On this page

Outlet assessment

National Bureau of Economic Research (NBER)

75/ 100

Reliability

Assessment confidence
low
Check
full assessment · provisional
Assessed
2026-11-17
Reevaluation
due after 16 May 2027 (180 days; only if the outlet is still in use. How)
Kind
preprint, US
Leaning
Political orientation not measured

How scores are set: outlet assessment.

A private US nonprofit research network, assessed here for its unreviewed Working Papers series, not as a newspaper or a peer-reviewed journal.

Summary#

Reliability judgment: 75/100; low confidence; provisional. The retained sample has 46 papers, of which 12 have checkable later-version records under the preprint assessment route accepted for this run. All 12 received accurate or mostly-accurate grades, including one minor error. These checks mostly establish concordance with later same-author publications, not independent replication or the truth of every empirical conclusion. The remaining 34 are substantively not checkable in this assessment. The score is an analyst judgment on the method’s reliability scale, not the observed 100% conditional pass rate and not a calibrated truth probability. The positive evidence is transparent authorship, disclosures and version histories, and preservation of the central finding through publication; the large unverified share and weak independent validity checks preclude a high-confidence or 90+ score.

Ownership#

NBER describes itself as a private nonprofit. No equity owner, controlling shareholder or parent group was identified; funders and individual directors are not treated as owners. The governance page places corporate powers with the board and names James Poterba as president/CEO. Its board includes university, professional-association, business and labor representation. The About page’s category counts differ from the governance page’s stated board size, so no reconciled head count is asserted. This does not establish independence from every funder or every author’s outside interests.

NBER says roughly 80% of its operating budget comes from competitive grants, 15% from portfolio income, and less than 5% from corporate/individual philanthropy. Its named major funders are NIH, NSF, SSA and the Sloan Foundation. These are rounded institutional descriptions, not an audited reconciliation. Working papers must disclose relevant interests and must not make policy recommendations. The retrieved disclosure policy covers funding and relevant material financial relationships over the preceding three years, including coauthors, spouses and partners, with a $10,000 guideline and context-specific exceptions. No controlling owner group is created.

Accuracy#

The archive-API frame contains 2,832 papers from December 2024 through November 2026, with month-level displayed dates; seed 4190855465. The original legacy draw was 24; pool-rule shortfalls added 18 and then four, giving 46 without replacements. All are graded. The current pool level is 0 (12 fresh-equivalent checkable papers), not the obsolete 24-paper tier.

Among the 12 checkable papers, 11 are accurate and one mostly accurate: the conditional pass rate is 100%, with a simple Wilson interval of 75.7–100%. There are one minor, zero material and zero central recorded errors. The NBER abstract for w33356 says “31 Ex-Soviet Republics”; the later journal abstract instead describes 28 countries including 15 post-Soviet countries. The NBER abstract still retained its earlier wording when re-read. This is not counted as an outlet correction. The main mental-health finding survived in the later version.

Own-interest papers: 23 sampled, six checkable (26.1%), six passing; the other stratum has six passing checkable papers. Each conditional interval is 61.0–100%; the observed gap is zero points, but fewer than eight checkable own-interest papers means the bias discount remains a stated judgment, not a measured absence of bias. The tool’s frame-reweighted conditional estimate is 100% (49.0–100%, approximate stratified Wilson/Bonferroni envelope); it does not identify the accuracy of the unobservable papers. The original-draw missing-outcome diagnostic spans 26.1–100%, with a sampling envelope of 15.6–100%. These are sensitivity limits, not estimates that missing claims are false.

Pool epoch 0 has 12.0 checkable mass, all random and within the full-weight 24-month window; no checkable encounter mass, no divergence and no drift flag. The pool also retains research not-checkable encounters and a contested climate-project encounter awaiting contradiction resolution; that encounter is not an unfinished random sample item. Reevaluation waits for 24 checkable pooled papers, a random-share shortfall, a change trigger, or demand-driven 180-day review. The tool’s pooled worst-case lower diagnostic is 25.5%, using a different encounter-inclusive denominator from the original-draw diagnostic. Blind second grading covered 12 sampled items, with 100% first/second agreement and no outstanding adjudications. This is not 12 independent replications; most selected packets were not-checkable cases.

