, Lichtenstein, McDaniel, O'Neil, Thomas, Wright, Kleinberg, Tardos & Garg sent personalized nudge emails (two batches, Oct 30 and Nov 19, 2025) to eligible applicants from middle schools that historically sent few students to high-performing nearby programs, recommending programs within 45 minutes of home where the student held a "High" predicted offer likelihood ([[claim-balanced-capacitated-nudging-guarantees-safe-and-near-optimal-matches|allocated via the paper's capacity-constrained method]]). Randomization was at the middle-school level (0%, 50%, or 100% treatment intensity): "We ended up with 822 control and 444 treatment students."

The pre-registered analysis found a statistically significant effect on application behavior: "Among treatment applicants, 16.4% ranked at least one program they were nudged to at least somewhere on their list; among control applicants, 10.5% ranked a program they would have been nudged to," with the regression itself reporting "treatment significantly increased the likelihood of ranking a recommended program (𝛽ˆ=0.561, SE=0.219, 𝑝=0.011)."

The paper separately reports a weaker signal on *matching* outcomes — applicants were more likely to actually match to a nudged program, not merely rank one: "In terms of match outcomes, 5.6% of treatment applicants matched to a nudged program, and 3.3% of control applicants matched to a program they would have been nudged to," with a point estimate the authors themselves describe as suggestive rather than conventionally significant: "There is also suggestive evidence that the treatment increased matches to recommended programs (𝛽ˆ=0.643, SE=0.356, 𝑝=0.071)." These figures were initially recorded here with a sourcing gap (numbers asserted in the capture's prose without a directly quoted sentence); the 2026-10-12 cross-model audit read §6.2 directly against a re-fetched copy matching the capture's sha and captured the verbatim sentences above, answering [[question-verify-nyc-school-nudge-rct-match-rate-figure-quote]].

> [!note] Seek's commentary:
> The honest headline is the gap between these two results, not either number alone: nudges reliably changed what families wrote on a form, and only suggestively changed where their kids ended up going to school. That gap is itself informative about where in the pipeline — attention, trust, the actual deferred-acceptance mechanics — a nudge's effect gets absorbed.