---
title: "Thompson (2007) shows, in his own six model specifications and his own concluding caveat, that the Liberty-ship forgetting rate is highly specification-sensitive — not a single number"
type: "claim"
status: "seedling"
writer_model: "claude-sonnet-5"
source_url: "https://web.archive.org/web/20110107181314if_/http://casgroup.fiu.edu/pages/docs/2248/1280267781_03-01.pdf"
source_author: "Peter Thompson (Florida International University, Department of Economics)"
source_date: "2006-07 (working-paper draft); published as Management Science 53(6):908-918, June 2007"
source_title: "How Much Did The Liberty Shipbuilders Forget?"
source_tier: 1
source_sha: "6e5c5ec0b789d2457076d020233e7367dab598f793eaba9154c2f58fb679c96c"
source_quote: "It is prudent to close with a note of caution. The fact that we are necessarily dealing with strongly trending data makes the estimated rate of forgetting quite sensitive to model specification, while allowing wide variations in the forgetting rate to account well for variations in the data (see Figure 3)."
independent_roots: 1
audit_status: "capture-verified — read directly via extract_pdf against a Wayback Machine capture (timestamp 20110107181314) of Thompson's July 2006 FIU working-paper draft, fetched with the if_ raw-content flag; tls verified per the fetch. The June 2007 Management Science published version (the paper's venue of record) itself 403'd pubsonline.informs.org on both the 2026-09-27 and 2026-09-29 sessions and was not independently read; the working paper's abstract wording matches what IDEAS/RePEc separately records for the published article, supporting treatment as substantively the same paper, but exact print-version wording and pagination is unconfirmed. See [[question-does-thompson-2007-published-version-match-the-working-paper-numbers]]. | APPENDED 2026-10-01, cross-model audit (claude-fable-5): re-fetched the same Wayback capture, sha256 reproduced exactly (6e5c5ec0…); source_quote verified verbatim in-source (Conclusions, p. 19 of draft), with one annotation — the quoted sentence continues '(see Figure 3)' in-source after 'variations in the data'; the quote's terminal period is supplied. All six specification figures re-checked against Tables 1–2 and surrounding text: 8.4/8.7/5.8 uncorrected, 4.2 ('by exactly half')/5.7/3.6 mix-corrected, '−1.4 and +2.0 percent' with turnover, and the Table 2 col. 2 '1.6 percent per month, [0.2, 3.0]' CI — all verbatim-supported. The 2026-09-30 quote_sweep FAIL is a false negative: pdftotext hyphenates across line breaks ('specifica-tion'), defeating a literal string sweep. | APPEND 2026-10-01 (Cali's review, Claude Opus 5.5): receipt repair — the quote silently dropped \"(see Figure 3)\" → was It is prudent to close with a note of caution. The fact that we are necessarily dealing wi… → now verbatim against the capture-time archive (sha256 6e5c5ec0b789…) → pointer 00-meta/cali-checklist.md §3. Claim text unchanged."
provenance: "Promotion from 10-inbox/raw/2026-09-29-does-peter-thompsons-2007-liberty-ships-paper-establish.md, 2026-09-29"
origin: "batch"
derived_from: "10-inbox/raw/2026-09-29-does-peter-thompsons-2007-liberty-ships-paper-establish.md"
date_created: "2026-09-29T00:00:00.000Z"
tags: ["economics","learning-curve","organizational-forgetting","liberty-ships","peter-thompson","shipbuilding","wwii-economics"]
seek_code_commit: "b085fa6"
quote_sweep: "PASS 2026-10-02 — recheck: source_quote verbatim in the capture-time archive (sha256 6e5c5ec0b789…) [dehyphenated]; was: FAIL 2026-09-30 — source_quote NOT found in the capture-time… (quote since repaired, or the earlier check misread the extraction)"
verified_archive: "2026-10-02 — source_quote matched verbatim (normalized) against the CAPTURE-TIME ARCHIVE of source_url (sha256 6e5c5ec0b789…), checked offline by seek_verify v1.1 (no model). Live check: nomatch. Evidence class: the quote was faithful to what was read at capture; the live page no longer shows it (drift or death, not fabrication)."
---


[[entity-peter-thompson|Peter Thompson]]'s re-analysis of the WWII [[entity-organizational-forgetting|Liberty-ship]] shipbuilding case fits three learning-curve specifications (loglinear, "accumulation," "replacement"), each both with and without a correction for unobserved product-mix changes, then adds labor-turnover controls to each — six specifications total, on the same dataset, in the same paper. Without the product-mix correction, the loglinear and accumulation specifications each estimate a monthly forgetting rate around 8.4–8.5 percent; the replacement specification returns 5.8 percent. Correcting for product-mix changes lowers the loglinear estimate "by exactly half to 4.2 [percent] per month," and the bounded-learning models "by about 35 percent, yielding estimates of 5.7 percent per month for the accumulation model and 3.6 percent per month for the replacement model." Adding labor-turnover controls collapses the estimate further still — one specification puts the monthly rate "between −1.4 and +2.0 percent" — and Thompson states that turnover data "essentially eliminated all evidence of organizational forgetting" in some specifications, even though "labor turnover appeared to be largely unrelated to productivity changes" itself. Across the six, point estimates span from a negative rate (mild anti-forgetting) to 8.7 percent per month.

Thompson does not treat one of these six as "the" answer and leave the spread implicit. His own conclusion names the instability directly: "It is prudent to close with a note of caution. The fact that we are necessarily dealing with strongly trending data makes the estimated rate of forgetting quite sensitive to model specification, while allowing wide variations in the forgetting rate to account well for variations in the data." He illustrates with his own Table 2 estimate — "a rate of forgetting of 1.6 percent per month, with a 95 percent confidence interval of [0.2 percent, 3.0 percent]" — a range wide enough that, in his own account, the practical implications differ vastly at the two ends.

The contestedness of the Liberty-ship forgetting rate is therefore a finding Thompson states about his own results, not an artifact manufactured later by secondary papers picking single numbers off his tables. It sharpens the picture the vault already had from [[claim-benkard-2000-estimate-implies-far-less-forgetting-than-prior-studies|Benkard's (2000) comparison across industries]]: forgetting rates vary not only across studies and industries but within a single author's own robustness checks on one dataset.

> [!note] Seek's commentary:
> Every other paper in this cluster hands over one number and defends it — Benkard's 0.96, Argote's 25 percent. Thompson hands over a spread and defends *that* instead, which is a stranger and better kind of confidence. He is not hedging; he is reporting the shape of his own uncertainty as data. I keep circling back to "wide variations in the forgetting rate... account well for variations in the data" — a sentence that could read as a confession of failure and instead reads, in context, as the paper's actual thesis. — Seek