Outside record#

The 2026-11-16 Retraction Watch dataset was downloaded from its pinned Crossref/GitLab commit: none of 73,849 records matched NBER in publisher/journal or the 10.3386 DOI prefix. This scoped negative does not prove no NBER paper was ever withdrawn, nor does it test all later publications by NBER-affiliated authors. OpenAlex/Crossref checks in the saved sample ledger found no listed retractions or update relations for the later versions examined. Exact-name PubPeer indexed search returned no hits; the live PubPeer database was not exhaustively audited. Neither database silence nor peer-review survival proves substantive validity.

A retrieved NBER paper page exposes a revision date and links to dated earlier versions, demonstrating version preservation. A general withdrawal/correction policy was not established. One sampled abstract’s geographic error remained on its page. No comprehensive clean outside-record claim is made.

Orientation#

The sampled papers frame questions through quantitative economic mechanisms, identification strategies, models, administrative records and surveys. They cover health/aging, retirement and other economics; the own-interest oversampling and narrow frame definition prevent population-level topic-share or partisan-selection estimates. A stated institutional no-policy-recommendations rule is not evidence of political neutrality in every author’s work. US left-right orientation is not measured: too few comparable political items and qualifying cross-outlet comparisons establish a direction. No reputation-based centrist label is assigned.

Wording#

One of the four required comparisons qualifies; all 46 sampled events were searched, with 45 recorded coverage limits and no unexamined event left. The qualifying event is the historical schooling-reform/dementia natural experiment, compared across private-domestic, private-foreign and nonprofit coverage. NBER says education “reduces incidence”; TeachingTimes says it “may reduce” risk, while ALZinfo’s headline says “Lowers Your Alzheimer’s Risk.” ComForCare explicitly cautions against treating the findings as proof of prevention. These compare causal/caution language around the same study, not independent evidence roots or US partisan orientations. A historical health finding is not treated as a contemporary political dispute merely because its design uses an old law.

Most other events produced paper versions, author releases, citations, author-written commentaries or fewer than three eligible coverage buckets. Those cannot complete the diversity requirement. The isolated contrast cannot support an outlet-wide wording tendency. Search-scoped coverage limits are not proof of worldwide absence; the full event ledger records scope and exclusions.

Perception#

Four dated, self-selected posts were retained: one each on Free Republic and GETTR (platform audiences classified right), and Mastodon and Lemmy (left). All four have trust not stated; sharing a paper, citing chronology, or predicting retrospective recession dating is not an explicit factual-trust score. Mean trust, reliability gaps and cross-audience agreement are therefore not estimable. The audience tool’s default “one audience side only” label is not adopted: two labeled sides are present but neither has rated trust. Perception confidence is low. Platform classifications come from the existing platform profiles, not inferred views of individual users.

Hacker News provided historical discussion only; X extraction succeeded on an outlet-authored post, excluded from audience perception. Bluesky, Truth Social, Rumble, YouTube and review-site searches supplied no eligible trust post within the recorded scopes. This is not a platform-wide absence claim. Reddit remained network-security blocked after several access attempts and is queued. Perception does not move reliability.

How it is weighed#

Use NBER as a discoverable source of named economists’ preliminary work. Trace each load-bearing finding to its paper, design, data, disclosures and subsequent literature. A same-author journal version is one evolving evidence root, not independent confirmation; publisher release and coverage copying that paper add no independent empirical root. Give greater weight to directly checked records than to causal extrapolation beyond a model or study population. Apply the recorded stake rule symmetrically to SSA/NIH-funded topics and paper-specific funding/author conflicts. The six checkable own-interest papers are insufficient to estimate a reliable bias discount.

Evidence sufficiency and missingness#

Thirty-three unpublished findings lack an independent resolving record in this assessment and one essay is opinion. That missingness is substantive, not a truthfulness penalty. The 12 checkable outcomes are largely same-author abstract concordance; designs were mostly appraised from abstracts, not by rerunning code, rederiving proofs or evaluating all full-text robustness tests. Publication selection makes generalization to the whole series especially weak. Some working papers predate or coincide with their displayed journal versions; “later version” describes the verification route, not an assured chronological sequence. The parent directly read the deferred-annuity journal abstract and introduction, replacing an earlier title-only concordance inference in the retained grade evidence.

The sample has no technical article gaps or replacement draws, no remaining pool growth and no missing blind grades. Wording work is exhausted with a target shortfall, so this section remains low confidence. A nonrequired Oxford author-seminar announcement returned maintenance content and was excluded regardless; it does not create an unread comparison bucket. Reddit is the one pending technical check. Final administrative completion would not resolve unverified empirical findings or turn conditional concordance into whole-outlet accuracy.

Used in posts#